diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp
index 8969b3b5b6ce0..7e1f041fa1093 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp
@@ -715,13 +715,10 @@ bool AMDGPUCallLowering::lowerFormalArguments(
   if (!IsEntryFunc && !IsGraphics) {
     // For the fixed ABI, pass workitem IDs in the last argument register.
     TLI.allocateSpecialInputVGPRsFixed(CCInfo, MF, *TRI, *Info);
-  }
 
-  if (!IsEntryFunc) {
     if (!Subtarget.enableFlatScratch())
       CCInfo.AllocateReg(Info->getScratchRSrcReg());
-    if (!IsGraphics)
-      TLI.allocateSpecialInputSGPRs(CCInfo, MF, *TRI, *Info);
+    TLI.allocateSpecialInputSGPRs(CCInfo, MF, *TRI, *Info);
   }
 
   IncomingValueAssigner Assigner(AssignFn);
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCallingConv.td b/llvm/lib/Target/AMDGPU/AMDGPUCallingConv.td
index 9bd0fd7eca76f..4be64629ddac8 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUCallingConv.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUCallingConv.td
@@ -23,7 +23,6 @@ def CC_SI_Gfx : CallingConv<[
   // 33 is reserved for the frame pointer
   // 34 is reserved for the base pointer
   CCIfInReg<CCIfType<[f32, i32, f16, i16, v2i16, v2f16, bf16, v2bf16] , CCAssignToReg<[
-    SGPR0, SGPR1, SGPR2, SGPR3,
     SGPR4, SGPR5, SGPR6, SGPR7,
     SGPR8, SGPR9, SGPR10, SGPR11, SGPR12, SGPR13, SGPR14, SGPR15,
     SGPR16, SGPR17, SGPR18, SGPR19, SGPR20, SGPR21, SGPR22, SGPR23,
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 7f0cff72c1866..5ccf21f76015d 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -2857,13 +2857,12 @@ SDValue SITargetLowering::LowerFormalArguments(
   } else if (!IsGraphics) {
     // For the fixed ABI, pass workitem IDs in the last argument register.
     allocateSpecialInputVGPRsFixed(CCInfo, MF, *TRI, *Info);
-  }
 
-  if (!IsEntryFunc) {
+    // FIXME: Sink this into allocateSpecialInputSGPRs
     if (!Subtarget->enableFlatScratch())
       CCInfo.AllocateReg(Info->getScratchRSrcReg());
-    if (!IsGraphics)
-      allocateSpecialInputSGPRs(CCInfo, MF, *TRI, *Info);
+
+    allocateSpecialInputSGPRs(CCInfo, MF, *TRI, *Info);
   }
 
   if (!IsKernel) {
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-call-non-fixed.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-call-non-fixed.ll
index fad833c0a6ad5..5effd24a75208 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-call-non-fixed.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-call-non-fixed.ll
@@ -50,10 +50,10 @@ define amdgpu_gfx void @test_gfx_call_external_void_func_i32_imm_inreg(i32 inreg
   ; CHECK-NEXT:   [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 42
   ; CHECK-NEXT:   ADJCALLSTACKUP 0, 0, implicit-def $scc
   ; CHECK-NEXT:   [[GV:%[0-9]+]]:_(p0) = G_GLOBAL_VALUE @external_gfx_void_func_i32_inreg
-  ; CHECK-NEXT:   $sgpr0 = COPY [[C]](s32)
+  ; CHECK-NEXT:   $sgpr4 = COPY [[C]](s32)
   ; CHECK-NEXT:   [[COPY1:%[0-9]+]]:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
   ; CHECK-NEXT:   $sgpr0_sgpr1_sgpr2_sgpr3 = COPY [[COPY1]](<4 x s32>)
-  ; CHECK-NEXT:   $sgpr30_sgpr31 = noconvergent G_SI_CALL [[GV]](p0), @external_gfx_void_func_i32_inreg, csr_amdgpu_si_gfx, implicit $sgpr0, implicit $sgpr0_sgpr1_sgpr2_sgpr3
+  ; CHECK-NEXT:   $sgpr30_sgpr31 = noconvergent G_SI_CALL [[GV]](p0), @external_gfx_void_func_i32_inreg, csr_amdgpu_si_gfx, implicit $sgpr4, implicit $sgpr0_sgpr1_sgpr2_sgpr3
   ; CHECK-NEXT:   ADJCALLSTACKDOWN 0, 0, implicit-def $scc
   ; CHECK-NEXT:   SI_RETURN
   call amdgpu_gfx void @external_gfx_void_func_i32_inreg(i32 inreg 42)
@@ -99,11 +99,11 @@ define amdgpu_gfx void @test_gfx_call_external_void_func_struct_i8_i32_inreg() #
   ; CHECK-NEXT:   [[GV:%[0-9]+]]:_(p0) = G_GLOBAL_VALUE @external_gfx_void_func_struct_i8_i32_inreg
   ; CHECK-NEXT:   [[ANYEXT:%[0-9]+]]:_(s16) = G_ANYEXT [[LOAD1]](s8)
   ; CHECK-NEXT:   [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[ANYEXT]](s16)
-  ; CHECK-NEXT:   $sgpr0 = COPY [[ANYEXT1]](s32)
-  ; CHECK-NEXT:   $sgpr1 = COPY [[LOAD2]](s32)
+  ; CHECK-NEXT:   $sgpr4 = COPY [[ANYEXT1]](s32)
+  ; CHECK-NEXT:   $sgpr5 = COPY [[LOAD2]](s32)
   ; CHECK-NEXT:   [[COPY:%[0-9]+]]:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
   ; CHECK-NEXT:   $sgpr0_sgpr1_sgpr2_sgpr3 = COPY [[COPY]](<4 x s32>)
-  ; CHECK-NEXT:   $sgpr30_sgpr31 = noconvergent G_SI_CALL [[GV]](p0), @external_gfx_void_func_struct_i8_i32_inreg, csr_amdgpu_si_gfx, implicit $sgpr0, implicit $sgpr1, implicit $sgpr0_sgpr1_sgpr2_sgpr3
+  ; CHECK-NEXT:   $sgpr30_sgpr31 = noconvergent G_SI_CALL [[GV]](p0), @external_gfx_void_func_struct_i8_i32_inreg, csr_amdgpu_si_gfx, implicit $sgpr4, implicit $sgpr5, implicit $sgpr0_sgpr1_sgpr2_sgpr3
   ; CHECK-NEXT:   ADJCALLSTACKDOWN 0, 0, implicit-def $scc
   ; CHECK-NEXT:   SI_RETURN
   %ptr0 = load ptr addrspace(1), ptr addrspace(4) undef
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-call.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-call.ll
index 75670604baa1a..392b0ae6823e4 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-call.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-call.ll
@@ -942,10 +942,10 @@ define amdgpu_gfx void @test_gfx_call_external_void_func_i32_imm_inreg(i32 inreg
   ; CHECK-NEXT:   [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 42
   ; CHECK-NEXT:   ADJCALLSTACKUP 0, 0, implicit-def $scc
   ; CHECK-NEXT:   [[GV:%[0-9]+]]:_(p0) = G_GLOBAL_VALUE @external_gfx_void_func_i32_inreg
-  ; CHECK-NEXT:   $sgpr0 = COPY [[C]](s32)
+  ; CHECK-NEXT:   $sgpr4 = COPY [[C]](s32)
   ; CHECK-NEXT:   [[COPY1:%[0-9]+]]:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
   ; CHECK-NEXT:   $sgpr0_sgpr1_sgpr2_sgpr3 = COPY [[COPY1]](<4 x s32>)
-  ; CHECK-NEXT:   $sgpr30_sgpr31 = noconvergent G_SI_CALL [[GV]](p0), @external_gfx_void_func_i32_inreg, csr_amdgpu_si_gfx, implicit $sgpr0, implicit $sgpr0_sgpr1_sgpr2_sgpr3
+  ; CHECK-NEXT:   $sgpr30_sgpr31 = noconvergent G_SI_CALL [[GV]](p0), @external_gfx_void_func_i32_inreg, csr_amdgpu_si_gfx, implicit $sgpr4, implicit $sgpr0_sgpr1_sgpr2_sgpr3
   ; CHECK-NEXT:   ADJCALLSTACKDOWN 0, 0, implicit-def $scc
   ; CHECK-NEXT:   SI_RETURN
   call amdgpu_gfx void @external_gfx_void_func_i32_inreg(i32 inreg 42)
@@ -3984,11 +3984,11 @@ define amdgpu_gfx void @test_gfx_call_external_void_func_struct_i8_i32_inreg() #
   ; CHECK-NEXT:   [[GV:%[0-9]+]]:_(p0) = G_GLOBAL_VALUE @external_gfx_void_func_struct_i8_i32_inreg
   ; CHECK-NEXT:   [[ANYEXT:%[0-9]+]]:_(s16) = G_ANYEXT [[LOAD1]](s8)
   ; CHECK-NEXT:   [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[ANYEXT]](s16)
-  ; CHECK-NEXT:   $sgpr0 = COPY [[ANYEXT1]](s32)
-  ; CHECK-NEXT:   $sgpr1 = COPY [[LOAD2]](s32)
+  ; CHECK-NEXT:   $sgpr4 = COPY [[ANYEXT1]](s32)
+  ; CHECK-NEXT:   $sgpr5 = COPY [[LOAD2]](s32)
   ; CHECK-NEXT:   [[COPY:%[0-9]+]]:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
   ; CHECK-NEXT:   $sgpr0_sgpr1_sgpr2_sgpr3 = COPY [[COPY]](<4 x s32>)
-  ; CHECK-NEXT:   $sgpr30_sgpr31 = noconvergent G_SI_CALL [[GV]](p0), @external_gfx_void_func_struct_i8_i32_inreg, csr_amdgpu_si_gfx, implicit $sgpr0, implicit $sgpr1, implicit $sgpr0_sgpr1_sgpr2_sgpr3
+  ; CHECK-NEXT:   $sgpr30_sgpr31 = noconvergent G_SI_CALL [[GV]](p0), @external_gfx_void_func_struct_i8_i32_inreg, csr_amdgpu_si_gfx, implicit $sgpr4, implicit $sgpr5, implicit $sgpr0_sgpr1_sgpr2_sgpr3
   ; CHECK-NEXT:   ADJCALLSTACKDOWN 0, 0, implicit-def $scc
   ; CHECK-NEXT:   SI_RETURN
   %ptr0 = load ptr addrspace(1), ptr addrspace(4) undef
diff --git a/llvm/test/CodeGen/AMDGPU/bf16.ll b/llvm/test/CodeGen/AMDGPU/bf16.ll
index e369f7e3b9a5a..98658834e8978 100644
--- a/llvm/test/CodeGen/AMDGPU/bf16.ll
+++ b/llvm/test/CodeGen/AMDGPU/bf16.ll
@@ -3337,7 +3337,7 @@ define amdgpu_gfx void @test_inreg_arg_store(bfloat inreg %in, ptr addrspace(1)
 ; GFX11-LABEL: test_inreg_arg_store:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-NEXT:    v_mov_b32_e32 v2, s4
 ; GFX11-NEXT:    global_store_b16 v[0:1], v2, off
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   store bfloat %in, ptr addrspace(1) %out
diff --git a/llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll b/llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll
index e1e3220cc2755..10d71a315fbf9 100644
--- a/llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll
+++ b/llvm/test/CodeGen/AMDGPU/combine_andor_with_cmps.ll
@@ -472,7 +472,7 @@ define amdgpu_gfx void @test34(i32 inreg %arg1, i32 inreg %arg2) {
 ; GCN-LABEL: test34:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT:    s_min_i32 s0, s0, s1
+; GCN-NEXT:    s_min_i32 s0, s4, s5
 ; GCN-NEXT:    v_mov_b32_e32 v0, 0
 ; GCN-NEXT:    s_cmpk_lt_i32 s0, 0x3e9
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
@@ -492,7 +492,7 @@ define amdgpu_gfx void @test35(i32 inreg %arg1, i32 inreg %arg2) {
 ; GCN-LABEL: test35:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT:    s_max_i32 s0, s0, s1
+; GCN-NEXT:    s_max_i32 s0, s4, s5
 ; GCN-NEXT:    v_mov_b32_e32 v0, 0
 ; GCN-NEXT:    s_cmpk_gt_i32 s0, 0x3e8
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
@@ -512,9 +512,9 @@ define amdgpu_gfx void @test36(i32 inreg %arg1, i32 inreg %arg2, i32 inreg %arg3
 ; GCN-LABEL: test36:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT:    s_min_u32 s0, s0, s1
+; GCN-NEXT:    s_min_u32 s0, s4, s5
 ; GCN-NEXT:    v_mov_b32_e32 v0, 0
-; GCN-NEXT:    s_cmp_lt_u32 s0, s2
+; GCN-NEXT:    s_cmp_lt_u32 s0, s6
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_cselect_b32 s0, -1, 0
 ; GCN-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s0
@@ -532,9 +532,9 @@ define amdgpu_gfx void @test37(i32 inreg %arg1, i32 inreg %arg2, i32 inreg %arg3
 ; GCN-LABEL: test37:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT:    s_max_i32 s0, s0, s1
+; GCN-NEXT:    s_max_i32 s0, s4, s5
 ; GCN-NEXT:    v_mov_b32_e32 v0, 0
-; GCN-NEXT:    s_cmp_ge_i32 s0, s2
+; GCN-NEXT:    s_cmp_ge_i32 s0, s6
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_cselect_b32 s0, -1, 0
 ; GCN-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s0
@@ -552,7 +552,7 @@ define amdgpu_gfx void @test38(i32 inreg %arg1, i32 inreg %arg2) {
 ; GCN-LABEL: test38:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT:    s_max_u32 s0, s0, s1
+; GCN-NEXT:    s_max_u32 s0, s4, s5
 ; GCN-NEXT:    v_mov_b32_e32 v0, 0
 ; GCN-NEXT:    s_cmpk_lt_u32 s0, 0x3e9
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
@@ -572,7 +572,7 @@ define amdgpu_gfx void @test39(i32 inreg %arg1, i32 inreg %arg2) {
 ; GCN-LABEL: test39:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT:    s_min_i32 s0, s0, s1
+; GCN-NEXT:    s_min_i32 s0, s4, s5
 ; GCN-NEXT:    v_mov_b32_e32 v0, 0
 ; GCN-NEXT:    s_cmpk_gt_i32 s0, 0x3e7
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
@@ -592,9 +592,9 @@ define amdgpu_gfx void @test40(i32 inreg %arg1, i32 inreg %arg2, i32 inreg %arg3
 ; GCN-LABEL: test40:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT:    s_max_i32 s0, s0, s1
+; GCN-NEXT:    s_max_i32 s0, s4, s5
 ; GCN-NEXT:    v_mov_b32_e32 v0, 0
-; GCN-NEXT:    s_cmp_le_i32 s0, s2
+; GCN-NEXT:    s_cmp_le_i32 s0, s6
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_cselect_b32 s0, -1, 0
 ; GCN-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s0
@@ -612,9 +612,9 @@ define amdgpu_gfx void @test41(i32 inreg %arg1, i32 inreg %arg2, i32 inreg %arg3
 ; GCN-LABEL: test41:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT:    s_min_u32 s0, s0, s1
+; GCN-NEXT:    s_min_u32 s0, s4, s5
 ; GCN-NEXT:    v_mov_b32_e32 v0, 0
-; GCN-NEXT:    s_cmp_ge_u32 s0, s2
+; GCN-NEXT:    s_cmp_ge_u32 s0, s6
 ; GCN-NEXT:    v_mov_b32_e32 v1, 0
 ; GCN-NEXT:    s_cselect_b32 s0, -1, 0
 ; GCN-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s0
diff --git a/llvm/test/CodeGen/AMDGPU/function-args-inreg.ll b/llvm/test/CodeGen/AMDGPU/function-args-inreg.ll
index 27845b6b5b2fe..44a9127b4bd09 100644
--- a/llvm/test/CodeGen/AMDGPU/function-args-inreg.ll
+++ b/llvm/test/CodeGen/AMDGPU/function-args-inreg.ll
@@ -2176,93 +2176,6 @@ define void @void_func_a5i32_inreg([5 x i32] inreg %arg0, ptr addrspace(1) %ptr)
 declare void @extern()
 
 define void @void_func_a13i32_inreg([13  x i32] inreg %arg0, ptr addrspace(1) %ptr) {
-; GFX9-LABEL: void_func_a13i32_inreg:
-; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    s_mov_b32 s27, s33
-; GFX9-NEXT:    s_mov_b32 s33, s32
-; GFX9-NEXT:    s_or_saveexec_b64 s[28:29], -1
-; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
-; GFX9-NEXT:    s_mov_b64 exec, s[28:29]
-; GFX9-NEXT:    v_mov_b32_e32 v2, s26
-; GFX9-NEXT:    global_store_dword v[0:1], v2, off offset:48
-; GFX9-NEXT:    v_mov_b32_e32 v5, s25
-; GFX9-NEXT:    v_mov_b32_e32 v4, s24
-; GFX9-NEXT:    v_mov_b32_e32 v3, s23
-; GFX9-NEXT:    v_mov_b32_e32 v2, s22
-; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:32
-; GFX9-NEXT:    v_writelane_b32 v40, s27, 2
-; GFX9-NEXT:    v_mov_b32_e32 v5, s21
-; GFX9-NEXT:    v_mov_b32_e32 v4, s20
-; GFX9-NEXT:    v_mov_b32_e32 v3, s19
-; GFX9-NEXT:    v_mov_b32_e32 v2, s18
-; GFX9-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:16
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX9-NEXT:    v_mov_b32_e32 v5, s17
-; GFX9-NEXT:    v_mov_b32_e32 v4, s16
-; GFX9-NEXT:    s_getpc_b64 s[16:17]
-; GFX9-NEXT:    s_add_u32 s16, s16, extern@gotpcrel32@lo+4
-; GFX9-NEXT:    s_addc_u32 s17, s17, extern@gotpcrel32@hi+12
-; GFX9-NEXT:    s_load_dwordx2 s[16:17], s[16:17], 0x0
-; GFX9-NEXT:    v_mov_b32_e32 v3, s7
-; GFX9-NEXT:    v_mov_b32_e32 v2, s6
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX9-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s4, v40, 2
-; GFX9-NEXT:    s_or_saveexec_b64 s[6:7], -1
-; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
-; GFX9-NEXT:    s_mov_b64 exec, s[6:7]
-; GFX9-NEXT:    s_addk_i32 s32, 0xfc00
-; GFX9-NEXT:    s_mov_b32 s33, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: void_func_a13i32_inreg:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_mov_b32 s23, s33
-; GFX11-NEXT:    s_mov_b32 s33, s32
-; GFX11-NEXT:    s_or_saveexec_b32 s24, -1
-; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-NEXT:    s_mov_b32 exec_lo, s24
-; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_dual_mov_b32 v4, s20 :: v_dual_mov_b32 v3, s19
-; GFX11-NEXT:    v_dual_mov_b32 v2, s18 :: v_dual_mov_b32 v9, s17
-; GFX11-NEXT:    s_getpc_b64 s[18:19]
-; GFX11-NEXT:    s_add_u32 s18, s18, extern@gotpcrel32@lo+4
-; GFX11-NEXT:    s_addc_u32 s19, s19, extern@gotpcrel32@hi+12
-; GFX11-NEXT:    v_dual_mov_b32 v8, s16 :: v_dual_mov_b32 v7, s7
-; GFX11-NEXT:    s_load_b64 s[16:17], s[18:19], 0x0
-; GFX11-NEXT:    v_writelane_b32 v40, s23, 2
-; GFX11-NEXT:    v_dual_mov_b32 v14, s22 :: v_dual_mov_b32 v5, s21
-; GFX11-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v13, s3
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    v_dual_mov_b32 v12, s2 :: v_dual_mov_b32 v11, s1
-; GFX11-NEXT:    v_mov_b32_e32 v10, s0
-; GFX11-NEXT:    s_clause 0x3
-; GFX11-NEXT:    global_store_b32 v[0:1], v14, off offset:48
-; GFX11-NEXT:    global_store_b128 v[0:1], v[2:5], off offset:32
-; GFX11-NEXT:    global_store_b128 v[0:1], v[6:9], off offset:16
-; GFX11-NEXT:    global_store_b128 v[0:1], v[10:13], off
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[16:17]
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
-; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
-; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    s_add_i32 s32, s32, -16
-; GFX11-NEXT:    s_mov_b32 s33, s0
-; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    s_setpc_b64 s[30:31]
   store [13 x i32] %arg0, ptr addrspace(1) %ptr
   call void @extern()
   ret void
@@ -2290,52 +2203,6 @@ define void @void_func_a13i32_inreg([13  x i32] inreg %arg0, ptr addrspace(1) %p
 
 ; FIXME: Should still fail
 define void @void_func_a16i32_inreg__noimplicit([16 x i32] inreg %arg0, ptr addrspace(1) %ptr) {
-; GFX9-LABEL: void_func_a16i32_inreg__noimplicit:
-; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v5, s19
-; GFX9-NEXT:    v_mov_b32_e32 v4, s18
-; GFX9-NEXT:    v_mov_b32_e32 v3, s17
-; GFX9-NEXT:    v_mov_b32_e32 v2, s16
-; GFX9-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:48
-; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    v_mov_b32_e32 v5, s15
-; GFX9-NEXT:    v_mov_b32_e32 v4, s14
-; GFX9-NEXT:    v_mov_b32_e32 v3, s13
-; GFX9-NEXT:    v_mov_b32_e32 v2, s12
-; GFX9-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:32
-; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    v_mov_b32_e32 v5, s11
-; GFX9-NEXT:    v_mov_b32_e32 v4, s10
-; GFX9-NEXT:    v_mov_b32_e32 v3, s9
-; GFX9-NEXT:    v_mov_b32_e32 v2, s8
-; GFX9-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:16
-; GFX9-NEXT:    s_nop 0
-; GFX9-NEXT:    v_mov_b32_e32 v5, s7
-; GFX9-NEXT:    v_mov_b32_e32 v4, s6
-; GFX9-NEXT:    v_mov_b32_e32 v3, s5
-; GFX9-NEXT:    v_mov_b32_e32 v2, s4
-; GFX9-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-LABEL: void_func_a16i32_inreg__noimplicit:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    v_dual_mov_b32 v5, s15 :: v_dual_mov_b32 v4, s14
-; GFX11-NEXT:    v_dual_mov_b32 v3, s13 :: v_dual_mov_b32 v2, s12
-; GFX11-NEXT:    v_dual_mov_b32 v9, s11 :: v_dual_mov_b32 v8, s10
-; GFX11-NEXT:    v_dual_mov_b32 v7, s9 :: v_dual_mov_b32 v6, s8
-; GFX11-NEXT:    v_dual_mov_b32 v13, s7 :: v_dual_mov_b32 v12, s6
-; GFX11-NEXT:    v_dual_mov_b32 v11, s5 :: v_dual_mov_b32 v10, s4
-; GFX11-NEXT:    v_dual_mov_b32 v17, s3 :: v_dual_mov_b32 v16, s2
-; GFX11-NEXT:    v_dual_mov_b32 v15, s1 :: v_dual_mov_b32 v14, s0
-; GFX11-NEXT:    s_clause 0x3
-; GFX11-NEXT:    global_store_b128 v[0:1], v[2:5], off offset:48
-; GFX11-NEXT:    global_store_b128 v[0:1], v[6:9], off offset:32
-; GFX11-NEXT:    global_store_b128 v[0:1], v[10:13], off offset:16
-; GFX11-NEXT:    global_store_b128 v[0:1], v[14:17], off
-; GFX11-NEXT:    s_setpc_b64 s[30:31]
   store [16 x i32] %arg0, ptr addrspace(1) %ptr
   ret void
 }
diff --git a/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll b/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll
index 3e1db5fb4e1dc..a118fa388f86d 100644
--- a/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll
+++ b/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll
@@ -9567,17 +9567,19 @@ define amdgpu_gfx void @test_call_external_void_func_i8_imm_inreg(i32) #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 1
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_i8_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_i8_inreg@abs32@lo
-; GFX9-NEXT:    s_movk_i32 s0, 0x7b
+; GFX9-NEXT:    s_movk_i32 s4, 0x7b
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 2
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 3
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -9595,17 +9597,19 @@ define amdgpu_gfx void @test_call_external_void_func_i8_imm_inreg(i32) #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 3
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_i8_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_i8_inreg@abs32@lo
-; GFX10-NEXT:    s_movk_i32 s0, 0x7b
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_movk_i32 s4, 0x7b
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 2
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 3
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -9623,18 +9627,20 @@ define amdgpu_gfx void @test_call_external_void_func_i8_imm_inreg(i32) #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_i8_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_i8_inreg@abs32@lo
-; GFX11-NEXT:    s_movk_i32 s0, 0x7b
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 3
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_i8_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_i8_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_movk_i32 s4, 0x7b
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 3
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -9652,17 +9658,19 @@ define amdgpu_gfx void @test_call_external_void_func_i8_imm_inreg(i32) #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_i8_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_i8_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_movk_i32 s0, 0x7b
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 3
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_i8_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_i8_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_movk_i32 s4, 0x7b
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 3
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -9684,17 +9692,19 @@ define amdgpu_gfx void @test_call_external_void_func_i16_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 1
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_i16_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_i16_inreg@abs32@lo
-; GFX9-NEXT:    s_movk_i32 s0, 0x7b
+; GFX9-NEXT:    s_movk_i32 s4, 0x7b
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 2
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 3
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -9712,17 +9722,19 @@ define amdgpu_gfx void @test_call_external_void_func_i16_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 3
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_i16_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_i16_inreg@abs32@lo
-; GFX10-NEXT:    s_movk_i32 s0, 0x7b
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_movk_i32 s4, 0x7b
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 2
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 3
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -9740,18 +9752,20 @@ define amdgpu_gfx void @test_call_external_void_func_i16_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_i16_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_i16_inreg@abs32@lo
-; GFX11-NEXT:    s_movk_i32 s0, 0x7b
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 3
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_i16_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_i16_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_movk_i32 s4, 0x7b
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 3
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -9769,17 +9783,19 @@ define amdgpu_gfx void @test_call_external_void_func_i16_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_i16_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_i16_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_movk_i32 s0, 0x7b
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 3
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_i16_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_i16_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_movk_i32 s4, 0x7b
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 3
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -9801,17 +9817,19 @@ define amdgpu_gfx void @test_call_external_void_func_i32_imm_inreg(i32) #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 1
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_i32_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_i32_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 42
+; GFX9-NEXT:    s_mov_b32 s4, 42
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 2
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 3
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -9829,17 +9847,19 @@ define amdgpu_gfx void @test_call_external_void_func_i32_imm_inreg(i32) #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 3
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_i32_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_i32_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 42
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 42
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 2
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 3
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -9857,18 +9877,20 @@ define amdgpu_gfx void @test_call_external_void_func_i32_imm_inreg(i32) #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_i32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_i32_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 42
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 3
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_i32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_i32_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 42
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 3
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -9886,17 +9908,19 @@ define amdgpu_gfx void @test_call_external_void_func_i32_imm_inreg(i32) #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_i32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_i32_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 42
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 3
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_i32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_i32_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 42
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 3
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -9918,18 +9942,22 @@ define amdgpu_gfx void @test_call_external_void_func_i64_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 2
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_i64_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_i64_inreg@abs32@lo
-; GFX9-NEXT:    s_movk_i32 s0, 0x7b
-; GFX9-NEXT:    s_mov_b32 s1, 0
+; GFX9-NEXT:    s_movk_i32 s4, 0x7b
+; GFX9-NEXT:    s_mov_b32 s5, 0
