Hi, Following our discussions about changing the frame direction on aarch64, I have written a small test case using a large frame which shows similar code generation with and without my patch. There are small differences in register allocation which lead to larger prologue/epilogue with my patch, but I think this could (should) be improved independently.
I have attached the sample C source and the code generated with (frane64.s.O2) and without (frame64.s.O2.ref) my patch when compiling at -O2. What do you think? Thanks, Christophe. On 28 June 2013 13:41, Christophe Lyon <christophe.l...@linaro.org> wrote: > Hi, > Following the discussion on > http://gcc.gnu.org/ml/gcc/2013-05/msg00208.html > here is a patch to change the frame direction. > > Passed 'make check' with no regression on aarch64-none-linux-gnu and > with a few UNSUPPORTED -> PASS on aarch64-none-elf (pr34225, pr38616, > pr38902, pr40971, pr46440, pr47766, pr49307, all of which are guarded > by dg-require-effective-target fstack_protector). > > OK for trunk? > > Thanks, > > Christophe. > 2013-06-28 Christophe Lyon <christophe.l...@linaro.org> > > * config/aarch64/aarch64.h (FRAME_GROWS_DOWNWARD): Define to 1. > * config/aarch64/aarch64.c (aarch64_initial_elimination_offset): > Add get_frame_size () when eliminating frame pointer.
#define SIZE1 100000
extern int func2(int);
int func1(int arg1, int arg2)
{
int var1;
int array1[SIZE1];
int var2 = 0;
for (var1 = arg1; var1 < arg2; var1++) {
array1[var1] = func2(var1);
}
for (var1 = arg1; var1 < arg2; var1++) {
var2 += array1[var1];
}
var2 += (array1[arg1] + array1[arg2-1]) / 2;
return var2;
}
frame64.s.O2
Description: Binary data
frame64.s.O2.ref
Description: Binary data
