Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[AMDGPU] 4-align TTMP triples #132759

Merged
merged 1 commit into from
Mar 24, 2025
Merged

[AMDGPU] 4-align TTMP triples #132759

merged 1 commit into from
Mar 24, 2025

Conversation

jayfoad
Copy link
Contributor

@jayfoad jayfoad commented Mar 24, 2025

Follow up to e4284a7 "[AMDGPU] 4-align SGPR triples".

Previously TTMP triples like ttmp[3:5] were aligned on a 3-TTMP boundary
which has no basis in hardware.

Aligning them on a 4-TTMP boundary matches what we do for SGPRs, which
reduces the number of extra register classes synthesized by TableGen,
bringing the total number down from 653 to 615.

@jayfoad
[AMDGPU] 4-align TTMP triples
009af66 
Follow up to e4284a7 "[AMDGPU] 4-align SGPR triples".

Previously TTMP triples like ttmp[3:5] were aligned on a 3-TTMP boundary
which has no basis in hardware.

Aligning them on a 4-TTMP boundary matches what we do for SGPRs, which
reduces the number of extra register classes synthesized by TableGen,
bringing the total number down from 653 to 615.
jayfoad
jayfoad commented Mar 24, 2025
View reviewed changes
llvm/test/CodeGen/AMDGPU/coalesce-copy-to-agpr-to-av-registers.mir
; CHECK-NEXT: SI_RETURN
%0:vreg_64 = COPY $vgpr0_vgpr1
%1:vreg_64 = COPY $vgpr2_vgpr3
undef %2.sub0_sub1:areg_128 = COPY %0
%2.sub2_sub3:areg_128 = COPY %1
INLINEASM &"; use $0", 0 /* attdialect */, 6225929 /* reguse:AReg_128 */, killed %2
INLINEASM &"; use $0", 0 /* attdialect */, 6029321 /* reguse:AReg_128 */, killed %2
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This was painful to update.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What is the rule behind this? It takes the register number with additional shift and mask right? If we know the rules, we could probably come up with some sort of semi-auto tool.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

or we should just fix this to have real syntax

Copy link
Contributor Author

@jayfoad jayfoad Mar 24, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What is the rule behind this? It takes the register number with additional shift and mask right? If we know the rules, we could probably come up with some sort of semi-auto tool.

Yes, if you know the tablegenerated register class index then the value here is (Idx + 1) << 16 | 9.

@llvmbot
Copy link
Member

llvmbot commented Mar 24, 2025

@llvm/pr-subscribers-backend-amdgpu

Author: Jay Foad (jayfoad)

Changes

Follow up to e4284a7 "[AMDGPU] 4-align SGPR triples".

Previously TTMP triples like ttmp[3:5] were aligned on a 3-TTMP boundary
which has no basis in hardware.

Aligning them on a 4-TTMP boundary matches what we do for SGPRs, which
reduces the number of extra register classes synthesized by TableGen,
bringing the total number down from 653 to 615.

Full diff: https://github.com/llvm/llvm-project/pull/132759.diff

4 Files Affected:

  • (modified) llvm/lib/Target/AMDGPU/SIRegisterInfo.td (+1-1)
  • (modified) llvm/test/CodeGen/AMDGPU/coalesce-copy-to-agpr-to-av-registers.mir (+8-8)
  • (modified) llvm/test/CodeGen/AMDGPU/inline-asm.i128.ll (+12-12)
  • (modified) llvm/test/CodeGen/AMDGPU/partial-regcopy-and-spill-missed-at-regalloc.ll (+6-6)
diff --git a/llvm/lib/Target/AMDGPU/SIRegisterInfo.td b/llvm/lib/Target/AMDGPU/SIRegisterInfo.td
index cd41b5e94902f..35c7b393a8ca4 100644
--- a/llvm/lib/Target/AMDGPU/SIRegisterInfo.td
+++ b/llvm/lib/Target/AMDGPU/SIRegisterInfo.td
@@ -431,7 +431,7 @@ def TTMP_LO16 : SIRegisterClass<"AMDGPU", [i16, f16, bf16], 16,
 def TTMP_64Regs : SIRegisterTuples<getSubRegs<2>.ret, TTMP_32, 15, 2, 2, "ttmp">;
 
 // Trap handler TMP 96-bit registers
-def TTMP_96Regs : SIRegisterTuples<getSubRegs<3>.ret, TTMP_32, 15, 3, 3, "ttmp">;
+def TTMP_96Regs : SIRegisterTuples<getSubRegs<3>.ret, TTMP_32, 15, 4, 3, "ttmp">;
 
