https://llvm.org/bugs/show_bug.cgi?id=27107
Bug ID: 27107
Summary: LLVM misses reciprocal estimate instructions in ISel
on ARMv7
Product: new-bugs
Version: trunk
Hardware: PC
OS: All
Status: NEW
Severity: normal
Priority: P
Component: new bugs
Assignee: unassignedb...@nondot.org
Reporter: ste...@uplinklabs.net
CC: llvm-bugs@lists.llvm.org
Classification: Unclassified
LLVM is missing the opportunity to use VRSQRTE/VRSQRTS (and the other
reciprocal estimate instructions):
$ cat rsqrt.c
#include <math.h>
float rsqrtf(float f)
{
return 1.0f / sqrtf(f);
}
$ clang -O3 -mcpu=native -mfpu=neon -mfloat-abi=hard -ffast-math -S -o -
rsqrt.c | showasm
rsqrtf: @ @rsqrtf
vsqrt.f32 s0, s0
vmov.f32 s2, #1.000000e+00
vdiv.f32 s0, s2, s0
bx lr
Conversely, on x86_64, LLVM does the right thing:
$ clang -O3 -march=core-avx2 -ffast-math -S -o - rsqrt.c | showasm
rsqrtf: # @rsqrtf
vrsqrtss %xmm0, %xmm0, %xmm1
vmulss %xmm1, %xmm1, %xmm2
vfmadd213ss .LCPI1_0(%rip), %xmm0, %xmm2
vmulss .LCPI1_1(%rip), %xmm1, %xmm0
vmulss %xmm0, %xmm2, %xmm0
retq
It will even apply this properly to packed vectors if the inputs make sense for
it.
Right now the lack of reciprocal square root estimates on ARM breaks
auto-vectorization for a silly program I wrote, and the hand-written NEON
intrinsics version is beating the auto-vectorized variants (because the
auto-vectorization fails and everything gets bottlenecked around vsqrt+vdiv).
I looked at implementing this myself but got confused trying to understand
tablegen syntax. It looks like there just needs to be an ARMTargetLowering
implementation for TargetLowering::getRsqrtEstimate and
TargetLowering::getRecipEstimate.
--
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
