| 1 |
/* { dg-do compile { target { ! riscv_abi_e } } } */ |
| 2 |
/* { dg-additional-options "-O1 -fno-schedule-insns -fno-schedule-insns2 -fno-pie" } */ |
| 3 |
/* { dg-add-options riscv_v } */ |
| 4 |
/* { dg-final { check-function-bodies "**" "" } } */ |
| 5 |
|
| 6 |
#if 0 /* Using include files when using a multilib-relevant -march option is dicey */ |
| 7 |
#include <string.h> |
| 8 |
#else |
| 9 |
extern void *memcpy(void *__restrict dest, const void *__restrict src, __SIZE_TYPE__ n); |
| 10 |
#endif |
| 11 |
|
| 12 |
#define MIN_VECTOR_BYTES (__riscv_v_min_vlen / 8) |
| 13 |
|
| 14 |
/* memcpy should be implemented using the cpymem pattern. |
| 15 |
** f1: |
| 16 |
XX \.L\d+: # local label is ignored |
| 17 |
** vsetvli\s+[ta][0-7],a2,e8,m1,ta,ma |
| 18 |
** vle8\.v\s+v\d+,0\(a1\) |
| 19 |
** vse8\.v\s+v\d+,0\(a0\) |
| 20 |
** add\s+a1,a1,[ta][0-7] |
| 21 |
** add\s+a0,a0,[ta][0-7] |
| 22 |
** sub\s+a2,a2,[ta][0-7] |
| 23 |
** bne\s+a2,zero,\.L\d+ |
| 24 |
** ret |
| 25 |
*/ |
| 26 |
|
| 27 |
void f1 (void *a, void *b, __SIZE_TYPE__ l) |
| 28 |
{ |
| 29 |
memcpy (a, b, l); |
| 30 |
} |
| 31 |
|
| 32 |
/* We should still use cpymem even with slightly different types, as signed |
| 33 |
overflow is undefined. |
| 34 |
** f2: |
| 35 |
XX \.L\d+: # local label is ignored |
| 36 |
** vsetvli\s+[ta][0-7],a2,e8,m1,ta,ma |
| 37 |
** vle8\.v\s+v\d+,0\(a1\) |
| 38 |
** vse8\.v\s+v\d+,0\(a0\) |
| 39 |
** add\s+a1,a1,[ta][0-7] |
| 40 |
** add\s+a0,a0,[ta][0-7] |
| 41 |
** sub\s+a2,a2,[ta][0-7] |
| 42 |
** bne\s+a2,zero,\.L\d+ |
| 43 |
** ret |
| 44 |
*/ |
| 45 |
void f2 (__INT32_TYPE__* a, __INT32_TYPE__* b, int l) |
| 46 |
{ |
| 47 |
memcpy (a, b, l); |
| 48 |
} |
| 49 |
|
| 50 |
/* If it is known that the pointer arguments to memcpy point |
| 51 |
to an aligned object, cpymem can use that alignment. |
| 52 |
Use extern here so that we get a known alignment, lest |
| 53 |
DATA_ALIGNMENT force us to make the scan pattern accomodate |
| 54 |
code for different alignments depending on word size. |
| 55 |
** f3: { target { no-opts "-mrvv-vector-bits=zvl" } } |
| 56 |
** ( |
| 57 |
** lui\s+[ta][0-7],%hi\(a_a\) |
| 58 |
** lui\s+[ta][0-7],%hi\(a_b\) |
| 59 |
** addi\s+[ta][0-7],[ta][0-7],%lo\(a_b\) |
| 60 |
** vsetivli\s+zero,4,e32,m1,ta,ma |
| 61 |
** | |
| 62 |
** lui\s+[ta][0-7],%hi\(a_a\) |
| 63 |
** lui\s+[ta][0-7],%hi\(a_b\) |
| 64 |
** li\s+[ta][0-7],\d+ |
| 65 |
** addi\s+[ta][0-7],[ta][0-7],%lo\(a_b\) |
| 66 |
** vsetvli\s+zero,[ta][0-7],e32,m1,ta,ma |
| 67 |
** | |
| 68 |
** lla\s+[ta][0-7],a_b |
| 69 |
** vsetivli\s+zero,4,e32,m1,ta,ma |
| 70 |
** | |
| 71 |
** li\s+[ta][0-7],\d+ |
| 72 |
** lla\s+[ta][0-7],a_b |
| 73 |
** vsetvli\s+zero,[ta][0-7],e32,m1,ta,ma |
| 74 |
** | |
| 75 |
** lla\s+[ta][0-7],a_b |
| 76 |
** li\s+[ta][0-7],32 |
| 77 |
** vsetvli\s+zero,[ta][0-7],e32,m1,ta,ma |
| 78 |
** ) |
| 79 |
** vle32.v\s+v\d+,0\([ta][0-7]\) |
| 80 |
** ( |
| 81 |
** addi\s+[ta][0-7],[ta][0-7],%lo\(a_a\) |
| 82 |
** | |
| 83 |
** lla\s+[ta][0-7],a_a |
| 84 |
** ) |
| 85 |
** vse32.v\s+v\d+,0\([ta][0-7]\) |
| 86 |
** ret |
| 87 |
*/ |
| 88 |
|
| 89 |
/* |
| 90 |
** f3: { target { any-opts "-mrvv-vector-bits=zvl" } } |
| 91 |
** ( |
| 92 |
** lui\s+[ta][0-7],%hi\(a_a\) |
| 93 |
** lui\s+[ta][0-7],%hi\(a_b\) |
| 94 |
** addi\s+[ta][0-7],[ta][0-7],%lo\(a_a\) |
| 95 |
** addi\s+[ta][0-7],[ta][0-7],%lo\(a_b\) |
| 96 |
** | |
| 97 |
** lla\s+[ta][0-7],a_a |
| 98 |
** lla\s+[ta][0-7],a_b |
| 99 |
** ) |
| 100 |
** vl1re32.v\s+v\d+,0\([ta][0-7]\) |
| 101 |
** vs1r.v\s+v\d+,0\([ta][0-7]\) |
| 102 |
** ret |
| 103 |
*/ |
| 104 |
|
| 105 |
extern struct { __INT32_TYPE__ a[MIN_VECTOR_BYTES / 4]; } a_a, a_b; |
| 106 |
|
| 107 |
void f3 () |
| 108 |
{ |
| 109 |
memcpy (&a_a, &a_b, sizeof a_a); |
| 110 |
} |
| 111 |
|
| 112 |
/* { dg-final { scan-assembler-not {\m(tail|call)\s+memcpy(?:@plt)?\M} } } */ |