http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50713
vincenzo Innocente <vincenzo.innocente at cern dot ch> changed:
What |Removed |Added
----------------------------------------------------------------------------
Summary|SLP vs loop: code generated |SLP vs loop: code generated
|differs |differs (SLP less
| |efficient)
--- Comment #5 from vincenzo Innocente <vincenzo.innocente at cern dot ch>
2012-10-25 13:19:23 UTC ---
I think that here the problem is similar
using 4.8 trunk
in the following code the explicit loop is way more efficient that the vector
notation
for all corei7, corei7-avx and bdver1 architectures.
(i spare you the assembler dumps. Inferring the number of instructions from nm
should enough
0000000000000170 T dfmav(double __vector, double __vector, double __vector,
double, double)
0000000000000370 T dfmal(double __vector, double __vector, double __vector,
double, double)
00000000000003c0 T dfma8v(double __vector, double __vector, double __vector,
double, double)
0000000000000810 T dfma8l(double __vector, double __vector, double __vector,
double, double)
00000000000008b0 short EH_frame1
)
typedef double __attribute__( ( vector_size( 32 ) ) ) float64x4_t;
typedef double __attribute__( ( vector_size( 64 ) ) ) float64x8_t;
float64x4_t dfmav(float64x4_t x, float64x4_t y, float64x4_t z, double a, double
b) {
return a*x*(y+b*z);
}
float64x4_t dfmal(float64x4_t x, float64x4_t y, float64x4_t z, double a, double
b) {
float64x4_t r;
for (int i=0; i!=4;++i)
r[i] = a*x[i]*(y[i]+b*z[i]);
return r;
}
float64x8_t dfma8v(float64x8_t x, float64x8_t y, float64x8_t z, double a,
double b) {
return a*x*(y+b*z);
}
float64x8_t dfma8l(float64x8_t x, float64x8_t y, float64x8_t z, double a,
double b) {
float64x8_t r;
for (int i=0; i!=8;++i)
r[i] = a*x[i]*(y[i]+b*z[i]);
return r;
}
