[Bug rtl-optimization/31455] New: [4.3 Regression] lower subreg causes a performance regression in the inner loop sometimes

[pinskia at gcc dot gnu dot org]

Testcase:
#include <stdio.h>
#include <stdlib.h>
#define SIZE 256*1024*1024
float   *data;
static inline
double one()
{
        int i;
        double sum;
        sum = 0;
        for (i=0; i<SIZE; i++)
                sum += data[i];
        return sum;
}
int f(double);
int main(int argc,char** argv)
{
        struct timeval  tv0,tv1;
        double                  s0,s1;
        int                             i;

        data = malloc(SIZE*sizeof(float));
        for (i=0; i<SIZE; i++)
                data[i] = 1;
        s0 = 0;
        for (i=0; i<SIZE; i++)
                s0 += data[i];
        printf("%f\n", s0);
                s1 = one();
        printf("%f\n", s1);
        free(data);
        return 0;
}



--------------
At -O2 -static on powerpc-darwin, we get an inner loop with:
L6:
        lfsx f0,r2,r9
        addi r2,r2,4
        lfd f13,56(r1)
        fadd f13,f13,f0
        stfd f13,56(r1)
        bdnz L6

That is storing to the stack, with -O2 -fno-split-wide-types, we get the
correct thing:
L6:
        lfsx f0,r2,r9
        addi r2,r2,4
        fadd f13,f13,f0
        bdnz L6


In a way this is caused by how we expand the var-args function call:
(insn 90 89 91 7 (set (subreg:DF (reg:DI 183) 0)
        (reg/v:DF 158 [ s0 ])) -1 (nil)
    (nil))

(insn 91 90 92 7 (set (reg:DI 184)
        (reg:DI 183)) -1 (nil)
    (nil))

(insn 92 91 93 7 (set (reg:DI 185)
        (reg:DI 183)) -1 (nil)
    (nil))

(insn 93 92 94 7 (set (reg:DF 186)
        (subreg:DF (reg:DI 184) 0)) -1 (nil)
    (nil))

(insn 94 93 95 7 (set (reg:DF 187)
        (subreg:DF (reg:DI 185) 0)) -1 (nil)
    (nil))

(insn 96 95 97 7 (set (reg:DF 4 r4)
        (reg:DF 186)) -1 (nil)
    (nil))
(insn 97 96 98 7 (set (reg:DF 33 f1)
        (reg:DF 187)) -1 (nil)
    (nil))



And how lower subreg comes a long and splits the sub regs of DI up.


-- 
           Summary: [4.3 Regression] lower subreg causes a performance
                    regression in the inner loop sometimes
           Product: gcc
           Version: 4.3.0
            Status: UNCONFIRMED
          Keywords: missed-optimization, ra
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: pinskia at gcc dot gnu dot org
GCC target triplet: powerpc-apple-darwin


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31455