[llvm-bugs] [Bug 34041] New: [X86] IR shufflevector instructions with undef values lowered to sub-optimal sequences

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=34041

            Bug ID: 34041
           Summary: [X86] IR shufflevector instructions with undef values
                    lowered to sub-optimal sequences
           Product: libraries
           Version: trunk
          Hardware: All
                OS: All
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: ayman.musa at intel.com
                CC: llvm-bugs at lists.llvm.org

shufflevector IR instruction gets lowered to sub-optimal sequences when its
mask includes undef values.

Small reproducers:

define <4 x double> @foo1(double* %p) {
   %s = load double, double* %p
   %vec = insertelement <2 x double> undef, double %s, i32 0
   %res = shufflevector <2 x double> %vec, <2 x double> undef, <4 x i32> <i32
1, i32 0, i32 0, i32 0>
   ret <4 x double> %res 
 }

Clang emits vmovsd + vpermpd instructions:
     vmovsd  (%rdi), %xmm0           # xmm0 = mem[0],zero
     vpermpd $1, %ymm0, %ymm0        # ymm0 = ymm0[1,0,0,0]
     retq 

while it can be replaced with simple vbroadcastsd instruction with better
throughput:
     vbroadcastsd  (%rdi), %ymm0


define <4 x double> @foo2(<2 x double>* %vp, <4 x double> %default) {
   %vec = load <2 x double>, <2 x double>* %vp
   %shuf = shufflevector <2 x double> %vec, <2 x double> undef, <4 x i32> <i32
0, i32 3, i32 undef, i32 1>
   %res = select <4 x i1> <i1 0, i1 1, i1 0, i1 1>, <4 x double> %shuf, <4 x
double> %default
   ret <4 x double> %res 
 }

Clang emits:
     vmovapd (%rdi), %xmm1
     movb    $10, %al
     kmovd   %eax, %k1
     vpermpd $100, %ymm1, %ymm0 {%k1} # ymm0 {%k1} = ymm1[0,1,2,1]
     retq

While it can be replaced with:
     movb    $10, %al 
     kmovd   %eax, %k1 
     vbroadcastf64x2 (%rdi), %ymm0 {%k1}


define <8 x float> @foo3(<2 x float>* %vp, <8 x float> %default) {
   %vec = load <2 x float>, <2 x float>* %vp
   %shuf = shufflevector <2 x float> %vec, <2 x float> undef, <8 x i32> <i32
undef, i32 1, i32 undef, i32 undef, i32 0, i32 2, i32 3, i32 undef>
   %res = select <8 x i1> <i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 0, i1 1>, <8
x float> %shuf, <8 x float> %default
   ret <8 x float> %res
 }

clang emits:
     vmovsd  (%rdi), %xmm1           # xmm1 = mem[0],zero
     vbroadcastsd    %xmm1, %ymm1
     movb    $-65, %al
     kmovd   %eax, %k1
     vmovaps %ymm1, %ymm0 {%k1}
     retq

While it can combine the mask with a 64-bit broadcast with 32-bit mask
granularity:
     movb    $-65, %al 
     kmovd   %eax, %k1 
     vbroadcastf32x2 (%rdi), %ymm1 {%k1}
     vmovaps %ymm1, %ymm0


define <8 x double> @foo4(<2 x double>* %vp) {
  %vec = load <2 x double>, <2 x double>* %vp
  %res = shufflevector <2 x double> %vec, <2 x double> undef, <8 x i32> <i32 3,
i32 1, i32 undef, i32 1, i32 0, i32 1, i32 0, i32 1>
  ret <8 x double> %res 
}

Clang does not recognize the broadcast template in the mask and emits a shuffle
instruction:
     vmovapd (%rdi), %xmm0
     vshuff64x2  $0, %zmm0, %zmm0, %zmm0 # zmm0 = zmm0[0,1,0,1,0,1,0,1]
     retq

While it can make use of a broadcast instruction which has better throughput
than shuffle:
     vbroadcastf64x2 (%rdi), %zmm0


define <8 x double> @foo5(<2 x double>* %vp) {
; CHECK-LABEL: test_2xdouble_to_8xdouble_mem_mask11:
; CHECK:       # BB#0:
; CHECK-NEXT:    vmovapd (%rdi), %xmm0
; CHECK-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm1
; CHECK-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
; CHECK-NEXT:    retq
  %vec = load <2 x double>, <2 x double>* %vp
  %res = shufflevector <2 x double> %vec, <2 x double> undef, <8 x i32> <i32 0,
i32 undef, i32 undef, i32 undef, i32 0, i32 1, i32 0, i32 1>
  ret <8 x double> %res
}

Clang emits:
     vmovapd (%rdi), %xmm0
     vinsertf128 $1, %xmm0, %ymm0, %ymm1
     vinsertf64x4    $1, %ymm1, %zmm0, %zmm0
     retq

While it can use the broadcast also here:
     vbroadcastf64x2 (%rdi), %zmm0

-- 
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/20170803/bb43d708/attachment-0001.html>