[InstCombine][SVE] Fix InstCombiner::visitAllocaInst for scalable vector.

[huihuiz]

Summary

DataLayout::getTypeAllocSize() return TypeSize. For cases where scalable
property doesn't matter (check for zero-sized alloca), we should explicitly
call getKnownMinSize() to avoid implicit type conversion to uint64_t, which is
invalid for scalable vector type.

[efriedma]

LGTM

llvm/test/Transforms/InstCombine/vscale_alloca.ll

24

"first"

[huihuiz]

llvm/test/Transforms/InstCombine/vscale_alloca.ll

24

Thanks Eli! Fixed in commit patch.

llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp

if (AI.getAllocatedType()->isSized()) {
     // If the alignment is 0 (unspecified), assign it the preferred alignment.
     if (AI.getAlignment() == 0)
       AI.setAlignment(
           MaybeAlign(DL.getPrefTypeAlignment(AI.getAllocatedType())));
 
     // Move all alloca's of zero byte objects to the entry block and merge them
     // together.  Note that we only do this for alloca's, because malloc should
     // allocate and return a unique pointer, even for a zero byte allocation.
-    if (DL.getTypeAllocSize(AI.getAllocatedType()) == 0) {
+    if (DL.getTypeAllocSize(AI.getAllocatedType()).getKnownMinSize() == 0) {
       // For a zero sized alloca there is no point in doing an array allocation.
       // This is helpful if the array size is a complicated expression not used
       // elsewhere.
       if (AI.isArrayAllocation())
         return replaceOperand(AI, 0,
             ConstantInt::get(AI.getArraySize()->getType(), 1));
 
       // Get the first instruction in the entry block.
       BasicBlock &EntryBlock = AI.getParent()->getParent()->getEntryBlock();
       Instruction *FirstInst = EntryBlock.getFirstNonPHIOrDbg();
       if (FirstInst != &AI) {
         // If the entry block doesn't start with a zero-size alloca then move
         // this one to the start of the entry block.  There is no problem with
         // dominance as the array size was forced to a constant earlier already.
         AllocaInst *EntryAI = dyn_cast<AllocaInst>(FirstInst);
         if (!EntryAI || !EntryAI->getAllocatedType()->isSized() ||
-            DL.getTypeAllocSize(EntryAI->getAllocatedType()) != 0) {
+            DL.getTypeAllocSize(EntryAI->getAllocatedType())
+                    .getKnownMinSize() != 0) {
           AI.moveBefore(FirstInst);
           return &AI;
         }
 
         // If the alignment of the entry block alloca is 0 (unspecified),
         // assign it the preferred alignment.
         if (EntryAI->getAlignment() == 0)
           EntryAI->setAlignment(

llvm/test/Transforms/InstCombine/vscale_alloca.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -instcombine -verify < %s | FileCheck %s
+
+define <vscale x 4 x i32> @alloca(<vscale x 4 x i32> %z) {
+; CHECK-LABEL: @alloca(
+; CHECK-NEXT:    ret <vscale x 4 x i32> [[Z:%.*]]
+;
+  %a = alloca <vscale x 4 x i32>
+  store <vscale x 4 x i32> %z, <vscale x 4 x i32>* %a
+  %load = load <vscale x 4 x i32>, <vscale x 4 x i32>* %a
+  ret <vscale x 4 x i32> %load
+}
+
+define void @alloca_dead_store(<vscale x 4 x i32> %z) {
+; CHECK-LABEL: @alloca_dead_store(
+; CHECK-NEXT:    ret void
+;
+  %a = alloca <vscale x 4 x i32>
+  store <vscale x 4 x i32> %z, <vscale x 4 x i32>* %a
+  ret void
+}
+
+declare void @use(...)
+define void @alloca_zero_byte_move_fist_inst() {
+; CHECK-LABEL: @alloca_zero_byte_move_fist_inst(
+; CHECK-NEXT:    [[B:%.*]] = alloca {}, align 8
+; CHECK-NEXT:    [[A:%.*]] = alloca <vscale x 16 x i8>, align 16
+; CHECK-NEXT:    call void (...) @use(<vscale x 16 x i8>* nonnull [[A]])
+; CHECK-NEXT:    call void (...) @use({}* nonnull [[B]])
+; CHECK-NEXT:    ret void
+;
+  %a = alloca <vscale x 16 x i8>
+  call void (...) @use( <vscale x 16 x i8>* %a )
+  %b = alloca {  }
+  call void (...) @use( {  }* %b )
+  ret void
+}