[AMDGPU] Allow narrowing muti-dword loads

Currently BE allows only a little load narrowing because
of the fear it will produce sub-dword ext loads. However,
we can always allow narrowing if we are shrinking one
multi-dword load to another multi-dword load.

In particular we were unable to reduce s_load_dwordx8 into
s_load_dwordx4 if identity shuffle was used to extract
low 4 dwords.

Differential Revision: https://reviews.llvm.org/D73133

llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp

@@ -641,8 +641,9 @@ bool AMDGPUTargetLowering::shouldReduceLoadWidth(SDNode *N,
 
   unsigned NewSize = NewVT.getStoreSizeInBits();
 
-  // If we are reducing to a 32-bit load, this is always better.
-  if (NewSize == 32)
+  // If we are reducing to a 32-bit load or a smaller multi-dword load,
+  // this is always better.
+  if (NewSize >= 32)
     return true;
 
   EVT OldVT = N->getValueType(0);

llvm/test/CodeGen/AMDGPU/vector_shuffle.packed.ll

@@ -629,7 +629,7 @@ define <4 x half> @shuffle_v8f16_0123(<8 x half> addrspace(1)* %arg0, <8 x half>
 ; GFX9-LABEL: shuffle_v8f16_0123:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_load_dwordx4 v[0:3], v[0:1], off
+; GFX9-NEXT:    global_load_dwordx2 v[0:1], v[0:1], off
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
   %val0 = load <8 x half>, <8 x half> addrspace(1)* %arg0
@@ -656,8 +656,10 @@ define <4 x half> @shuffle_v8f16_10_11_2_3(<8 x half> addrspace(1)* %arg0, <8 x
 ; GFX9-LABEL: shuffle_v8f16_10_11_2_3:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v2, v[2:3], off offset:4
 ; GFX9-NEXT:    global_load_dword v1, v[0:1], off offset:4
-; GFX9-NEXT:    global_load_dword v0, v[2:3], off offset:4
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_mov_b32_e32 v0, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
   %val0 = load <8 x half>, <8 x half> addrspace(1)* %arg0
@@ -670,12 +672,12 @@ define <4 x half> @shuffle_v8f16_13_14_2_3(<8 x half> addrspace(1)* %arg0, <8 x
 ; GFX9-LABEL: shuffle_v8f16_13_14_2_3:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_load_dwordx4 v[2:5], v[2:3], off
+; GFX9-NEXT:    global_load_dwordx2 v[2:3], v[2:3], off offset:8
 ; GFX9-NEXT:    global_load_dword v1, v[0:1], off offset:4
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0xffff
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_and_b32_sdwa v0, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; GFX9-NEXT:    v_lshl_or_b32 v0, v5, 16, v0
+; GFX9-NEXT:    v_and_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX9-NEXT:    v_lshl_or_b32 v0, v3, 16, v0
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
   %val0 = load <8 x half>, <8 x half> addrspace(1)* %arg0
@@ -809,6 +811,27 @@ define <4 x half> @shuffle_v4f16_0456(<4 x half> addrspace(1)* %arg0, <4 x half>
   ret <4 x half> %shuffle
 }
 
+define amdgpu_kernel void @shuffle_scalar_load_v8i32_0123(<8 x i32> addrspace(4)* %in, <4 x i32> addrspace(1)* %out)  {
+; GFX9-LABEL: shuffle_scalar_load_v8i32_0123:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x0
+; GFX9-NEXT:    v_mov_b32_e32 v4, s2
+; GFX9-NEXT:    v_mov_b32_e32 v5, s3
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:    global_store_dwordx4 v[4:5], v[0:3], off
+; GFX9-NEXT:    s_endpgm
+  %ld8 = load <8 x i32>, <8 x i32> addrspace(4)* %in, align 16
+  %id = shufflevector <8 x i32> %ld8, <8 x i32> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  store <4 x i32> %id, <4 x i32> addrspace(1)* %out, align 8
+  ret void
+}
+
 declare <2 x half> @llvm.fma.v2f16(<2 x half>, <2 x half>, <2 x half>) #0
 declare i32 @llvm.amdgcn.workitem.id.x() #0