[PATCH] D123694: [AMDGPU][FIX] Let load-store-vectorizer is working with opaque pointers

[Johannes Doerfert via Phabricator via llvm-commits]

jdoerfert created this revision.
jdoerfert added a reviewer: arsenm.
Herald added subscribers: dexonsmith, kerbowa, bollu, hiraditya, t-tye, tpr, dstuttard, yaxunl, jvesely, kzhuravl.
Herald added a project: All.
jdoerfert requested review of this revision.
Herald added a subscriber: wdng.
Herald added a project: LLVM.

The original code relied on the fact that we needed a bitcast
instruction (for non constant base objects). With opaque pointers there
might not be a bitcast. Always check if reordering is required instead.

Fixes: https://github.com/llvm/llvm-project/issues/54896


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D123694

Files:
  llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
  llvm/test/Transforms/LoadStoreVectorizer/AMDGPU/opaque_ptr.ll


Index: llvm/test/Transforms/LoadStoreVectorizer/AMDGPU/opaque_ptr.ll
===================================================================
--- /dev/null
+++ llvm/test/Transforms/LoadStoreVectorizer/AMDGPU/opaque_ptr.ll
@@ -0,0 +1,24 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -mtriple=amdgcn-amd-amdhsa -basic-aa -load-store-vectorizer -S -o - %s | FileCheck %s
+
+; Vectorize and emit valid code (Issue #54896).
+
+%S = type { i64, i64 }
+ at S = external global %S
+
+define i64 @order() {
+; CHECK-LABEL: @order(
+; CHECK-NEXT:    [[IDX0:%.*]] = getelementptr inbounds [[S:%.*]], ptr @S, i32 0, i32 0
+; CHECK-NEXT:    [[TMP1:%.*]] = load <2 x i64>, ptr [[IDX0]], align 8
+; CHECK-NEXT:    [[L01:%.*]] = extractelement <2 x i64> [[TMP1]], i32 0
+; CHECK-NEXT:    [[L12:%.*]] = extractelement <2 x i64> [[TMP1]], i32 1
+; CHECK-NEXT:    [[ADD:%.*]] = add i64 [[L01]], [[L12]]
+; CHECK-NEXT:    ret i64 [[ADD]]
+;
+  %idx1 = getelementptr inbounds %S, ptr @S, i32 0, i32 1
+  %l1 = load i64, i64* %idx1, align 8
+  %idx0 = getelementptr inbounds %S, ptr @S, i32 0, i32 0
+  %l0 = load i64, i64* %idx0, align 8
+  %add = add i64 %l0, %l1
+  ret i64 %add
+}
Index: llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
===================================================================
--- llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
+++ llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
@@ -1297,10 +1297,8 @@
     CV->replaceAllUsesWith(V);
   }
 
-  // Bitcast might not be an Instruction, if the value being loaded is a
-  // constant. In that case, no need to reorder anything.
-  if (Instruction *BitcastInst = dyn_cast<Instruction>(Bitcast))
-    reorder(BitcastInst);
+  // Since we might have opaque pointers we might end up using the pointer operand of the first load (wrt. memory loaded) for the vector load. Since this first load might not be the first in the block we potentially need to reorder the pointer operand (and its operands).
+  reorder(LI);
 
   eraseInstructions(Chain);
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D123694.422561.patch
Type: text/x-patch
Size: 2054 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220413/fd550e33/attachment.bin>