[SIFoldOperands] Folding immediate into a copy invalidates candidates in the fold list

Fixes SWDEV-542372

Author: jmmartinez

llvm/lib/Target/AMDGPU/SIFoldOperands.cpp

@@ -1761,7 +1761,9 @@ bool SIFoldOperandsImpl::foldInstOperand(MachineInstr &MI,
   for (MachineInstr *Copy : CopiesToReplace)
     Copy->addImplicitDefUseOperands(*MF);
 
-  for (FoldCandidate &Fold : FoldList) {
+  for (auto FoldIt = FoldList.begin(), End = FoldList.end(); FoldIt != End;
+       ++FoldIt) {
+    FoldCandidate &Fold = *FoldIt;
     assert(!Fold.isReg() || Fold.Def.OpToFold);
     if (Fold.isReg() && Fold.getReg().isVirtual()) {
       Register Reg = Fold.getReg();
@@ -1785,9 +1787,13 @@ bool SIFoldOperandsImpl::foldInstOperand(MachineInstr &MI,
 
       if (Fold.isImm() && tryConstantFoldOp(Fold.UseMI)) {
         LLVM_DEBUG(dbgs() << "Constant folded " << *Fold.UseMI);
+        // The instruction was folded into a copy, we have to skip any other
+        // occurence of UseMI in the fold list
+        End = std::remove_if(FoldIt + 1, End, [&](const FoldCandidate &F) {
+          return F.UseMI == Fold.UseMI;
+        });
         Changed = true;
       }
-
     } else if (Fold.Commuted) {
       // Restoring instruction's original operand order if fold has failed.
       TII->commuteInstruction(*Fold.UseMI, false);

llvm/test/CodeGen/AMDGPU/si-fold-operands-swdev-542372.ll

@@ -1,6 +1,29 @@
-; RUN: not --crash llc -O3 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a %s -o /dev/null
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -O3 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a %s -o - | FileCheck %s
 
 define amdgpu_kernel void @kernel() {
+; CHECK-LABEL: kernel:
+; CHECK:       ; %bb.0: ; %entry
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
+; CHECK-NEXT:    ds_read_b64 v[2:3], v0
+; CHECK-NEXT:    ds_write_b32 v0, v0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(1)
+; CHECK-NEXT:    v_cmp_ne_u32_e32 vcc, 2, v2
+; CHECK-NEXT:    s_cbranch_vccnz .LBB0_3
+; CHECK-NEXT:  ; %bb.1: ; %land.rhs49
+; CHECK-NEXT:    ds_read_b64 v[0:1], v0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    v_cmp_ne_u32_e32 vcc, 1, v0
+; CHECK-NEXT:    s_cbranch_vccnz .LBB0_3
+; CHECK-NEXT:  ; %bb.2: ; %land.rhs57
+; CHECK-NEXT:    s_mov_b32 s0, 0
+; CHECK-NEXT:    s_mov_b32 s1, s0
+; CHECK-NEXT:    v_mov_b32_e32 v2, 0
+; CHECK-NEXT:    v_pk_mov_b32 v[0:1], s[0:1], s[0:1] op_sel:[0,1]
+; CHECK-NEXT:    s_cmp_lg_u32 s0, 0
+; CHECK-NEXT:    ds_write_b64 v2, v[0:1]
+; CHECK-NEXT:  .LBB0_3: ; %land.end59
+; CHECK-NEXT:    s_endpgm
 entry:
   %0 = load <2 x i32>, ptr addrspace(3) null, align 8
   %vecext.i55 = extractelement <2 x i32> %0, i64 0