| 1 |
/* { dg-do compile } */ |
| 2 |
/* { dg-options "-O2 -mavx2 -fdump-tree-optimized -Wno-psabi -mno-avx512f" } */ |
| 3 |
|
| 4 |
typedef unsigned int u32v4 __attribute__((vector_size(16))); |
| 5 |
typedef unsigned short u16v16 __attribute__((vector_size(32))); |
| 6 |
typedef unsigned char u8v16 __attribute__((vector_size(16))); |
| 7 |
|
| 8 |
union vec128 { |
| 9 |
u8v16 u8; |
| 10 |
u32v4 u32; |
| 11 |
}; |
| 12 |
|
| 13 |
#define memcpy __builtin_memcpy |
| 14 |
|
| 15 |
static u16v16 zxt(u8v16 x) |
| 16 |
{ |
| 17 |
return (u16v16) { |
| 18 |
x[0], x[1], x[2], x[3], x[4], x[5], x[6], x[7], |
| 19 |
x[8], x[9], x[10], x[11], x[12], x[13], x[14], x[15] |
| 20 |
}; |
| 21 |
} |
| 22 |
|
| 23 |
static u8v16 narrow(u16v16 x) |
| 24 |
{ |
| 25 |
return (u8v16) { |
| 26 |
x[0], x[1], x[2], x[3], x[4], x[5], x[6], x[7], |
| 27 |
x[8], x[9], x[10], x[11], x[12], x[13], x[14], x[15] |
| 28 |
}; |
| 29 |
} |
| 30 |
|
| 31 |
void f(char *dst, char *src, unsigned long n, unsigned c) |
| 32 |
{ |
| 33 |
unsigned ia = 255 - (c >> 24); |
| 34 |
ia += ia >> 7; |
| 35 |
|
| 36 |
union vec128 c4 = {0}, ia16 = {0}; |
| 37 |
c4.u32 += c; |
| 38 |
ia16.u8 += (unsigned char)ia; |
| 39 |
|
| 40 |
u16v16 c16 = (zxt(c4.u8) << 8) + 128; |
| 41 |
|
| 42 |
for (; n; src += 16, dst += 16, n -= 4) { |
| 43 |
union vec128 s; |
| 44 |
memcpy(&s, src, sizeof s); |
| 45 |
s.u8 = narrow((zxt(s.u8)*zxt(ia16.u8) + c16) >> 8); |
| 46 |
memcpy(dst, &s, sizeof s); |
| 47 |
} |
| 48 |
} |
| 49 |
|
| 50 |
/* { dg-final { scan-tree-dump-times "\\(vector\\(16\\) short unsigned int\\)" 3 "optimized" } } */ |
| 51 |
/* We're missing an opportunity to, after later optimizations, combine |
| 52 |
a uniform CTOR with a vector promotion to a CTOR on a promoted |
| 53 |
element. */ |
| 54 |
/* { dg-final { scan-tree-dump-times "\\(vector\\(16\\) short unsigned int\\)" 2 "optimized" { xfail *-*-* } } } */ |