[llvm-bugs] [Bug 31032] New: [X86] Extra scalar move in scalar intrinsic sequences
via llvm-bugs
llvm-bugs at lists.llvm.org
Tue Nov 15 21:34:31 PST 2016
https://llvm.org/bugs/show_bug.cgi?id=31032
Bug ID: 31032
Summary: [X86] Extra scalar move in scalar intrinsic sequences
Product: libraries
Version: trunk
Hardware: PC
OS: All
Status: NEW
Severity: normal
Priority: P
Component: Backend: X86
Assignee: unassignedbugs at nondot.org
Reporter: craig.topper at gmail.com
CC: llvm-bugs at lists.llvm.org
Classification: Unclassified
This IR sequence produces an extra move or blend of zero into the upper bits
just before the minss instruction.
; Function Attrs: nounwind
define i16 @test1(float %f) #0 {
%tmp = insertelement <4 x float> undef, float %f, i32 0
%tmp10 = insertelement <4 x float> %tmp, float 0.000000e+00, i32 1
%tmp11 = insertelement <4 x float> %tmp10, float 0.000000e+00, i32 2
%tmp12 = insertelement <4 x float> %tmp11, float 0.000000e+00, i32 3
%1 = extractelement <4 x float> %tmp12, i32 0
%2 = fsub float %1, 1.000000e+00
%3 = insertelement <4 x float> %tmp12, float %2, i32 0
%4 = extractelement <4 x float> %3, i32 0
%5 = fmul float %4, 5.000000e-01
%6 = insertelement <4 x float> %3, float %5, i32 0
%tmp48 = tail call <4 x float> @llvm.x86.sse.min.ss(<4 x float> %6, <4 x
float> <float 6.553500e+04, float 0.000000e+00, float 0.000000e+00, float
0.000000e+00>)
%tmp59 = tail call <4 x float> @llvm.x86.sse.max.ss(<4 x float> %tmp48, <4 x
float> zeroinitializer)
%tmp.upgrd.1 = tail call i32 @llvm.x86.sse.cvttss2si(<4 x float> %tmp59)
%tmp69 = trunc i32 %tmp.upgrd.1 to i16
ret i16 %tmp69
}
Output for the AVX target:
vxorps %xmm1, %xmm1, %xmm1
vaddss LCPI0_0(%rip), %xmm0, %xmm0
vmulss LCPI0_1(%rip), %xmm0, %xmm0
vblendps $1, %xmm0, %xmm1, %xmm0 ## xmm0 = xmm0[0],xmm1[1,2,3]
vminss LCPI0_2(%rip), %xmm0, %xmm0
vmaxss %xmm1, %xmm0, %xmm0
vcvttss2si %xmm0, %eax
## kill: %AX<def> %AX<kill> %EAX<kill>
retq
This occurs because the floats constant fold forward to the final insertelement
at %6 creating a vzmovl SDNode that is lowered as a blendps or movss depending
on which features are enabled. Ideally we'd realize that the min, max, and
vcvtss2si don't need the bits are move them. I suspect running this code
through InstCombine first might figure that out.
--
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/20161116/4db74236/attachment.html>
More information about the llvm-bugs
mailing list