[llvm] [AMDGPU] Pre-commit tests for folding SGPR->VGPR copies escaping a divergent loop (NFC) (PR #203502)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Jun 12 03:56:55 PDT 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: Arseniy Obolenskiy (aobolensk)
<details>
<summary>Changes</summary>
pre-commit for https://github.com/llvm/llvm-project/pull/203256
---
Full diff: https://github.com/llvm/llvm-project/pull/203502.diff
1 Files Affected:
- (modified) llvm/test/CodeGen/AMDGPU/do-not-fold-copy.mir (+128)
``````````diff
diff --git a/llvm/test/CodeGen/AMDGPU/do-not-fold-copy.mir b/llvm/test/CodeGen/AMDGPU/do-not-fold-copy.mir
index 5c206da8c544f..2beaf9a2bb2e4 100644
--- a/llvm/test/CodeGen/AMDGPU/do-not-fold-copy.mir
+++ b/llvm/test/CodeGen/AMDGPU/do-not-fold-copy.mir
@@ -54,3 +54,131 @@ body: |
%11:vgpr_32 = V_SET_INACTIVE_B32 0, %9, 0, 0, killed %10, implicit $exec
S_ENDPGM 0
...
+
+# An SGPR->VGPR copy with no implicit $exec read, inserted in a divergent loop
+# to latch a per-lane value, is read after the loop. SIFoldOperands currently
+# folds the scalar source into that exit use, dropping the per-lane snapshot;
+# this is wrong because the value escapes the loop. FIXME: should not fold.
+---
+name: do_not_fold_sgpr_to_vgpr_copy_escaping_loop
+tracksRegLiveness: true
+body: |
+ ; CHECK-LABEL: name: do_not_fold_sgpr_to_vgpr_copy_escaping_loop
+ ; CHECK: bb.0:
+ ; CHECK-NEXT: successors: %bb.1(0x80000000)
+ ; CHECK-NEXT: liveins: $sgpr0, $vgpr0
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+ ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+ ; CHECK-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64 = S_MOV_B64 0
+ ; CHECK-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 0
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: bb.1:
+ ; CHECK-NEXT: successors: %bb.2(0x04000000), %bb.1(0x7c000000)
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: [[PHI:%[0-9]+]]:sreg_64 = PHI [[S_MOV_B64_]], %bb.0, %5, %bb.1
+ ; CHECK-NEXT: [[PHI1:%[0-9]+]]:sreg_32 = PHI [[S_MOV_B32_]], %bb.0, %7, %bb.1
+ ; CHECK-NEXT: [[S_XOR_B32_:%[0-9]+]]:sreg_32 = S_XOR_B32 [[COPY]], [[PHI1]], implicit-def dead $scc
+ ; CHECK-NEXT: [[S_ADD_I32_:%[0-9]+]]:sreg_32 = S_ADD_I32 [[PHI1]], 1, implicit-def dead $scc
+ ; CHECK-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64 = V_CMP_EQ_U32_e64 [[COPY1]], [[S_ADD_I32_]], implicit $exec
+ ; CHECK-NEXT: [[SI_IF_BREAK:%[0-9]+]]:sreg_64 = SI_IF_BREAK [[V_CMP_EQ_U32_e64_]], [[PHI]], implicit-def dead $scc
+ ; CHECK-NEXT: SI_LOOP [[SI_IF_BREAK]], %bb.1, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ ; CHECK-NEXT: S_BRANCH %bb.2
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: bb.2:
+ ; CHECK-NEXT: SI_END_CF [[SI_IF_BREAK]], implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ ; CHECK-NEXT: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[S_XOR_B32_]], 1, 0, implicit $exec
+ ; CHECK-NEXT: $vgpr0 = COPY [[V_ADD_U32_e64_]]
+ ; CHECK-NEXT: SI_RETURN implicit $vgpr0
+ bb.0:
+ successors: %bb.1
+ liveins: $sgpr0, $vgpr0
+
+ %0:sreg_32 = COPY $sgpr0
+ %1:vgpr_32 = COPY $vgpr0
+ %2:sreg_64 = S_MOV_B64 0
+ %3:sreg_32 = S_MOV_B32 0
+
+ bb.1:
+ successors: %bb.2(0x04000000), %bb.1(0x7c000000)
+
+ %4:sreg_64 = PHI %2, %bb.0, %5, %bb.1
+ %6:sreg_32 = PHI %3, %bb.0, %7, %bb.1
+ %8:sreg_32 = S_XOR_B32 %0, %6, implicit-def dead $scc
+ %9:vgpr_32 = COPY %8
+ %7:sreg_32 = S_ADD_I32 %6, 1, implicit-def dead $scc
+ %10:sreg_64 = V_CMP_EQ_U32_e64 %1, %7, implicit $exec
+ %5:sreg_64 = SI_IF_BREAK %10, %4, implicit-def dead $scc
+ SI_LOOP %5, %bb.1, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ S_BRANCH %bb.2
+
+ bb.2:
+ SI_END_CF %5, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ %11:vgpr_32 = V_ADD_U32_e64 %9, 1, 0, implicit $exec
+ $vgpr0 = COPY %11
+ SI_RETURN implicit $vgpr0
+...
