[PATCH] D121787: [VectorCombine] Avoid crossing address space boundaries

[Fraser Cormack via Phabricator via llvm-commits]

frasercrmck created this revision.
frasercrmck added reviewers: spatel, tra, yaxunl, lebedev.ri.
Herald added a subscriber: hiraditya.
Herald added a project: All.
frasercrmck requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

We can not bitcast pointers across different address spaces. This was
previously fixed in D89577 <https://reviews.llvm.org/D89577> but then in D93229 <https://reviews.llvm.org/D93229> an enhancement was added
which peeks further through the ponter operand, opening up the
possibility that address-space violations could be introduced.

It's unclear to me whether we need to bail in this situation as I've
done, or whether we can issue an addrspace cast, at least under certain
circumstances.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D121787

Files:
  llvm/lib/Transforms/Vectorize/VectorCombine.cpp
  llvm/test/Transforms/VectorCombine/X86/load-inselt-addrspace.ll


Index: llvm/test/Transforms/VectorCombine/X86/load-inselt-addrspace.ll
===================================================================
--- /dev/null
+++ llvm/test/Transforms/VectorCombine/X86/load-inselt-addrspace.ll
@@ -0,0 +1,20 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=vector-combine -S -mtriple=x86_64-- -mattr=sse2 | FileCheck %s
+; RUN: opt < %s -passes=vector-combine -S -mtriple=x86_64-- -mattr=avx2 | FileCheck %s
+
+define <4 x i32> @unsafe_load_i32_insert_v4i32_addrspace(i32* align 16 dereferenceable(16) %v3) {
+; CHECK-LABEL: @unsafe_load_i32_insert_v4i32_addrspace(
+; CHECK-NEXT:    [[T0:%.*]] = getelementptr inbounds i32, i32* [[V3:%.*]], i32 1
+; CHECK-NEXT:    [[T1:%.*]] = addrspacecast i32* [[T0]] to i32 addrspace(42)*
+; CHECK-NEXT:    [[T2:%.*]] = getelementptr inbounds i32, i32 addrspace(42)* [[T1]], i64 1
+; CHECK-NEXT:    [[VAL:%.*]] = load i32, i32 addrspace(42)* [[T2]], align 4
+; CHECK-NEXT:    [[INSELT:%.*]] = insertelement <4 x i32> poison, i32 [[VAL]], i32 0
+; CHECK-NEXT:    ret <4 x i32> [[INSELT]]
+;
+  %t0 = getelementptr inbounds i32, i32* %v3, i32 1
+  %t1 = addrspacecast i32* %t0 to i32 addrspace(42)*
+  %t2 = getelementptr inbounds i32, i32 addrspace(42)* %t1, i64 1
+  %val = load i32, i32 addrspace(42)* %t2, align 4
+  %inselt = insertelement <4 x i32> poison, i32 %val, i32 0
+  ret <4 x i32> %inselt
+}
Index: llvm/lib/Transforms/Vectorize/VectorCombine.cpp
===================================================================
--- llvm/lib/Transforms/Vectorize/VectorCombine.cpp
+++ llvm/lib/Transforms/Vectorize/VectorCombine.cpp
@@ -190,6 +190,10 @@
     if (Offset.isNegative())
       return false;
 
+    // We can't bitcast this pointer due to differing address spaces.
+    if (AS != SrcPtr->getType()->getPointerAddressSpace())
+      return false;
+
     // The offset must be a multiple of the scalar element to shuffle cleanly
     // in the element's size.
     uint64_t ScalarSizeInBytes = ScalarSize / 8;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D121787.415762.patch
Type: text/x-patch
Size: 2039 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220316/252c5ed9/attachment.bin>