 // Trap handler TMP 128-bit registers
 def TTMP_128Regs : SIRegisterTuples<getSubRegs<4>.ret, TTMP_32, 15, 4, 4, "ttmp">;
diff --git a/llvm/test/CodeGen/AMDGPU/coalesce-copy-to-agpr-to-av-registers.mir b/llvm/test/CodeGen/AMDGPU/coalesce-copy-to-agpr-to-av-registers.mir
index e3346918f8fee..8ea5c3ea73e77 100644
--- a/llvm/test/CodeGen/AMDGPU/coalesce-copy-to-agpr-to-av-registers.mir
+++ b/llvm/test/CodeGen/AMDGPU/coalesce-copy-to-agpr-to-av-registers.mir
@@ -129,13 +129,13 @@ body:             |
     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr2_vgpr3
     ; CHECK-NEXT: undef [[COPY2:%[0-9]+]].sub0_sub1:areg_128 = COPY [[COPY]]
     ; CHECK-NEXT: [[COPY2:%[0-9]+]].sub2_sub3:areg_128 = COPY [[COPY1]]
-    ; CHECK-NEXT: INLINEASM &"; use $0", 0 /* attdialect */, 6225929 /* reguse:AReg_128 */, [[COPY2]]
+    ; CHECK-NEXT: INLINEASM &"; use $0", 0 /* attdialect */, 6029321 /* reguse:AReg_128 */, [[COPY2]]
     ; CHECK-NEXT: SI_RETURN
     %0:vreg_64 = COPY $vgpr0_vgpr1
     %1:vreg_64 = COPY $vgpr2_vgpr3
     undef %2.sub0_sub1:areg_128 = COPY %0
     %2.sub2_sub3:areg_128 = COPY %1
-    INLINEASM &"; use $0", 0 /* attdialect */, 6225929 /* reguse:AReg_128 */, killed %2
+    INLINEASM &"; use $0", 0 /* attdialect */, 6029321 /* reguse:AReg_128 */, killed %2
     SI_RETURN
 
 ...
@@ -154,13 +154,13 @@ body:             |
     ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr2_vgpr3
     ; CHECK-NEXT: undef [[COPY2:%[0-9]+]].sub0_sub1:areg_128_align2 = COPY [[COPY]]
     ; CHECK-NEXT: [[COPY2:%[0-9]+]].sub2_sub3:areg_128_align2 = COPY [[COPY1]]
-    ; CHECK-NEXT: INLINEASM &"; use $0", 0 /* attdialect */, 6488073 /* reguse:AReg_128_Align2 */, [[COPY2]]
+    ; CHECK-NEXT: INLINEASM &"; use $0", 0 /* attdialect */, 6291465 /* reguse:AReg_128_Align2 */, [[COPY2]]
     ; CHECK-NEXT: SI_RETURN
     %0:vreg_64 = COPY $vgpr0_vgpr1
     %1:vreg_64 = COPY $vgpr2_vgpr3
     undef %2.sub0_sub1:areg_128_align2 = COPY %0
     %2.sub2_sub3:areg_128_align2 = COPY %1
-    INLINEASM &"; use $0", 0 /* attdialect */, 6488073 /* reguse:AReg_128_Align2 */, %2
+    INLINEASM &"; use $0", 0 /* attdialect */, 6291465 /* reguse:AReg_128_Align2 */, %2
     SI_RETURN
 
 ...
@@ -399,14 +399,14 @@ body:             |
     ; CHECK-NEXT: [[COPY1:%[0-9]+]].sub1:areg_128 = COPY [[COPY]]
     ; CHECK-NEXT: [[COPY1:%[0-9]+]].sub2:areg_128 = COPY [[COPY]]
     ; CHECK-NEXT: [[COPY1:%[0-9]+]].sub3:areg_128 = COPY [[COPY]]
-    ; CHECK-NEXT: INLINEASM &"; use $0", 0 /* attdialect */, 6225929 /* reguse:AReg_128 */, [[COPY1]]
+    ; CHECK-NEXT: INLINEASM &"; use $0", 0 /* attdialect */, 6029321 /* reguse:AReg_128 */, [[COPY1]]
     ; CHECK-NEXT: SI_RETURN
     %0:vgpr_32 = COPY $vgpr0
     undef %1.sub0:areg_128 = COPY %0
     %1.sub1:areg_128 = COPY %0
     %1.sub2:areg_128 = COPY %0
     %1.sub3:areg_128 = COPY %0
-    INLINEASM &"; use $0", 0 /* attdialect */, 6225929 /* reguse:AReg_128 */, killed %1
+    INLINEASM &"; use $0", 0 /* attdialect */, 6029321 /* reguse:AReg_128 */, killed %1
     SI_RETURN
 
