Re: Fix tree-optimization/20920
On Sat, 2005-04-09 at 23:27 -0400, Diego Novillo wrote: > Sigh. Another SSA_NAME flowing through EH edges getting in the > way. > > I'm starting to think that it would be much better to make these > SSA_NAME_OCCURS_IN_ABNORMAL_PHI into virtuals. We would avoid > all the monkeying around that optimizers have to do. It would > involve a bit of trickery because we only detect them when they > first go into SSA form, but it's doable. Comments from EH folks? I think that would have a rather bad effect on C++ and Java code. One of the nice things about the current approach is that it is very fine grained -- ie, the flag bit appears on SSA_NAMEs and thus has a relatively limited affect on optimization. Your approach would in effect penalize all other instances of a _DECL just because one instance happened to flow through an EH edge. Jeff
Re: gcc-2.95.3, but not gcc-2.95.2, requires gas on hppa2.0-hp-hpux10.20
In message <[EMAIL PROTECTED] com>you write: > It turns out it isn't an assembler problem, but a compiler problem. I was > suprised to notice that gcc emits different code depending on whether it > uses gas or not; the problem only occurs if not using gas. That's because the HP assembler can't handle certain constructs we'd like to use. For example it can't handle a label as a field in an "ldo" instrction. > The problem is that we have an apparently unused exception handler which is > optimized out during flow, but a reference to its label survives in a > constant pool. It's a bit odd that the label got deleted, since the load > from the constant pool has a REG_LABEL note and you'd think flow would coun > that as a use. Definitely odd. Seems to me that the use in the constant pool should have counted for the purposes of LABEL_NUSES. jeff
