[llvm] f223411 - [InstCombine]PtrReplacer: Correctly handle select with unavailable operands (#148829)

[via llvm-commits]

Author: Pierre van Houtryve
Date: 2025-07-16T09:32:05+02:00
New Revision: f223411e2e005ecd54523801db603b54d69f3159

URL: https://github.com/llvm/llvm-project/commit/f223411e2e005ecd54523801db603b54d69f3159
DIFF: https://github.com/llvm/llvm-project/commit/f223411e2e005ecd54523801db603b54d69f3159.diff

LOG: [InstCombine]PtrReplacer: Correctly handle select with unavailable operands (#148829)

The testcase I added previously failed because a SelectInst with invalid
operands was created (one side `addrspace(4)`, the other
`addrspace(5)`).
PointerReplacer needs to dig deeper if the true and/or false
instructions of the select are not available.

Fixes SWDEV-542957

Added: 
    

Modified: 
    llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
    llvm/test/Transforms/InstCombine/AMDGPU/ptr-replace-alloca.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp b/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
index e791bc4b56e52..2cc1bc9fa4a67 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp
@@ -277,6 +277,15 @@ bool PointerReplacer::collectUsers() {
           Worklist.emplace_back(I);
   };
 
+  auto TryPushInstOperand = [&](Instruction *InstOp) {
+    if (!UsersToReplace.contains(InstOp)) {
+      if (!ValuesToRevisit.insert(InstOp))
+        return false;
+      Worklist.emplace_back(InstOp);
+    }
+    return true;
+  };
+
   PushUsersToWorklist(&Root);
   while (!Worklist.empty()) {
     Instruction *Inst = Worklist.pop_back_val();
@@ -309,12 +318,8 @@ bool PointerReplacer::collectUsers() {
       // incoming values.
       Worklist.emplace_back(PHI);
       for (unsigned Idx = 0; Idx < PHI->getNumIncomingValues(); ++Idx) {
-        auto *IncomingValue = cast<Instruction>(PHI->getIncomingValue(Idx));
-        if (UsersToReplace.contains(IncomingValue))
-          continue;
-        if (!ValuesToRevisit.insert(IncomingValue))
+        if (!TryPushInstOperand(cast<Instruction>(PHI->getIncomingValue(Idx))))
           return false;
-        Worklist.emplace_back(IncomingValue);
       }
     } else if (auto *SI = dyn_cast<SelectInst>(Inst)) {
       auto *TrueInst = dyn_cast<Instruction>(SI->getTrueValue());
@@ -322,8 +327,17 @@ bool PointerReplacer::collectUsers() {
       if (!TrueInst || !FalseInst)
         return false;
 
-      UsersToReplace.insert(SI);
-      PushUsersToWorklist(SI);
+      if (isAvailable(TrueInst) && isAvailable(FalseInst)) {
+        UsersToReplace.insert(SI);
+        PushUsersToWorklist(SI);
+        continue;
+      }
+
+      // Push select back onto the stack, followed by unavailable true/false
+      // value.
+      Worklist.emplace_back(SI);
+      if (!TryPushInstOperand(TrueInst) || !TryPushInstOperand(FalseInst))
+        return false;
     } else if (auto *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
       UsersToReplace.insert(GEP);
       PushUsersToWorklist(GEP);

diff  --git a/llvm/test/Transforms/InstCombine/AMDGPU/ptr-replace-alloca.ll b/llvm/test/Transforms/InstCombine/AMDGPU/ptr-replace-alloca.ll
index 538cc19f9722e..beb84362b7f92 100644
--- a/llvm/test/Transforms/InstCombine/AMDGPU/ptr-replace-alloca.ll
+++ b/llvm/test/Transforms/InstCombine/AMDGPU/ptr-replace-alloca.ll
@@ -76,4 +76,37 @@ sink:
   ret <2 x i64> %val.sink
 }
 
+; Crashed in IC PtrReplacer because an invalid select was generated with addrspace(4) and addrspace(5)
+; operands.
+define amdgpu_kernel void @select_addr4_addr5(ptr addrspace(4) byref([12 x i8]) align 16 %arg) {
+; CHECK-LABEL: define amdgpu_kernel void @select_addr4_addr5(
+; CHECK-SAME: ptr addrspace(4) byref([12 x i8]) align 16 [[ARG:%.*]]) {
+; CHECK-NEXT:  [[BB:.*:]]
+; CHECK-NEXT:    ret void
+;
+bb:
+  %alloca = alloca i32, i32 0, align 8, addrspace(5)
+  %alloca1 = alloca [12 x i8], align 16, addrspace(5)
+  call void @llvm.memcpy.p5.p4.i64(ptr addrspace(5) %alloca1, ptr addrspace(4) %arg, i64 0, i1 false)
+  %select = select i1 false, ptr addrspace(5) %alloca1, ptr addrspace(5) %alloca
+  call void @llvm.memcpy.p0.p5.i64(ptr null, ptr addrspace(5) %select, i64 0, i1 false)
+  ret void
+}
+
+; Same as above but with swapped operands on the select.
+define amdgpu_kernel void @select_addr4_addr5_swapped(ptr addrspace(4) byref([12 x i8]) align 16 %arg) {
+; CHECK-LABEL: define amdgpu_kernel void @select_addr4_addr5_swapped(
+; CHECK-SAME: ptr addrspace(4) byref([12 x i8]) align 16 [[ARG:%.*]]) {
+; CHECK-NEXT:  [[BB:.*:]]
+; CHECK-NEXT:    ret void
+;
+bb:
+  %alloca = alloca i32, i32 0, align 8, addrspace(5)
+  %alloca1 = alloca [12 x i8], align 16, addrspace(5)
+  call void @llvm.memcpy.p5.p4.i64(ptr addrspace(5) %alloca1, ptr addrspace(4) %arg, i64 0, i1 false)
+  %select = select i1 false, ptr addrspace(5) %alloca, ptr addrspace(5) %alloca1
+  call void @llvm.memcpy.p0.p5.i64(ptr null, ptr addrspace(5) %select, i64 0, i1 false)
+  ret void
+}
+
 declare void @llvm.memcpy.p5.p4.i64(ptr addrspace(5) noalias writeonly captures(none), ptr addrspace(4) noalias readonly captures(none), i64, i1 immarg) #0