[llvm] AMDGPU: Mark the SCC def dead when expanding 64-bit scalar add/sub (PR #226269)

Matt Arsenault via llvm-commits llvm-commits at lists.llvm.org
Fri Sep 25 03:47:28 PDT 2026


================
@@ -68,3 +68,37 @@ body: |
     %0:sreg_64 = IMPLICIT_DEF
     %1:sreg_64 = S_ADD_U64_PSEUDO %0, 6565, implicit-def $scc
 ...
+
+---
+name: preserve_dead
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; GFX11-LABEL: name: preserve_dead
+    ; GFX11: [[DEF:%[0-9]+]]:sreg_64 = IMPLICIT_DEF
+    ; GFX11-NEXT: [[DEF1:%[0-9]+]]:sreg_64 = IMPLICIT_DEF
+    ; GFX11-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY [[DEF]].sub0
+    ; GFX11-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY [[DEF]].sub1
+    ; GFX11-NEXT: [[COPY2:%[0-9]+]]:sreg_32 = COPY [[DEF1]].sub0
+    ; GFX11-NEXT: [[COPY3:%[0-9]+]]:sreg_32 = COPY [[DEF1]].sub1
+    ; GFX11-NEXT: [[S_ADD_U32_:%[0-9]+]]:sreg_32 = S_ADD_U32 [[COPY]], [[COPY2]], implicit-def $scc
+    ; GFX11-NEXT: [[S_ADDC_U32_:%[0-9]+]]:sreg_32 = S_ADDC_U32 [[COPY1]], [[COPY3]], implicit-def dead $scc, implicit $scc
+    ; GFX11-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[S_ADD_U32_]], %subreg.sub0, [[S_ADDC_U32_]], %subreg.sub1
+    ; GFX11-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY [[DEF]].sub0
+    ; GFX11-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY [[DEF]].sub1
+    ; GFX11-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY [[DEF1]].sub0
+    ; GFX11-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY [[DEF1]].sub1
+    ; GFX11-NEXT: [[S_SUB_U32_:%[0-9]+]]:sreg_32 = S_SUB_U32 [[COPY4]], [[COPY6]], implicit-def $scc
+    ; GFX11-NEXT: [[S_SUBB_U32_:%[0-9]+]]:sreg_32 = S_SUBB_U32 [[COPY5]], [[COPY7]], implicit-def dead $scc, implicit $scc
+    ; GFX11-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[S_SUB_U32_]], %subreg.sub0, [[S_SUBB_U32_]], %subreg.sub1
+    ;
+    ; GFX12-LABEL: name: preserve_dead
+    ; GFX12: [[DEF:%[0-9]+]]:sreg_64 = IMPLICIT_DEF
+    ; GFX12-NEXT: [[DEF1:%[0-9]+]]:sreg_64 = IMPLICIT_DEF
+    ; GFX12-NEXT: [[S_ADD_U64_:%[0-9]+]]:sreg_64 = S_ADD_U64 [[DEF]], [[DEF1]]
+    ; GFX12-NEXT: [[S_SUB_U64_:%[0-9]+]]:sreg_64 = S_SUB_U64 [[DEF]], [[DEF1]]
----------------
arsenm wrote:

Yes, but the pseudos do. Ideally we would only use the pseudos when when the instruction isn't available 

https://github.com/llvm/llvm-project/pull/226269


More information about the llvm-commits mailing list