[PATCH] D90997: [AMDGPU] SIWholeQuadMode fix mode insertion when SCC always defined
Carl Ritson via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Fri Nov 6 23:18:08 PST 2020
critson created this revision.
critson added reviewers: nhaehnle, mceier.
Herald added subscribers: llvm-commits, kerbowa, hiraditya, t-tye, tpr, dstuttard, yaxunl, jvesely, kzhuravl, arsenm.
Herald added a project: LLVM.
critson requested review of this revision.
Herald added a subscriber: wdng.
Fix a crash when SCC is defined until end of block and mode change
must be inserted in SCC live region.
Repository:
rG LLVM Github Monorepo
https://reviews.llvm.org/D90997
Files:
llvm/lib/Target/AMDGPU/SIWholeQuadMode.cpp
llvm/test/CodeGen/AMDGPU/wqm.mir
Index: llvm/test/CodeGen/AMDGPU/wqm.mir
===================================================================
--- llvm/test/CodeGen/AMDGPU/wqm.mir
+++ llvm/test/CodeGen/AMDGPU/wqm.mir
@@ -142,3 +142,60 @@
S_ENDPGM 0
...
+
+---
+# Check exit of WQM is still inserted correctly when SCC is live until block end.
+# Critially this tests that compilation does not fail.
+#CHECK-LABEL: name: scc_always_live
+#CHECK: %8:vreg_128 = IMAGE_SAMPLE_V4_V2 %7
+#CHECK-NEXT: S_CMP_EQ_U32 %2, 0, implicit-def $scc
+#CHECK-NEXT: undef %9.sub0:vreg_64 = nsz arcp nofpexcept V_ADD_F32_e64
+#CHECK-NEXT: %9.sub1:vreg_64 = nsz arcp nofpexcept V_MUL_F32_e32
+#CHECK-NEXT: %14:sreg_32_xm0 = COPY $scc
+#CHECK-NEXT: $exec = S_AND_B64 $exec, %13, implicit-def $scc
+#CHECK-NEXT: $scc = COPY %14
+#CHECK-NEXT: %10:vgpr_32 = nsz arcp nofpexcept V_ADD_F32_e64
+#CHECK-NEXT: %11:vreg_128 = IMAGE_SAMPLE_V4_V2
+#CHECK-NEXT: S_CBRANCH_SCC0 %bb.2
+name: scc_always_live
+tracksRegLiveness: true
+body: |
+ bb.0:
+ liveins: $sgpr1, $sgpr2, $vgpr1, $vgpr2
+
+ $m0 = COPY $sgpr1
+ %0:vgpr_32 = COPY $vgpr1
+ %1:vgpr_32 = COPY $vgpr2
+ %8:sgpr_32 = COPY $sgpr2
+ %100:sgpr_256 = IMPLICIT_DEF
+ %101:sgpr_128 = IMPLICIT_DEF
+
+ %2:vgpr_32 = V_INTERP_P1_F32 %0:vgpr_32, 3, 2, implicit $mode, implicit $m0, implicit $exec
+ %3:vgpr_32 = V_INTERP_P1_F32 %1:vgpr_32, 3, 2, implicit $mode, implicit $m0, implicit $exec
+
+ undef %7.sub0:vreg_64 = COPY %2:vgpr_32
+ %7.sub1:vreg_64 = COPY %3:vgpr_32
+
+ %4:vreg_128 = IMAGE_SAMPLE_V4_V2 %7:vreg_64, %100:sgpr_256, %101:sgpr_128, 15, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16, align 4, addrspace 4)
+ S_CMP_EQ_U32 %8:sgpr_32, 0, implicit-def $scc
+
+ undef %5.sub0:vreg_64 = nsz arcp nofpexcept V_ADD_F32_e64 0, %4.sub0:vreg_128, 0, %3:vgpr_32, 1, 0, implicit $mode, implicit $exec
+ %5.sub1:vreg_64 = nsz arcp nofpexcept V_MUL_F32_e32 %2, %3, implicit $mode, implicit $exec
+ %6:vgpr_32 = nsz arcp nofpexcept V_ADD_F32_e64 0, %2:vgpr_32, 0, %3:vgpr_32, 1, 0, implicit $mode, implicit $exec
+
+ %9:vreg_128 = IMAGE_SAMPLE_V4_V2 %5:vreg_64, %100:sgpr_256, %101:sgpr_128, 15, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16, align 4, addrspace 4)
+
+ S_CBRANCH_SCC0 %bb.2, implicit $scc
+
+ bb.1:
+ %10:sreg_32 = S_MOV_B32 0
+ BUFFER_STORE_DWORD_OFFSET_exact %6:vgpr_32, %101:sgpr_128, %10:sreg_32, 4, 0, 0, 0, 0, 0, implicit $exec
+ S_ENDPGM 0
+
+ bb.2:
+ $vgpr0 = COPY %4.sub0:vreg_128
+ $vgpr1 = COPY %4.sub1:vreg_128
+ $vgpr2 = COPY %9.sub0:vreg_128
+ $vgpr3 = COPY %9.sub1:vreg_128
+ SI_RETURN_TO_EPILOG $vgpr0, $vgpr1, $vgpr2, $vgpr3
+...
Index: llvm/lib/Target/AMDGPU/SIWholeQuadMode.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/SIWholeQuadMode.cpp
+++ llvm/lib/Target/AMDGPU/SIWholeQuadMode.cpp
@@ -653,6 +653,8 @@
MachineInstr *EndMI = LIS->getInstructionFromIndex(S->end.getBaseIndex());
assert(EndMI && "Segment does not end on valid instruction");
auto NextI = std::next(EndMI->getIterator());
+ if (NextI == MBB.end())
+ break;
SlotIndex Next = LIS->getInstructionIndex(*NextI);
if (Next > LastIdx)
break;
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D90997.303618.patch
Type: text/x-patch
Size: 3289 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20201107/564c5aeb/attachment.bin>
More information about the llvm-commits
mailing list