https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98143
Bug ID: 98143
Summary: arm: missed vectorization with MVE compared to Neon
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: clyon at gcc dot gnu.org
Target Milestone: ---
While working on enabling auto-vectorization for MVE, I noticed a missed
optimization compared to Neon:
#include <stdint.h>
uint16_t *dest;
void func()
{
int i;
for (i=0;i<16;i++)
dest[i]=3;
}
Compiled with -O3 -S -dp -mfloat-abi=hard -mfpu=auto -mcpu=cortex-a9 -mthumb:
func:
movw r3, #:lower16:.LANCHOR0 @ 15 [c=4 l=4] *thumb2_movsi_vfp/4
vmov.i16 q8, #3 @ v8hi @ 7 [c=4 l=4] *neon_movv8hi/2
movt r3, #:upper16:.LANCHOR0 @ 16 [c=4 l=4] *arm_movt/0
ldr r3, [r3] @ 14 [c=12 l=4] *thumb2_movsi_vfp/5
vst1.16 {q8}, [r3]! @ 8 [c=8 l=4] *movmisalignv8hi_neon_store
vst1.16 {q8}, [r3] @ 11 [c=8 l=4] *movmisalignv8hi_neon_store
bx lr @ 44 [c=8 l=4] *thumb2_return
Compiled with -O3 -S -dp -mfloat-abi=hard -mfpu=auto -march=armv8.1-m.main+mve
-mthumb:
func:
movs r2, #3 @ 7 [c=4 l=2] *thumb2_movsi_shortim
ldr r3, .L3 @ 5 [c=12 l=4] *thumb2_movsi_vfp/5
ldr r3, [r3] @ 6 [c=12 l=4] *thumb2_movsi_vfp/5
strh r2, [r3] @ movhi @ 9 [c=4 l=4] *thumb2_movhi_vfp/4
strh r2, [r3, #2] @ movhi @ 12 [c=4 l=4] *thumb2_movhi_vfp/4
strh r2, [r3, #4] @ movhi @ 15 [c=4 l=4] *thumb2_movhi_vfp/4
strh r2, [r3, #6] @ movhi @ 18 [c=4 l=4] *thumb2_movhi_vfp/4
strh r2, [r3, #8] @ movhi @ 21 [c=4 l=4] *thumb2_movhi_vfp/4
strh r2, [r3, #10] @ movhi @ 24 [c=4 l=4] *thumb2_movhi_vfp/4
strh r2, [r3, #12] @ movhi @ 27 [c=4 l=4] *thumb2_movhi_vfp/4
strh r2, [r3, #14] @ movhi @ 30 [c=4 l=4] *thumb2_movhi_vfp/4
strh r2, [r3, #16] @ movhi @ 33 [c=4 l=4] *thumb2_movhi_vfp/4
strh r2, [r3, #18] @ movhi @ 36 [c=4 l=4] *thumb2_movhi_vfp/4
strh r2, [r3, #20] @ movhi @ 39 [c=4 l=4] *thumb2_movhi_vfp/4
strh r2, [r3, #22] @ movhi @ 42 [c=4 l=4] *thumb2_movhi_vfp/4
strh r2, [r3, #24] @ movhi @ 45 [c=4 l=4] *thumb2_movhi_vfp/4
strh r2, [r3, #26] @ movhi @ 48 [c=4 l=4] *thumb2_movhi_vfp/4
strh r2, [r3, #28] @ movhi @ 51 [c=4 l=4] *thumb2_movhi_vfp/4
strh r2, [r3, #30] @ movhi @ 54 [c=4 l=4] *thumb2_movhi_vfp/4
bx lr @ 84 [c=8 l=4] *thumb2_return
This PR is about building the const, as the problems with stores are probably
part of PR97875.
In summry, with Neon we build the constant vector with:
vmov.i16 q8, #3 @ v8hi @ 7 [c=4 l=4] *neon_movv8hi/2
but with MVE:
movs r2, #3 @ 7 [c=4 l=2] *thumb2_movsi_shortim
and then store it as 16-bits value as many times as needed.
I haven't managed to understand why we can't make use of mve.md's mve_mov<mode>
where there is an alternative with "Dm", which should work?