+
+# Same latch, but the loop-exit use is itself a COPY. The scalar source is
+# currently propagated through that exit copy too (separate fold path).
+# FIXME: should not fold.
+---
+name: do_not_fold_sgpr_to_vgpr_copy_escaping_loop_via_copy_use
+tracksRegLiveness: true
+body: |
+ ; CHECK-LABEL: name: do_not_fold_sgpr_to_vgpr_copy_escaping_loop_via_copy_use
+ ; CHECK: bb.0:
+ ; CHECK-NEXT: successors: %bb.1(0x80000000)
+ ; CHECK-NEXT: liveins: $sgpr0, $vgpr0
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+ ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+ ; CHECK-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64 = S_MOV_B64 0
+ ; CHECK-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 0
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: bb.1:
+ ; CHECK-NEXT: successors: %bb.2(0x04000000), %bb.1(0x7c000000)
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: [[PHI:%[0-9]+]]:sreg_64 = PHI [[S_MOV_B64_]], %bb.0, %5, %bb.1
+ ; CHECK-NEXT: [[PHI1:%[0-9]+]]:sreg_32 = PHI [[S_MOV_B32_]], %bb.0, %7, %bb.1
+ ; CHECK-NEXT: [[S_XOR_B32_:%[0-9]+]]:sreg_32 = S_XOR_B32 [[COPY]], [[PHI1]], implicit-def dead $scc
+ ; CHECK-NEXT: [[S_ADD_I32_:%[0-9]+]]:sreg_32 = S_ADD_I32 [[PHI1]], 1, implicit-def dead $scc
+ ; CHECK-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64 = V_CMP_EQ_U32_e64 [[COPY1]], [[S_ADD_I32_]], implicit $exec
+ ; CHECK-NEXT: [[SI_IF_BREAK:%[0-9]+]]:sreg_64 = SI_IF_BREAK [[V_CMP_EQ_U32_e64_]], [[PHI]], implicit-def dead $scc
+ ; CHECK-NEXT: SI_LOOP [[SI_IF_BREAK]], %bb.1, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ ; CHECK-NEXT: S_BRANCH %bb.2
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: bb.2:
+ ; CHECK-NEXT: SI_END_CF [[SI_IF_BREAK]], implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ ; CHECK-NEXT: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[S_XOR_B32_]], 1, 0, implicit $exec
+ ; CHECK-NEXT: $vgpr0 = COPY [[V_ADD_U32_e64_]]
+ ; CHECK-NEXT: SI_RETURN implicit $vgpr0
+ bb.0:
+ successors: %bb.1
+ liveins: $sgpr0, $vgpr0
+
+ %0:sreg_32 = COPY $sgpr0
+ %1:vgpr_32 = COPY $vgpr0
+ %2:sreg_64 = S_MOV_B64 0
+ %3:sreg_32 = S_MOV_B32 0
+
+ bb.1:
+ successors: %bb.2(0x04000000), %bb.1(0x7c000000)
+
+ %4:sreg_64 = PHI %2, %bb.0, %5, %bb.1
+ %6:sreg_32 = PHI %3, %bb.0, %7, %bb.1
+ %8:sreg_32 = S_XOR_B32 %0, %6, implicit-def dead $scc
+ %9:vgpr_32 = COPY %8
+ %7:sreg_32 = S_ADD_I32 %6, 1, implicit-def dead $scc
+ %10:sreg_64 = V_CMP_EQ_U32_e64 %1, %7, implicit $exec
+ %5:sreg_64 = SI_IF_BREAK %10, %4, implicit-def dead $scc
+ SI_LOOP %5, %bb.1, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ S_BRANCH %bb.2
+
+ bb.2:
+ SI_END_CF %5, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+ %11:vgpr_32 = COPY %9
+ %12:vgpr_32 = V_ADD_U32_e64 %11, 1, 0, implicit $exec
+ $vgpr0 = COPY %12
+ SI_RETURN implicit $vgpr0
+...
``````````
</details>
https://github.com/llvm/llvm-project/pull/203502
More information about the llvm-commits
mailing list