[llvm-bugs] [Bug 46621] New: [X86][AVX] Avoid vpsrlq $32, %xmm, %xmm split in v4u64 uint2fp v4f64

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=46621

            Bug ID: 46621
           Summary: [X86][AVX] Avoid vpsrlq $32, %xmm, %xmm split in v4u64
                    uint2fp v4f64
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: llvm-dev at redking.me.uk
                CC: craig.topper at gmail.com, llvm-bugs at lists.llvm.org,
                    llvm-dev at redking.me.uk, spatel+llvm at rotateright.com

Codegen: https://godbolt.org/z/Kz72eo

For AVX1 only targets, the v4u64 -> v4f64 uint2fp conversion:

define <4 x double> @cvt_v4u64_v4f64(4 x i64> %0) {
  %2 = uitofp <4 x i64> %0 to <4 x double>
  ret <4 x double> %2
}

cvt_v4u64_v4f64:
  vxorps %xmm1, %xmm1, %xmm1
  vblendps $170, %ymm1, %ymm0, %ymm1 # ymm1 =
ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
  vpsrlq $32, %xmm0, %xmm2
  vextractf128 $1, %ymm0, %xmm0
  vorps .LCPI0_0(%rip), %ymm1, %ymm1
  vpsrlq $32, %xmm0, %xmm0
  vinsertf128 $1, %xmm0, %ymm2, %ymm0
  vorpd .LCPI0_1(%rip), %ymm0, %ymm0
  vsubpd .LCPI0_2(%rip), %ymm0, %ymm0
  vaddpd %ymm0, %ymm1, %ymm0
  retq

we should be able to avoid the extractf128+2*vpsrlq and perform this as 256-bit
vshufps*2 instead, reducing register pressure.

I'm not sure yet whether this is better to handle in shuffle combining, in
TargetLowering::expandUINT_TO_FP or shift lowering.

-- 
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/20200707/686f583e/attachment.html>