Commit 53c8de3f authored by Tanya Lattner's avatar Tanya Lattner
Browse files

Merge from mainline. 

Fix a nasty miscompilation of 176.gcc on linux/x86 where we synthesized
a memset using 16-byte XMM stores, but where the stack realignment code
didn't work.  Until it does (PR2962) disable use of xmm regs in memcpy
and memset formation for linux and other targets with insufficiently
aligned stacks.

This is part of PR2888

llvm-svn: 58423
parent c46ec58f

llvm/lib/Target/X86/X86ISelLowering.cpp

+9 −4
Original line number Diff line number Diff line
@@ -837,10 +837,15 @@ unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const { 
MVT

X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,

                                       bool isSrcConst, bool isSrcStr) const {

  // FIXME: This turns off use of xmm stores for memset/memcpy on targets like

  // linux.  This is because the stack realignment code can't handle certain

  // cases like PR2962.  This should be removed when PR2962 is fixed.

  if (Subtarget->getStackAlignment() >= 16) {

    if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)

      return MVT::v4i32;

    if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)

      return MVT::v4f32;

  }

  if (Subtarget->is64Bit() && Size >= 8)

    return MVT::i64;

  return MVT::i32;

llvm/test/CodeGen/X86/2008-10-27-StackRealignment.ll

0 → 100644
+22 −0
Original line number Diff line number Diff line 
; Linux doesn't support stack realignment for functions with allocas (PR2888).

; Until it does, we shouldn't use movaps to access the stack.  On targets with

; sufficiently aligned stack (e.g. darwin) we should.



; RUN: llvm-as < %s | llc -mtriple=i386-pc-linux-gnu -mcpu=yonah | not grep movaps

; RUN: llvm-as < %s | llc -mtriple=i686-apple-darwin9 -mcpu=yonah | grep movaps | count 2





target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:32:32"

target triple = "i386-pc-linux-gnu"

  

define void @foo(i32 %t) nounwind {

  %tmp1210 = alloca i8, i32 32, align 4

  call void @llvm.memset.i64(i8* %tmp1210, i8 0, i64 32, i32 4)

  

  %x = alloca i8, i32 %t

  call void @dummy(i8* %x)

  ret void

}



declare void @dummy(i8* %x)

declare void @llvm.memset.i64(i8*, i8, i64, i32) nounwind