[Bug tree-optimization/121766] New: [16 regression] Regressions in Snappy workload for -mcpu=neoverse-v2 after modifying costing hooks

[soumyaa at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121766

            Bug ID: 121766
           Summary: [16 regression] Regressions in Snappy workload for
                    -mcpu=neoverse-v2 after modifying costing hooks
           Product: gcc
           Version: 16.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: soumyaa at gcc dot gnu.org
                CC: tamar.christina at arm dot com
  Target Milestone: ---

Hi, 

After the following commit:
https://gcc.gnu.org/cgit/gcc/commit/?id=08cdd61e


For the following code, compiled with:
-O3 -mcpu=neoverse-v2 -msve-vector-bits=128

#define SIZE 1024
void test (char* data) {

    for (int i = 0; i < SIZE; i++) {
        data[i] = i;
    }

}

NOW:

test:
        index   z31.s, #0, #1
        movi    v30.4s, 0x4
        add     x1, x0, 1024
        movi    v29.4s, 0x8
        movi    v28.4s, 0xc
        movi    v27.4s, 0x10
        .p2align 5,,15
.L2:
        add     v26.4s, v31.4s, v30.4s # 4-7
        add     v25.4s, v31.4s, v29.4s # 8-11
        add     v24.4s, v31.4s, v28.4s # 12-15
        uzp1    v26.8h, v31.8h, v26.8h # 0-3, 4-7
        uzp1    v24.8h, v25.8h, v24.8h # 8-11, 12-15
        add     v31.4s, v31.4s, v27.4s # 16-19
        uzp1    v24.16b, v26.16b, v24.16b # 0-16
        str     q24, [x0], 16
        cmp     x1, x0
        bne     .L2
        ret

BEFORE:

test:
.LFB11:
        .cfi_startproc
        add     x1, x0, 1024
        index   z31.s, #0, #1
        ptrue   p7.b, vl16
        .p2align 5,,15
.L2:
        movprfx z30, z31
        add     z30.s, z30.s, #4
        add     x0, x0, 8
        st1b    z31.s, p7, [x0, #-2, mul vl]
        st1b    z30.s, p7, [x0, #-1, mul vl]
        add     z31.s, z31.s, #8
        cmp     x1, x0
        bne     .L2
        ret

Seems like we no longer capture these as possible SVE candidates after
modifying the target hook. 

This is seen in BM_UIOVecSink/0 in Snappy, which has had a ~16% regression post
this commit.