[PATCH] D87139: [SVE][CodeGen] Fix InlineFunction for scalable vectors

[David Sherwood via Phabricator via llvm-commits]

david-arm created this revision.
david-arm added reviewers: sdesmalen, kmclaughlin, ctetreau.
Herald added subscribers: llvm-commits, psnobl, hiraditya, kristof.beyls, tschuett.
Herald added a reviewer: rengolin.
Herald added a reviewer: efriedma.
Herald added a project: LLVM.
david-arm requested review of this revision.

When inlining functions containing allocas of scalable vectors we
cannot specify the size in the lifetime markers, since we don't
know this at compile time.

Added new test here:

  test/Transforms/Inline/AArch64/sve-alloca-merge.ll


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D87139

Files:
  llvm/lib/Transforms/Utils/InlineFunction.cpp
  llvm/test/Transforms/Inline/AArch64/sve-alloca-merge.ll


Index: llvm/test/Transforms/Inline/AArch64/sve-alloca-merge.ll
===================================================================
--- /dev/null
+++ llvm/test/Transforms/Inline/AArch64/sve-alloca-merge.ll
@@ -0,0 +1,29 @@
+; RUN: opt -mtriple=aarch64--linux-gnu -mattr=+sve < %s -inline -S | FileCheck %s
+
+define void @bar(<vscale x 2 x i64>* %a) {
+entry:
+  %b = alloca <vscale x 2 x i64>, align 16
+  store <vscale x 2 x i64> zeroinitializer, <vscale x 2 x i64>* %b, align 16
+  %c = load <vscale x 2 x i64>, <vscale x 2 x i64>* %a, align 16
+  %d = load <vscale x 2 x i64>, <vscale x 2 x i64>* %b, align 16
+  %e = add <vscale x 2 x i64> %c, %d
+  %f = add <vscale x 2 x i64> %e, %c
+  store <vscale x 2 x i64> %f, <vscale x 2 x i64>* %a, align 16
+  ret void
+}
+
+define i64 @foo() {
+; CHECK-LABEL: @foo(
+; CHECK: %0 = bitcast <vscale x 2 x i64>* %{{.*}} to i8*
+; CHECK-NEXT: call void @llvm.lifetime.start.p0i8(i64 -1, i8* %0)
+; CHECK: %1 = bitcast <vscale x 2 x i64>* %{{.*}} to i8*
+; CHECK-NEXT: call void @llvm.lifetime.end.p0i8(i64 -1, i8* %1)
+entry:
+  %a = alloca <vscale x 2 x i64>, align 16
+  store <vscale x 2 x i64> zeroinitializer, <vscale x 2 x i64>* %a, align 16
+  %a1 = bitcast <vscale x 2 x i64>* %a to i64*
+  store i64 1, i64* %a1, align 8
+  call void @bar(<vscale x 2 x i64>* %a)
+  %el = load i64, i64* %a1
+  ret i64 %el
+}
Index: llvm/lib/Transforms/Utils/InlineFunction.cpp
===================================================================
--- llvm/lib/Transforms/Utils/InlineFunction.cpp
+++ llvm/lib/Transforms/Utils/InlineFunction.cpp
@@ -2061,7 +2061,7 @@
           dyn_cast<ConstantInt>(AI->getArraySize())) {
         auto &DL = Caller->getParent()->getDataLayout();
         Type *AllocaType = AI->getAllocatedType();
-        uint64_t AllocaTypeSize = DL.getTypeAllocSize(AllocaType);
+        TypeSize AllocaTypeSize = DL.getTypeAllocSize(AllocaType);
         uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
 
         // Don't add markers for zero-sized allocas.
@@ -2070,9 +2070,10 @@
 
         // Check that array size doesn't saturate uint64_t and doesn't
         // overflow when it's multiplied by type size.
-        if (AllocaArraySize != std::numeric_limits<uint64_t>::max() &&
+        if (!AllocaTypeSize.isScalable() &&
+            AllocaArraySize != std::numeric_limits<uint64_t>::max() &&
             std::numeric_limits<uint64_t>::max() / AllocaArraySize >=
-                AllocaTypeSize) {
+                AllocaTypeSize.getFixedSize()) {
           AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
                                         AllocaArraySize * AllocaTypeSize);
         }


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D87139.289929.patch
Type: text/x-patch
Size: 2690 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200904/24626e15/attachment.bin>