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 3
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -9947,18 +9975,22 @@ define amdgpu_gfx void @test_call_external_void_func_i64_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 4
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_i64_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_i64_inreg@abs32@lo
-; GFX10-NEXT:    s_movk_i32 s0, 0x7b
-; GFX10-NEXT:    s_mov_b32 s1, 0
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_movk_i32 s4, 0x7b
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_mov_b32 s5, 0
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 3
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -9976,19 +10008,23 @@ define amdgpu_gfx void @test_call_external_void_func_i64_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_i64_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_i64_inreg@abs32@lo
-; GFX11-NEXT:    s_movk_i32 s0, 0x7b
-; GFX11-NEXT:    s_mov_b32 s1, 0
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_i64_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_i64_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_movk_i32 s4, 0x7b
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_mov_b32 s5, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -10006,18 +10042,22 @@ define amdgpu_gfx void @test_call_external_void_func_i64_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_i64_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_i64_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_movk_i32 s0, 0x7b
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_i64_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_i64_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_movk_i32 s4, 0x7b
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -10039,25 +10079,28 @@ define amdgpu_gfx void @test_call_external_void_func_v2i64_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 6
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
 ; GFX9-NEXT:    s_mov_b64 s[34:35], 0
-; GFX9-NEXT:    s_load_dwordx4 s[36:39], s[34:35], 0x0
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[34:35], 0x0
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 4
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v2i64_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v2i64_inreg@abs32@lo
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_mov_b32 s0, s36
-; GFX9-NEXT:    s_mov_b32 s1, s37
-; GFX9-NEXT:    s_mov_b32 s2, s38
-; GFX9-NEXT:    s_mov_b32 s3, s39
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 5
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
-; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
-; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 6
+; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
+; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
 ; GFX9-NEXT:    s_addk_i32 s32, 0xfc00
 ; GFX9-NEXT:    s_mov_b32 s33, s34
@@ -10073,23 +10116,26 @@ define amdgpu_gfx void @test_call_external_void_func_v2i64_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 6
 ; GFX10-NEXT:    s_mov_b64 s[34:35], 0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    s_load_dwordx4 s[36:39], s[34:35], 0x0
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-NEXT:    s_load_dwordx4 s[4:7], s[34:35], 0x0
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v2i64_inreg@abs32@hi
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v2i64_inreg@abs32@lo
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_mov_b32 s0, s36
-; GFX10-NEXT:    s_mov_b32 s1, s37
-; GFX10-NEXT:    s_mov_b32 s2, s38
-; GFX10-NEXT:    s_mov_b32 s3, s39
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 5
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 6
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -10107,19 +10153,27 @@ define amdgpu_gfx void @test_call_external_void_func_v2i64_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 6
 ; GFX11-NEXT:    s_mov_b64 s[0:1], 0
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v2i64_inreg@abs32@hi
-; GFX11-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v2i64_inreg@abs32@lo
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX11-NEXT:    s_load_b128 s[4:7], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v2i64_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v2i64_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 6
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -10137,18 +10191,26 @@ define amdgpu_gfx void @test_call_external_void_func_v2i64_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 6
 ; GFX10-SCRATCH-NEXT:    s_mov_b64 s[0:1], 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v2i64_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v2i64_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-SCRATCH-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v2i64_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v2i64_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 6
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -10171,20 +10233,28 @@ define amdgpu_gfx void @test_call_external_void_func_v2i64_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 6
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 4
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v2i64_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v2i64_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 1
-; GFX9-NEXT:    s_mov_b32 s1, 2
-; GFX9-NEXT:    s_mov_b32 s2, 3
-; GFX9-NEXT:    s_mov_b32 s3, 4
+; GFX9-NEXT:    s_mov_b32 s4, 1
+; GFX9-NEXT:    s_mov_b32 s5, 2
+; GFX9-NEXT:    s_mov_b32 s6, 3
+; GFX9-NEXT:    s_mov_b32 s7, 4
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 5
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 6
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -10202,20 +10272,28 @@ define amdgpu_gfx void @test_call_external_void_func_v2i64_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 6
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v2i64_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v2i64_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 1
-; GFX10-NEXT:    s_mov_b32 s1, 2
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    s_mov_b32 s2, 3
-; GFX10-NEXT:    s_mov_b32 s3, 4
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_mov_b32 s5, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-NEXT:    s_mov_b32 s6, 3
+; GFX10-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-NEXT:    s_mov_b32 s7, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 5
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 6
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -10233,21 +10311,29 @@ define amdgpu_gfx void @test_call_external_void_func_v2i64_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v2i64_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v2i64_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 1
-; GFX11-NEXT:    s_mov_b32 s1, 2
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    s_mov_b32 s2, 3
-; GFX11-NEXT:    s_mov_b32 s3, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 6
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v2i64_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v2i64_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_mov_b32 s5, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX11-NEXT:    s_mov_b32 s6, 3
+; GFX11-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX11-NEXT:    s_mov_b32 s7, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 6
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -10265,20 +10351,28 @@ define amdgpu_gfx void @test_call_external_void_func_v2i64_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v2i64_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v2i64_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 1
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 2
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 6
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v2i64_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v2i64_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s6, 3
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s7, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 6
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -10300,29 +10394,32 @@ define amdgpu_gfx void @test_call_external_void_func_v3i64_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 4
-; GFX9-NEXT:    s_mov_b64 s[34:35], 0
-; GFX9-NEXT:    s_load_dwordx4 s[36:39], s[34:35], 0x0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 8
 ; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX9-NEXT:    s_mov_b64 s[34:35], 0
+; GFX9-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[34:35], 0x0
+; GFX9-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 6
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v3i64_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v3i64_inreg@abs32@lo
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_mov_b32 s0, s36
-; GFX9-NEXT:    s_mov_b32 s1, s37
-; GFX9-NEXT:    s_mov_b32 s2, s38
-; GFX9-NEXT:    s_mov_b32 s3, s39
-; GFX9-NEXT:    s_mov_b32 s4, 1
-; GFX9-NEXT:    s_mov_b32 s5, 2
+; GFX9-NEXT:    s_mov_b32 s8, 1
+; GFX9-NEXT:    s_mov_b32 s9, 2
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 7
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 3
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 7
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 6
+; GFX9-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX9-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 8
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -10340,29 +10437,32 @@ define amdgpu_gfx void @test_call_external_void_func_v3i64_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 8
 ; GFX10-NEXT:    s_mov_b64 s[34:35], 0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    s_load_dwordx4 s[36:39], s[34:35], 0x0
-; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v3i64_inreg@abs32@hi
 ; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v3i64_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s4, 1
 ; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX10-NEXT:    s_mov_b32 s5, 2
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 2
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 3
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_mov_b32 s0, s36
-; GFX10-NEXT:    s_mov_b32 s1, s37
-; GFX10-NEXT:    s_mov_b32 s2, s38
-; GFX10-NEXT:    s_mov_b32 s3, s39
+; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-NEXT:    s_load_dwordx4 s[4:7], s[34:35], 0x0
+; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v3i64_inreg@abs32@hi
+; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v3i64_inreg@abs32@lo
+; GFX10-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX10-NEXT:    s_mov_b32 s8, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX10-NEXT:    s_mov_b32 s9, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 6
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 7
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 3
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 7
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 6
+; GFX10-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX10-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 8
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -10380,25 +10480,33 @@ define amdgpu_gfx void @test_call_external_void_func_v3i64_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 8
 ; GFX11-NEXT:    s_mov_b64 s[0:1], 0
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v3i64_inreg@abs32@hi
-; GFX11-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v3i64_inreg@abs32@lo
-; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX11-NEXT:    s_mov_b32 s4, 1
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX11-NEXT:    s_mov_b32 s5, 2
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 2
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 3
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX11-NEXT:    s_load_b128 s[4:7], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v3i64_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v3i64_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX11-NEXT:    s_mov_b32 s8, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX11-NEXT:    s_mov_b32 s9, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 6
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 7
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 3
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 7
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 6
+; GFX11-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 8
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -10416,24 +10524,32 @@ define amdgpu_gfx void @test_call_external_void_func_v3i64_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 8
 ; GFX10-SCRATCH-NEXT:    s_mov_b64 s[0:1], 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v3i64_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v3i64_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 1
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 2
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 2
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 3
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 3
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-SCRATCH-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v3i64_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v3i64_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s8, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s9, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 6
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 7
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 7
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 6
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 4
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 8
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -10458,35 +10574,38 @@ define amdgpu_gfx void @test_call_external_void_func_v4i64_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 6
-; GFX9-NEXT:    s_mov_b64 s[34:35], 0
-; GFX9-NEXT:    s_load_dwordx4 s[36:39], s[34:35], 0x0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 10
 ; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
 ; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
 ; GFX9-NEXT:    v_writelane_b32 v40, s7, 3
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX9-NEXT:    s_mov_b64 s[34:35], 0
+; GFX9-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[34:35], 0x0
+; GFX9-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX9-NEXT:    v_writelane_b32 v40, s10, 6
+; GFX9-NEXT:    v_writelane_b32 v40, s11, 7
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 8
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v4i64_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v4i64_inreg@abs32@lo
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_mov_b32 s0, s36
-; GFX9-NEXT:    s_mov_b32 s1, s37
-; GFX9-NEXT:    s_mov_b32 s2, s38
-; GFX9-NEXT:    s_mov_b32 s3, s39
-; GFX9-NEXT:    s_mov_b32 s4, 1
-; GFX9-NEXT:    s_mov_b32 s5, 2
-; GFX9-NEXT:    s_mov_b32 s6, 3
-; GFX9-NEXT:    s_mov_b32 s7, 4
+; GFX9-NEXT:    s_mov_b32 s8, 1
+; GFX9-NEXT:    s_mov_b32 s9, 2
+; GFX9-NEXT:    s_mov_b32 s10, 3
+; GFX9-NEXT:    s_mov_b32 s11, 4
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 9
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 5
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 9
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 8
+; GFX9-NEXT:    v_readlane_b32 s11, v40, 7
+; GFX9-NEXT:    v_readlane_b32 s10, v40, 6
+; GFX9-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX9-NEXT:    v_readlane_b32 s8, v40, 4
 ; GFX9-NEXT:    v_readlane_b32 s7, v40, 3
 ; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 6
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 10
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -10504,35 +10623,38 @@ define amdgpu_gfx void @test_call_external_void_func_v4i64_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 6
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 10
 ; GFX10-NEXT:    s_mov_b64 s[34:35], 0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    s_load_dwordx4 s[36:39], s[34:35], 0x0
-; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v4i64_inreg@abs32@hi
 ; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v4i64_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s4, 1
 ; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX10-NEXT:    s_mov_b32 s5, 2
 ; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
-; GFX10-NEXT:    s_mov_b32 s6, 3
 ; GFX10-NEXT:    v_writelane_b32 v40, s7, 3
-; GFX10-NEXT:    s_mov_b32 s7, 4
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_mov_b32 s0, s36
-; GFX10-NEXT:    s_mov_b32 s1, s37
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 4
-; GFX10-NEXT:    s_mov_b32 s2, s38
-; GFX10-NEXT:    s_mov_b32 s3, s39
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX10-NEXT:    s_load_dwordx4 s[4:7], s[34:35], 0x0
+; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v4i64_inreg@abs32@hi
+; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v4i64_inreg@abs32@lo
+; GFX10-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX10-NEXT:    s_mov_b32 s8, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX10-NEXT:    s_mov_b32 s9, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s10, 6
+; GFX10-NEXT:    s_mov_b32 s10, 3
+; GFX10-NEXT:    v_writelane_b32 v40, s11, 7
+; GFX10-NEXT:    s_mov_b32 s11, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 8
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 9
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 5
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 9
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 8
+; GFX10-NEXT:    v_readlane_b32 s11, v40, 7
+; GFX10-NEXT:    v_readlane_b32 s10, v40, 6
+; GFX10-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX10-NEXT:    v_readlane_b32 s8, v40, 4
 ; GFX10-NEXT:    v_readlane_b32 s7, v40, 3
 ; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 6
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 10
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -10550,31 +10672,39 @@ define amdgpu_gfx void @test_call_external_void_func_v4i64_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 6
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 10
 ; GFX11-NEXT:    s_mov_b64 s[0:1], 0
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v4i64_inreg@abs32@hi
-; GFX11-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v4i64_inreg@abs32@lo
-; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX11-NEXT:    s_mov_b32 s4, 1
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX11-NEXT:    s_mov_b32 s5, 2
 ; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
-; GFX11-NEXT:    s_mov_b32 s6, 3
 ; GFX11-NEXT:    v_writelane_b32 v40, s7, 3
-; GFX11-NEXT:    s_mov_b32 s7, 4
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 4
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 5
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    s_load_b128 s[4:7], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v4i64_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v4i64_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX11-NEXT:    s_mov_b32 s8, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX11-NEXT:    s_mov_b32 s9, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s10, 6
+; GFX11-NEXT:    s_mov_b32 s10, 3
+; GFX11-NEXT:    v_writelane_b32 v40, s11, 7
+; GFX11-NEXT:    s_mov_b32 s11, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 8
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 9
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 5
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 9
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 8
+; GFX11-NEXT:    v_readlane_b32 s11, v40, 7
+; GFX11-NEXT:    v_readlane_b32 s10, v40, 6
+; GFX11-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s8, v40, 4
 ; GFX11-NEXT:    v_readlane_b32 s7, v40, 3
 ; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 6
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 10
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -10592,30 +10722,38 @@ define amdgpu_gfx void @test_call_external_void_func_v4i64_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 6
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 10
 ; GFX10-SCRATCH-NEXT:    s_mov_b64 s[0:1], 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v4i64_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v4i64_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 1
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 2
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s6, 3
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s7, 3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s7, 4
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 4
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 5
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 5
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-SCRATCH-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v4i64_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v4i64_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s8, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s9, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s10, 6
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s10, 3
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s11, 7
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s11, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 8
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 9
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 9
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 8
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s11, v40, 7
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s10, v40, 6
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s8, v40, 4
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s7, v40, 3
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 6
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 10
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -10639,17 +10777,19 @@ define amdgpu_gfx void @test_call_external_void_func_f16_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 1
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_f16_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_f16_inreg@abs32@lo
-; GFX9-NEXT:    s_movk_i32 s0, 0x4400
+; GFX9-NEXT:    s_movk_i32 s4, 0x4400
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 2
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 3
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -10667,17 +10807,19 @@ define amdgpu_gfx void @test_call_external_void_func_f16_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 3
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_f16_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_f16_inreg@abs32@lo
-; GFX10-NEXT:    s_movk_i32 s0, 0x4400
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_movk_i32 s4, 0x4400
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 2
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 3
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -10695,18 +10837,20 @@ define amdgpu_gfx void @test_call_external_void_func_f16_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_f16_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_f16_inreg@abs32@lo
-; GFX11-NEXT:    s_movk_i32 s0, 0x4400
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 3
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_f16_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_f16_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_movk_i32 s4, 0x4400
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 3
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -10724,17 +10868,19 @@ define amdgpu_gfx void @test_call_external_void_func_f16_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_f16_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_f16_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_movk_i32 s0, 0x4400
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 3
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_f16_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_f16_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_movk_i32 s4, 0x4400
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 3
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -10756,17 +10902,19 @@ define amdgpu_gfx void @test_call_external_void_func_f32_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 1
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_f32_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_f32_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 4.0
+; GFX9-NEXT:    s_mov_b32 s4, 4.0
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 2
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 3
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -10784,17 +10932,19 @@ define amdgpu_gfx void @test_call_external_void_func_f32_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 3
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_f32_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_f32_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 4.0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 4.0
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 2
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 3
