| 1 |
/* { dg-do compile { target { powerpc64le-*-* } } } */ |
| 2 |
/* { dg-options "-mdejagnu-cpu=power8 -O3" } */ |
| 3 |
|
| 4 |
/* Verify that swap optimization properly removes swaps for unaligned |
| 5 |
vector stores. See PR65456. */ |
| 6 |
|
| 7 |
typedef unsigned char UChar; |
| 8 |
typedef unsigned short UShort; |
| 9 |
typedef unsigned int UWord; |
| 10 |
|
| 11 |
typedef unsigned long SizeT; |
| 12 |
typedef unsigned long Addr; |
| 13 |
|
| 14 |
void *memmove(void *dst, const void *src, SizeT len) |
| 15 |
{ |
| 16 |
const Addr WS = sizeof(UWord);/* 8 or 4 */ |
| 17 |
const Addr WM = WS - 1;/* 7 or 3 */ |
| 18 |
|
| 19 |
/* Copying backwards. */ |
| 20 |
SizeT n = len; |
| 21 |
Addr d = (Addr) dst; |
| 22 |
Addr s = (Addr) src; |
| 23 |
|
| 24 |
if (((s ^ d) & WM) == 0) { |
| 25 |
/* s and d have same UWord alignment. */ |
| 26 |
/* Pull up to a UWord boundary. */ |
| 27 |
while ((s & WM) != 0 && n >= 1) { |
| 28 |
*(UChar *) d = *(UChar *) s; |
| 29 |
s += 1; |
| 30 |
d += 1; |
| 31 |
n -= 1; |
| 32 |
} |
| 33 |
/* Copy UWords. */ |
| 34 |
while (n >= WS) { |
| 35 |
*(UWord *) d = *(UWord *) s; |
| 36 |
s += WS; |
| 37 |
d += WS; |
| 38 |
n -= WS; |
| 39 |
} |
| 40 |
if (n == 0) |
| 41 |
return dst; |
| 42 |
} |
| 43 |
if (((s | d) & 1) == 0) { |
| 44 |
/* Both are 16-aligned; copy what we can thusly. */ |
| 45 |
while (n >= 2) { |
| 46 |
*(UShort *) d = *(UShort *) s; |
| 47 |
s += 2; |
| 48 |
d += 2; |
| 49 |
n -= 2; |
| 50 |
} |
| 51 |
} |
| 52 |
/* Copy leftovers, or everything if misaligned. */ |
| 53 |
while (n >= 1) { |
| 54 |
*(UChar *) d = *(UChar *) s; |
| 55 |
s += 1; |
| 56 |
d += 1; |
| 57 |
n -= 1; |
| 58 |
} |
| 59 |
|
| 60 |
return dst; |
| 61 |
} |
| 62 |
|
| 63 |
/* { dg-final { scan-assembler-not "xxpermdi" } } */ |
| 64 |
/* { dg-final { scan-assembler-not "xxswapd" } } */ |