https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106484
Bug ID: 106484
Summary: Failure to optimize uint64_t/constant division on
ARM32
Product: gcc
Version: 12.1.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: rsaxvc at gmail dot com
Target Milestone: ---
Target: arm
The following test function compiles into a call to __aeabi_uldivmod, even
though the divisor is a constant. Here's an example function:
#include <stdint.h>
uint64_t ns_to_s( uint64_t ns64 )
{
return ns64 / 1000000000ULL;
}
CortexM4(-O3 -Wall -Wextra -mcpu=cortex-m4) assembly:
ns_to_s(unsigned long long):
push {r3, lr}
adr r3, .L4
ldrd r2, [r3]
bl __aeabi_uldivmod
pop {r3, pc}
.L4:
.word 1000000000
.word 0
Interestingly, gcc 12.1 for aarch64 compiles the above C function by
implementing division by a constant with scaled multiplication by the inverse
using the umulh instruction(not present on 32-bit ARM). (-O3 -Wall -Wextra):
ns_to_s(unsigned long):
mov x1, 23123
lsr x0, x0, 9
movk x1, 0xa09b, lsl 16
movk x1, 0xb82f, lsl 32
movk x1, 0x44, lsl 48
umulh x0, x0, x1
lsr x0, x0, 11
ret
Instead, if something like __umulh could be added to libgcc, then GCC could use
the constants generated in the aarch64 logic to implement uint64_t/constant
division. Example umulh approach is attached. On Cortex-M4, the umulh-based
approach is significantly faster, although this depends on the specific libc
__aeabi_uldivmod linked against as well as the numerator.