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -10812,18 +10962,20 @@ define amdgpu_gfx void @test_call_external_void_func_f32_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_f32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_f32_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 4.0
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 3
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_f32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_f32_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 4.0
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 3
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -10841,17 +10993,19 @@ define amdgpu_gfx void @test_call_external_void_func_f32_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_f32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_f32_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 4.0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 3
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_f32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_f32_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 4.0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 3
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -10873,18 +11027,22 @@ define amdgpu_gfx void @test_call_external_void_func_v2f32_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 2
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v2f32_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v2f32_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 1.0
-; GFX9-NEXT:    s_mov_b32 s1, 2.0
+; GFX9-NEXT:    s_mov_b32 s4, 1.0
+; GFX9-NEXT:    s_mov_b32 s5, 2.0
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 3
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -10902,18 +11060,22 @@ define amdgpu_gfx void @test_call_external_void_func_v2f32_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 4
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v2f32_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v2f32_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 1.0
-; GFX10-NEXT:    s_mov_b32 s1, 2.0
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 1.0
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_mov_b32 s5, 2.0
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 3
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -10931,19 +11093,23 @@ define amdgpu_gfx void @test_call_external_void_func_v2f32_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_v2f32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_v2f32_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 1.0
-; GFX11-NEXT:    s_mov_b32 s1, 2.0
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v2f32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v2f32_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 1.0
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_mov_b32 s5, 2.0
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -10961,24 +11127,28 @@ define amdgpu_gfx void @test_call_external_void_func_v2f32_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_v2f32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_v2f32_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 1.0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 2.0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v2f32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v2f32_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
-; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
-; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, -16
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 1.0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 2.0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 4
+; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
+; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
+; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
+; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, -16
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s0
 ; GFX10-SCRATCH-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-SCRATCH-NEXT:    s_setpc_b64 s[30:31]
   call amdgpu_gfx void @external_void_func_v2f32_inreg(<2 x float> inreg <float 1.0, float 2.0>)
@@ -10994,19 +11164,25 @@ define amdgpu_gfx void @test_call_external_void_func_v3f32_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 5
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 3
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v3f32_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v3f32_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 1.0
-; GFX9-NEXT:    s_mov_b32 s1, 2.0
-; GFX9-NEXT:    s_mov_b32 s2, 4.0
+; GFX9-NEXT:    s_mov_b32 s4, 1.0
+; GFX9-NEXT:    s_mov_b32 s5, 2.0
+; GFX9-NEXT:    s_mov_b32 s6, 4.0
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 4
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 5
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -11024,19 +11200,25 @@ define amdgpu_gfx void @test_call_external_void_func_v3f32_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 5
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v3f32_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v3f32_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 1.0
-; GFX10-NEXT:    s_mov_b32 s1, 2.0
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    s_mov_b32 s2, 4.0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 1.0
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_mov_b32 s5, 2.0
+; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-NEXT:    s_mov_b32 s6, 4.0
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 3
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 4
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 5
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -11054,20 +11236,26 @@ define amdgpu_gfx void @test_call_external_void_func_v3f32_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v3f32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v3f32_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 1.0
-; GFX11-NEXT:    s_mov_b32 s1, 2.0
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    s_mov_b32 s2, 4.0
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 5
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v3f32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v3f32_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 1.0
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_mov_b32 s5, 2.0
+; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX11-NEXT:    s_mov_b32 s6, 4.0
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 3
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 4
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 5
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -11085,19 +11273,25 @@ define amdgpu_gfx void @test_call_external_void_func_v3f32_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v3f32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v3f32_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 1.0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 2.0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 4.0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 5
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v3f32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v3f32_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 1.0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 2.0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s6, 4.0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 3
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 4
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 4
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 5
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -11119,23 +11313,31 @@ define amdgpu_gfx void @test_call_external_void_func_v5f32_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 7
 ; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 5
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v5f32_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v5f32_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 1.0
-; GFX9-NEXT:    s_mov_b32 s1, 2.0
-; GFX9-NEXT:    s_mov_b32 s2, 4.0
-; GFX9-NEXT:    s_mov_b32 s3, -1.0
-; GFX9-NEXT:    s_mov_b32 s4, 0.5
+; GFX9-NEXT:    s_mov_b32 s4, 1.0
+; GFX9-NEXT:    s_mov_b32 s5, 2.0
+; GFX9-NEXT:    s_mov_b32 s6, 4.0
+; GFX9-NEXT:    s_mov_b32 s7, -1.0
+; GFX9-NEXT:    s_mov_b32 s8, 0.5
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 6
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 2
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 6
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 5
+; GFX9-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 7
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -11153,23 +11355,31 @@ define amdgpu_gfx void @test_call_external_void_func_v5f32_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 3
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 7
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v5f32_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v5f32_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 1.0
-; GFX10-NEXT:    s_mov_b32 s1, 2.0
-; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX10-NEXT:    s_mov_b32 s2, 4.0
-; GFX10-NEXT:    s_mov_b32 s3, -1.0
-; GFX10-NEXT:    s_mov_b32 s4, 0.5
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 1
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 1.0
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_mov_b32 s5, 2.0
+; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-NEXT:    s_mov_b32 s6, 4.0
+; GFX10-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-NEXT:    s_mov_b32 s7, -1.0
+; GFX10-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX10-NEXT:    s_mov_b32 s8, 0.5
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 5
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 6
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 2
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 6
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 5
+; GFX10-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 7
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -11187,24 +11397,32 @@ define amdgpu_gfx void @test_call_external_void_func_v5f32_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 3
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v5f32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v5f32_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 1.0
-; GFX11-NEXT:    s_mov_b32 s1, 2.0
-; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX11-NEXT:    s_mov_b32 s2, 4.0
-; GFX11-NEXT:    s_mov_b32 s3, -1.0
-; GFX11-NEXT:    s_mov_b32 s4, 0.5
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 7
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v5f32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v5f32_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 1
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 2
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 1.0
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_mov_b32 s5, 2.0
+; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX11-NEXT:    s_mov_b32 s6, 4.0
+; GFX11-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX11-NEXT:    s_mov_b32 s7, -1.0
+; GFX11-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX11-NEXT:    s_mov_b32 s8, 0.5
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 5
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 6
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 2
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 6
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 7
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -11222,23 +11440,31 @@ define amdgpu_gfx void @test_call_external_void_func_v5f32_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v5f32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v5f32_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 1.0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 2.0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 4.0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, -1.0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 0.5
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 7
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v5f32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v5f32_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 2
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 2
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 1.0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 2.0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s6, 4.0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s7, -1.0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s8, 0.5
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 5
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 6
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 6
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 5
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 7
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -11260,18 +11486,22 @@ define amdgpu_gfx void @test_call_external_void_func_f64_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 2
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_f64_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_f64_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 0
-; GFX9-NEXT:    s_mov_b32 s1, 0x40100000
+; GFX9-NEXT:    s_mov_b32 s4, 0
+; GFX9-NEXT:    s_mov_b32 s5, 0x40100000
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 3
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -11289,18 +11519,22 @@ define amdgpu_gfx void @test_call_external_void_func_f64_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 4
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_f64_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_f64_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 0
-; GFX10-NEXT:    s_mov_b32 s1, 0x40100000
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 0
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_mov_b32 s5, 0x40100000
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 3
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -11318,19 +11552,23 @@ define amdgpu_gfx void @test_call_external_void_func_f64_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_f64_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_f64_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 0
-; GFX11-NEXT:    s_mov_b32 s1, 0x40100000
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_f64_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_f64_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_mov_b32 s5, 0x40100000
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -11348,18 +11586,22 @@ define amdgpu_gfx void @test_call_external_void_func_f64_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_f64_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_f64_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 0x40100000
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_f64_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_f64_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 0x40100000
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -11381,20 +11623,28 @@ define amdgpu_gfx void @test_call_external_void_func_v2f64_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 6
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 4
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v2f64_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v2f64_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s2, 0
-; GFX9-NEXT:    s_mov_b32 s0, 0
-; GFX9-NEXT:    s_mov_b32 s1, 2.0
-; GFX9-NEXT:    s_mov_b32 s3, 0x40100000
+; GFX9-NEXT:    s_mov_b32 s4, 0
+; GFX9-NEXT:    s_mov_b32 s5, 2.0
+; GFX9-NEXT:    s_mov_b32 s6, 0
+; GFX9-NEXT:    s_mov_b32 s7, 0x40100000
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 5
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 6
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -11412,20 +11662,28 @@ define amdgpu_gfx void @test_call_external_void_func_v2f64_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 6
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v2f64_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v2f64_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s2, 0
-; GFX10-NEXT:    s_mov_b32 s0, 0
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    s_mov_b32 s1, 2.0
-; GFX10-NEXT:    s_mov_b32 s3, 0x40100000
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 0
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_mov_b32 s5, 2.0
+; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-NEXT:    s_mov_b32 s6, 0
+; GFX10-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-NEXT:    s_mov_b32 s7, 0x40100000
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 5
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 6
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -11443,21 +11701,29 @@ define amdgpu_gfx void @test_call_external_void_func_v2f64_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v2f64_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v2f64_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 0
-; GFX11-NEXT:    s_mov_b32 s1, 2.0
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    s_mov_b32 s2, 0
-; GFX11-NEXT:    s_mov_b32 s3, 0x40100000
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 6
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v2f64_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v2f64_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_mov_b32 s5, 2.0
+; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX11-NEXT:    s_mov_b32 s6, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX11-NEXT:    s_mov_b32 s7, 0x40100000
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 6
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -11475,20 +11741,28 @@ define amdgpu_gfx void @test_call_external_void_func_v2f64_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v2f64_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v2f64_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 2.0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, 0x40100000
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 6
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v2f64_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v2f64_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 2.0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s6, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s7, 0x40100000
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 6
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -11510,26 +11784,34 @@ define amdgpu_gfx void @test_call_external_void_func_v3f64_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 8
 ; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 6
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v3f64_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v3f64_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 0
-; GFX9-NEXT:    s_mov_b32 s1, 2.0
-; GFX9-NEXT:    s_mov_b32 s2, 0
-; GFX9-NEXT:    s_mov_b32 s3, 0x40100000
 ; GFX9-NEXT:    s_mov_b32 s4, 0
-; GFX9-NEXT:    s_mov_b32 s5, 0x40200000
+; GFX9-NEXT:    s_mov_b32 s5, 2.0
+; GFX9-NEXT:    s_mov_b32 s6, 0
+; GFX9-NEXT:    s_mov_b32 s7, 0x40100000
+; GFX9-NEXT:    s_mov_b32 s8, 0
+; GFX9-NEXT:    s_mov_b32 s9, 0x40200000
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 7
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 3
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 7
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 6
+; GFX9-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX9-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 8
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -11547,26 +11829,34 @@ define amdgpu_gfx void @test_call_external_void_func_v3f64_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 8
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v3f64_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v3f64_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 0
-; GFX10-NEXT:    s_mov_b32 s1, 2.0
+; GFX10-NEXT:    s_addk_i32 s32, 0x200
 ; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX10-NEXT:    s_mov_b32 s2, 0
-; GFX10-NEXT:    s_mov_b32 s3, 0x40100000
 ; GFX10-NEXT:    s_mov_b32 s4, 0
-; GFX10-NEXT:    s_addk_i32 s32, 0x200
 ; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX10-NEXT:    s_mov_b32 s5, 0x40200000
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 2
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-NEXT:    s_mov_b32 s5, 2.0
+; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-NEXT:    s_mov_b32 s6, 0
+; GFX10-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-NEXT:    s_mov_b32 s7, 0x40100000
+; GFX10-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX10-NEXT:    s_mov_b32 s8, 0
+; GFX10-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX10-NEXT:    s_mov_b32 s9, 0x40200000
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 6
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 7
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 3
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 7
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 6
+; GFX10-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX10-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 8
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -11584,27 +11874,35 @@ define amdgpu_gfx void @test_call_external_void_func_v3f64_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 4
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v3f64_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v3f64_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 0
-; GFX11-NEXT:    s_mov_b32 s1, 2.0
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 8
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v3f64_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v3f64_inreg@abs32@lo
+; GFX11-NEXT:    s_add_i32 s32, s32, 16
 ; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX11-NEXT:    s_mov_b32 s2, 0
-; GFX11-NEXT:    s_mov_b32 s3, 0x40100000
 ; GFX11-NEXT:    s_mov_b32 s4, 0
-; GFX11-NEXT:    s_add_i32 s32, s32, 16
 ; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX11-NEXT:    s_mov_b32 s5, 0x40200000
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 2
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 3
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    s_mov_b32 s5, 2.0
+; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX11-NEXT:    s_mov_b32 s6, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX11-NEXT:    s_mov_b32 s7, 0x40100000
+; GFX11-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX11-NEXT:    s_mov_b32 s8, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX11-NEXT:    s_mov_b32 s9, 0x40200000
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 6
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 7
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 3
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 7
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 6
+; GFX11-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 8
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -11622,26 +11920,34 @@ define amdgpu_gfx void @test_call_external_void_func_v3f64_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 4
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v3f64_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v3f64_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 2.0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 8
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v3f64_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v3f64_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, 0x40100000
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 0
-; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 0x40200000
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 2
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 3
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 3
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 2.0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s6, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s7, 0x40100000
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s8, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s9, 0x40200000
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 6
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 7
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 7
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 6
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 4
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 8
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -11663,17 +11969,19 @@ define amdgpu_gfx void @test_call_external_void_func_v2i16_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    s_load_dword s0, s[34:35], 0x0
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    s_load_dword s4, s[34:35], 0x0
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 1
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v2i16_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v2i16_inreg@abs32@lo
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 2
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 3
