https://llvm.org/bugs/show_bug.cgi?id=30371
Bug ID: 30371
Summary: [x86, AVX] poor codegen for shufflevector with
constants
Product: libraries
Version: trunk
Hardware: PC
OS: All
Status: NEW
Severity: normal
Priority: P
Component: Backend: X86
Assignee: unassignedb...@nondot.org
Reporter: spatel+l...@rotateright.com
CC: llvm-bugs@lists.llvm.org
Classification: Unclassified
Testing with llc built from r281342:
define <4 x float> @does_it_blend(float %f) {
%a0 = insertelement <4 x float> undef, float %f, i32 0
%ret = shufflevector <4 x float> %a0, <4 x float> <float undef, float 4.0,
float 5.0, float 3.0>, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
ret <4 x float> %ret
}
This should be a blendps or insertps with a vector constant, but we scalarized
the constant loads, so what could have been 2 ops (load+blend) became 6 ops:
$ ./llc blender.ll -o - -mattr=avx
vinsertps $16, LCPI0_0(%rip), %xmm0, %xmm0 ## xmm0 =
xmm0[0],mem[0],xmm0[2,3]
vinsertps $32, LCPI0_1(%rip), %xmm0, %xmm0 ## xmm0 =
xmm0[0,1],mem[0],xmm0[3]
vinsertps $48, LCPI0_2(%rip), %xmm0, %xmm0 ## xmm0 = xmm0[0,1,2],mem[0]
retq
--
You are receiving this mail because:
You are on the CC list for the bug.
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
