[clang-tools-extra] [llvm] [AMDGPU] Quit PromoteAllocaToVector if intrinsic is used (PR #68744)

[Mariusz Sikora via llvm-commits]

https://github.com/mariusz-sikora-at-amd updated https://github.com/llvm/llvm-project/pull/68744

>From 3dc3a43193247015933392b7da76c4ef77268231 Mon Sep 17 00:00:00 2001
From: Mariusz Sikora <mariusz.sikora at amd.com>
Date: Tue, 10 Oct 2023 21:50:48 +0200
Subject: [PATCH 1/2] [AMDGPU] Bail if assume-like intrinsic is used in
 PromoteAllocaToVector

Attached test will cause crash without this change.
---
 llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp           | 2 ++
 .../test/CodeGen/AMDGPU/promote-alloca-mem-intrinsics.ll | 9 +++++++++
 2 files changed, 11 insertions(+)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp b/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
index 3707a960211eb4..8ec7d29e00c939 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
@@ -772,6 +772,8 @@ bool AMDGPUPromoteAllocaImpl::tryPromoteAllocaToVector(AllocaInst &Alloca) {
 
     // Ignore assume-like intrinsics and comparisons used in assumes.
     if (isAssumeLikeIntrinsic(Inst)) {
+      if (!Inst->use_empty())
+        return RejectUser(Inst, "assume-like intrinsic cannot have any users");
       UsersToRemove.push_back(Inst);
       continue;
     }
diff --git a/llvm/test/CodeGen/AMDGPU/promote-alloca-mem-intrinsics.ll b/llvm/test/CodeGen/AMDGPU/promote-alloca-mem-intrinsics.ll
index 0bba1bdce95655..5616bc0f5ef3c1 100644
--- a/llvm/test/CodeGen/AMDGPU/promote-alloca-mem-intrinsics.ll
+++ b/llvm/test/CodeGen/AMDGPU/promote-alloca-mem-intrinsics.ll
@@ -53,6 +53,15 @@ define amdgpu_kernel void @promote_with_objectsize(ptr addrspace(1) %out) #0 {
   ret void
 }
 
+; CHECK-LABEL: @promote_with_objectsize_8(
+; CHECK: [[PTR:%[0-9]+]] = getelementptr inbounds [64 x [8 x i32]], ptr addrspace(3) @promote_with_objectsize_8.alloca, i32 0, i32 %{{[0-9]+}}
+; CHECK: call i32 @llvm.objectsize.i32.p3(ptr addrspace(3) [[PTR]], i1 false, i1 false, i1 false)
+define amdgpu_kernel void @promote_with_objectsize_8(ptr addrspace(1) %out) #0 {
+  %alloca = alloca [8 x i32], align 4, addrspace(5)
+  %size = call i32 @llvm.objectsize.i32.p5(ptr addrspace(5) %alloca, i1 false, i1 false, i1 false)
+  store i32 %size, ptr addrspace(1) %out
+  ret void
+}
 ; CHECK-LABEL: @promote_alloca_used_twice_in_memcpy(
 ; CHECK: call void @llvm.memcpy.p3.p3.i64(ptr addrspace(3) align 8 dereferenceable(16) %arrayidx1, ptr addrspace(3) align 8 dereferenceable(16) %arrayidx2, i64 16, i1 false)
 define amdgpu_kernel void @promote_alloca_used_twice_in_memcpy(i32 %c) {

>From a85ab29fe8f338ff740456ea6b8890dd1212ca6c Mon Sep 17 00:00:00 2001
From: Mariusz Sikora <mariusz.sikora at amd.com>
Date: Tue, 21 Nov 2023 09:18:45 +0100
Subject: [PATCH 2/2] Fixup - handle objectsize in PromoteAlloca

---
 llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp   | 16 ++++++++++++++++
 .../AMDGPU/promote-alloca-mem-intrinsics.ll      |  3 +--
 2 files changed, 17 insertions(+), 2 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp b/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
index 8ec7d29e00c939..90d0dc96898758 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
@@ -528,6 +528,15 @@ static Value *promoteAllocaUserToVector(
       return Builder.CreateVectorSplat(VectorTy->getElementCount(), Elt);
     }
 
+    if (auto *Intr = dyn_cast<IntrinsicInst>(Inst)) {
+      if (Intr->getIntrinsicID() == Intrinsic::objectsize) {
+        Intr->replaceAllUsesWith(
+            Builder.getIntN(Intr->getType()->getIntegerBitWidth(),
+                            DL.getTypeAllocSize(VectorTy)));
+        return nullptr;
+      }
+    }
+
     llvm_unreachable("Unsupported call when promoting alloca to vector");
   }
 
@@ -770,6 +779,13 @@ bool AMDGPUPromoteAllocaImpl::tryPromoteAllocaToVector(AllocaInst &Alloca) {
       continue;
     }
 
+    if (auto *Intr = dyn_cast<IntrinsicInst>(Inst)) {
+      if (Intr->getIntrinsicID() == Intrinsic::objectsize) {
+        WorkList.push_back(Inst);
+        continue;
+      }
+    }
+
     // Ignore assume-like intrinsics and comparisons used in assumes.
     if (isAssumeLikeIntrinsic(Inst)) {
       if (!Inst->use_empty())
diff --git a/llvm/test/CodeGen/AMDGPU/promote-alloca-mem-intrinsics.ll b/llvm/test/CodeGen/AMDGPU/promote-alloca-mem-intrinsics.ll
index 5616bc0f5ef3c1..aabd5df9568370 100644
--- a/llvm/test/CodeGen/AMDGPU/promote-alloca-mem-intrinsics.ll
+++ b/llvm/test/CodeGen/AMDGPU/promote-alloca-mem-intrinsics.ll
@@ -54,8 +54,7 @@ define amdgpu_kernel void @promote_with_objectsize(ptr addrspace(1) %out) #0 {
 }
 
 ; CHECK-LABEL: @promote_with_objectsize_8(
-; CHECK: [[PTR:%[0-9]+]] = getelementptr inbounds [64 x [8 x i32]], ptr addrspace(3) @promote_with_objectsize_8.alloca, i32 0, i32 %{{[0-9]+}}
-; CHECK: call i32 @llvm.objectsize.i32.p3(ptr addrspace(3) [[PTR]], i1 false, i1 false, i1 false)
+; CHECK: store i32 32, ptr addrspace(1) %out, align 4
 define amdgpu_kernel void @promote_with_objectsize_8(ptr addrspace(1) %out) #0 {
   %alloca = alloca [8 x i32], align 4, addrspace(5)
   %size = call i32 @llvm.objectsize.i32.p5(ptr addrspace(5) %alloca, i1 false, i1 false, i1 false)