[Loads] Fix crash in isSafeToLoadUnconditionally with scalable accessed type

[lukel97]

This fixes #82606 by updating isSafeToLoadUnconditionally to handle fixed sized loads from a scalable accessed type.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

@llvm/pr-subscribers-llvm-analysis

Author: Luke Lau (lukel97)

Changes

This fixes #82606 by updating isSafeToLoadUnconditionally to handle fixed sized loads from a scalable accessed type.

---

Full diff: https://github.com/llvm/llvm-project/pull/82650.diff

2 Files Affected:

• (modified) llvm/lib/Analysis/Loads.cpp (+3-3)
• (added) llvm/test/Transforms/VectorCombine/RISCV/load-widening.ll (+22)

diff --git a/llvm/lib/Analysis/Loads.cpp b/llvm/lib/Analysis/Loads.cpp
index 6bf0d2f56eb4eb..5916d2ab48ecec 100644
--- a/llvm/lib/Analysis/Loads.cpp
+++ b/llvm/lib/Analysis/Loads.cpp
@@ -364,7 +364,7 @@ bool llvm::isSafeToLoadUnconditionally(Value *V, Align Alignment, APInt &Size,
 
   if (Size.getBitWidth() > 64)
     return false;
-  const uint64_t LoadSize = Size.getZExtValue();
+  const TypeSize LoadSize = TypeSize::getFixed(Size.getZExtValue());
 
   // Otherwise, be a little bit aggressive by scanning the local block where we
   // want to check to see if the pointer is already being loaded or stored
@@ -414,11 +414,11 @@ bool llvm::isSafeToLoadUnconditionally(Value *V, Align Alignment, APInt &Size,
 
     // Handle trivial cases.
     if (AccessedPtr == V &&
-        LoadSize <= DL.getTypeStoreSize(AccessedTy))
+        TypeSize::isKnownLE(LoadSize, DL.getTypeStoreSize(AccessedTy)))
       return true;
 
     if (AreEquivalentAddressValues(AccessedPtr->stripPointerCasts(), V) &&
-        LoadSize <= DL.getTypeStoreSize(AccessedTy))
+        TypeSize::isKnownLE(LoadSize, DL.getTypeStoreSize(AccessedTy)))
       return true;
   }
   return false;
diff --git a/llvm/test/Transforms/VectorCombine/RISCV/load-widening.ll b/llvm/test/Transforms/VectorCombine/RISCV/load-widening.ll
new file mode 100644
index 00000000000000..49d0d4216c52b2
--- /dev/null
+++ b/llvm/test/Transforms/VectorCombine/RISCV/load-widening.ll
@@ -0,0 +1,22 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
+; RUN: opt < %s -passes=vector-combine -S -mtriple=riscv32 -mattr=+v | FileCheck %s -check-prefixes=CHECK,RV32
+; RUN: opt < %s -passes=vector-combine -S -mtriple=riscv64 -mattr=+v | FileCheck %s -check-prefixes=CHECK,RV64
+
+define void @fixed_load_scalable_src() {
+; CHECK-LABEL: define void @fixed_load_scalable_src(
+; CHECK-SAME: ) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    store <vscale x 4 x i16> zeroinitializer, ptr null, align 8
+; CHECK-NEXT:    [[TMP0:%.*]] = load <4 x i16>, ptr null, align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <4 x i16> [[TMP0]], <4 x i16> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison>
+; CHECK-NEXT:    ret void
+;
+entry:
+  store <vscale x 4 x i16> zeroinitializer, ptr null
+  %0 = load <4 x i16>, ptr null
+  %1 = shufflevector <4 x i16> %0, <4 x i16> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison>
+  ret void
+}
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; RV32: {{.*}}
+; RV64: {{.*}}

[lukel97]

VectorCombine::widenSubvectorLoad triggers the crash here when calling isSafeToLoadUnconditionally. Let me know if there's a better place to test for this!

[nikic]

LG

[nikic]

Drop the RV32/RV64 prefixes.

[lukel97]

I didn't have them initially but UTC complained about differing CHECK lines?

WARNING: Found conflicting asm under the same prefix: 'CHECK'!

Inspecting the actual check lines it seems to be the same though bar the attributes, which is why I presume it's complaining. Dropped them anyway.