| Message ID | or39no70tr.fsf@livre.localdomain |
|---|---|
| State | New |
| Headers | show |
On Wed, Feb 16, 2011 at 11:04 AM, Alexandre Oliva <aoliva@redhat.com> wrote: > Before the latest SRA rewrite, we had a pass that attempted to group > multiple bitfields into single variables. It attempted to extend these > variables to cover padding bits, but there was a bug that got it > confused when it found padding bits before the range it intended to > cover. > > This patch changes the tests so that the tentatively-extended range will > always shrink towards the range of interest, instead of getting stuck at > a padding range. > > Regstrapped on x86_64-linux-gnu. Ok to install? Please add a runtime testcase for such wrong-code bugs. Ok with that, please also install the testcase on the 4.5 branch and trunk. Thanks, Richard.
for gcc/ChangeLog from Alexandre Oliva <aoliva@redhat.com> PR tree-optimization/46620 * tree-sra.c (try_instantiate_multiple_fields): Don't get stuck at padding within accessed words. Index: gcc/tree-sra.c =================================================================== --- gcc/tree-sra.c.orig 2011-02-15 22:32:20.659134294 -0200 +++ gcc/tree-sra.c 2011-02-15 22:37:41.632435631 -0200 @@ -1661,13 +1661,13 @@ try_instantiate_multiple_fields (struct || mbit + msize <= fbit) continue; - if (fbit <= mbit) + if (fbit < bit) { unsigned HOST_WIDE_INT diff = fbit + fsize - mbit; mbit += diff; msize -= diff; } - else if (fbit > mbit) + else if (fbit > bit) msize -= (mbit + msize - fbit); else gcc_unreachable ();