[llvm-bugs] [Bug 28131] New: Register allocator unnecessarily spilling in vector code

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=28131

            Bug ID: 28131
           Summary: Register allocator unnecessarily spilling in vector
                    code
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: Register Allocator
          Assignee: unassignedbugs at nondot.org
          Reporter: mkuper at google.com
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

Created attachment 16543
  --> https://llvm.org/bugs/attachment.cgi?id=16543&action=edit
Reproducer

For the attached IR, we get:

.LBB0_1:                                # %vector.body
                                        # =>This Inner Loop Header: Depth=1
    vmovdqa    c+976(%rax), %xmm0
    vmovdqa    c+960(%rax), %xmm1
    vmovdqa    c+1008(%rax), %xmm2
    vmovdqa    c+992(%rax), %xmm3
    vpaddb    b+992(%rax), %xmm3, %xmm3
    vpaddb    b+1008(%rax), %xmm2, %xmm2
    vpaddb    b+960(%rax), %xmm1, %xmm1
    vpaddb    b+976(%rax), %xmm0, %xmm0
    vmovdqa    %xmm0, a+976(%rax)
    vmovdqa    %xmm1, a+960(%rax)
    vmovdqa    %xmm2, a+1008(%rax)
    vmovdqa    %xmm3, a+992(%rax)
    vmovdqa    w+4048(,%rax,4), %xmm8
    vmovdqa    w+4032(,%rax,4), %xmm9
    vmovdqa    w+4080(,%rax,4), %xmm10
    vmovdqa    w+4064(,%rax,4), %xmm11
    vmovdqa    w+3984(,%rax,4), %xmm12
    vmovdqa    w+3968(,%rax,4), %xmm13
    vmovdqa    w+4016(,%rax,4), %xmm14
    vmovdqa    w+4000(,%rax,4), %xmm15
    vmovdqa    w+3920(,%rax,4), %xmm0
    vmovdqa    w+3904(,%rax,4), %xmm1
    vmovdqa    w+3952(,%rax,4), %xmm2
    vmovdqa    w+3936(,%rax,4), %xmm3
    vmovdqa    w+3856(,%rax,4), %xmm4
    vmovdqa    w+3840(,%rax,4), %xmm5
    vmovdqa    w+3888(,%rax,4), %xmm6
    vmovdqa    w+3872(,%rax,4), %xmm7
    vpaddd    v+3872(,%rax,4), %xmm7, %xmm7
    vmovdqa    %xmm7, -24(%rsp)        # 16-byte Spill
    vpaddd    v+3888(,%rax,4), %xmm6, %xmm6
    vpaddd    v+3840(,%rax,4), %xmm5, %xmm5
    vpaddd    v+3856(,%rax,4), %xmm4, %xmm4
    vpaddd    v+3936(,%rax,4), %xmm3, %xmm3
    vpaddd    v+3952(,%rax,4), %xmm2, %xmm2
    vpaddd    v+3904(,%rax,4), %xmm1, %xmm1
    vpaddd    v+3920(,%rax,4), %xmm0, %xmm0
    vpaddd    v+4000(,%rax,4), %xmm15, %xmm15
    vpaddd    v+4016(,%rax,4), %xmm14, %xmm14
    vpaddd    v+3968(,%rax,4), %xmm13, %xmm13
    vpaddd    v+3984(,%rax,4), %xmm12, %xmm12
    vpaddd    v+4064(,%rax,4), %xmm11, %xmm11
    vpaddd    v+4080(,%rax,4), %xmm10, %xmm10
    vpaddd    v+4032(,%rax,4), %xmm9, %xmm9
    vpaddd    v+4048(,%rax,4), %xmm8, %xmm7
    vmovdqa    %xmm7, u+4048(,%rax,4)
    vmovdqa    %xmm9, u+4032(,%rax,4)
    vmovdqa    %xmm10, u+4080(,%rax,4)
    vmovdqa    %xmm11, u+4064(,%rax,4)
    vmovdqa    %xmm12, u+3984(,%rax,4)
    vmovdqa    %xmm13, u+3968(,%rax,4)
    vmovdqa    %xmm14, u+4016(,%rax,4)
    vmovdqa    %xmm15, u+4000(,%rax,4)
    vmovdqa    %xmm0, u+3920(,%rax,4)
    vmovdqa    %xmm1, u+3904(,%rax,4)
    vmovdqa    %xmm2, u+3952(,%rax,4)
    vmovdqa    %xmm3, u+3936(,%rax,4)
    vmovdqa    %xmm4, u+3856(,%rax,4)
    vmovdqa    %xmm5, u+3840(,%rax,4)
    vmovdqa    %xmm6, u+3888(,%rax,4)
    vmovaps    -24(%rsp), %xmm0        # 16-byte Reload
    vmovaps    %xmm0, u+3872(,%rax,4)
    addq    $64, %rax
    jne    .LBB0_1

There's absolutely no reason to spill xmm7 - the last vpadd could clobber xmm8
just like the rest of the adds do (and that's regardless of the fact we could
get around it by scheduling.)

-- 
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/20160614/961b3d0d/attachment-0001.html>