| Message ID | 20190116224901.GB30353@tucnak |
|---|---|
| State | New |
| Headers | show |
| Series | Don't add unnecessary var self-conflicts (PR tree-optimization/86214) | expand |
On 1/16/19 3:49 PM, Jakub Jelinek wrote: > Hi! > > While looking at this PR (I've just started) I've noticed that > add_stack_var_conflict is quite often called with x == y. > We don't really need to record that a variable conflicts with itself, > the only reader of the conflicts bitmaps, stack_var_conflict_p, > starts with > if (x == y) > return false; > conflicts bitmap are set either by this function, or by the > EXECUTE_IF_SET_IN_BITMAP (work, 0, i, bi) > { > struct stack_var *a = &stack_vars[i]; > if (!a->conflicts) > a->conflicts = BITMAP_ALLOC (&stack_var_bitmap_obstack); > bitmap_ior_into (a->conflicts, work); > } > code (where work isn't derived from any conflicts bitmap though, so > doesn't care if we've added those self-conflicts or not). The above > bitmap_ior_into stuff actually always sets self-conflicts (if you think > bitmap_clear_bit is worth it, I can add it afterwards though). > > But I think the following patch is helpful, don't create the conflicts > bitmaps at all if all we'd record is just self-conflict which we'll ignore. > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? > > 2019-01-16 Jakub Jelinek <jakub@redhat.com> > > PR tree-optimization/86214 > * cfgexpand.c (add_stack_var_conflict): Don't add any conflicts > if x == y. Yea. Seems reasonable. jeff
--- gcc/cfgexpand.c.jj 2019-01-16 09:35:09.131247513 +0100 +++ gcc/cfgexpand.c 2019-01-16 20:14:11.445467399 +0100 @@ -470,6 +470,8 @@ add_stack_var_conflict (size_t x, size_t { struct stack_var *a = &stack_vars[x]; struct stack_var *b = &stack_vars[y]; + if (x == y) + return; if (!a->conflicts) a->conflicts = BITMAP_ALLOC (&stack_var_bitmap_obstack); if (!b->conflicts)
Hi! While looking at this PR (I've just started) I've noticed that add_stack_var_conflict is quite often called with x == y. We don't really need to record that a variable conflicts with itself, the only reader of the conflicts bitmaps, stack_var_conflict_p, starts with if (x == y) return false; conflicts bitmap are set either by this function, or by the EXECUTE_IF_SET_IN_BITMAP (work, 0, i, bi) { struct stack_var *a = &stack_vars[i]; if (!a->conflicts) a->conflicts = BITMAP_ALLOC (&stack_var_bitmap_obstack); bitmap_ior_into (a->conflicts, work); } code (where work isn't derived from any conflicts bitmap though, so doesn't care if we've added those self-conflicts or not). The above bitmap_ior_into stuff actually always sets self-conflicts (if you think bitmap_clear_bit is worth it, I can add it afterwards though). But I think the following patch is helpful, don't create the conflicts bitmaps at all if all we'd record is just self-conflict which we'll ignore. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2019-01-16 Jakub Jelinek <jakub@redhat.com> PR tree-optimization/86214 * cfgexpand.c (add_stack_var_conflict): Don't add any conflicts if x == y. Jakub