[LSV] Skip all non-byte sizes, not only less than eight bits

[bjope]

Summary

The code comments indicate that no effort has been spent on
handling load/stores when the size isn't a multiple of the
byte size correctly. However, the code only avoided types
smaller than 8 bits. So for example a load of an i28 could
still be considered as a candidate for vectorization.

This patch adjusts the code to behave according to the code
comment.

The test case used to hit the following assert when
trying to use "cast" an i32 to i28 using CreateBitOrPointerCast:

opt: ../lib/IR/Instructions.cpp:2565: Assertion `castIsValid(op, S, Ty) && "Invalid cast!"' failed.
#0 PrintStackTraceSignalHandler(void*)
#1 SignalHandler(int)
#2 restore_rt
#3 GI_raise
#4 GI_abort
#5 GI___assert_fail
#6 llvm::CastInst::Create(llvm::Instruction::CastOps, llvm::Value*, llvm::Type*, llvm::Twine const&, llvm::Instruction*)
#7 llvm::IRBuilder<llvm::ConstantFolder, llvm::IRBuilderDefaultInserter>::CreateBitOrPointerCast(llvm::Value*, llvm::Type*, llvm::Twine const&)
#8 (anonymous namespace)::Vectorizer::vectorizeLoadChain(llvm::ArrayRef<llvm::Instruction*>, llvm::SmallPtrSet<llvm::Instruction*, 16u>*)

[arsenm]

LGTM

include/llvm/IR/DataLayout.h

382–385 ↗

(On Diff #120270)

This should probably be a separate patch

[bjope]

include/llvm/IR/DataLayout.h

382–385 ↗

(On Diff #120270)

Ok, I'll skip doing the refactoring in this patch (even though my hidden agenda was to move the byte size dependency to DataLayout, enhancing readability and at the same time making it easier for out-of-tree targets with different byte sizes to simply patch DataLayout).

llvm/trunk/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp

if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
 
       Type *Ty = LI->getType();
       if (!VectorType::isValidElementType(Ty->getScalarType()))
         continue;
 
       // Skip weird non-byte sizes. They probably aren't worth the effort of
       // handling correctly.
       unsigned TySize = DL.getTypeSizeInBits(Ty);
-      if (TySize < 8)
+      if ((TySize % 8) != 0)
         continue;
 
       Value *Ptr = LI->getPointerOperand();
       unsigned AS = Ptr->getType()->getPointerAddressSpace();
       unsigned VecRegSize = TTI.getLoadStoreVecRegBitWidth(AS);
 
       // No point in looking at these if they're too big to vectorize.
       if (TySize > VecRegSize / 2)
if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
 
       Type *Ty = SI->getValueOperand()->getType();
       if (!VectorType::isValidElementType(Ty->getScalarType()))
         continue;
 
       // Skip weird non-byte sizes. They probably aren't worth the effort of
       // handling correctly.
       unsigned TySize = DL.getTypeSizeInBits(Ty);
-      if (TySize < 8)
+      if ((TySize % 8) != 0)
         continue;
 
       Value *Ptr = SI->getPointerOperand();
       unsigned AS = Ptr->getType()->getPointerAddressSpace();
       unsigned VecRegSize = TTI.getLoadStoreVecRegBitWidth(AS);
+
+      // No point in looking at these if they're too big to vectorize.
       if (TySize > VecRegSize / 2)
         continue;
 
       if (isa<VectorType>(Ty) && !llvm::all_of(SI->users(), [](const User *U) {
             const ExtractElementInst *EEI = dyn_cast<ExtractElementInst>(U);
             return EEI && isa<ConstantInt>(EEI->getOperand(1));
           }))
         continue;

llvm/trunk/test/Transforms/LoadStoreVectorizer/X86/non-byte-size.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -load-store-vectorizer -S -mtriple=x86_64-unknown-linux-gnu | FileCheck %s
+
+%rec = type { i32, i28 }
+
+; We currently do not optimize this scenario.
+; But we verify that we no longer crash when compiling this.
+define void @test1(%rec* %out, %rec* %in) {
+; CHECK-LABEL: @test1(
+; CHECK-NEXT:    [[IN1:%.*]] = getelementptr [[REC:%.*]], %rec* [[IN:%.*]], i16 0, i32 0
+; CHECK-NEXT:    [[IN2:%.*]] = getelementptr [[REC]], %rec* [[IN]], i16 0, i32 1
+; CHECK-NEXT:    [[VAL1:%.*]] = load i32, i32* [[IN1]], align 8
+; CHECK-NEXT:    [[VAL2:%.*]] = load i28, i28* [[IN2]]
+; CHECK-NEXT:    [[OUT1:%.*]] = getelementptr [[REC]], %rec* [[OUT:%.*]], i16 0, i32 0
+; CHECK-NEXT:    [[OUT2:%.*]] = getelementptr [[REC]], %rec* [[OUT]], i16 0, i32 1
+; CHECK-NEXT:    store i32 [[VAL1]], i32* [[OUT1]], align 8
+; CHECK-NEXT:    store i28 [[VAL2]], i28* [[OUT2]]
+; CHECK-NEXT:    ret void
+;
+  %in1 = getelementptr %rec, %rec* %in, i16 0, i32 0
+  %in2 = getelementptr %rec, %rec* %in, i16 0, i32 1
+  %val1 = load i32, i32* %in1, align 8
+  %val2 = load i28, i28* %in2
+  %out1 = getelementptr %rec, %rec* %out, i16 0, i32 0
+  %out2 = getelementptr %rec, %rec* %out, i16 0, i32 1
+  store i32 %val1, i32* %out1, align 8
+  store i28 %val2, i28* %out2
+  ret void
+}
+