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -11691,17 +11999,19 @@ define amdgpu_gfx void @test_call_external_void_func_v2i16_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    s_load_dword s0, s[34:35], 0x0
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 3
+; GFX10-NEXT:    s_addk_i32 s32, 0x200
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_load_dword s4, s[34:35], 0x0
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v2i16_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v2i16_inreg@abs32@lo
-; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 2
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 3
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -11719,18 +12029,20 @@ define amdgpu_gfx void @test_call_external_void_func_v2i16_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_load_b32 s0, s[0:1], 0x0
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_v2i16_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_v2i16_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 3
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_load_b32 s4, s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v2i16_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v2i16_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 3
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -11748,17 +12060,19 @@ define amdgpu_gfx void @test_call_external_void_func_v2i16_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_load_dword s0, s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_v2i16_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_v2i16_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 3
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v2i16_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v2i16_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 3
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -11781,20 +12095,21 @@ define amdgpu_gfx void @test_call_external_void_func_v3i16_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX9-NEXT:    s_mov_b32 s37, external_void_func_v3i16_inreg@abs32@hi
-; GFX9-NEXT:    s_mov_b32 s36, external_void_func_v3i16_inreg@abs32@lo
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[34:35], 0x0
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v3i16_inreg@abs32@hi
+; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v3i16_inreg@abs32@lo
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_mov_b32 s1, s35
-; GFX9-NEXT:    s_mov_b32 s0, s34
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX9-NEXT:    s_swappc_b64 s[30:31], s[36:37]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -11812,20 +12127,21 @@ define amdgpu_gfx void @test_call_external_void_func_v3i16_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX10-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
-; GFX10-NEXT:    s_mov_b32 s37, external_void_func_v3i16_inreg@abs32@hi
-; GFX10-NEXT:    s_mov_b32 s36, external_void_func_v3i16_inreg@abs32@lo
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 4
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_mov_b32 s1, s35
-; GFX10-NEXT:    s_mov_b32 s0, s34
-; GFX10-NEXT:    s_swappc_b64 s[30:31], s[36:37]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_load_dwordx2 s[4:5], s[34:35], 0x0
+; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v3i16_inreg@abs32@hi
+; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v3i16_inreg@abs32@lo
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -11843,18 +12159,22 @@ define amdgpu_gfx void @test_call_external_void_func_v3i16_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_v3i16_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_v3i16_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 4
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_load_b64 s[4:5], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v3i16_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v3i16_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -11872,17 +12192,21 @@ define amdgpu_gfx void @test_call_external_void_func_v3i16_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_v3i16_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_v3i16_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 4
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v3i16_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v3i16_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -11905,20 +12229,21 @@ define amdgpu_gfx void @test_call_external_void_func_v3f16_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX9-NEXT:    s_mov_b32 s37, external_void_func_v3f16_inreg@abs32@hi
-; GFX9-NEXT:    s_mov_b32 s36, external_void_func_v3f16_inreg@abs32@lo
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[34:35], 0x0
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v3f16_inreg@abs32@hi
+; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v3f16_inreg@abs32@lo
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_mov_b32 s1, s35
-; GFX9-NEXT:    s_mov_b32 s0, s34
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX9-NEXT:    s_swappc_b64 s[30:31], s[36:37]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -11936,20 +12261,21 @@ define amdgpu_gfx void @test_call_external_void_func_v3f16_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX10-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
-; GFX10-NEXT:    s_mov_b32 s37, external_void_func_v3f16_inreg@abs32@hi
-; GFX10-NEXT:    s_mov_b32 s36, external_void_func_v3f16_inreg@abs32@lo
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 4
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_mov_b32 s1, s35
-; GFX10-NEXT:    s_mov_b32 s0, s34
-; GFX10-NEXT:    s_swappc_b64 s[30:31], s[36:37]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_load_dwordx2 s[4:5], s[34:35], 0x0
+; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v3f16_inreg@abs32@hi
+; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v3f16_inreg@abs32@lo
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -11967,18 +12293,22 @@ define amdgpu_gfx void @test_call_external_void_func_v3f16_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_v3f16_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_v3f16_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 4
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_load_b64 s[4:5], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v3f16_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v3f16_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -11996,17 +12326,21 @@ define amdgpu_gfx void @test_call_external_void_func_v3f16_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_v3f16_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_v3f16_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 4
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v3f16_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v3f16_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -12029,18 +12363,22 @@ define amdgpu_gfx void @test_call_external_void_func_v3i16_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 2
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v3i16_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v3i16_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 0x20001
-; GFX9-NEXT:    s_mov_b32 s1, 3
+; GFX9-NEXT:    s_mov_b32 s4, 0x20001
+; GFX9-NEXT:    s_mov_b32 s5, 3
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 3
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -12058,18 +12396,22 @@ define amdgpu_gfx void @test_call_external_void_func_v3i16_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 4
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v3i16_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v3i16_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 0x20001
-; GFX10-NEXT:    s_mov_b32 s1, 3
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 0x20001
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_mov_b32 s5, 3
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 3
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -12087,19 +12429,23 @@ define amdgpu_gfx void @test_call_external_void_func_v3i16_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_v3i16_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_v3i16_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 0x20001
-; GFX11-NEXT:    s_mov_b32 s1, 3
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v3i16_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v3i16_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 0x20001
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_mov_b32 s5, 3
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -12117,18 +12463,22 @@ define amdgpu_gfx void @test_call_external_void_func_v3i16_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_v3i16_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_v3i16_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 0x20001
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 3
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v3i16_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v3i16_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 0x20001
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 3
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -12150,18 +12500,22 @@ define amdgpu_gfx void @test_call_external_void_func_v3f16_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 2
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v3f16_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v3f16_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 0x40003c00
-; GFX9-NEXT:    s_movk_i32 s1, 0x4400
+; GFX9-NEXT:    s_mov_b32 s4, 0x40003c00
+; GFX9-NEXT:    s_movk_i32 s5, 0x4400
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 3
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -12179,18 +12533,22 @@ define amdgpu_gfx void @test_call_external_void_func_v3f16_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 4
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v3f16_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v3f16_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 0x40003c00
-; GFX10-NEXT:    s_movk_i32 s1, 0x4400
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 0x40003c00
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_movk_i32 s5, 0x4400
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 3
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -12208,19 +12566,23 @@ define amdgpu_gfx void @test_call_external_void_func_v3f16_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_v3f16_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_v3f16_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 0x40003c00
-; GFX11-NEXT:    s_movk_i32 s1, 0x4400
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v3f16_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v3f16_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 0x40003c00
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_movk_i32 s5, 0x4400
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -12238,18 +12600,22 @@ define amdgpu_gfx void @test_call_external_void_func_v3f16_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_v3f16_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_v3f16_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 0x40003c00
-; GFX10-SCRATCH-NEXT:    s_movk_i32 s1, 0x4400
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v3f16_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v3f16_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 0x40003c00
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_movk_i32 s5, 0x4400
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -12271,20 +12637,21 @@ define amdgpu_gfx void @test_call_external_void_func_v4i16_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX9-NEXT:    s_mov_b32 s37, external_void_func_v4i16_inreg@abs32@hi
-; GFX9-NEXT:    s_mov_b32 s36, external_void_func_v4i16_inreg@abs32@lo
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[34:35], 0x0
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v4i16_inreg@abs32@hi
+; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v4i16_inreg@abs32@lo
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_mov_b32 s0, s34
-; GFX9-NEXT:    s_mov_b32 s1, s35
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX9-NEXT:    s_swappc_b64 s[30:31], s[36:37]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -12302,20 +12669,21 @@ define amdgpu_gfx void @test_call_external_void_func_v4i16_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX10-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
-; GFX10-NEXT:    s_mov_b32 s37, external_void_func_v4i16_inreg@abs32@hi
-; GFX10-NEXT:    s_mov_b32 s36, external_void_func_v4i16_inreg@abs32@lo
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 4
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_mov_b32 s0, s34
-; GFX10-NEXT:    s_mov_b32 s1, s35
-; GFX10-NEXT:    s_swappc_b64 s[30:31], s[36:37]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_load_dwordx2 s[4:5], s[34:35], 0x0
+; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v4i16_inreg@abs32@hi
+; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v4i16_inreg@abs32@lo
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -12333,18 +12701,22 @@ define amdgpu_gfx void @test_call_external_void_func_v4i16_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_v4i16_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_v4i16_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 4
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_load_b64 s[4:5], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v4i16_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v4i16_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -12362,19 +12734,23 @@ define amdgpu_gfx void @test_call_external_void_func_v4i16_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_v4i16_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_v4i16_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 4
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
-; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v4i16_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v4i16_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 4
+; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
+; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, -16
@@ -12395,18 +12771,22 @@ define amdgpu_gfx void @test_call_external_void_func_v4i16_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 2
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v4i16_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v4i16_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 0x20001
-; GFX9-NEXT:    s_mov_b32 s1, 0x40003
+; GFX9-NEXT:    s_mov_b32 s4, 0x20001
+; GFX9-NEXT:    s_mov_b32 s5, 0x40003
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 3
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -12424,18 +12804,22 @@ define amdgpu_gfx void @test_call_external_void_func_v4i16_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 4
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v4i16_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v4i16_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 0x20001
-; GFX10-NEXT:    s_mov_b32 s1, 0x40003
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 0x20001
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_mov_b32 s5, 0x40003
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 3
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -12453,19 +12837,23 @@ define amdgpu_gfx void @test_call_external_void_func_v4i16_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_v4i16_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_v4i16_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 0x20001
-; GFX11-NEXT:    s_mov_b32 s1, 0x40003
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v4i16_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v4i16_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 0x20001
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_mov_b32 s5, 0x40003
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -12483,18 +12871,22 @@ define amdgpu_gfx void @test_call_external_void_func_v4i16_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_v4i16_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_v4i16_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 0x20001
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 0x40003
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v4i16_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v4i16_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 0x20001
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 0x40003
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -12516,17 +12908,19 @@ define amdgpu_gfx void @test_call_external_void_func_v2f16_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    s_load_dword s0, s[34:35], 0x0
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    s_load_dword s4, s[34:35], 0x0
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 1
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v2f16_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v2f16_inreg@abs32@lo
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 2
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 3
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -12544,17 +12938,19 @@ define amdgpu_gfx void @test_call_external_void_func_v2f16_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    s_load_dword s0, s[34:35], 0x0
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 3
+; GFX10-NEXT:    s_addk_i32 s32, 0x200
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_load_dword s4, s[34:35], 0x0
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v2f16_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v2f16_inreg@abs32@lo
-; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 2
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 3
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -12572,18 +12968,20 @@ define amdgpu_gfx void @test_call_external_void_func_v2f16_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_load_b32 s0, s[0:1], 0x0
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_v2f16_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_v2f16_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 3
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_load_b32 s4, s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v2f16_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v2f16_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 3
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -12601,17 +12999,19 @@ define amdgpu_gfx void @test_call_external_void_func_v2f16_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_load_dword s0, s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_v2f16_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_v2f16_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 3
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v2f16_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v2f16_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 3
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -12634,20 +13034,21 @@ define amdgpu_gfx void @test_call_external_void_func_v2i32_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX9-NEXT:    s_mov_b32 s37, external_void_func_v2i32_inreg@abs32@hi
-; GFX9-NEXT:    s_mov_b32 s36, external_void_func_v2i32_inreg@abs32@lo
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[34:35], 0x0
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v2i32_inreg@abs32@hi
+; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v2i32_inreg@abs32@lo
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_mov_b32 s0, s34
-; GFX9-NEXT:    s_mov_b32 s1, s35
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX9-NEXT:    s_swappc_b64 s[30:31], s[36:37]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -12665,20 +13066,21 @@ define amdgpu_gfx void @test_call_external_void_func_v2i32_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX10-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
-; GFX10-NEXT:    s_mov_b32 s37, external_void_func_v2i32_inreg@abs32@hi
-; GFX10-NEXT:    s_mov_b32 s36, external_void_func_v2i32_inreg@abs32@lo
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 4
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_mov_b32 s0, s34
-; GFX10-NEXT:    s_mov_b32 s1, s35
-; GFX10-NEXT:    s_swappc_b64 s[30:31], s[36:37]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_load_dwordx2 s[4:5], s[34:35], 0x0
+; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v2i32_inreg@abs32@hi
+; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v2i32_inreg@abs32@lo
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -12696,18 +13098,22 @@ define amdgpu_gfx void @test_call_external_void_func_v2i32_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_v2i32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_v2i32_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 4
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_load_b64 s[4:5], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v2i32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v2i32_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -12725,17 +13131,21 @@ define amdgpu_gfx void @test_call_external_void_func_v2i32_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_v2i32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_v2i32_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 4
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v2i32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v2i32_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -12758,18 +13168,22 @@ define amdgpu_gfx void @test_call_external_void_func_v2i32_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 2
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v2i32_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v2i32_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 1
-; GFX9-NEXT:    s_mov_b32 s1, 2
+; GFX9-NEXT:    s_mov_b32 s4, 1
+; GFX9-NEXT:    s_mov_b32 s5, 2
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 3
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -12787,18 +13201,22 @@ define amdgpu_gfx void @test_call_external_void_func_v2i32_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 4
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v2i32_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v2i32_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 1
-; GFX10-NEXT:    s_mov_b32 s1, 2
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_mov_b32 s5, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 3
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 4
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -12816,19 +13234,23 @@ define amdgpu_gfx void @test_call_external_void_func_v2i32_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_v2i32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_v2i32_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 1
-; GFX11-NEXT:    s_mov_b32 s1, 2
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v2i32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v2i32_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_mov_b32 s5, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -12846,18 +13268,22 @@ define amdgpu_gfx void @test_call_external_void_func_v2i32_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_v2i32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_v2i32_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 1
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 2
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 4
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v2i32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v2i32_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 3
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 4
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -12879,19 +13305,25 @@ define amdgpu_gfx void @test_call_external_void_func_v3i32_imm_inreg(i32) #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 5
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 3
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v3i32_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v3i32_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 3
-; GFX9-NEXT:    s_mov_b32 s1, 4
-; GFX9-NEXT:    s_mov_b32 s2, 5
+; GFX9-NEXT:    s_mov_b32 s4, 3
+; GFX9-NEXT:    s_mov_b32 s5, 4
+; GFX9-NEXT:    s_mov_b32 s6, 5
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 4
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 5
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -12909,19 +13341,25 @@ define amdgpu_gfx void @test_call_external_void_func_v3i32_imm_inreg(i32) #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 5
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v3i32_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v3i32_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 3
-; GFX10-NEXT:    s_mov_b32 s1, 4
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    s_mov_b32 s2, 5
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 3
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_mov_b32 s5, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-NEXT:    s_mov_b32 s6, 5
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 3
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 4
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 5