 ...
@@ -426,14 +426,14 @@ body:             |
     ; CHECK-NEXT: [[COPY1:%[0-9]+]].sub1:areg_128_align2 = COPY [[COPY]]
     ; CHECK-NEXT: [[COPY1:%[0-9]+]].sub2:areg_128_align2 = COPY [[COPY]]
     ; CHECK-NEXT: [[COPY1:%[0-9]+]].sub3:areg_128_align2 = COPY [[COPY]]
-    ; CHECK-NEXT: INLINEASM &"; use $0", 0 /* attdialect */, 6488073 /* reguse:AReg_128_Align2 */, [[COPY1]]
+    ; CHECK-NEXT: INLINEASM &"; use $0", 0 /* attdialect */, 6291465 /* reguse:AReg_128_Align2 */, [[COPY1]]
     ; CHECK-NEXT: SI_RETURN
     %0:vgpr_32 = COPY $vgpr0
     undef %1.sub0:areg_128_align2 = COPY %0
     %1.sub1:areg_128_align2 = COPY %0
     %1.sub2:areg_128_align2 = COPY %0
     %1.sub3:areg_128_align2 = COPY %0
-    INLINEASM &"; use $0", 0 /* attdialect */, 6488073 /* reguse:AReg_128_Align2 */, %1
+    INLINEASM &"; use $0", 0 /* attdialect */, 6291465 /* reguse:AReg_128_Align2 */, %1
     SI_RETURN
 
