On Wed, 28 Aug 2013, Paulo Matos wrote:
Hello,
I have the feeling this is a problem with my understand of restrict so bear with me.
Consider the following:
typedef short int16_t;
typedef int16_t c16mat4by4[10/*rows*/][10/*columns*/];
void nofmat4x4mul_v6(c16mat4by4 C_out[], c16mat4by4 D_out[], c16mat4by4 A[], c16mat4by4 B[])
{
c16mat4by4 * restrict arow_p = (c16mat4by4 *)&(A[0][0][0]);
c16mat4by4 * restrict brow_p = (c16mat4by4 *)&(B[0][0][0]);
c16mat4by4 * restrict crow_p = (c16mat4by4 *)&(C_out[0][0][0]);
c16mat4by4 * restrict drow_p = (c16mat4by4 *)&(D_out[0][0][0]);
for (int i = 0; i < 4; i++)
for (int j = 0; j < 5; j++)
for (int z = 0; z < 5; z++)
{
crow_p[i][j][z] = brow_p[i][j][z] + arow_p[i][j][z];
drow_p[i][j][z] = brow_p[i][j][z];
}
}
I would assume that at each cycle brow_p[i][j][z] would only be loaded once due to brow_p and crow_p being marked as restrict. However, this doesn't happen under my port on 4.8.1. I confirmed it doesn't happen either on v850 (compiled with `gcc-v850/gcc/cc1 -O2 -std=c99 -o- restrict.c -fpreprocessed'):
.L5:
lh %r2,0(%r4,%r1)
ah %r2,0(%r12,%r1)
sth %r2,0(%r11,%r1)
lh %r2,0(%r4,%r1)
sth %r2,0(%r5,%r1)
I would have thought the second lh would have been unnecessary.
This means either alias analysis is wrong or my understanding of restrict in this case is wrong. I bet on the latter and that's why this is in gcc-help.
You are using int16_t***restrict, and the restrict only goes one layer of
pointer deep, maybe? (I didn't look very closely)
--
Marc Glisse
