[RISCV] Suboptimal code generation for memset with RVV.

[tclin914]

The case there are only 128-bit width memset. The code generation of scalar version looks like better than RVV.

#include <string.h>

void foo(long *ptr) {
    memset(ptr, 0, 16);
}

-O3 -march=rv64gc

foo:
        sd      zero, 0(a0)
        sd      zero, 8(a0)
        ret

-O3 -march=rv64gcv

foo:
        vsetivli        zero, 2, e64, m1, ta, ma
        vmv.v.i v8, 0
        vse64.v v8, (a0)
        ret

https://godbolt.org/z/er7drbEcb

Should we define a size threshold to determine when memset should be translated to RVV instructions in getOptimalMemOpType?

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: Jim Lin (tclin914)

The case there are only 128-bit width memset. The code generation of scalar version looks like better than RVV.

#include <string.h>

void foo(long *ptr) {
    memset(ptr, 0, 16);
}

-O3 -march=rv64gc

foo:
        sd      zero, 0(a0)
        sd      zero, 8(a0)
        ret

-O3 -march=rv64gcv

foo:
        vsetivli        zero, 2, e64, m1, ta, ma
        vmv.v.i v8, 0
        vse64.v v8, (a0)
        ret

Should we define a size threshold to determine when memset should be translated to RVV instructions in getOptimalMemOpType?

[wangpc-pp]

#146673 may be related. cc @BoyaoWang430

[tclin914]

#146673 may be related. cc @BoyaoWang430

I didn't realize there are related issue #144562. Close this duplicated issue.