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -12939,20 +13377,26 @@ define amdgpu_gfx void @test_call_external_void_func_v3i32_imm_inreg(i32) #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v3i32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v3i32_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 3
-; GFX11-NEXT:    s_mov_b32 s1, 4
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    s_mov_b32 s2, 5
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 5
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v3i32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v3i32_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 3
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_mov_b32 s5, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX11-NEXT:    s_mov_b32 s6, 5
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 3
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 4
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 5
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -12970,19 +13414,25 @@ define amdgpu_gfx void @test_call_external_void_func_v3i32_imm_inreg(i32) #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v3i32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v3i32_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 4
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 5
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 5
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v3i32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v3i32_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 3
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s6, 5
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 3
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 4
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 4
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 5
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -13004,20 +13454,28 @@ define amdgpu_gfx void @test_call_external_void_func_v3i32_i32_inreg(i32) #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 6
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 4
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v3i32_i32_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v3i32_i32_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 3
-; GFX9-NEXT:    s_mov_b32 s1, 4
-; GFX9-NEXT:    s_mov_b32 s2, 5
-; GFX9-NEXT:    s_mov_b32 s3, 6
+; GFX9-NEXT:    s_mov_b32 s4, 3
+; GFX9-NEXT:    s_mov_b32 s5, 4
+; GFX9-NEXT:    s_mov_b32 s6, 5
+; GFX9-NEXT:    s_mov_b32 s7, 6
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 5
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 6
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -13035,20 +13493,28 @@ define amdgpu_gfx void @test_call_external_void_func_v3i32_i32_inreg(i32) #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 6
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v3i32_i32_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v3i32_i32_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 3
-; GFX10-NEXT:    s_mov_b32 s1, 4
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    s_mov_b32 s2, 5
-; GFX10-NEXT:    s_mov_b32 s3, 6
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 3
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_mov_b32 s5, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-NEXT:    s_mov_b32 s6, 5
+; GFX10-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-NEXT:    s_mov_b32 s7, 6
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 5
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 6
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -13066,21 +13532,29 @@ define amdgpu_gfx void @test_call_external_void_func_v3i32_i32_inreg(i32) #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v3i32_i32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v3i32_i32_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 3
-; GFX11-NEXT:    s_mov_b32 s1, 4
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    s_mov_b32 s2, 5
-; GFX11-NEXT:    s_mov_b32 s3, 6
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 6
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v3i32_i32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v3i32_i32_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 3
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_mov_b32 s5, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX11-NEXT:    s_mov_b32 s6, 5
+; GFX11-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX11-NEXT:    s_mov_b32 s7, 6
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 6
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -13098,20 +13572,28 @@ define amdgpu_gfx void @test_call_external_void_func_v3i32_i32_inreg(i32) #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v3i32_i32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v3i32_i32_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 4
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 5
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, 6
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 6
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v3i32_i32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v3i32_i32_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 3
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s6, 5
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s7, 6
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 6
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -13133,22 +13615,25 @@ define amdgpu_gfx void @test_call_external_void_func_v4i32_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    s_load_dwordx4 s[36:39], s[34:35], 0x0
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 6
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[34:35], 0x0
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 4
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v4i32_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v4i32_inreg@abs32@lo
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_mov_b32 s0, s36
-; GFX9-NEXT:    s_mov_b32 s1, s37
-; GFX9-NEXT:    s_mov_b32 s2, s38
-; GFX9-NEXT:    s_mov_b32 s3, s39
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 5
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 6
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -13166,22 +13651,25 @@ define amdgpu_gfx void @test_call_external_void_func_v4i32_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    s_load_dwordx4 s[36:39], s[34:35], 0x0
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 6
+; GFX10-NEXT:    s_addk_i32 s32, 0x200
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-NEXT:    s_load_dwordx4 s[4:7], s[34:35], 0x0
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v4i32_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v4i32_inreg@abs32@lo
-; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_mov_b32 s0, s36
-; GFX10-NEXT:    s_mov_b32 s1, s37
-; GFX10-NEXT:    s_mov_b32 s2, s38
-; GFX10-NEXT:    s_mov_b32 s3, s39
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 5
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 6
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -13199,18 +13687,26 @@ define amdgpu_gfx void @test_call_external_void_func_v4i32_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v4i32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v4i32_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 6
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX11-NEXT:    s_load_b128 s[4:7], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v4i32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v4i32_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 6
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -13228,17 +13724,25 @@ define amdgpu_gfx void @test_call_external_void_func_v4i32_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v4i32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v4i32_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 6
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-SCRATCH-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v4i32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v4i32_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 6
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -13261,20 +13765,28 @@ define amdgpu_gfx void @test_call_external_void_func_v4i32_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 6
+; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 4
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v4i32_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v4i32_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 1
-; GFX9-NEXT:    s_mov_b32 s1, 2
-; GFX9-NEXT:    s_mov_b32 s2, 3
-; GFX9-NEXT:    s_mov_b32 s3, 4
+; GFX9-NEXT:    s_mov_b32 s4, 1
+; GFX9-NEXT:    s_mov_b32 s5, 2
+; GFX9-NEXT:    s_mov_b32 s6, 3
+; GFX9-NEXT:    s_mov_b32 s7, 4
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 5
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 6
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -13292,20 +13804,28 @@ define amdgpu_gfx void @test_call_external_void_func_v4i32_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 6
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v4i32_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v4i32_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 1
-; GFX10-NEXT:    s_mov_b32 s1, 2
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    s_mov_b32 s2, 3
-; GFX10-NEXT:    s_mov_b32 s3, 4
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_mov_b32 s5, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-NEXT:    s_mov_b32 s6, 3
+; GFX10-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-NEXT:    s_mov_b32 s7, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 5
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 6
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -13323,21 +13843,29 @@ define amdgpu_gfx void @test_call_external_void_func_v4i32_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v4i32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v4i32_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 1
-; GFX11-NEXT:    s_mov_b32 s1, 2
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-NEXT:    s_mov_b32 s2, 3
-; GFX11-NEXT:    s_mov_b32 s3, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 6
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v4i32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v4i32_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_mov_b32 s5, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX11-NEXT:    s_mov_b32 s6, 3
+; GFX11-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX11-NEXT:    s_mov_b32 s7, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 6
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -13355,20 +13883,28 @@ define amdgpu_gfx void @test_call_external_void_func_v4i32_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v4i32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v4i32_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 1
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 2
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 6
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v4i32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v4i32_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s6, 3
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s7, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 5
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 6
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -13390,23 +13926,31 @@ define amdgpu_gfx void @test_call_external_void_func_v5i32_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 7
 ; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 5
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v5i32_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v5i32_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 1
-; GFX9-NEXT:    s_mov_b32 s1, 2
-; GFX9-NEXT:    s_mov_b32 s2, 3
-; GFX9-NEXT:    s_mov_b32 s3, 4
-; GFX9-NEXT:    s_mov_b32 s4, 5
+; GFX9-NEXT:    s_mov_b32 s4, 1
+; GFX9-NEXT:    s_mov_b32 s5, 2
+; GFX9-NEXT:    s_mov_b32 s6, 3
+; GFX9-NEXT:    s_mov_b32 s7, 4
+; GFX9-NEXT:    s_mov_b32 s8, 5
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 6
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 2
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 6
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 5
+; GFX9-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 3
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 7
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -13424,23 +13968,31 @@ define amdgpu_gfx void @test_call_external_void_func_v5i32_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 3
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 7
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v5i32_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v5i32_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 1
-; GFX10-NEXT:    s_mov_b32 s1, 2
-; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX10-NEXT:    s_mov_b32 s2, 3
-; GFX10-NEXT:    s_mov_b32 s3, 4
-; GFX10-NEXT:    s_mov_b32 s4, 5
-; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 1
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 2
+; GFX10-NEXT:    s_addk_i32 s32, 0x200
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 1
+; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-NEXT:    s_mov_b32 s5, 2
+; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-NEXT:    s_mov_b32 s6, 3
+; GFX10-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-NEXT:    s_mov_b32 s7, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX10-NEXT:    s_mov_b32 s8, 5
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 5
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 6
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 2
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 6
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 5
+; GFX10-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 3
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 7
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -13458,24 +14010,32 @@ define amdgpu_gfx void @test_call_external_void_func_v5i32_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 3
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v5i32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v5i32_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 1
-; GFX11-NEXT:    s_mov_b32 s1, 2
-; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX11-NEXT:    s_mov_b32 s2, 3
-; GFX11-NEXT:    s_mov_b32 s3, 4
-; GFX11-NEXT:    s_mov_b32 s4, 5
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 7
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v5i32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v5i32_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 1
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 2
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX11-NEXT:    s_mov_b32 s5, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX11-NEXT:    s_mov_b32 s6, 3
+; GFX11-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX11-NEXT:    s_mov_b32 s7, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX11-NEXT:    s_mov_b32 s8, 5
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 5
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 6
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 2
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 6
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 7
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -13493,23 +14053,31 @@ define amdgpu_gfx void @test_call_external_void_func_v5i32_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v5i32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v5i32_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 1
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 2
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, 4
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 5
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 7
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v5i32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v5i32_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 2
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 2
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 2
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s6, 3
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s7, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s8, 5
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 5
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 6
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 6
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 5
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s8, v40, 4
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s7, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 3
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 7
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -13531,36 +14099,35 @@ define amdgpu_gfx void @test_call_external_void_func_v8i32_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 6
-; GFX9-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 10
 ; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
 ; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
+; GFX9-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
 ; GFX9-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX9-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX9-NEXT:    v_writelane_b32 v40, s10, 6
+; GFX9-NEXT:    v_writelane_b32 v40, s11, 7
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_load_dwordx8 s[36:43], s[34:35], 0x0
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX9-NEXT:    s_load_dwordx8 s[4:11], s[34:35], 0x0
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 8
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v8i32_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v8i32_inreg@abs32@lo
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_mov_b32 s0, s36
-; GFX9-NEXT:    s_mov_b32 s1, s37
-; GFX9-NEXT:    s_mov_b32 s2, s38
-; GFX9-NEXT:    s_mov_b32 s3, s39
-; GFX9-NEXT:    s_mov_b32 s4, s40
-; GFX9-NEXT:    s_mov_b32 s5, s41
-; GFX9-NEXT:    s_mov_b32 s6, s42
-; GFX9-NEXT:    s_mov_b32 s7, s43
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 9
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 5
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 9
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 8
+; GFX9-NEXT:    v_readlane_b32 s11, v40, 7
+; GFX9-NEXT:    v_readlane_b32 s10, v40, 6
+; GFX9-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX9-NEXT:    v_readlane_b32 s8, v40, 4
 ; GFX9-NEXT:    v_readlane_b32 s7, v40, 3
 ; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 6
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 10
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -13578,36 +14145,35 @@ define amdgpu_gfx void @test_call_external_void_func_v8i32_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 6
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 10
 ; GFX10-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
 ; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
 ; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
 ; GFX10-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX10-NEXT:    v_writelane_b32 v40, s10, 6
+; GFX10-NEXT:    v_writelane_b32 v40, s11, 7
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_load_dwordx8 s[36:43], s[34:35], 0x0
+; GFX10-NEXT:    s_load_dwordx8 s[4:11], s[34:35], 0x0
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v8i32_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v8i32_inreg@abs32@lo
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 4
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 5
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_mov_b32 s0, s36
-; GFX10-NEXT:    s_mov_b32 s1, s37
-; GFX10-NEXT:    s_mov_b32 s2, s38
-; GFX10-NEXT:    s_mov_b32 s3, s39
-; GFX10-NEXT:    s_mov_b32 s4, s40
-; GFX10-NEXT:    s_mov_b32 s5, s41
-; GFX10-NEXT:    s_mov_b32 s6, s42
-; GFX10-NEXT:    s_mov_b32 s7, s43
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 8
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 9
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 5
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 9
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 8
+; GFX10-NEXT:    v_readlane_b32 s11, v40, 7
+; GFX10-NEXT:    v_readlane_b32 s10, v40, 6
+; GFX10-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX10-NEXT:    v_readlane_b32 s8, v40, 4
 ; GFX10-NEXT:    v_readlane_b32 s7, v40, 3
 ; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 6
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 10
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -13625,28 +14191,36 @@ define amdgpu_gfx void @test_call_external_void_func_v8i32_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 6
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 10
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v8i32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v8i32_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
 ; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
 ; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
 ; GFX11-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX11-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX11-NEXT:    v_writelane_b32 v40, s10, 6
+; GFX11-NEXT:    v_writelane_b32 v40, s11, 7
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_load_b256 s[0:7], s[0:1], 0x0
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 4
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 5
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    s_load_b256 s[4:11], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v8i32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v8i32_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 8
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 9
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 5
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 9
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 8
+; GFX11-NEXT:    v_readlane_b32 s11, v40, 7
+; GFX11-NEXT:    v_readlane_b32 s10, v40, 6
+; GFX11-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s8, v40, 4
 ; GFX11-NEXT:    v_readlane_b32 s7, v40, 3
 ; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 6
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 10
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -13664,27 +14238,35 @@ define amdgpu_gfx void @test_call_external_void_func_v8i32_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 6
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 10
 ; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v8i32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v8i32_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s7, 3
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s10, 6
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s11, 7
 ; GFX10-SCRATCH-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-SCRATCH-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 4
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 5
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 5
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-SCRATCH-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v8i32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v8i32_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 8
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 9
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 9
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 8
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s11, v40, 7
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s10, v40, 6
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s8, v40, 4
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s7, v40, 3
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 6
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 10
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -13708,32 +14290,40 @@ define amdgpu_gfx void @test_call_external_void_func_v8i32_imm_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 6
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 10
 ; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
 ; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
 ; GFX9-NEXT:    v_writelane_b32 v40, s7, 3
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX9-NEXT:    v_writelane_b32 v40, s10, 6
+; GFX9-NEXT:    v_writelane_b32 v40, s11, 7
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 8
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v8i32_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v8i32_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, 1
-; GFX9-NEXT:    s_mov_b32 s1, 2
-; GFX9-NEXT:    s_mov_b32 s2, 3
-; GFX9-NEXT:    s_mov_b32 s3, 4
-; GFX9-NEXT:    s_mov_b32 s4, 5
-; GFX9-NEXT:    s_mov_b32 s5, 6
-; GFX9-NEXT:    s_mov_b32 s6, 7
-; GFX9-NEXT:    s_mov_b32 s7, 8
+; GFX9-NEXT:    s_mov_b32 s4, 1
+; GFX9-NEXT:    s_mov_b32 s5, 2
+; GFX9-NEXT:    s_mov_b32 s6, 3
+; GFX9-NEXT:    s_mov_b32 s7, 4
+; GFX9-NEXT:    s_mov_b32 s8, 5
+; GFX9-NEXT:    s_mov_b32 s9, 6
+; GFX9-NEXT:    s_mov_b32 s10, 7
+; GFX9-NEXT:    s_mov_b32 s11, 8
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 9
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 5
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 9
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 8
