[LLVMbugs] [Bug 17631] New: llc inserts vzeroupper before a call and spoils avx registers, which lives across the call

[bugzilla-daemon at llvm.org]

http://llvm.org/bugs/show_bug.cgi?id=17631

            Bug ID: 17631
           Summary: llc inserts vzeroupper before a call and spoils avx
                    registers, which lives across the call
           Product: tools
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: llc
          Assignee: unassignedbugs at nondot.org
          Reporter: ili.filippov at gmail.com
                CC: llvmbugs at cs.uiuc.edu
    Classification: Unclassified

This bug reproduces on 32 bit both Linux and Windows. Though it seems to be
more frequently appears in real life on Windows.
Reproducer:

; ModuleID = 'a.bc'
target datalayout =
"e-p:32:32:32-S32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f16:16:16-f32:32:32-f64:64:64-f128:128:128-v64:64:64-v128:128:128-a0:0:64-f80:32:32-n8:16:32"
target triple = "i386-pc-win32"

%struct_type = type { [64 x <8 x float>], <8 x float> }

; Function Attrs: nounwind readnone
declare i32 @llvm.x86.avx.movmsk.ps.256(<8 x float>) #0

; Function Attrs: nounwind
define i32 @equal(<8 x i32> %A) #1 {
allocas:
  %first_alloc  = alloca [64 x <8 x i32>]
  %second_alloc = alloca %struct_type

  %A1 = bitcast <8 x i32> %A to <8 x float>
  %A2 = call i32 @llvm.x86.avx.movmsk.ps.256(<8 x float> %A1)
  ret i32 %A2
}

; Function Attrs: nounwind
define i32 @before_equal() #1 {
allocas:
  %A = alloca <8 x i32>
  store <8 x i32> <i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32
-1>, <8 x i32>* %A
  %A1 = load <8 x i32>* %A
  %A2 = call i32 @equal(<8 x i32> %A1)
  ret i32 %A2
}

attributes #0 = { nounwind readnone }
attributes #1 = { nounwind "target-features"="+avx,+popcnt,+cmov" }

Compilation command:
> llc a.ll -o a.s -mcpu=corei7-avx
> cat a.s

_equal:
# BB#0:                                 # %allocas
        pushl   %ebp
        movl    %esp, %ebp
        andl    $-32, %esp
        movl    $4160, %eax             # imm = 0x1040
        vzeroupper
        calll   __chkstk
        vmovmskps       %ymm0, %eax
        movl    %ebp, %esp
        popl    %ebp
        vzeroupper
        ret

        .def     _before_equal;
        .scl    2;
        .type   32;
        .endef
        .globl  _before_equal
        .align  16, 0x90
_before_equal:                          # @before_equal
# BB#0:                                 # %allocas
        pushl   %ebp
        movl    %esp, %ebp
        andl    $-32, %esp
        subl    $64, %esp
        vpcmpeqd        %xmm0, %xmm0, %xmm0
        vinsertf128     $1, %xmm0, %ymm0, %ymm0
        vmovaps %ymm0, (%esp)
        calll   _equal
        movl    %ebp, %esp
        popl    %ebp
        vzeroupper
        ret

ymm0 is initialized by all 1s (vpcmpeqd + vinsertf128). Inside _equal function
it's assumed to be live across __chkstk call, but vzeroupper zeros upper half
of the register. Right after the call ymm0 is used (vmovmskps), which yields
incorrect value in eax (0xf instead of 0xff).

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20131021/923ff07b/attachment.html>