https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119187

--- Comment #16 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
Friday vectorizer codegen fun.

```
#include <arm_neon.h>
#include <arm_sve.h>

void foo (uint8_t *a, uint8_t *b, uint8_t *c, int n)
{
  for (int i = 0; i < n; i+=8, a+=8, b+=8, c+=8)
    {
        uint8x8_t av1 = vld1_u8 (a);
        uint8x8_t bv1 = vld1_u8 (b);
        uint8x8_t cmp = vcgt_u8 (bv1, vdup_n_u8 (0));
        uint8x8_t add = vadd_u8 (bv1, vdup_n_u8 (1));
        uint8x8_t sel = vbsl_u8 (cmp, add, bv1);
        vst1_u8 (a, sel);
    }
}
```
now when compiled with `-O3 -march=armv9-a` instead of
```
.L3:
        ldr     d28, [x1, x2]
        cmtst   v27.8b, v28.8b, v28.8b
        add     v26.8b, v28.8b, v29.8b
        bif     v26.8b, v28.8b, v27.8b
        str     d26, [x0, x2]
        add     x2, x2, 8
        cmp     w3, w2
        bgt     .L3
```
generates
```        mov     z30.b, #1
.L4:
        ld1b    z31.b, p7/z, [x1, x2]
        cmpne   p6.b, p7/z, z31.b, #0
        add     z31.b, p6/m, z31.b, z30.b
        st1b    z31.b, p7, [x0, x2]
        incb    x2
        whilelo p7.b, x2, x3
        b.any   .L4```
with the vectorizer reporting
```
 note:   ------>vectorizing SLP node starting from: _14 =
__builtin_aarch64_simd_bslv8qi_uuuu (_11, _13, _10);
 note:   vect_is_simple_use: operand patt_22 ? -1 : 0, type of def: internal
 note:   vect_is_simple_use: operand _10 + { 1, 1, 1, 1, 1, 1, 1, 1 }, type of
def: internal
 note:   vect_is_simple_use: operand # VUSE <.MEM_26>
MEM <__Uint8x8_tD.2485> [(unsigned charD.21 * {ref-all})b_24], type of def:
internal
 note:   transform vector call.
 note:   add new stmt: vect_vect_patt_21.19_63 =
VIEW_CONVERT_EXPR<svuint8_t>(vect_patt_21.16_60);
 note:   inserted type shim between: svuint8_t and vector([16,16]) unsigned
char due to mismatch between generic type and vector call.
 note:   add new stmt: vect_vect__13.20_64 =
VIEW_CONVERT_EXPR<svuint8_t>(vect__13.17_61);
 note:   inserted type shim between: svuint8_t and vector([16,16]) unsigned
char due to mismatch between generic type and vector call.
 note:   add new stmt: vect_vect__10.21_65 =
VIEW_CONVERT_EXPR<svuint8_t>(vect__10.14_58);
 note:   inserted type shim between: svuint8_t and vector([16,16]) unsigned
char due to mismatch between generic type and vector call.
 note:   add new stmt: vect__14.22_66 = svbsl_u8 (vect_vect_patt_21.19_63,
vect_vect__13.20_64, vect_vect__10.21_65);
 note:   add new stmt: vect__14.18_62 = VIEW_CONVERT_EXPR<vector([16,16])
unsigned char>(vect__14.22_66);
 note:   inserted type shim between: svuint8_t and vector([16,16]) unsigned
char due to mismatch between generic type and vector call.
```
and at -O3 -march=armv8-a
```
.L4:
        ldr     q30, [x1, x2]
        cmtst   v29.16b, v30.16b, v30.16b
        add     v28.16b, v30.16b, v31.16b
        bif     v28.16b, v30.16b, v29.16b
        str     q28, [x0, x2]
        add     x2, x2, 16
        cmp     x2, x5
        bne     .L4
        cmp     w4, w7, lsl 4
        beq     .L1
        lsl     x4, x6, 7
        movi    v25.8b, 0x1
        lsl     w2, w7, 7
        add     x6, x1, x4
        add     w7, w2, 8
        add     x5, x0, x4
        ldr     d31, [x1, x4]
        cmtst   v30.8b, v31.8b, v31.8b
        add     v29.8b, v31.8b, v25.8b
        bif     v29.8b, v31.8b, v30.8b
        str     d29, [x0, x4]
        cmp     w3, w7
        ble     .L1
```
with
```
 note:   ------>vectorizing SLP node starting from: _14 =
__builtin_aarch64_simd_bslv8qi_uuuu (_11, _13, _10);
 note:   vect_is_simple_use: operand patt_22 ? -1 : 0, type of def: internal
 note:   vect_is_simple_use: operand _10 + { 1, 1, 1, 1, 1, 1, 1, 1 }, type of
def: internal
 note:   vect_is_simple_use: operand # VUSE <.MEM_26>
MEM <__Uint8x8_tD.2485> [(unsigned charD.21 * {ref-all})b_24], type of def:
internal
 note:   transform vector call.
 note:   add new stmt: vect_vect_patt_21.23_90 =
VIEW_CONVERT_EXPR<__Uint8x16_t>(vect_patt_21.20_87);
 note:   inserted type shim between: __Uint8x16_t and vector(16) unsigned char
due to mismatch between generic type and vector call.
 note:   add new stmt: vect_vect__13.24_91 =
VIEW_CONVERT_EXPR<__Uint8x16_t>(vect__13.21_88);
 note:   inserted type shim between: __Uint8x16_t and vector(16) unsigned char
due to mismatch between generic type and vector call.
 note:   add new stmt: vect_vect__10.25_92 =
VIEW_CONVERT_EXPR<__Uint8x16_t>(vect__10.18_85);
 note:   inserted type shim between: __Uint8x16_t and vector(16) unsigned char
due to mismatch between generic type and vector call.
 note:   add new stmt: vect__14.26_93 = __builtin_aarch64_simd_bslv16qi_uuuu
(vect_vect_patt_21.23_90, vect_vect__13.24_91, vect_vect__10.25_92);
 note:   add new stmt: vect__14.22_89 = VIEW_CONVERT_EXPR<vector(16) unsigned
char>(vect__14.26_93);
 note:   inserted type shim between: __Uint8x16_t and vector(16) unsigned char
due to mismatch between generic type and vector call.
```

Things left before I send an RFC:

* vector patterns
* reductions (including accum splits)
* inductions
* unpacking (matching _hi/_lo)
* permutes
* alias analysis
* if-convert

Reply via email to