https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116560
Bug ID: 116560
Summary: RISC-V : rv32 code optimization , big code difference
between 8/9.x and 10.x
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: [email protected]
Target Milestone: ---
C code (mostly swap big-endian to u16 variable + a test) :
#include <stdint.h>
uint8_t test_swap(uint8_t *ptr, uint32_t *res)
{
uint16_t temp=(ptr[0]<<8)|ptr[1];
*res=temp;
if (temp==0) return 1;
return 0;
}
Up to 9.x, with -O2, the generated code is great , with 1 OR:
test_swap:
lbu a5,0(a0)
slli a5,a5,8
lbu a0,1(a0)
or a0,a5,a0
sw a0,0(a1)
seqz a0,a0
ret
After 9.x (tested on godbolt.org - https://godbolt.org/z/f38EaosxM ),
useless slli/srli, and 2 OR...
test_swap:
lbu a5,1(a0)
lbu a4,0(a0)
slli a0,a5,8
or a0,a0,a4 // ptr[1]<<8 | ptr[0] ??? why ? want the opposite
slli a5,a0,8 // ptr[1]<<16 | ptr[0]<<8 | 0
srli a0,a0,8 // ptr[1]
or a0,a5,a0 // ptr[1]<<16 | ptr[0]<<8 | ptr[1]
slli a0,a0,16 // and 0xFFFF
srli a0,a0,16
sw a0,0(a1) // ptr[0]<<8 | ptr[1]
seqz a0,a0
ret
Actually, up to 12.x with -O1, the code is the same as 8.x/9.x
After that, with 13.x and -O1, a pair of useless slli/srli (i.e.
and 0xFFFF) is added :
test_swap:
lbu a5,0(a0)
slli a5,a5,8
lbu a4,1(a0)
or a0,a5,a4
slli a0,a0,16 // not necessary
srli a0,a0,16 // not necessary
sw a0,0(a1)
seqz a0,a0
ret