 ...
diff --git a/llvm/test/CodeGen/AMDGPU/inline-asm.i128.ll b/llvm/test/CodeGen/AMDGPU/inline-asm.i128.ll
index 2ceaca3497ece..9eb966b4e2a94 100644
--- a/llvm/test/CodeGen/AMDGPU/inline-asm.i128.ll
+++ b/llvm/test/CodeGen/AMDGPU/inline-asm.i128.ll
@@ -8,16 +8,16 @@
 define amdgpu_kernel void @s_input_output_i128() {
   ; GFX908-LABEL: name: s_input_output_i128
   ; GFX908: bb.0 (%ir-block.0):
-  ; GFX908-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 7471114 /* regdef:SGPR_128 */, def %12
+  ; GFX908-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 7208970 /* regdef:SGPR_128 */, def %12
   ; GFX908-NEXT:   [[COPY:%[0-9]+]]:sgpr_128 = COPY %12
-  ; GFX908-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 7471113 /* reguse:SGPR_128 */, [[COPY]]
+  ; GFX908-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 7208969 /* reguse:SGPR_128 */, [[COPY]]
   ; GFX908-NEXT:   S_ENDPGM 0
   ;
   ; GFX90A-LABEL: name: s_input_output_i128
   ; GFX90A: bb.0 (%ir-block.0):
-  ; GFX90A-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 7471114 /* regdef:SGPR_128 */, def %10
+  ; GFX90A-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 7208970 /* regdef:SGPR_128 */, def %10
   ; GFX90A-NEXT:   [[COPY:%[0-9]+]]:sgpr_128 = COPY %10
-  ; GFX90A-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 7471113 /* reguse:SGPR_128 */, [[COPY]]
+  ; GFX90A-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 7208969 /* reguse:SGPR_128 */, [[COPY]]
   ; GFX90A-NEXT:   S_ENDPGM 0
   %val = tail call i128 asm sideeffect "; def $0", "=s"()
   call void asm sideeffect "; use $0", "s"(i128 %val)
@@ -27,16 +27,16 @@ define amdgpu_kernel void @s_input_output_i128() {
 define amdgpu_kernel void @v_input_output_i128() {
   ; GFX908-LABEL: name: v_input_output_i128
   ; GFX908: bb.0 (%ir-block.0):
-  ; GFX908-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6291466 /* regdef:VReg_128 */, def %12
+  ; GFX908-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6094858 /* regdef:VReg_128 */, def %12
   ; GFX908-NEXT:   [[COPY:%[0-9]+]]:vreg_128 = COPY %12
-  ; GFX908-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 6291465 /* reguse:VReg_128 */, [[COPY]]
+  ; GFX908-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 6094857 /* reguse:VReg_128 */, [[COPY]]
   ; GFX908-NEXT:   S_ENDPGM 0
   ;
   ; GFX90A-LABEL: name: v_input_output_i128
   ; GFX90A: bb.0 (%ir-block.0):
-  ; GFX90A-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6619146 /* regdef:VReg_128_Align2 */, def %10
+  ; GFX90A-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6422538 /* regdef:VReg_128_Align2 */, def %10
   ; GFX90A-NEXT:   [[COPY:%[0-9]+]]:vreg_128_align2 = COPY %10
-  ; GFX90A-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 6619145 /* reguse:VReg_128_Align2 */, [[COPY]]
+  ; GFX90A-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 6422537 /* reguse:VReg_128_Align2 */, [[COPY]]
   ; GFX90A-NEXT:   S_ENDPGM 0
   %val = tail call i128 asm sideeffect "; def $0", "=v"()
   call void asm sideeffect "; use $0", "v"(i128 %val)
@@ -46,16 +46,16 @@ define amdgpu_kernel void @v_input_output_i128() {
 define amdgpu_kernel void @a_input_output_i128() {
   ; GFX908-LABEL: name: a_input_output_i128
   ; GFX908: bb.0 (%ir-block.0):
-  ; GFX908-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6225930 /* regdef:AReg_128 */, def %12
+  ; GFX908-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6029322 /* regdef:AReg_128 */, def %12
   ; GFX908-NEXT:   [[COPY:%[0-9]+]]:areg_128 = COPY %12
-  ; GFX908-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 6225929 /* reguse:AReg_128 */, [[COPY]]
+  ; GFX908-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 6029321 /* reguse:AReg_128 */, [[COPY]]
   ; GFX908-NEXT:   S_ENDPGM 0
   ;
   ; GFX90A-LABEL: name: a_input_output_i128
   ; GFX90A: bb.0 (%ir-block.0):
-  ; GFX90A-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6488074 /* regdef:AReg_128_Align2 */, def %10
+  ; GFX90A-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6291466 /* regdef:AReg_128_Align2 */, def %10
   ; GFX90A-NEXT:   [[COPY:%[0-9]+]]:areg_128_align2 = COPY %10
-  ; GFX90A-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 6488073 /* reguse:AReg_128_Align2 */, [[COPY]]
+  ; GFX90A-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 6291465 /* reguse:AReg_128_Align2 */, [[COPY]]
   ; GFX90A-NEXT:   S_ENDPGM 0
   %val = call i128 asm sideeffect "; def $0", "=a"()
   call void asm sideeffect "; use $0", "a"(i128 %val)
diff --git a/llvm/test/CodeGen/AMDGPU/partial-regcopy-and-spill-missed-at-regalloc.ll b/llvm/test/CodeGen/AMDGPU/partial-regcopy-and-spill-missed-at-regalloc.ll
index 106824a085b42..1a34fa3bbbf4d 100644
--- a/llvm/test/CodeGen/AMDGPU/partial-regcopy-and-spill-missed-at-regalloc.ll
+++ b/llvm/test/CodeGen/AMDGPU/partial-regcopy-and-spill-missed-at-regalloc.ll
@@ -11,7 +11,7 @@ define amdgpu_kernel void @partial_copy(<4 x i32> %arg) #0 {
   ; REGALLOC-GFX908-NEXT:   liveins: $sgpr4_sgpr5
   ; REGALLOC-GFX908-NEXT: {{  $}}
   ; REGALLOC-GFX908-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 2162697 /* reguse:AGPR_32 */, undef %5:agpr_32
-  ; REGALLOC-GFX908-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6291466 /* regdef:VReg_128 */, def %6
