| Message ID | 1515790698-6489-3-git-send-email-lists@coryfields.com |
|---|---|
| State | New |
| Headers | show |
| Series | Fix unstable sorts affecting codegen | expand |
On 01/12/2018 01:58 PM, lists@coryfields.com wrote: > From: Cory Fields <cory-nospam-@coryfields.com> > > 2018-01-12 Cory Fields <cory-nospam-@coryfields.com> > * tree-ira.c (allocno_hard_regs_compare): stabilize sort > --- > gcc/ChangeLog | 3 +++ > gcc/ira-color.c | 3 +-- > 2 files changed, 4 insertions(+), 2 deletions(-) > > diff --git a/gcc/ChangeLog b/gcc/ChangeLog > index ab96bd6..546e84c 100644 > --- a/gcc/ChangeLog > +++ b/gcc/ChangeLog > @@ -1,4 +1,7 @@ > 2018-01-12 Cory Fields <cory-nospam-@coryfields.com> > + * tree-ira.c (allocno_hard_regs_compare): stabilize sort Note I'm not sure this is sufficient to stabilize the sort. Given two allocnos with the same cost and the same potential hard reg set ought to hash to the same value. Similarly if two sets with the same cost had a hash collision. But it's still more stable than doing nothing. jeff
On Sat, Jan 13, 2018 at 2:48 PM, Jeff Law <law@redhat.com> wrote: > On 01/12/2018 01:58 PM, lists@coryfields.com wrote: >> From: Cory Fields <cory-nospam-@coryfields.com> >> >> 2018-01-12 Cory Fields <cory-nospam-@coryfields.com> >> * tree-ira.c (allocno_hard_regs_compare): stabilize sort >> --- >> gcc/ChangeLog | 3 +++ >> gcc/ira-color.c | 3 +-- >> 2 files changed, 4 insertions(+), 2 deletions(-) >> >> diff --git a/gcc/ChangeLog b/gcc/ChangeLog >> index ab96bd6..546e84c 100644 >> --- a/gcc/ChangeLog >> +++ b/gcc/ChangeLog >> @@ -1,4 +1,7 @@ >> 2018-01-12 Cory Fields <cory-nospam-@coryfields.com> >> + * tree-ira.c (allocno_hard_regs_compare): stabilize sort > Note I'm not sure this is sufficient to stabilize the sort. Given two > allocnos with the same cost and the same potential hard reg set ought to > hash to the same value. > If the set and cost are the same, then the two allocno_hard_regs are bit-for-bit identical, so returning 0 in that case is fine. Or have I misunderstood you? > Similarly if two sets with the same cost had a hash collision. A hash collision would cause breakage here, yes. I'd be happy to change it to a full memcmp-like compare, it just wasn't obvious to me how to do so. > > But it's still more stable than doing nothing. Agreed, it's enough to fix my bootstrap issues, at least. Side-note: I just noticed that I wrote the wrong filename in the changelog/commit msg. I assume those entries will likely change anyway if committed, but please let me know if I should re-send. Cory
On 01/12/2018 01:58 PM, lists@coryfields.com wrote: > From: Cory Fields <cory-nospam-@coryfields.com> > > 2018-01-12 Cory Fields <cory-nospam-@coryfields.com> > * tree-ira.c (allocno_hard_regs_compare): stabilize sort Thanks. I fixed the ChangeLog entry and installed hte patch on the trunk. jeff
Thanks! Cory On Jan 15, 2018 1:06 AM, "Jeff Law" <law@redhat.com> wrote: > On 01/12/2018 01:58 PM, lists@coryfields.com wrote: > > From: Cory Fields <cory-nospam-@coryfields.com> > > > > 2018-01-12 Cory Fields <cory-nospam-@coryfields.com> > > * tree-ira.c (allocno_hard_regs_compare): stabilize sort > Thanks. I fixed the ChangeLog entry and installed hte patch on the trunk. > > jeff >
diff --git a/gcc/ChangeLog b/gcc/ChangeLog index ab96bd6..546e84c 100644 --- a/gcc/ChangeLog +++ b/gcc/ChangeLog @@ -1,4 +1,7 @@ 2018-01-12 Cory Fields <cory-nospam-@coryfields.com> + * tree-ira.c (allocno_hard_regs_compare): stabilize sort + +2018-01-12 Cory Fields <cory-nospam-@coryfields.com> * tree-ssa-loop-im.c (sort_bbs_in_loop_postorder_cmp): stabilize sort 2018-01-12 Jakub Jelinek <jakub@redhat.com> diff --git a/gcc/ira-color.c b/gcc/ira-color.c index c8b6ab4..8c02c27 100644 --- a/gcc/ira-color.c +++ b/gcc/ira-color.c @@ -300,8 +300,7 @@ allocno_hard_regs_compare (const void *v1p, const void *v2p) return 1; else if (hv2->cost < hv1->cost) return -1; - else - return 0; + return SORTGT(allocno_hard_regs_hasher::hash(hv2), allocno_hard_regs_hasher::hash(hv1)); }
From: Cory Fields <cory-nospam-@coryfields.com> 2018-01-12 Cory Fields <cory-nospam-@coryfields.com> * tree-ira.c (allocno_hard_regs_compare): stabilize sort --- gcc/ChangeLog | 3 +++ gcc/ira-color.c | 3 +-- 2 files changed, 4 insertions(+), 2 deletions(-)