AMDGPU: Fix assert when multi operands to update after folding imm (llvm#148205)

In the original motivating test case,
[FoldList](https://github.com/llvm/llvm-project/blob/d8a2141ff98ee35cd1886f536ccc3548b012820b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp#L1764)
had entries:
```
  #0: UseMI: %224:sreg_32 = S_OR_B32 %219.sub0:sreg_64, %219.sub1:sreg_64, implicit-def dead $scc
      UseOpNo: 1

  #1: UseMI: %224:sreg_32 = S_OR_B32 %219.sub0:sreg_64, %219.sub1:sreg_64, implicit-def dead $scc
      UseOpNo: 2
```
After calling
[updateOperand(#0)](https://github.com/llvm/llvm-project/blob/d8a2141ff98ee35cd1886f536ccc3548b012820b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp#L1773),
[tryConstantFoldOp(#0.UseMI)](https://github.com/llvm/llvm-project/blob/d8a2141ff98ee35cd1886f536ccc3548b012820b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp#L1786)
removed operand 1, and entry #&llvm#8203;1.UseOpNo was no longer valid,
resulting in an
[assert](https://github.com/llvm/llvm-project/blob/4a35214bddbb67f9597a500d48ab8c4fb25af150/llvm/include/llvm/ADT/ArrayRef.h#L452).

This change defers constant folding until all operands have been updated
so that UseOpNo values remain stable.

Author: macurtis-amd

llvm/lib/Target/AMDGPU/SIFoldOperands.cpp

@@ -1761,6 +1761,7 @@ bool SIFoldOperandsImpl::foldInstOperand(MachineInstr &MI,
   for (MachineInstr *Copy : CopiesToReplace)
     Copy->addImplicitDefUseOperands(*MF);
 
+  SetVector<MachineInstr *> ConstantFoldCandidates;
   for (FoldCandidate &Fold : FoldList) {
     assert(!Fold.isReg() || Fold.Def.OpToFold);
     if (Fold.isReg() && Fold.getReg().isVirtual()) {
@@ -1783,16 +1784,21 @@ bool SIFoldOperandsImpl::foldInstOperand(MachineInstr &MI,
                         << static_cast<int>(Fold.UseOpNo) << " of "
                         << *Fold.UseMI);
 
-      if (Fold.isImm() && tryConstantFoldOp(Fold.UseMI)) {
-        LLVM_DEBUG(dbgs() << "Constant folded " << *Fold.UseMI);
-        Changed = true;
-      }
+      if (Fold.isImm())
+        ConstantFoldCandidates.insert(Fold.UseMI);
 
     } else if (Fold.Commuted) {
       // Restoring instruction's original operand order if fold has failed.
       TII->commuteInstruction(*Fold.UseMI, false);
     }
   }
+
+  for (MachineInstr *MI : ConstantFoldCandidates) {
+    if (tryConstantFoldOp(MI)) {
+      LLVM_DEBUG(dbgs() << "Constant folded " << *MI);
+      Changed = true;
+    }
+  }
   return true;
 }
 

llvm/test/CodeGen/AMDGPU/bug-multi-operands-to-update-after-fold.mir

@@ -0,0 +1,15 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -mtriple=amdgcn-amd-hsa -mcpu=gfx1031 -run-pass=si-fold-operands -o - %s | FileCheck %s
+---
+name: snork
+body:  |
+  bb.0:
+    ; CHECK-LABEL: name: snork
+    ; CHECK: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 0
+    ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[S_MOV_B32_]], %subreg.sub0, [[S_MOV_B32_]], %subreg.sub1, [[S_MOV_B32_]], %subreg.sub2, [[S_MOV_B32_]], %subreg.sub3
+    ; CHECK-NEXT: SI_RETURN
+    %0:sreg_32 = S_MOV_B32 0
+    %1:sgpr_128 = REG_SEQUENCE %0, %subreg.sub0, %0, %subreg.sub1, %0, %subreg.sub2, %0, %subreg.sub3
+    %2:sreg_32 = S_OR_B32 %1.sub0, %1.sub3, implicit-def dead $scc
+    SI_RETURN
+...