+  ; REGALLOC-GFX908-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6094858 /* regdef:VReg_128 */, def %6
   ; REGALLOC-GFX908-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 3538954 /* regdef:VReg_64 */, def %7
   ; REGALLOC-GFX908-NEXT:   GLOBAL_STORE_DWORDX4 undef %14:vreg_64, %6, 0, 0, implicit $exec :: (volatile store (s128) into `ptr addrspace(1) poison`, addrspace 1)
   ; REGALLOC-GFX908-NEXT:   renamable $sgpr0_sgpr1_sgpr2_sgpr3 = S_LOAD_DWORDX4_IMM killed renamable $sgpr4_sgpr5, 0, 0 :: (dereferenceable invariant load (s128) from %ir.arg.kernarg.offset1, addrspace 4)
@@ -20,8 +20,8 @@ define amdgpu_kernel void @partial_copy(<4 x i32> %arg) #0 {
   ; REGALLOC-GFX908-NEXT:   [[V_MOV_B32_e32_1:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 2, implicit $exec
   ; REGALLOC-GFX908-NEXT:   [[V_MFMA_I32_4X4X4I8_e64_:%[0-9]+]]:areg_128 = V_MFMA_I32_4X4X4I8_e64 [[V_MOV_B32_e32_]], [[V_MOV_B32_e32_1]], [[COPY]], 0, 0, 0, implicit $mode, implicit $exec
   ; REGALLOC-GFX908-NEXT:   GLOBAL_STORE_DWORDX2 undef %16:vreg_64, %7, 0, 0, implicit $exec :: (volatile store (s64) into `ptr addrspace(1) poison`, addrspace 1)
-  ; REGALLOC-GFX908-NEXT:   [[COPY3:%[0-9]+]]:vreg_128 = COPY [[V_MFMA_I32_4X4X4I8_e64_]]
-  ; REGALLOC-GFX908-NEXT:   GLOBAL_STORE_DWORDX4 undef %18:vreg_64, [[COPY3]], 0, 0, implicit $exec :: (volatile store (s128) into `ptr addrspace(1) poison`, addrspace 1)
+  ; REGALLOC-GFX908-NEXT:   [[COPY1:%[0-9]+]]:vreg_128 = COPY [[V_MFMA_I32_4X4X4I8_e64_]]
+  ; REGALLOC-GFX908-NEXT:   GLOBAL_STORE_DWORDX4 undef %18:vreg_64, [[COPY1]], 0, 0, implicit $exec :: (volatile store (s128) into `ptr addrspace(1) poison`, addrspace 1)
   ; REGALLOC-GFX908-NEXT:   S_ENDPGM 0
   ;
   ; PEI-GFX908-LABEL: name: partial_copy
@@ -32,7 +32,7 @@ define amdgpu_kernel void @partial_copy(<4 x i32> %arg) #0 {
   ; PEI-GFX908-NEXT:   $sgpr8 = S_ADD_U32 $sgpr8, $sgpr7, implicit-def $scc, implicit-def $sgpr8_sgpr9_sgpr10_sgpr11
   ; PEI-GFX908-NEXT:   $sgpr9 = S_ADDC_U32 $sgpr9, 0, implicit-def dead $scc, implicit $scc, implicit-def $sgpr8_sgpr9_sgpr10_sgpr11
   ; PEI-GFX908-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 2162697 /* reguse:AGPR_32 */, undef renamable $agpr0
-  ; PEI-GFX908-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6291466 /* regdef:VReg_128 */, def renamable $vgpr0_vgpr1_vgpr2_vgpr3
+  ; PEI-GFX908-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6094858 /* regdef:VReg_128 */, def renamable $vgpr0_vgpr1_vgpr2_vgpr3
   ; PEI-GFX908-NEXT:   renamable $agpr0_agpr1_agpr2_agpr3 = COPY killed renamable $vgpr0_vgpr1_vgpr2_vgpr3, implicit $exec
   ; PEI-GFX908-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 3538954 /* regdef:VReg_64 */, def renamable $vgpr0_vgpr1
   ; PEI-GFX908-NEXT:   BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr8_sgpr9_sgpr10_sgpr11, 0, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1, implicit $vgpr0_vgpr1 :: (store (s32) into %stack.0, addrspace 5)
@@ -56,7 +56,7 @@ define amdgpu_kernel void @partial_copy(<4 x i32> %arg) #0 {
   ; REGALLOC-GFX90A-NEXT:   liveins: $sgpr4_sgpr5
   ; REGALLOC-GFX90A-NEXT: {{  $}}
   ; REGALLOC-GFX90A-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 2162697 /* reguse:AGPR_32 */, undef %5:agpr_32
-  ; REGALLOC-GFX90A-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6619146 /* regdef:VReg_128_Align2 */, def %6
+  ; REGALLOC-GFX90A-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6422538 /* regdef:VReg_128_Align2 */, def %6
   ; REGALLOC-GFX90A-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 3866634 /* regdef:VReg_64_Align2 */, def %7
   ; REGALLOC-GFX90A-NEXT:   GLOBAL_STORE_DWORDX4 undef %14:vreg_64_align2, %6, 0, 0, implicit $exec :: (volatile store (s128) into `ptr addrspace(1) poison`, addrspace 1)
   ; REGALLOC-GFX90A-NEXT:   renamable $sgpr0_sgpr1_sgpr2_sgpr3 = S_LOAD_DWORDX4_IMM killed renamable $sgpr4_sgpr5, 0, 0 :: (dereferenceable invariant load (s128) from %ir.arg.kernarg.offset1, addrspace 4)
@@ -76,7 +76,7 @@ define amdgpu_kernel void @partial_copy(<4 x i32> %arg) #0 {
   ; PEI-GFX90A-NEXT:   $sgpr8 = S_ADD_U32 $sgpr8, $sgpr7, implicit-def $scc, implicit-def $sgpr8_sgpr9_sgpr10_sgpr11
   ; PEI-GFX90A-NEXT:   $sgpr9 = S_ADDC_U32 $sgpr9, 0, implicit-def dead $scc, implicit $scc, implicit-def $sgpr8_sgpr9_sgpr10_sgpr11
   ; PEI-GFX90A-NEXT:   INLINEASM &"; use $0", 1 /* sideeffect attdialect */, 2162697 /* reguse:AGPR_32 */, undef renamable $agpr0
-  ; PEI-GFX90A-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6619146 /* regdef:VReg_128_Align2 */, def renamable $vgpr0_vgpr1_vgpr2_vgpr3
+  ; PEI-GFX90A-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 6422538 /* regdef:VReg_128_Align2 */, def renamable $vgpr0_vgpr1_vgpr2_vgpr3
   ; PEI-GFX90A-NEXT:   renamable $agpr0_agpr1_agpr2_agpr3 = COPY killed renamable $vgpr0_vgpr1_vgpr2_vgpr3, implicit $exec
   ; PEI-GFX90A-NEXT:   INLINEASM &"; def $0", 1 /* sideeffect attdialect */, 3866634 /* regdef:VReg_64_Align2 */, def renamable $vgpr0_vgpr1
   ; PEI-GFX90A-NEXT:   BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr8_sgpr9_sgpr10_sgpr11, 0, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1, implicit $vgpr0_vgpr1 :: (store (s32) into %stack.0, addrspace 5)