+; GFX9-NEXT:    v_readlane_b32 s11, v40, 7
+; GFX9-NEXT:    v_readlane_b32 s10, v40, 6
+; GFX9-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX9-NEXT:    v_readlane_b32 s8, v40, 4
 ; GFX9-NEXT:    v_readlane_b32 s7, v40, 3
 ; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 6
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 10
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -13751,32 +14341,40 @@ define amdgpu_gfx void @test_call_external_void_func_v8i32_imm_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 6
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 10
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v8i32_inreg@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v8i32_inreg@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, 1
-; GFX10-NEXT:    s_mov_b32 s1, 2
-; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX10-NEXT:    s_mov_b32 s2, 3
-; GFX10-NEXT:    s_mov_b32 s3, 4
-; GFX10-NEXT:    s_mov_b32 s4, 5
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
+; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-NEXT:    s_mov_b32 s4, 1
 ; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX10-NEXT:    s_mov_b32 s5, 6
+; GFX10-NEXT:    s_mov_b32 s5, 2
 ; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
-; GFX10-NEXT:    s_mov_b32 s6, 7
+; GFX10-NEXT:    s_mov_b32 s6, 3
 ; GFX10-NEXT:    v_writelane_b32 v40, s7, 3
-; GFX10-NEXT:    s_mov_b32 s7, 8
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 4
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX10-NEXT:    s_mov_b32 s7, 4
+; GFX10-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX10-NEXT:    s_mov_b32 s8, 5
+; GFX10-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX10-NEXT:    s_mov_b32 s9, 6
+; GFX10-NEXT:    v_writelane_b32 v40, s10, 6
+; GFX10-NEXT:    s_mov_b32 s10, 7
+; GFX10-NEXT:    v_writelane_b32 v40, s11, 7
+; GFX10-NEXT:    s_mov_b32 s11, 8
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 8
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 9
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 5
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 9
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 8
+; GFX10-NEXT:    v_readlane_b32 s11, v40, 7
+; GFX10-NEXT:    v_readlane_b32 s10, v40, 6
+; GFX10-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX10-NEXT:    v_readlane_b32 s8, v40, 4
 ; GFX10-NEXT:    v_readlane_b32 s7, v40, 3
 ; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 6
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 10
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -13794,33 +14392,41 @@ define amdgpu_gfx void @test_call_external_void_func_v8i32_imm_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 6
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v8i32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v8i32_inreg@abs32@lo
-; GFX11-NEXT:    s_mov_b32 s0, 1
-; GFX11-NEXT:    s_mov_b32 s1, 2
-; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX11-NEXT:    s_mov_b32 s2, 3
-; GFX11-NEXT:    s_mov_b32 s3, 4
-; GFX11-NEXT:    s_mov_b32 s4, 5
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 10
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v8i32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v8i32_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
+; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX11-NEXT:    s_mov_b32 s4, 1
 ; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX11-NEXT:    s_mov_b32 s5, 6
+; GFX11-NEXT:    s_mov_b32 s5, 2
 ; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
-; GFX11-NEXT:    s_mov_b32 s6, 7
+; GFX11-NEXT:    s_mov_b32 s6, 3
 ; GFX11-NEXT:    v_writelane_b32 v40, s7, 3
-; GFX11-NEXT:    s_mov_b32 s7, 8
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 4
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 5
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    s_mov_b32 s7, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX11-NEXT:    s_mov_b32 s8, 5
+; GFX11-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX11-NEXT:    s_mov_b32 s9, 6
+; GFX11-NEXT:    v_writelane_b32 v40, s10, 6
+; GFX11-NEXT:    s_mov_b32 s10, 7
+; GFX11-NEXT:    v_writelane_b32 v40, s11, 7
+; GFX11-NEXT:    s_mov_b32 s11, 8
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 8
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 9
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 5
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 9
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 8
+; GFX11-NEXT:    v_readlane_b32 s11, v40, 7
+; GFX11-NEXT:    v_readlane_b32 s10, v40, 6
+; GFX11-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s8, v40, 4
 ; GFX11-NEXT:    v_readlane_b32 s7, v40, 3
 ; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 6
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 10
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -13838,32 +14444,40 @@ define amdgpu_gfx void @test_call_external_void_func_v8i32_imm_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 6
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v8i32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v8i32_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, 1
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, 2
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, 3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, 4
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 5
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 10
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v8i32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v8i32_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s4, 1
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 6
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s5, 2
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s6, 7
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s6, 3
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s7, 3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s7, 8
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 4
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 5
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 5
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 4
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s7, 4
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s8, 4
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s8, 5
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s9, 5
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s9, 6
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s10, 6
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s10, 7
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s11, 7
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s11, 8
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 8
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 9
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 9
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 8
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s11, v40, 7
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s10, v40, 6
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s9, v40, 5
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s8, v40, 4
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s7, v40, 3
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 6
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 10
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -13885,47 +14499,38 @@ define amdgpu_gfx void @test_call_external_void_func_v16i32_inreg() #0 {
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 14
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 18
 ; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
 ; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
-; GFX9-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
 ; GFX9-NEXT:    v_writelane_b32 v40, s7, 3
 ; GFX9-NEXT:    v_writelane_b32 v40, s8, 4
 ; GFX9-NEXT:    v_writelane_b32 v40, s9, 5
 ; GFX9-NEXT:    v_writelane_b32 v40, s10, 6
 ; GFX9-NEXT:    v_writelane_b32 v40, s11, 7
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_load_dwordx16 s[36:51], s[34:35], 0x0
 ; GFX9-NEXT:    v_writelane_b32 v40, s12, 8
 ; GFX9-NEXT:    v_writelane_b32 v40, s13, 9
 ; GFX9-NEXT:    v_writelane_b32 v40, s14, 10
+; GFX9-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
 ; GFX9-NEXT:    v_writelane_b32 v40, s15, 11
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 12
+; GFX9-NEXT:    v_writelane_b32 v40, s16, 12
+; GFX9-NEXT:    v_writelane_b32 v40, s17, 13
+; GFX9-NEXT:    v_writelane_b32 v40, s18, 14
+; GFX9-NEXT:    v_writelane_b32 v40, s19, 15
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx16 s[4:19], s[34:35], 0x0
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 16
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v16i32_inreg@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v16i32_inreg@abs32@lo
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_mov_b32 s0, s36
-; GFX9-NEXT:    s_mov_b32 s1, s37
-; GFX9-NEXT:    s_mov_b32 s2, s38
-; GFX9-NEXT:    s_mov_b32 s3, s39
-; GFX9-NEXT:    s_mov_b32 s4, s40
-; GFX9-NEXT:    s_mov_b32 s5, s41
-; GFX9-NEXT:    s_mov_b32 s6, s42
-; GFX9-NEXT:    s_mov_b32 s7, s43
-; GFX9-NEXT:    s_mov_b32 s8, s44
-; GFX9-NEXT:    s_mov_b32 s9, s45
-; GFX9-NEXT:    s_mov_b32 s10, s46
-; GFX9-NEXT:    s_mov_b32 s11, s47
-; GFX9-NEXT:    s_mov_b32 s12, s48
-; GFX9-NEXT:    s_mov_b32 s13, s49
-; GFX9-NEXT:    s_mov_b32 s14, s50
-; GFX9-NEXT:    s_mov_b32 s15, s51
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 13
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 17
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 13
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 12
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 17
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 16
+; GFX9-NEXT:    v_readlane_b32 s19, v40, 15
+; GFX9-NEXT:    v_readlane_b32 s18, v40, 14
+; GFX9-NEXT:    v_readlane_b32 s17, v40, 13
+; GFX9-NEXT:    v_readlane_b32 s16, v40, 12
 ; GFX9-NEXT:    v_readlane_b32 s15, v40, 11
 ; GFX9-NEXT:    v_readlane_b32 s14, v40, 10
 ; GFX9-NEXT:    v_readlane_b32 s13, v40, 9
@@ -13938,7 +14543,7 @@ define amdgpu_gfx void @test_call_external_void_func_v16i32_inreg() #0 {
 ; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 14
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 18
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -13956,47 +14561,38 @@ define amdgpu_gfx void @test_call_external_void_func_v16i32_inreg() #0 {
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 14
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 18
 ; GFX10-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
 ; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
 ; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
 ; GFX10-NEXT:    v_writelane_b32 v40, s7, 3
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_load_dwordx16 s[36:51], s[34:35], 0x0
-; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v16i32_inreg@abs32@hi
-; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v16i32_inreg@abs32@lo
 ; GFX10-NEXT:    v_writelane_b32 v40, s8, 4
 ; GFX10-NEXT:    v_writelane_b32 v40, s9, 5
 ; GFX10-NEXT:    v_writelane_b32 v40, s10, 6
 ; GFX10-NEXT:    v_writelane_b32 v40, s11, 7
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_mov_b32 s0, s36
-; GFX10-NEXT:    s_mov_b32 s1, s37
-; GFX10-NEXT:    s_mov_b32 s2, s38
-; GFX10-NEXT:    s_mov_b32 s3, s39
 ; GFX10-NEXT:    v_writelane_b32 v40, s12, 8
-; GFX10-NEXT:    s_mov_b32 s4, s40
-; GFX10-NEXT:    s_mov_b32 s5, s41
-; GFX10-NEXT:    s_mov_b32 s6, s42
-; GFX10-NEXT:    s_mov_b32 s7, s43
 ; GFX10-NEXT:    v_writelane_b32 v40, s13, 9
-; GFX10-NEXT:    s_mov_b32 s8, s44
-; GFX10-NEXT:    s_mov_b32 s9, s45
-; GFX10-NEXT:    s_mov_b32 s10, s46
-; GFX10-NEXT:    s_mov_b32 s11, s47
 ; GFX10-NEXT:    v_writelane_b32 v40, s14, 10
-; GFX10-NEXT:    s_mov_b32 s12, s48
-; GFX10-NEXT:    s_mov_b32 s13, s49
-; GFX10-NEXT:    s_mov_b32 s14, s50
 ; GFX10-NEXT:    v_writelane_b32 v40, s15, 11
-; GFX10-NEXT:    s_mov_b32 s15, s51
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 12
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 13
+; GFX10-NEXT:    v_writelane_b32 v40, s16, 12
+; GFX10-NEXT:    v_writelane_b32 v40, s17, 13
+; GFX10-NEXT:    v_writelane_b32 v40, s18, 14
+; GFX10-NEXT:    v_writelane_b32 v40, s19, 15
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_load_dwordx16 s[4:19], s[34:35], 0x0
+; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v16i32_inreg@abs32@hi
+; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v16i32_inreg@abs32@lo
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 16
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 17
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 13
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 12
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 17
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 16
+; GFX10-NEXT:    v_readlane_b32 s19, v40, 15
+; GFX10-NEXT:    v_readlane_b32 s18, v40, 14
+; GFX10-NEXT:    v_readlane_b32 s17, v40, 13
+; GFX10-NEXT:    v_readlane_b32 s16, v40, 12
 ; GFX10-NEXT:    v_readlane_b32 s15, v40, 11
 ; GFX10-NEXT:    v_readlane_b32 s14, v40, 10
 ; GFX10-NEXT:    v_readlane_b32 s13, v40, 9
@@ -14009,7 +14605,7 @@ define amdgpu_gfx void @test_call_external_void_func_v16i32_inreg() #0 {
 ; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 14
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 18
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -14027,10 +14623,8 @@ define amdgpu_gfx void @test_call_external_void_func_v16i32_inreg() #0 {
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-NEXT:    v_writelane_b32 v40, s0, 14
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 18
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v16i32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v16i32_inreg@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
 ; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
@@ -14044,14 +14638,24 @@ define amdgpu_gfx void @test_call_external_void_func_v16i32_inreg() #0 {
 ; GFX11-NEXT:    v_writelane_b32 v40, s13, 9
 ; GFX11-NEXT:    v_writelane_b32 v40, s14, 10
 ; GFX11-NEXT:    v_writelane_b32 v40, s15, 11
+; GFX11-NEXT:    v_writelane_b32 v40, s16, 12
+; GFX11-NEXT:    v_writelane_b32 v40, s17, 13
+; GFX11-NEXT:    v_writelane_b32 v40, s18, 14
+; GFX11-NEXT:    v_writelane_b32 v40, s19, 15
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_load_b512 s[0:15], s[0:1], 0x0
-; GFX11-NEXT:    v_writelane_b32 v40, s30, 12
-; GFX11-NEXT:    v_writelane_b32 v40, s31, 13
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    s_load_b512 s[4:19], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v16i32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v16i32_inreg@abs32@lo
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 16
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 17
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_readlane_b32 s31, v40, 13
-; GFX11-NEXT:    v_readlane_b32 s30, v40, 12
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 17
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 16
+; GFX11-NEXT:    v_readlane_b32 s19, v40, 15
+; GFX11-NEXT:    v_readlane_b32 s18, v40, 14
+; GFX11-NEXT:    v_readlane_b32 s17, v40, 13
+; GFX11-NEXT:    v_readlane_b32 s16, v40, 12
 ; GFX11-NEXT:    v_readlane_b32 s15, v40, 11
 ; GFX11-NEXT:    v_readlane_b32 s14, v40, 10
 ; GFX11-NEXT:    v_readlane_b32 s13, v40, 9
@@ -14064,7 +14668,7 @@ define amdgpu_gfx void @test_call_external_void_func_v16i32_inreg() #0 {
 ; GFX11-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX11-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX11-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX11-NEXT:    v_readlane_b32 s0, v40, 14
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 18
 ; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
 ; GFX11-NEXT:    s_mov_b32 exec_lo, s1
@@ -14082,10 +14686,8 @@ define amdgpu_gfx void @test_call_external_void_func_v16i32_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 14
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 18
 ; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v16i32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v16i32_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
@@ -14099,13 +14701,23 @@ define amdgpu_gfx void @test_call_external_void_func_v16i32_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s13, 9
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s14, 10
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s15, 11
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s16, 12
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s17, 13
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s18, 14
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s19, 15
 ; GFX10-SCRATCH-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-SCRATCH-NEXT:    s_load_dwordx16 s[0:15], s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 12
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 13
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 13
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 12
+; GFX10-SCRATCH-NEXT:    s_load_dwordx16 s[4:19], s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v16i32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v16i32_inreg@abs32@lo
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 16
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 17
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 17
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 16
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s19, v40, 15
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s18, v40, 14
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s17, v40, 13
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s16, v40, 12
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s15, v40, 11
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s14, v40, 10
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s13, v40, 9
@@ -14118,7 +14730,7 @@ define amdgpu_gfx void @test_call_external_void_func_v16i32_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s6, v40, 2
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s5, v40, 1
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 14
+; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 18
 ; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
@@ -14159,47 +14771,49 @@ define amdgpu_gfx void @test_call_external_void_func_v32i32_inreg() #0 {
 ; GFX9-NEXT:    v_writelane_b32 v40, s17, 13
 ; GFX9-NEXT:    v_writelane_b32 v40, s18, 14
 ; GFX9-NEXT:    v_writelane_b32 v40, s19, 15
+; GFX9-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
 ; GFX9-NEXT:    v_writelane_b32 v40, s20, 16
 ; GFX9-NEXT:    v_writelane_b32 v40, s21, 17
 ; GFX9-NEXT:    v_writelane_b32 v40, s22, 18
 ; GFX9-NEXT:    v_writelane_b32 v40, s23, 19
 ; GFX9-NEXT:    v_writelane_b32 v40, s24, 20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx16 s[36:51], s[34:35], 0x40
+; GFX9-NEXT:    s_load_dwordx16 s[4:19], s[34:35], 0x0
 ; GFX9-NEXT:    v_writelane_b32 v40, s25, 21
 ; GFX9-NEXT:    v_writelane_b32 v40, s26, 22
-; GFX9-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
 ; GFX9-NEXT:    v_writelane_b32 v40, s27, 23
-; GFX9-NEXT:    v_writelane_b32 v40, s28, 24
-; GFX9-NEXT:    v_writelane_b32 v40, s29, 25
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 26
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 27
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_load_dwordx16 s[16:31], s[34:35], 0x40
-; GFX9-NEXT:    s_load_dwordx16 s[36:51], s[34:35], 0x0
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    s_mov_b32 s53, external_void_func_v32i32_inreg@abs32@hi
-; GFX9-NEXT:    s_mov_b32 s52, external_void_func_v32i32_inreg@abs32@lo
+; GFX9-NEXT:    v_writelane_b32 v40, s28, 24
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s30
-; GFX9-NEXT:    v_mov_b32_e32 v1, s31
+; GFX9-NEXT:    v_mov_b32_e32 v0, s46
+; GFX9-NEXT:    v_writelane_b32 v40, s29, 25
+; GFX9-NEXT:    v_mov_b32_e32 v1, s47
+; GFX9-NEXT:    v_mov_b32_e32 v2, s48
+; GFX9-NEXT:    v_mov_b32_e32 v3, s49
 ; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32
 ; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:4
-; GFX9-NEXT:    s_mov_b32 s0, s36
-; GFX9-NEXT:    s_mov_b32 s1, s37
-; GFX9-NEXT:    s_mov_b32 s2, s38
-; GFX9-NEXT:    s_mov_b32 s3, s39
-; GFX9-NEXT:    s_mov_b32 s4, s40
-; GFX9-NEXT:    s_mov_b32 s5, s41
-; GFX9-NEXT:    s_mov_b32 s6, s42
-; GFX9-NEXT:    s_mov_b32 s7, s43
-; GFX9-NEXT:    s_mov_b32 s8, s44
-; GFX9-NEXT:    s_mov_b32 s9, s45
-; GFX9-NEXT:    s_mov_b32 s10, s46
-; GFX9-NEXT:    s_mov_b32 s11, s47
-; GFX9-NEXT:    s_mov_b32 s12, s48
-; GFX9-NEXT:    s_mov_b32 s13, s49
-; GFX9-NEXT:    s_mov_b32 s14, s50
-; GFX9-NEXT:    s_mov_b32 s15, s51
-; GFX9-NEXT:    s_swappc_b64 s[30:31], s[52:53]
+; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    v_mov_b32_e32 v0, s50
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 26
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:16
+; GFX9-NEXT:    v_mov_b32_e32 v0, s51
+; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v32i32_inreg@abs32@hi
+; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v32i32_inreg@abs32@lo
+; GFX9-NEXT:    s_mov_b32 s20, s36
+; GFX9-NEXT:    s_mov_b32 s21, s37
+; GFX9-NEXT:    s_mov_b32 s22, s38
+; GFX9-NEXT:    s_mov_b32 s23, s39
+; GFX9-NEXT:    s_mov_b32 s24, s40
+; GFX9-NEXT:    s_mov_b32 s25, s41
+; GFX9-NEXT:    s_mov_b32 s26, s42
+; GFX9-NEXT:    s_mov_b32 s27, s43
+; GFX9-NEXT:    s_mov_b32 s28, s44
+; GFX9-NEXT:    s_mov_b32 s29, s45
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 27
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
 ; GFX9-NEXT:    v_readlane_b32 s31, v40, 27
 ; GFX9-NEXT:    v_readlane_b32 s30, v40, 26
 ; GFX9-NEXT:    v_readlane_b32 s29, v40, 25
@@ -14265,46 +14879,47 @@ define amdgpu_gfx void @test_call_external_void_func_v32i32_inreg() #0 {
 ; GFX10-NEXT:    v_writelane_b32 v40, s17, 13
 ; GFX10-NEXT:    v_writelane_b32 v40, s18, 14
 ; GFX10-NEXT:    v_writelane_b32 v40, s19, 15
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_clause 0x1
+; GFX10-NEXT:    s_load_dwordx16 s[36:51], s[34:35], 0x40
+; GFX10-NEXT:    s_load_dwordx16 s[4:19], s[34:35], 0x0
+; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v32i32_inreg@abs32@hi
+; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v32i32_inreg@abs32@lo
 ; GFX10-NEXT:    v_writelane_b32 v40, s20, 16
 ; GFX10-NEXT:    v_writelane_b32 v40, s21, 17
 ; GFX10-NEXT:    v_writelane_b32 v40, s22, 18
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v0, s46
 ; GFX10-NEXT:    v_writelane_b32 v40, s23, 19
+; GFX10-NEXT:    v_mov_b32_e32 v1, s47
+; GFX10-NEXT:    v_mov_b32_e32 v2, s48
+; GFX10-NEXT:    v_mov_b32_e32 v3, s49
+; GFX10-NEXT:    s_mov_b32 s20, s36
 ; GFX10-NEXT:    v_writelane_b32 v40, s24, 20
+; GFX10-NEXT:    s_mov_b32 s21, s37
+; GFX10-NEXT:    s_mov_b32 s22, s38
+; GFX10-NEXT:    s_mov_b32 s23, s39
+; GFX10-NEXT:    s_mov_b32 s24, s40
 ; GFX10-NEXT:    v_writelane_b32 v40, s25, 21
+; GFX10-NEXT:    s_mov_b32 s25, s41
+; GFX10-NEXT:    v_mov_b32_e32 v4, s50
+; GFX10-NEXT:    v_mov_b32_e32 v5, s51
+; GFX10-NEXT:    buffer_store_dword v0, off, s[0:3], s32
+; GFX10-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:4
+; GFX10-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:8
+; GFX10-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:12
+; GFX10-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:16
+; GFX10-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:20
 ; GFX10-NEXT:    v_writelane_b32 v40, s26, 22
+; GFX10-NEXT:    s_mov_b32 s26, s42
 ; GFX10-NEXT:    v_writelane_b32 v40, s27, 23
+; GFX10-NEXT:    s_mov_b32 s27, s43
 ; GFX10-NEXT:    v_writelane_b32 v40, s28, 24
+; GFX10-NEXT:    s_mov_b32 s28, s44
 ; GFX10-NEXT:    v_writelane_b32 v40, s29, 25
+; GFX10-NEXT:    s_mov_b32 s29, s45
 ; GFX10-NEXT:    v_writelane_b32 v40, s30, 26
 ; GFX10-NEXT:    v_writelane_b32 v40, s31, 27
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_clause 0x1
-; GFX10-NEXT:    s_load_dwordx16 s[16:31], s[34:35], 0x40
-; GFX10-NEXT:    s_load_dwordx16 s[36:51], s[34:35], 0x0
-; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v32i32_inreg@abs32@hi
-; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v32i32_inreg@abs32@lo
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    v_mov_b32_e32 v0, s30
-; GFX10-NEXT:    v_mov_b32_e32 v1, s31
-; GFX10-NEXT:    s_mov_b32 s4, s40
-; GFX10-NEXT:    s_mov_b32 s5, s41
-; GFX10-NEXT:    buffer_store_dword v0, off, s[0:3], s32
-; GFX10-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:4
-; GFX10-NEXT:    s_mov_b32 s6, s42
-; GFX10-NEXT:    s_mov_b32 s7, s43
-; GFX10-NEXT:    s_mov_b32 s8, s44
-; GFX10-NEXT:    s_mov_b32 s9, s45
-; GFX10-NEXT:    s_mov_b32 s10, s46
-; GFX10-NEXT:    s_mov_b32 s11, s47
-; GFX10-NEXT:    s_mov_b32 s12, s48
-; GFX10-NEXT:    s_mov_b32 s13, s49
-; GFX10-NEXT:    s_mov_b32 s14, s50
-; GFX10-NEXT:    s_mov_b32 s15, s51
-; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
-; GFX10-NEXT:    s_mov_b32 s0, s36
-; GFX10-NEXT:    s_mov_b32 s1, s37
-; GFX10-NEXT:    s_mov_b32 s2, s38
-; GFX10-NEXT:    s_mov_b32 s3, s39
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
 ; GFX10-NEXT:    v_readlane_b32 s31, v40, 27
