[LLVMbugs] [Bug 17479] New: Vector conversion problem with x86 SSE/AVX

bugzilla-daemon at llvm.org bugzilla-daemon at llvm.org
Fri Oct 4 11:50:10 PDT 2013


            Bug ID: 17479
           Summary: Vector conversion problem with x86 SSE/AVX
           Product: tools
           Version: 3.2
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: llc
          Assignee: unassignedbugs at nondot.org
          Reporter: arsene.perard at laposte.net
                CC: llvmbugs at cs.uiuc.edu
    Classification: Unclassified

I cannot get llc to generate correct code for a special sequence of
instructions. I actually compare two vectors of floats, get a vector of
booleans, and I want to convert this into an integer bitmask. To do so, I use a
bitcast to convert the boolean vector into a i4 value, and then a zero extend
operation to get an integer. I actually would expect the compiler to output a
movmskps instruction. Here is the code :

@num = private unnamed_addr constant [4 x i8] c"%d\0A\00"
@a = private unnamed_addr constant <4 x float> <float 0.0, float 1.0, float
2.0, float 3.0>
@b = private unnamed_addr constant <4 x float> <float 3.0, float 2.0, float
1.0, float 0.0>

declare i32 @printf(i8* nocapture, i32) nounwind

define i32 @test(<4 x float>* %a, <4 x float>* %b)
    %1 = load <4 x float>* %a, align 16
    %2 = load <4 x float>* %b, align 16
    %3 = fcmp ogt <4 x float> %1, %2
    %4 = bitcast <4 x i1> %3 to i4
    %5 = zext i4 %4 to i32
    ret i32 %5

define i32 @main() {
    %res = call i32 @test(<4 x float>* @a, <4 x float>* @b)
    %str = getelementptr [4 x i8]* @num, i64 0, i64 0
    call i32 @printf(i8* %str, i32 %res)
    ret i32 0

And here is the output I get :

test:                                   # @test
# BB#0:
    vmovaps    (%rsi), %xmm0
    vcmpltps    (%rdi), %xmm0, %xmm0
    vpextrd    $2, %xmm0, %ecx
    vpextrd    $1, %xmm0, %eax
    andl    $1, %ecx
    vpextrd    $3, %xmm0, %edx
    andl    $1, %edx
    movb    %dl, -4(%rsp)
    movb    %cl, -4(%rsp)
    andl    $1, %eax
    movb    %al, -4(%rsp)
    vmovd    %xmm0, %eax
    andl    $1, %eax
    movb    %al, -4(%rsp)
    movzbl    -4(%rsp), %eax

Clearly, this code is not correct : it only shows the value of the first

(Code :
    vmovd    %xmm0, %eax
    andl    $1, %eax

not the expected bitmask. I'm compiling with "lcc test.ll" and my machine is an
opensuse distribution (12.3) using the lastest packages (llvm 3.2).

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/20131004/04afa2cb/attachment.html>

More information about the llvm-bugs mailing list