@arsenm
Copy link
Contributor

arsenm commented Mar 24, 2025

Aligning them on a 4-TTMP boundary matches what we do for SGPRs

We probably should have the unaligned SGPR classes

@jayfoad
Copy link
Contributor Author

jayfoad commented Mar 24, 2025

Aligning them on a 4-TTMP boundary matches what we do for SGPRs

We probably should have the unaligned SGPR classes

There is no hardware use for them. To quote the (RDNA4) ISA doc, "Quad-alignment of SGPRs is required for operations on more than 64-bits".

@arsenm
Copy link
Contributor

arsenm commented Mar 24, 2025

There is no hardware use for them. To quote the (RDNA4) ISA doc, "Quad-alignment of SGPRs is required for operations on more than 64-bits".

Yes, there is no hardware use. But we can do an unaligned copy from a real value to a synthetic unaligned class and potentially avoid a spill

@jayfoad
Copy link
Contributor Author

jayfoad commented Mar 24, 2025

There is no hardware use for them. To quote the (RDNA4) ISA doc, "Quad-alignment of SGPRs is required for operations on more than 64-bits".

Yes, there is no hardware use. But we can do an unaligned copy from a real value to a synthetic unaligned class and potentially avoid a spill

I see. I'm not thrilled about the idea of creating artificial register classes like TTMP_96_Unaligned just for the benefit of artificial instructions like S_MOV_B96_pseudo. It seems like there should be a better way.

@jayfoad jayfoad merged commit 02ed659 into llvm:main Mar 24, 2025
13 checks passed
@jayfoad jayfoad deleted the ttmp-triple branch March 24, 2025 17:13
@arsenm
Copy link
Contributor

arsenm commented Mar 25, 2025

for the benefit of artificial instructions like S_MOV_B96_pseudo. It seems like there should be a better way.

It's for COPY and spills

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@shiltian shiltian

@arsenm arsenm

Assignees
No one assigned
Labels
backend:AMDGPU
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
4 participants
@jayfoad @llvmbot @arsenm @shiltian