[LLVMbugs] [Bug 17185] New: match broadcast instruction to data type (int or FP)

[bugzilla-daemon at llvm.org]

http://llvm.org/bugs/show_bug.cgi?id=17185

            Bug ID: 17185
           Summary: match broadcast instruction to data type (int or FP)
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: normal
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: kkhoo at perfwizard.com
                CC: llvmbugs at cs.uiuc.edu
    Classification: Unclassified

Here's a function with codegen for an AVX2-equipped machine. I think LLVM does
a decent job generating the main loop to store using AVX instructions 16-bytes
at a time (could have used 32-byte stores?), but there appears to be an
instruction selection problem because 'vbroadcastsd' is used to splat the
integer values into a ymm register. 

Intel added integer variants of the broadcast instructions with AVX2 because
there's a hardware penalty for mixing integer and FP data in the CPU (
explanation from Intel here:
http://software.intel.com/en-us/forums/topic/280799 ). It would be better if
LLVM used 'vpbroadcastd' in this code.

$ ./clang -v
clang version 3.4 (trunk 189776)
Target: x86_64-apple-darwin11.4.2
Thread model: posix

$ cat broadcast.c
void foo(int *x, int n, int val) {
   int i;
   for(i=0; i<n; i++) {
      x[i] = val + i;
   }
}

$ ./clang -S -O3 -fomit-frame-pointer -march=core-avx2 -o /dev/stdout
broadcast.c 
    .section    __TEXT,__text,regular,pure_instructions
    .section    __TEXT,__const
    .align    5
LCPI0_0:
    .quad    4                       ## 0x4
    .quad    5                       ## 0x5
    .quad    6                       ## 0x6
    .quad    7                       ## 0x7
LCPI0_1:
    .quad    0                       ## 0x0
    .quad    1                       ## 0x1
    .quad    2                       ## 0x2
    .quad    3                       ## 0x3
LCPI0_2:
    .long    0                       ## 0x0
    .long    2                       ## 0x2
    .long    4                       ## 0x4
    .long    6                       ## 0x6
    .long    0                       ## 0x0
    .long    0                       ## 0x0
    .long    0                       ## 0x0
    .long    0                       ## 0x0
LCPI0_3:
    .quad    8                       ## 0x8
    .quad    9                       ## 0x9
    .quad    10                      ## 0xa
    .quad    11                      ## 0xb
    .section    __TEXT,__text,regular,pure_instructions
    .globl    _foo
    .align    4, 0x90
_foo:                                   ## @foo
    .cfi_startproc
## BB#0:                                ## %entry
    testl    %esi, %esi
    jle    LBB0_7
## BB#1:                                ## %for.body.lr.ph
    movl    $2863311531, %eax       ## imm = 0xAAAAAAAB
    movl    %esi, %r8d
    imulq    %r8, %rax
    shrq    $35, %rax
    imull    $12, %eax, %ecx
    xorl    %r9d, %r9d
    movl    %esi, %eax
    subl    %ecx, %eax
    testl    %ecx, %ecx
    je    LBB0_4
## BB#2:                                ## %vector.ph
    movslq    %edx, %r10
    movq    %r8, %r9
    subq    %rax, %r9
    movl    %eax, %ecx
    movq    %r8, %r11
    subq    %rcx, %r11
    xorl    %ecx, %ecx
    vmovd    %r10, %xmm0
    vbroadcastsd    %xmm0, %ymm0
    vmovdqa    LCPI0_0(%rip), %ymm1
    vmovdqa    LCPI0_1(%rip), %ymm2
    vmovdqa    LCPI0_2(%rip), %ymm3
    vmovdqa    LCPI0_3(%rip), %ymm4
    .align    4, 0x90
LBB0_3:                                 ## %vector.body
                                        ## =>This Inner Loop Header: Depth=1
    vmovd    %rcx, %xmm5
    vpermq    $0, %ymm5, %ymm5        ## ymm5 = ymm5[0,0,0,0]
    vpaddq    %ymm1, %ymm5, %ymm6
    vpaddq    %ymm2, %ymm5, %ymm7
    vpaddq    %ymm0, %ymm7, %ymm7
    vpermd    %ymm7, %ymm3, %ymm7
    vmovdqu    %xmm7, (%rdi,%rcx,4)
    vpaddq    %ymm0, %ymm6, %ymm6
    vpaddq    %ymm4, %ymm5, %ymm5
    vpaddq    %ymm0, %ymm5, %ymm5
    vpermd    %ymm5, %ymm3, %ymm5
    vpermd    %ymm6, %ymm3, %ymm6
    leaq    12(%rcx), %rax
    cmpq    %rax, %r11
    vmovdqu    %xmm6, 16(%rdi,%rcx,4)
    vmovdqu    %xmm5, 32(%rdi,%rcx,4)
    movq    %rax, %rcx
    jne    LBB0_3
LBB0_4:                                 ## %middle.block
    cmpq    %r9, %r8
    je    LBB0_7
## BB#5:                                ## %for.body.preheader
    leaq    (%rdi,%r9,4), %rax
    subl    %r9d, %esi
    addl    %r9d, %edx
    .align    4, 0x90
LBB0_6:                                 ## %for.body
                                        ## =>This Inner Loop Header: Depth=1
    movl    %edx, (%rax)
    incl    %edx
    addq    $4, %rax
    decl    %esi
    jne    LBB0_6
LBB0_7:                                 ## %for.end
    vzeroupper
    ret

-- 
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/20130910/84fdb079/attachment.html>