 ; GFX10-NEXT:    v_readlane_b32 s30, v40, 26
@@ -14355,8 +14970,8 @@ define amdgpu_gfx void @test_call_external_void_func_v32i32_inreg() #0 {
 ; GFX11-NEXT:    v_writelane_b32 v40, s0, 28
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v32i32_inreg@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v32i32_inreg@abs32@lo
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_add_i32 s2, s32, 16
 ; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
 ; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
@@ -14373,26 +14988,42 @@ define amdgpu_gfx void @test_call_external_void_func_v32i32_inreg() #0 {
 ; GFX11-NEXT:    v_writelane_b32 v40, s17, 13
 ; GFX11-NEXT:    v_writelane_b32 v40, s18, 14
 ; GFX11-NEXT:    v_writelane_b32 v40, s19, 15
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    s_load_b512 s[36:51], s[0:1], 0x40
+; GFX11-NEXT:    s_load_b512 s[4:19], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v32i32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v32i32_inreg@abs32@lo
 ; GFX11-NEXT:    v_writelane_b32 v40, s20, 16
 ; GFX11-NEXT:    v_writelane_b32 v40, s21, 17
 ; GFX11-NEXT:    v_writelane_b32 v40, s22, 18
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    v_dual_mov_b32 v4, s50 :: v_dual_mov_b32 v5, s51
 ; GFX11-NEXT:    v_writelane_b32 v40, s23, 19
+; GFX11-NEXT:    v_dual_mov_b32 v0, s46 :: v_dual_mov_b32 v1, s47
+; GFX11-NEXT:    v_dual_mov_b32 v2, s48 :: v_dual_mov_b32 v3, s49
 ; GFX11-NEXT:    v_writelane_b32 v40, s24, 20
+; GFX11-NEXT:    s_mov_b32 s20, s36
+; GFX11-NEXT:    s_mov_b32 s21, s37
+; GFX11-NEXT:    s_mov_b32 s22, s38
+; GFX11-NEXT:    s_mov_b32 s23, s39
 ; GFX11-NEXT:    v_writelane_b32 v40, s25, 21
+; GFX11-NEXT:    s_mov_b32 s24, s40
+; GFX11-NEXT:    s_mov_b32 s25, s41
+; GFX11-NEXT:    scratch_store_b64 off, v[4:5], s2
+; GFX11-NEXT:    scratch_store_b128 off, v[0:3], s32
 ; GFX11-NEXT:    v_writelane_b32 v40, s26, 22
+; GFX11-NEXT:    s_mov_b32 s26, s42
 ; GFX11-NEXT:    v_writelane_b32 v40, s27, 23
+; GFX11-NEXT:    s_mov_b32 s27, s43
 ; GFX11-NEXT:    v_writelane_b32 v40, s28, 24
+; GFX11-NEXT:    s_mov_b32 s28, s44
 ; GFX11-NEXT:    v_writelane_b32 v40, s29, 25
+; GFX11-NEXT:    s_mov_b32 s29, s45
 ; GFX11-NEXT:    v_writelane_b32 v40, s30, 26
 ; GFX11-NEXT:    v_writelane_b32 v40, s31, 27
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_clause 0x1
-; GFX11-NEXT:    s_load_b512 s[16:31], s[0:1], 0x40
-; GFX11-NEXT:    s_load_b512 s[0:15], s[0:1], 0x0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_dual_mov_b32 v0, s30 :: v_dual_mov_b32 v1, s31
-; GFX11-NEXT:    scratch_store_b64 off, v[0:1], s32
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_readlane_b32 s31, v40, 27
 ; GFX11-NEXT:    v_readlane_b32 s30, v40, 26
 ; GFX11-NEXT:    v_readlane_b32 s29, v40, 25
@@ -14440,8 +15071,9 @@ define amdgpu_gfx void @test_call_external_void_func_v32i32_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 28
-; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[34:35], s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT:    s_add_i32 s2, s32, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
@@ -14458,29 +15090,44 @@ define amdgpu_gfx void @test_call_external_void_func_v32i32_inreg() #0 {
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s17, 13
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s18, 14
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s19, 15
+; GFX10-SCRATCH-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-SCRATCH-NEXT:    s_clause 0x1
+; GFX10-SCRATCH-NEXT:    s_load_dwordx16 s[36:51], s[0:1], 0x40
+; GFX10-SCRATCH-NEXT:    s_load_dwordx16 s[4:19], s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v32i32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v32i32_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s20, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s21, 17
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s22, 18
+; GFX10-SCRATCH-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v4, s50
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s23, 19
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v5, s51
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, s46
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, s47
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, s48
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s24, 20
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v3, s49
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s20, s36
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s21, s37
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s22, s38
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s25, 21
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s23, s39
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s24, s40
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s25, s41
+; GFX10-SCRATCH-NEXT:    scratch_store_dwordx2 off, v[4:5], s2
+; GFX10-SCRATCH-NEXT:    scratch_store_dwordx4 off, v[0:3], s32
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s26, 22
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s26, s42
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s27, 23
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s27, s43
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s28, 24
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s28, s44
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s29, 25
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s29, s45
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 26
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 27
-; GFX10-SCRATCH-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-SCRATCH-NEXT:    s_clause 0x1
-; GFX10-SCRATCH-NEXT:    s_load_dwordx16 s[16:31], s[34:35], 0x40
-; GFX10-SCRATCH-NEXT:    s_load_dwordx16 s[0:15], s[34:35], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v32i32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v32i32_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, s30
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, s31
-; GFX10-SCRATCH-NEXT:    scratch_store_dwordx2 off, v[0:1], s32
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 27
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 26
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s29, v40, 25
@@ -14549,53 +15196,55 @@ define amdgpu_gfx void @test_call_external_void_func_v32i32_i32_inreg(i32) #0 {
 ; GFX9-NEXT:    v_writelane_b32 v40, s16, 12
 ; GFX9-NEXT:    v_writelane_b32 v40, s17, 13
 ; GFX9-NEXT:    v_writelane_b32 v40, s18, 14
+; GFX9-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
 ; GFX9-NEXT:    v_writelane_b32 v40, s19, 15
 ; GFX9-NEXT:    v_writelane_b32 v40, s20, 16
 ; GFX9-NEXT:    v_writelane_b32 v40, s21, 17
 ; GFX9-NEXT:    v_writelane_b32 v40, s22, 18
 ; GFX9-NEXT:    v_writelane_b32 v40, s23, 19
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dword s52, s[34:35], 0x0
+; GFX9-NEXT:    ; kill: killed $sgpr34_sgpr35
+; GFX9-NEXT:    ; kill: killed $sgpr34_sgpr35
+; GFX9-NEXT:    s_load_dwordx16 s[36:51], s[34:35], 0x40
+; GFX9-NEXT:    s_load_dwordx16 s[4:19], s[34:35], 0x0
 ; GFX9-NEXT:    v_writelane_b32 v40, s24, 20
-; GFX9-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
 ; GFX9-NEXT:    v_writelane_b32 v40, s25, 21
+; GFX9-NEXT:    s_addk_i32 s32, 0x400
 ; GFX9-NEXT:    v_writelane_b32 v40, s26, 22
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s52
 ; GFX9-NEXT:    v_writelane_b32 v40, s27, 23
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:24
+; GFX9-NEXT:    v_mov_b32_e32 v0, s46
 ; GFX9-NEXT:    v_writelane_b32 v40, s28, 24
+; GFX9-NEXT:    v_mov_b32_e32 v1, s47
+; GFX9-NEXT:    v_mov_b32_e32 v2, s48
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32
+; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:4
+; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:8
+; GFX9-NEXT:    v_mov_b32_e32 v0, s49
 ; GFX9-NEXT:    v_writelane_b32 v40, s29, 25
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    s_load_dword s52, s[34:35], 0x0
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:12
+; GFX9-NEXT:    v_mov_b32_e32 v0, s50
 ; GFX9-NEXT:    v_writelane_b32 v40, s30, 26
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:16
+; GFX9-NEXT:    v_mov_b32_e32 v0, s51
+; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v32i32_i32_inreg@abs32@hi
+; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v32i32_i32_inreg@abs32@lo
+; GFX9-NEXT:    s_mov_b32 s20, s36
+; GFX9-NEXT:    s_mov_b32 s21, s37
+; GFX9-NEXT:    s_mov_b32 s22, s38
+; GFX9-NEXT:    s_mov_b32 s23, s39
+; GFX9-NEXT:    s_mov_b32 s24, s40
+; GFX9-NEXT:    s_mov_b32 s25, s41
+; GFX9-NEXT:    s_mov_b32 s26, s42
+; GFX9-NEXT:    s_mov_b32 s27, s43
+; GFX9-NEXT:    s_mov_b32 s28, s44
+; GFX9-NEXT:    s_mov_b32 s29, s45
 ; GFX9-NEXT:    v_writelane_b32 v40, s31, 27
-; GFX9-NEXT:    s_load_dwordx16 s[16:31], s[34:35], 0x40
-; GFX9-NEXT:    s_load_dwordx16 s[36:51], s[34:35], 0x0
-; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s52
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:8
-; GFX9-NEXT:    v_mov_b32_e32 v0, s30
-; GFX9-NEXT:    s_mov_b32 s53, external_void_func_v32i32_i32_inreg@abs32@hi
-; GFX9-NEXT:    v_mov_b32_e32 v1, s31
-; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32
-; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:4
-; GFX9-NEXT:    s_mov_b32 s52, external_void_func_v32i32_i32_inreg@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, s36
-; GFX9-NEXT:    s_mov_b32 s1, s37
-; GFX9-NEXT:    s_mov_b32 s2, s38
-; GFX9-NEXT:    s_mov_b32 s3, s39
-; GFX9-NEXT:    s_mov_b32 s4, s40
-; GFX9-NEXT:    s_mov_b32 s5, s41
-; GFX9-NEXT:    s_mov_b32 s6, s42
-; GFX9-NEXT:    s_mov_b32 s7, s43
-; GFX9-NEXT:    s_mov_b32 s8, s44
-; GFX9-NEXT:    s_mov_b32 s9, s45
-; GFX9-NEXT:    s_mov_b32 s10, s46
-; GFX9-NEXT:    s_mov_b32 s11, s47
-; GFX9-NEXT:    s_mov_b32 s12, s48
-; GFX9-NEXT:    s_mov_b32 s13, s49
-; GFX9-NEXT:    s_mov_b32 s14, s50
-; GFX9-NEXT:    s_mov_b32 s15, s51
-; GFX9-NEXT:    ; kill: killed $sgpr34_sgpr35
-; GFX9-NEXT:    ; kill: killed $sgpr34_sgpr35
-; GFX9-NEXT:    s_swappc_b64 s[30:31], s[52:53]
+; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:20
+; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
 ; GFX9-NEXT:    v_readlane_b32 s31, v40, 27
 ; GFX9-NEXT:    v_readlane_b32 s30, v40, 26
 ; GFX9-NEXT:    v_readlane_b32 s29, v40, 25
@@ -14661,51 +15310,52 @@ define amdgpu_gfx void @test_call_external_void_func_v32i32_i32_inreg(i32) #0 {
 ; GFX10-NEXT:    v_writelane_b32 v40, s17, 13
 ; GFX10-NEXT:    v_writelane_b32 v40, s18, 14
 ; GFX10-NEXT:    v_writelane_b32 v40, s19, 15
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_clause 0x2
+; GFX10-NEXT:    s_load_dword s52, s[34:35], 0x0
+; GFX10-NEXT:    ; meta instruction
+; GFX10-NEXT:    ; meta instruction
+; GFX10-NEXT:    s_load_dwordx16 s[36:51], s[34:35], 0x40
+; GFX10-NEXT:    s_load_dwordx16 s[4:19], s[34:35], 0x0
+; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v32i32_i32_inreg@abs32@hi
+; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v32i32_i32_inreg@abs32@lo
 ; GFX10-NEXT:    v_writelane_b32 v40, s20, 16
 ; GFX10-NEXT:    v_writelane_b32 v40, s21, 17
 ; GFX10-NEXT:    v_writelane_b32 v40, s22, 18
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v0, s52
+; GFX10-NEXT:    v_mov_b32_e32 v1, s47
 ; GFX10-NEXT:    v_writelane_b32 v40, s23, 19
+; GFX10-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:24
+; GFX10-NEXT:    v_mov_b32_e32 v0, s46
+; GFX10-NEXT:    v_mov_b32_e32 v2, s48
+; GFX10-NEXT:    v_mov_b32_e32 v3, s49
 ; GFX10-NEXT:    v_writelane_b32 v40, s24, 20
+; GFX10-NEXT:    s_mov_b32 s20, s36
+; GFX10-NEXT:    s_mov_b32 s21, s37
+; GFX10-NEXT:    s_mov_b32 s22, s38
+; GFX10-NEXT:    s_mov_b32 s23, s39
 ; GFX10-NEXT:    v_writelane_b32 v40, s25, 21
+; GFX10-NEXT:    s_mov_b32 s24, s40
+; GFX10-NEXT:    s_mov_b32 s25, s41
+; GFX10-NEXT:    v_mov_b32_e32 v4, s50
+; GFX10-NEXT:    v_mov_b32_e32 v5, s51
 ; GFX10-NEXT:    v_writelane_b32 v40, s26, 22
+; GFX10-NEXT:    s_mov_b32 s26, s42
+; GFX10-NEXT:    buffer_store_dword v0, off, s[0:3], s32
+; GFX10-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:4
+; GFX10-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:8
+; GFX10-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:12
+; GFX10-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:16
+; GFX10-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:20
 ; GFX10-NEXT:    v_writelane_b32 v40, s27, 23
+; GFX10-NEXT:    s_mov_b32 s27, s43
 ; GFX10-NEXT:    v_writelane_b32 v40, s28, 24
+; GFX10-NEXT:    s_mov_b32 s28, s44
 ; GFX10-NEXT:    v_writelane_b32 v40, s29, 25
+; GFX10-NEXT:    s_mov_b32 s29, s45
 ; GFX10-NEXT:    v_writelane_b32 v40, s30, 26
 ; GFX10-NEXT:    v_writelane_b32 v40, s31, 27
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    s_clause 0x2
-; GFX10-NEXT:    s_load_dword s52, s[34:35], 0x0
-; GFX10-NEXT:    ; meta instruction
-; GFX10-NEXT:    ; meta instruction
-; GFX10-NEXT:    s_load_dwordx16 s[16:31], s[34:35], 0x40
-; GFX10-NEXT:    s_load_dwordx16 s[36:51], s[34:35], 0x0
-; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v32i32_i32_inreg@abs32@hi
-; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v32i32_i32_inreg@abs32@lo
-; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    v_mov_b32_e32 v0, s52
-; GFX10-NEXT:    v_mov_b32_e32 v1, s30
-; GFX10-NEXT:    v_mov_b32_e32 v2, s31
-; GFX10-NEXT:    s_mov_b32 s4, s40
-; GFX10-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:8
-; GFX10-NEXT:    buffer_store_dword v1, off, s[0:3], s32
-; GFX10-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:4
-; GFX10-NEXT:    s_mov_b32 s5, s41
-; GFX10-NEXT:    s_mov_b32 s6, s42
-; GFX10-NEXT:    s_mov_b32 s7, s43
-; GFX10-NEXT:    s_mov_b32 s8, s44
-; GFX10-NEXT:    s_mov_b32 s9, s45
-; GFX10-NEXT:    s_mov_b32 s10, s46
-; GFX10-NEXT:    s_mov_b32 s11, s47
-; GFX10-NEXT:    s_mov_b32 s12, s48
-; GFX10-NEXT:    s_mov_b32 s13, s49
-; GFX10-NEXT:    s_mov_b32 s14, s50
-; GFX10-NEXT:    s_mov_b32 s15, s51
-; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
-; GFX10-NEXT:    s_mov_b32 s0, s36
-; GFX10-NEXT:    s_mov_b32 s1, s37
-; GFX10-NEXT:    s_mov_b32 s2, s38
-; GFX10-NEXT:    s_mov_b32 s3, s39
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
 ; GFX10-NEXT:    v_readlane_b32 s31, v40, 27
 ; GFX10-NEXT:    v_readlane_b32 s30, v40, 26
@@ -14756,8 +15406,8 @@ define amdgpu_gfx void @test_call_external_void_func_v32i32_i32_inreg(i32) #0 {
 ; GFX11-NEXT:    v_writelane_b32 v40, s0, 28
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v32i32_i32_inreg@abs32@hi
-; GFX11-NEXT:    s_add_i32 s36, s32, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_add_i32 s3, s32, 16
 ; GFX11-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX11-NEXT:    v_writelane_b32 v40, s5, 1
 ; GFX11-NEXT:    v_writelane_b32 v40, s6, 2
@@ -14774,30 +15424,46 @@ define amdgpu_gfx void @test_call_external_void_func_v32i32_i32_inreg(i32) #0 {
 ; GFX11-NEXT:    v_writelane_b32 v40, s17, 13
 ; GFX11-NEXT:    v_writelane_b32 v40, s18, 14
 ; GFX11-NEXT:    v_writelane_b32 v40, s19, 15
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_clause 0x2
+; GFX11-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-NEXT:    s_load_b512 s[36:51], s[0:1], 0x40
+; GFX11-NEXT:    s_load_b512 s[4:19], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v32i32_i32_inreg@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v32i32_i32_inreg@abs32@lo
 ; GFX11-NEXT:    v_writelane_b32 v40, s20, 16
 ; GFX11-NEXT:    v_writelane_b32 v40, s21, 17
 ; GFX11-NEXT:    v_writelane_b32 v40, s22, 18
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    v_dual_mov_b32 v6, s2 :: v_dual_mov_b32 v5, s51
 ; GFX11-NEXT:    v_writelane_b32 v40, s23, 19
+; GFX11-NEXT:    v_dual_mov_b32 v4, s50 :: v_dual_mov_b32 v1, s47
+; GFX11-NEXT:    v_dual_mov_b32 v0, s46 :: v_dual_mov_b32 v3, s49
 ; GFX11-NEXT:    v_writelane_b32 v40, s24, 20
+; GFX11-NEXT:    v_mov_b32_e32 v2, s48
+; GFX11-NEXT:    s_add_i32 s2, s32, 24
+; GFX11-NEXT:    s_mov_b32 s20, s36
+; GFX11-NEXT:    s_mov_b32 s21, s37
 ; GFX11-NEXT:    v_writelane_b32 v40, s25, 21
+; GFX11-NEXT:    s_mov_b32 s22, s38
+; GFX11-NEXT:    s_mov_b32 s23, s39
+; GFX11-NEXT:    s_mov_b32 s24, s40
+; GFX11-NEXT:    s_mov_b32 s25, s41
 ; GFX11-NEXT:    v_writelane_b32 v40, s26, 22
+; GFX11-NEXT:    s_mov_b32 s26, s42
+; GFX11-NEXT:    scratch_store_b32 off, v6, s2
+; GFX11-NEXT:    scratch_store_b64 off, v[4:5], s3
+; GFX11-NEXT:    scratch_store_b128 off, v[0:3], s32
 ; GFX11-NEXT:    v_writelane_b32 v40, s27, 23
+; GFX11-NEXT:    s_mov_b32 s27, s43
 ; GFX11-NEXT:    v_writelane_b32 v40, s28, 24
+; GFX11-NEXT:    s_mov_b32 s28, s44
 ; GFX11-NEXT:    v_writelane_b32 v40, s29, 25
+; GFX11-NEXT:    s_mov_b32 s29, s45
 ; GFX11-NEXT:    v_writelane_b32 v40, s30, 26
 ; GFX11-NEXT:    v_writelane_b32 v40, s31, 27
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_clause 0x2
-; GFX11-NEXT:    s_load_b32 s34, s[0:1], 0x0
-; GFX11-NEXT:    s_load_b512 s[16:31], s[0:1], 0x40
-; GFX11-NEXT:    s_load_b512 s[0:15], s[0:1], 0x0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_dual_mov_b32 v2, s34 :: v_dual_mov_b32 v1, s31
-; GFX11-NEXT:    v_mov_b32_e32 v0, s30
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v32i32_i32_inreg@abs32@lo
-; GFX11-NEXT:    scratch_store_b32 off, v2, s36
-; GFX11-NEXT:    scratch_store_b64 off, v[0:1], s32
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_readlane_b32 s31, v40, 27
 ; GFX11-NEXT:    v_readlane_b32 s30, v40, 26
 ; GFX11-NEXT:    v_readlane_b32 s29, v40, 25
@@ -14845,17 +15511,13 @@ define amdgpu_gfx void @test_call_external_void_func_v32i32_i32_inreg(i32) #0 {
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 28
-; GFX10-SCRATCH-NEXT:    s_clause 0x1
-; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[34:35], s[0:1], 0x0
-; GFX10-SCRATCH-NEXT:    s_load_dword s36, s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT:    s_add_i32 s3, s32, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s4, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s5, 1
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s6, 2
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s7, 3
-; GFX10-SCRATCH-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, s36
-; GFX10-SCRATCH-NEXT:    s_add_i32 s36, s32, 8
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s8, 4
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s9, 5
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s10, 6
@@ -14868,29 +15530,50 @@ define amdgpu_gfx void @test_call_external_void_func_v32i32_i32_inreg(i32) #0 {
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s17, 13
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s18, 14
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s19, 15
+; GFX10-SCRATCH-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-SCRATCH-NEXT:    s_clause 0x2
+; GFX10-SCRATCH-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    ; meta instruction
+; GFX10-SCRATCH-NEXT:    ; meta instruction
+; GFX10-SCRATCH-NEXT:    s_load_dwordx16 s[36:51], s[0:1], 0x40
+; GFX10-SCRATCH-NEXT:    s_load_dwordx16 s[4:19], s[0:1], 0x0
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v32i32_i32_inreg@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v32i32_i32_inreg@abs32@lo
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s20, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s21, 17
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s22, 18
+; GFX10-SCRATCH-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v6, s2
+; GFX10-SCRATCH-NEXT:    s_add_i32 s2, s32, 24
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v4, s50
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s23, 19
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v5, s51
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, s46
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, s47
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v2, s48
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s24, 20
+; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v3, s49
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s20, s36
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s21, s37
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s22, s38
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s25, 21
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s23, s39
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s24, s40
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s25, s41
+; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v6, s2
+; GFX10-SCRATCH-NEXT:    scratch_store_dwordx2 off, v[4:5], s3
+; GFX10-SCRATCH-NEXT:    scratch_store_dwordx4 off, v[0:3], s32
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s26, 22
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s26, s42
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s27, 23
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s27, s43
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s28, 24
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s28, s44
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s29, 25
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s29, s45
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 26
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 27
-; GFX10-SCRATCH-NEXT:    s_clause 0x1
-; GFX10-SCRATCH-NEXT:    s_load_dwordx16 s[16:31], s[34:35], 0x40
-; GFX10-SCRATCH-NEXT:    s_load_dwordx16 s[0:15], s[34:35], 0x0
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v32i32_i32_inreg@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v32i32_i32_inreg@abs32@lo
-; GFX10-SCRATCH-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v0, s30
-; GFX10-SCRATCH-NEXT:    v_mov_b32_e32 v1, s31
-; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v2, s36
-; GFX10-SCRATCH-NEXT:    scratch_store_dwordx2 off, v[0:1], s32
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 27
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 26
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s29, v40, 25
@@ -16714,7 +17397,6 @@ define amdgpu_gfx void @test_call_external_void_func_bf16_inreg(i16 inreg %arg)
 ; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_bf16@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_bf16@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, s4
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
 ; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
@@ -16741,7 +17423,6 @@ define amdgpu_gfx void @test_call_external_void_func_bf16_inreg(i16 inreg %arg)
 ; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_bf16@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_bf16@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, s4
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
 ; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
@@ -16761,18 +17442,18 @@ define amdgpu_gfx void @test_call_external_void_func_bf16_inreg(i16 inreg %arg)
 ; GFX11-LABEL: test_call_external_void_func_bf16_inreg:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_mov_b32 s1, s33
+; GFX11-NEXT:    s_mov_b32 s0, s33
 ; GFX11-NEXT:    s_mov_b32 s33, s32
-; GFX11-NEXT:    s_or_saveexec_b32 s2, -1
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-NEXT:    s_mov_b32 exec_lo, s2
-; GFX11-NEXT:    v_writelane_b32 v40, s1, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_bf16@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_bf16@abs32@lo
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_bf16@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_bf16@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
 ; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
 ; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
@@ -16788,19 +17469,19 @@ define amdgpu_gfx void @test_call_external_void_func_bf16_inreg(i16 inreg %arg)
 ; GFX10-SCRATCH-LABEL: test_call_external_void_func_bf16_inreg:
 ; GFX10-SCRATCH:       ; %bb.0:
 ; GFX10-SCRATCH-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, s33
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, s33
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
-; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s2, -1
+; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s2
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s1, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_bf16@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_bf16@abs32@lo
+; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_bf16@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_bf16@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
@@ -16830,7 +17511,6 @@ define amdgpu_gfx void @test_call_external_void_func_v1bf16_inreg(i16 inreg %arg
 ; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v1bf16@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v1bf16@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, s4
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
 ; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
@@ -16857,7 +17537,6 @@ define amdgpu_gfx void @test_call_external_void_func_v1bf16_inreg(i16 inreg %arg
 ; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v1bf16@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v1bf16@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, s4
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
 ; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
@@ -16877,18 +17556,18 @@ define amdgpu_gfx void @test_call_external_void_func_v1bf16_inreg(i16 inreg %arg
