https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111931
Bug ID: 111931
Summary: RISC-V: Trivial optimization of VSETVL PASS
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: juzhe.zhong at rivai dot ai
Target Milestone: ---
#include "riscv_vector.h"
void foo9 (int8_t *base, int8_t* out, size_t vl, size_t m)
{
vint8mf8_t v0;
size_t avl = __riscv_vsetvl_e8mf8 (vl);
for (size_t i = 0; i < m; i++)
{
v0 = __riscv_vle8_v_i8mf8 (base + i, avl);
__riscv_vse8_v_i8mf8 (out + i, v0, avl);
}
}
ASM:
vsetvli a2,a2,e8,mf8,ta,ma
beq a3,zero,.L8
add a3,a0,a3
.L3:
vle8.v v1,0(a0)
addi a0,a0,1
vse8.v v1,0(a1)
addi a1,a1,1
bne a0,a3,.L3
.L8:
ret
The vsetvl should be optimized into "vsetvl zero, a2" instead of "vsetvl a2,a2"
The reason we failed to optimize it is because we set the entry block as
unknown block when computing reaching_def. So preds_has_same_avl_p is false.
As long as there are predecessors block before the user vsetvl, we can optimize
it.
This is an trivial issue. I will fix it if I find the time.