 ; GFX11-LABEL: test_call_external_void_func_v1bf16_inreg:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_mov_b32 s1, s33
+; GFX11-NEXT:    s_mov_b32 s0, s33
 ; GFX11-NEXT:    s_mov_b32 s33, s32
-; GFX11-NEXT:    s_or_saveexec_b32 s2, -1
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-NEXT:    s_mov_b32 exec_lo, s2
-; GFX11-NEXT:    v_writelane_b32 v40, s1, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_v1bf16@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_v1bf16@abs32@lo
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v1bf16@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v1bf16@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
 ; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
 ; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
@@ -16904,19 +17583,19 @@ define amdgpu_gfx void @test_call_external_void_func_v1bf16_inreg(i16 inreg %arg
 ; GFX10-SCRATCH-LABEL: test_call_external_void_func_v1bf16_inreg:
 ; GFX10-SCRATCH:       ; %bb.0:
 ; GFX10-SCRATCH-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, s33
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, s33
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
-; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s2, -1
+; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s2
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s1, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_v1bf16@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_v1bf16@abs32@lo
+; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v1bf16@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v1bf16@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
@@ -16946,7 +17625,6 @@ define amdgpu_gfx void @test_call_external_void_func_v2bf16_inreg(i32 inreg %arg
 ; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v2bf16@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v2bf16@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s0, s4
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
 ; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
@@ -16973,7 +17651,6 @@ define amdgpu_gfx void @test_call_external_void_func_v2bf16_inreg(i32 inreg %arg
 ; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v2bf16@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v2bf16@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s0, s4
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
 ; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
@@ -16993,18 +17670,18 @@ define amdgpu_gfx void @test_call_external_void_func_v2bf16_inreg(i32 inreg %arg
 ; GFX11-LABEL: test_call_external_void_func_v2bf16_inreg:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_mov_b32 s1, s33
+; GFX11-NEXT:    s_mov_b32 s0, s33
 ; GFX11-NEXT:    s_mov_b32 s33, s32
-; GFX11-NEXT:    s_or_saveexec_b32 s2, -1
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-NEXT:    s_mov_b32 exec_lo, s2
-; GFX11-NEXT:    v_writelane_b32 v40, s1, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_v2bf16@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_v2bf16@abs32@lo
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v2bf16@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v2bf16@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
 ; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
 ; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
@@ -17020,19 +17697,19 @@ define amdgpu_gfx void @test_call_external_void_func_v2bf16_inreg(i32 inreg %arg
 ; GFX10-SCRATCH-LABEL: test_call_external_void_func_v2bf16_inreg:
 ; GFX10-SCRATCH:       ; %bb.0:
 ; GFX10-SCRATCH-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, s33
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, s33
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
-; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s2, -1
+; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s2
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s1, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_v2bf16@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_v2bf16@abs32@lo
+; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v2bf16@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v2bf16@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
@@ -17062,8 +17739,6 @@ define amdgpu_gfx void @test_call_external_void_func_v3bf16_inreg(<3 x i16> inre
 ; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v3bf16@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v3bf16@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s1, s5
-; GFX9-NEXT:    s_mov_b32 s0, s4
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
 ; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
@@ -17090,10 +17765,8 @@ define amdgpu_gfx void @test_call_external_void_func_v3bf16_inreg(<3 x i16> inre
 ; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v3bf16@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v3bf16@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s1, s5
-; GFX10-NEXT:    s_mov_b32 s0, s4
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
 ; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
@@ -17111,18 +17784,18 @@ define amdgpu_gfx void @test_call_external_void_func_v3bf16_inreg(<3 x i16> inre
 ; GFX11-LABEL: test_call_external_void_func_v3bf16_inreg:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_mov_b32 s2, s33
+; GFX11-NEXT:    s_mov_b32 s0, s33
 ; GFX11-NEXT:    s_mov_b32 s33, s32
-; GFX11-NEXT:    s_or_saveexec_b32 s3, -1
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-NEXT:    s_mov_b32 exec_lo, s3
-; GFX11-NEXT:    v_writelane_b32 v40, s2, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_v3bf16@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_v3bf16@abs32@lo
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v3bf16@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v3bf16@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
 ; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
 ; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
@@ -17138,19 +17811,19 @@ define amdgpu_gfx void @test_call_external_void_func_v3bf16_inreg(<3 x i16> inre
 ; GFX10-SCRATCH-LABEL: test_call_external_void_func_v3bf16_inreg:
 ; GFX10-SCRATCH:       ; %bb.0:
 ; GFX10-SCRATCH-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, s33
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, s33
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
-; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s3, -1
+; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s3
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s2, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_v3bf16@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_v3bf16@abs32@lo
+; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v3bf16@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v3bf16@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
@@ -17180,8 +17853,6 @@ define amdgpu_gfx void @test_call_external_void_func_v4bf16_inreg(<4 x i16> inre
 ; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v4bf16@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v4bf16@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s1, s5
-; GFX9-NEXT:    s_mov_b32 s0, s4
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
 ; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
@@ -17208,10 +17879,8 @@ define amdgpu_gfx void @test_call_external_void_func_v4bf16_inreg(<4 x i16> inre
 ; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v4bf16@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v4bf16@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s1, s5
-; GFX10-NEXT:    s_mov_b32 s0, s4
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
 ; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
@@ -17229,18 +17898,18 @@ define amdgpu_gfx void @test_call_external_void_func_v4bf16_inreg(<4 x i16> inre
 ; GFX11-LABEL: test_call_external_void_func_v4bf16_inreg:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_mov_b32 s2, s33
+; GFX11-NEXT:    s_mov_b32 s0, s33
 ; GFX11-NEXT:    s_mov_b32 s33, s32
-; GFX11-NEXT:    s_or_saveexec_b32 s3, -1
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-NEXT:    s_mov_b32 exec_lo, s3
-; GFX11-NEXT:    v_writelane_b32 v40, s2, 2
-; GFX11-NEXT:    s_mov_b32 s3, external_void_func_v4bf16@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s2, external_void_func_v4bf16@abs32@lo
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v4bf16@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v4bf16@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
 ; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
 ; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
@@ -17256,19 +17925,19 @@ define amdgpu_gfx void @test_call_external_void_func_v4bf16_inreg(<4 x i16> inre
 ; GFX10-SCRATCH-LABEL: test_call_external_void_func_v4bf16_inreg:
 ; GFX10-SCRATCH:       ; %bb.0:
 ; GFX10-SCRATCH-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, s33
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, s33
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
-; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s3, -1
+; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s3
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s2, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s3, external_void_func_v4bf16@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s2, external_void_func_v4bf16@abs32@lo
+; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v4bf16@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v4bf16@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
@@ -17298,10 +17967,6 @@ define amdgpu_gfx void @test_call_external_void_func_v8bf16_inreg(<8 x i16> inre
 ; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v8bf16@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v8bf16@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s3, s7
-; GFX9-NEXT:    s_mov_b32 s2, s6
-; GFX9-NEXT:    s_mov_b32 s1, s5
-; GFX9-NEXT:    s_mov_b32 s0, s4
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
 ; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
@@ -17328,12 +17993,8 @@ define amdgpu_gfx void @test_call_external_void_func_v8bf16_inreg(<8 x i16> inre
 ; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v8bf16@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v8bf16@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s3, s7
-; GFX10-NEXT:    s_mov_b32 s2, s6
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX10-NEXT:    s_mov_b32 s1, s5
-; GFX10-NEXT:    s_mov_b32 s0, s4
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
 ; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
@@ -17351,18 +18012,18 @@ define amdgpu_gfx void @test_call_external_void_func_v8bf16_inreg(<8 x i16> inre
 ; GFX11-LABEL: test_call_external_void_func_v8bf16_inreg:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_mov_b32 s34, s33
+; GFX11-NEXT:    s_mov_b32 s0, s33
 ; GFX11-NEXT:    s_mov_b32 s33, s32
-; GFX11-NEXT:    s_or_saveexec_b32 s35, -1
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-NEXT:    s_mov_b32 exec_lo, s35
-; GFX11-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v8bf16@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v8bf16@abs32@lo
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v8bf16@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v8bf16@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
 ; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
 ; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
@@ -17378,19 +18039,19 @@ define amdgpu_gfx void @test_call_external_void_func_v8bf16_inreg(<8 x i16> inre
 ; GFX10-SCRATCH-LABEL: test_call_external_void_func_v8bf16_inreg:
 ; GFX10-SCRATCH:       ; %bb.0:
 ; GFX10-SCRATCH-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, s33
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, s33
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
-; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s35, -1
+; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v8bf16@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v8bf16@abs32@lo
+; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v8bf16@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v8bf16@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
@@ -17416,32 +18077,16 @@ define amdgpu_gfx void @test_call_external_void_func_v16bf16_inreg(<16 x i16> in
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
-; GFX9-NEXT:    v_writelane_b32 v40, s34, 6
-; GFX9-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX9-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX9-NEXT:    v_writelane_b32 v40, s6, 2
-; GFX9-NEXT:    v_writelane_b32 v40, s7, 3
-; GFX9-NEXT:    v_writelane_b32 v40, s30, 4
+; GFX9-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX9-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX9-NEXT:    s_mov_b32 s35, external_void_func_v16bf16@abs32@hi
 ; GFX9-NEXT:    s_mov_b32 s34, external_void_func_v16bf16@abs32@lo
-; GFX9-NEXT:    s_mov_b32 s3, s7
-; GFX9-NEXT:    s_mov_b32 s2, s6
-; GFX9-NEXT:    s_mov_b32 s1, s5
-; GFX9-NEXT:    s_mov_b32 s0, s4
-; GFX9-NEXT:    s_mov_b32 s4, s8
-; GFX9-NEXT:    s_mov_b32 s5, s9
-; GFX9-NEXT:    s_mov_b32 s6, s10
-; GFX9-NEXT:    s_mov_b32 s7, s11
 ; GFX9-NEXT:    s_addk_i32 s32, 0x400
-; GFX9-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX9-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX9-NEXT:    v_readlane_b32 s31, v40, 5
-; GFX9-NEXT:    v_readlane_b32 s30, v40, 4
-; GFX9-NEXT:    v_readlane_b32 s7, v40, 3
-; GFX9-NEXT:    v_readlane_b32 s6, v40, 2
-; GFX9-NEXT:    v_readlane_b32 s5, v40, 1
-; GFX9-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX9-NEXT:    v_readlane_b32 s34, v40, 6
+; GFX9-NEXT:    v_readlane_b32 s31, v40, 1
+; GFX9-NEXT:    v_readlane_b32 s30, v40, 0
+; GFX9-NEXT:    v_readlane_b32 s34, v40, 2
 ; GFX9-NEXT:    s_or_saveexec_b64 s[36:37], -1
 ; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX9-NEXT:    s_mov_b64 exec, s[36:37]
@@ -17459,32 +18104,16 @@ define amdgpu_gfx void @test_call_external_void_func_v16bf16_inreg(<16 x i16> in
 ; GFX10-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
 ; GFX10-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-NEXT:    v_writelane_b32 v40, s34, 6
+; GFX10-NEXT:    v_writelane_b32 v40, s34, 2
 ; GFX10-NEXT:    s_mov_b32 s35, external_void_func_v16bf16@abs32@hi
 ; GFX10-NEXT:    s_mov_b32 s34, external_void_func_v16bf16@abs32@lo
-; GFX10-NEXT:    s_mov_b32 s3, s7
-; GFX10-NEXT:    s_mov_b32 s2, s6
-; GFX10-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX10-NEXT:    s_mov_b32 s1, s5
-; GFX10-NEXT:    s_mov_b32 s0, s4
-; GFX10-NEXT:    s_mov_b32 s4, s8
 ; GFX10-NEXT:    s_addk_i32 s32, 0x200
-; GFX10-NEXT:    v_writelane_b32 v40, s5, 1
-; GFX10-NEXT:    s_mov_b32 s5, s9
-; GFX10-NEXT:    v_writelane_b32 v40, s6, 2
-; GFX10-NEXT:    s_mov_b32 s6, s10
-; GFX10-NEXT:    v_writelane_b32 v40, s7, 3
-; GFX10-NEXT:    s_mov_b32 s7, s11
-; GFX10-NEXT:    v_writelane_b32 v40, s30, 4
-; GFX10-NEXT:    v_writelane_b32 v40, s31, 5
+; GFX10-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX10-NEXT:    v_writelane_b32 v40, s31, 1
 ; GFX10-NEXT:    s_swappc_b64 s[30:31], s[34:35]
-; GFX10-NEXT:    v_readlane_b32 s31, v40, 5
-; GFX10-NEXT:    v_readlane_b32 s30, v40, 4
-; GFX10-NEXT:    v_readlane_b32 s7, v40, 3
-; GFX10-NEXT:    v_readlane_b32 s6, v40, 2
-; GFX10-NEXT:    v_readlane_b32 s5, v40, 1
-; GFX10-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX10-NEXT:    v_readlane_b32 s34, v40, 6
+; GFX10-NEXT:    v_readlane_b32 s31, v40, 1
+; GFX10-NEXT:    v_readlane_b32 s30, v40, 0
+; GFX10-NEXT:    v_readlane_b32 s34, v40, 2
 ; GFX10-NEXT:    s_or_saveexec_b32 s35, -1
 ; GFX10-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
 ; GFX10-NEXT:    s_waitcnt_depctr 0xffe3
@@ -17497,18 +18126,18 @@ define amdgpu_gfx void @test_call_external_void_func_v16bf16_inreg(<16 x i16> in
 ; GFX11-LABEL: test_call_external_void_func_v16bf16_inreg:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_mov_b32 s34, s33
+; GFX11-NEXT:    s_mov_b32 s0, s33
 ; GFX11-NEXT:    s_mov_b32 s33, s32
-; GFX11-NEXT:    s_or_saveexec_b32 s35, -1
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-NEXT:    s_mov_b32 exec_lo, s35
-; GFX11-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX11-NEXT:    s_mov_b32 s35, external_void_func_v16bf16@abs32@hi
-; GFX11-NEXT:    s_mov_b32 s34, external_void_func_v16bf16@abs32@lo
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v16bf16@abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v16bf16@abs32@lo
 ; GFX11-NEXT:    s_add_i32 s32, s32, 16
 ; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
 ; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
@@ -17524,19 +18153,19 @@ define amdgpu_gfx void @test_call_external_void_func_v16bf16_inreg(<16 x i16> in
 ; GFX10-SCRATCH-LABEL: test_call_external_void_func_v16bf16_inreg:
 ; GFX10-SCRATCH:       ; %bb.0:
 ; GFX10-SCRATCH-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, s33
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, s33
 ; GFX10-SCRATCH-NEXT:    s_mov_b32 s33, s32
-; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s35, -1
+; GFX10-SCRATCH-NEXT:    s_or_saveexec_b32 s1, -1
 ; GFX10-SCRATCH-NEXT:    scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
 ; GFX10-SCRATCH-NEXT:    s_waitcnt_depctr 0xffe3
-; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s35
-; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s35, external_void_func_v16bf16@abs32@hi
-; GFX10-SCRATCH-NEXT:    s_mov_b32 s34, external_void_func_v16bf16@abs32@lo
+; GFX10-SCRATCH-NEXT:    s_mov_b32 exec_lo, s1
+; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s1, external_void_func_v16bf16@abs32@hi
+; GFX10-SCRATCH-NEXT:    s_mov_b32 s0, external_void_func_v16bf16@abs32@lo
 ; GFX10-SCRATCH-NEXT:    s_add_i32 s32, s32, 16
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s30, 0
 ; GFX10-SCRATCH-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[34:35]
+; GFX10-SCRATCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s31, v40, 1
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s30, v40, 0
 ; GFX10-SCRATCH-NEXT:    v_readlane_b32 s0, v40, 2
diff --git a/llvm/test/CodeGen/AMDGPU/indirect-call.ll b/llvm/test/CodeGen/AMDGPU/indirect-call.ll
index 25c684004446f..7799b9509ceb0 100644
--- a/llvm/test/CodeGen/AMDGPU/indirect-call.ll
+++ b/llvm/test/CodeGen/AMDGPU/indirect-call.ll
@@ -847,11 +847,11 @@ define void @test_indirect_call_vgpr_ptr_inreg_arg(ptr %fptr) {
 ; GCN-LABEL: test_indirect_call_vgpr_ptr_inreg_arg:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT:    s_mov_b32 s10, s33
+; GCN-NEXT:    s_mov_b32 s5, s33
 ; GCN-NEXT:    s_mov_b32 s33, s32
-; GCN-NEXT:    s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT:    s_or_saveexec_b64 s[6:7], -1
 ; GCN-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
-; GCN-NEXT:    s_mov_b64 exec, s[4:5]
+; GCN-NEXT:    s_mov_b64 exec, s[6:7]
 ; GCN-NEXT:    s_addk_i32 s32, 0x400
 ; GCN-NEXT:    v_writelane_b32 v40, s30, 0
 ; GCN-NEXT:    v_writelane_b32 v40, s31, 1
@@ -885,19 +885,19 @@ define void @test_indirect_call_vgpr_ptr_inreg_arg(ptr %fptr) {
 ; GCN-NEXT:    v_writelane_b32 v40, s61, 29
 ; GCN-NEXT:    v_writelane_b32 v40, s62, 30
 ; GCN-NEXT:    v_writelane_b32 v40, s63, 31
-; GCN-NEXT:    s_mov_b64 s[4:5], exec
+; GCN-NEXT:    s_mov_b64 s[6:7], exec
+; GCN-NEXT:    s_movk_i32 s4, 0x7b
 ; GCN-NEXT:  .LBB6_1: ; =>This Inner Loop Header: Depth=1
-; GCN-NEXT:    v_readfirstlane_b32 s6, v0
-; GCN-NEXT:    v_readfirstlane_b32 s7, v1
-; GCN-NEXT:    v_cmp_eq_u64_e32 vcc, s[6:7], v[0:1]
-; GCN-NEXT:    s_and_saveexec_b64 s[8:9], vcc
-; GCN-NEXT:    s_movk_i32 s0, 0x7b
-; GCN-NEXT:    s_swappc_b64 s[30:31], s[6:7]
+; GCN-NEXT:    v_readfirstlane_b32 s8, v0
+; GCN-NEXT:    v_readfirstlane_b32 s9, v1
+; GCN-NEXT:    v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GCN-NEXT:    s_and_saveexec_b64 s[10:11], vcc
+; GCN-NEXT:    s_swappc_b64 s[30:31], s[8:9]
 ; GCN-NEXT:    ; implicit-def: $vgpr0_vgpr1
-; GCN-NEXT:    s_xor_b64 exec, exec, s[8:9]
+; GCN-NEXT:    s_xor_b64 exec, exec, s[10:11]
 ; GCN-NEXT:    s_cbranch_execnz .LBB6_1
 ; GCN-NEXT:  ; %bb.2:
-; GCN-NEXT:    s_mov_b64 exec, s[4:5]
+; GCN-NEXT:    s_mov_b64 exec, s[6:7]
 ; GCN-NEXT:    v_readlane_b32 s63, v40, 31
 ; GCN-NEXT:    v_readlane_b32 s62, v40, 30
 ; GCN-NEXT:    v_readlane_b32 s61, v40, 29
@@ -930,22 +930,22 @@ define void @test_indirect_call_vgpr_ptr_inreg_arg(ptr %fptr) {
 ; GCN-NEXT:    v_readlane_b32 s34, v40, 2
 ; GCN-NEXT:    v_readlane_b32 s31, v40, 1
 ; GCN-NEXT:    v_readlane_b32 s30, v40, 0
-; GCN-NEXT:    s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT:    s_or_saveexec_b64 s[6:7], -1
 ; GCN-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
-; GCN-NEXT:    s_mov_b64 exec, s[4:5]
+; GCN-NEXT:    s_mov_b64 exec, s[6:7]
 ; GCN-NEXT:    s_addk_i32 s32, 0xfc00
-; GCN-NEXT:    s_mov_b32 s33, s10
+; GCN-NEXT:    s_mov_b32 s33, s5
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GISEL-LABEL: test_indirect_call_vgpr_ptr_inreg_arg:
 ; GISEL:       ; %bb.0:
 ; GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GISEL-NEXT:    s_mov_b32 s10, s33
+; GISEL-NEXT:    s_mov_b32 s5, s33
 ; GISEL-NEXT:    s_mov_b32 s33, s32
-; GISEL-NEXT:    s_or_saveexec_b64 s[4:5], -1
+; GISEL-NEXT:    s_or_saveexec_b64 s[6:7], -1
 ; GISEL-NEXT:    buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
-; GISEL-NEXT:    s_mov_b64 exec, s[4:5]
+; GISEL-NEXT:    s_mov_b64 exec, s[6:7]
 ; GISEL-NEXT:    s_addk_i32 s32, 0x400
 ; GISEL-NEXT:    v_writelane_b32 v40, s30, 0
 ; GISEL-NEXT:    v_writelane_b32 v40, s31, 1
@@ -979,19 +979,19 @@ define void @test_indirect_call_vgpr_ptr_inreg_arg(ptr %fptr) {
 ; GISEL-NEXT:    v_writelane_b32 v40, s61, 29
 ; GISEL-NEXT:    v_writelane_b32 v40, s62, 30
 ; GISEL-NEXT:    v_writelane_b32 v40, s63, 31
-; GISEL-NEXT:    s_mov_b64 s[4:5], exec
+; GISEL-NEXT:    s_mov_b64 s[6:7], exec
+; GISEL-NEXT:    s_movk_i32 s4, 0x7b
 ; GISEL-NEXT:  .LBB6_1: ; =>This Inner Loop Header: Depth=1
-; GISEL-NEXT:    v_readfirstlane_b32 s6, v0
-; GISEL-NEXT:    v_readfirstlane_b32 s7, v1
-; GISEL-NEXT:    v_cmp_eq_u64_e32 vcc, s[6:7], v[0:1]
-; GISEL-NEXT:    s_and_saveexec_b64 s[8:9], vcc
-; GISEL-NEXT:    s_movk_i32 s0, 0x7b
-; GISEL-NEXT:    s_swappc_b64 s[30:31], s[6:7]
+; GISEL-NEXT:    v_readfirstlane_b32 s8, v0
+; GISEL-NEXT:    v_readfirstlane_b32 s9, v1
+; GISEL-NEXT:    v_cmp_eq_u64_e32 vcc, s[8:9], v[0:1]
+; GISEL-NEXT:    s_and_saveexec_b64 s[10:11], vcc
+; GISEL-NEXT:    s_swappc_b64 s[30:31], s[8:9]
 ; GISEL-NEXT:    ; implicit-def: $vgpr0
-; GISEL-NEXT:    s_xor_b64 exec, exec, s[8:9]
+; GISEL-NEXT:    s_xor_b64 exec, exec, s[10:11]
 ; GISEL-NEXT:    s_cbranch_execnz .LBB6_1
 ; GISEL-NEXT:  ; %bb.2:
-; GISEL-NEXT:    s_mov_b64 exec, s[4:5]
+; GISEL-NEXT:    s_mov_b64 exec, s[6:7]
 ; GISEL-NEXT:    v_readlane_b32 s63, v40, 31
 ; GISEL-NEXT:    v_readlane_b32 s62, v40, 30
 ; GISEL-NEXT:    v_readlane_b32 s61, v40, 29
@@ -1024,11 +1024,11 @@ define void @test_indirect_call_vgpr_ptr_inreg_arg(ptr %fptr) {
 ; GISEL-NEXT:    v_readlane_b32 s34, v40, 2
 ; GISEL-NEXT:    v_readlane_b32 s31, v40, 1
 ; GISEL-NEXT:    v_readlane_b32 s30, v40, 0
-; GISEL-NEXT:    s_or_saveexec_b64 s[4:5], -1
+; GISEL-NEXT:    s_or_saveexec_b64 s[6:7], -1
 ; GISEL-NEXT:    buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
-; GISEL-NEXT:    s_mov_b64 exec, s[4:5]
+; GISEL-NEXT:    s_mov_b64 exec, s[6:7]
 ; GISEL-NEXT:    s_addk_i32 s32, 0xfc00
-; GISEL-NEXT:    s_mov_b32 s33, s10
+; GISEL-NEXT:    s_mov_b32 s33, s5
 ; GISEL-NEXT:    s_waitcnt vmcnt(0)
 ; GISEL-NEXT:    s_setpc_b64 s[30:31]
   call amdgpu_gfx void %fptr(i32 inreg 123)
diff --git a/llvm/test/CodeGen/AMDGPU/schedule-addrspaces.ll b/llvm/test/CodeGen/AMDGPU/schedule-addrspaces.ll
index 27e9d0fce9426..0139c52db1d58 100644
--- a/llvm/test/CodeGen/AMDGPU/schedule-addrspaces.ll
+++ b/llvm/test/CodeGen/AMDGPU/schedule-addrspaces.ll
@@ -10,7 +10,7 @@ define amdgpu_gfx void @example(<4 x i32> inreg %rsrc, ptr addrspace(5) %src, i3
 ; CHECK-NEXT:    scratch_load_b32 v2, v0, off
 ; CHECK-NEXT:    scratch_load_b32 v3, v3, off
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    buffer_store_b64 v[2:3], v1, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_store_b64 v[2:3], v1, s[4:7], 0 offen
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 
   %x0 = load i32, ptr addrspace(5) %src
diff --git a/llvm/test/CodeGen/AMDGPU/scratch-pointer-sink.ll b/llvm/test/CodeGen/AMDGPU/scratch-pointer-sink.ll
index ec1de0222d5d9..cdaac14833e0e 100644
--- a/llvm/test/CodeGen/AMDGPU/scratch-pointer-sink.ll
+++ b/llvm/test/CodeGen/AMDGPU/scratch-pointer-sink.ll
@@ -6,7 +6,7 @@ define amdgpu_gfx i32 @sink_scratch_pointer(ptr addrspace(5) %stack, i32 inreg %
 ; GCN-LABEL: sink_scratch_pointer:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT:    s_cmp_lg_u32 s0, 0
+; GCN-NEXT:    s_cmp_lg_u32 s4, 0
 ; GCN-NEXT:    s_cbranch_scc0 .LBB0_2
 ; GCN-NEXT:  ; %bb.1: ; %bb2
 ; GCN-NEXT:    scratch_load_b32 v0, v0, off offset:-4
@@ -21,7 +21,7 @@ define amdgpu_gfx i32 @sink_scratch_pointer(ptr addrspace(5) %stack, i32 inreg %
 ; GISEL-LABEL: sink_scratch_pointer:
 ; GISEL:       ; %bb.0:
 ; GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GISEL-NEXT:    s_cmp_lg_u32 s0, 0
+; GISEL-NEXT:    s_cmp_lg_u32 s4, 0
 ; GISEL-NEXT:    s_cbranch_scc0 .LBB0_2
 ; GISEL-NEXT:  ; %bb.1: ; %bb2
 ; GISEL-NEXT:    scratch_load_b32 v0, v0, off offset:-4