On Mon, Jan 9, 2012 at 12:45 PM, Eric Botcazou <ebotca...@adacore.com> wrote: >> Hmm, it seems to be because we do not value-number loads that >> satisfy stmt_could_throw_p (for whatever reason ...). Seems to date >> back to rev. 131610, the fix for PR34648. Looking at that bug it >> seems that we could at least allow stmts that only throw externally >> (but generally CSE should still work, so value-number them when >> we are doing FRE but not PRE). > > OK. The annoying thing is that it's optimized if you have y_1(D)->i2: > > t$F$i2_9 = y_1(D)->i2; > D.1579_3 = y_1(D)->i2; > > and this is equivalent wrt throwing (in particular TREE_THIS_NOTRAP is not set > anywhere).
Yeah, there is no good reason to disallow CSEing throwing loads (or even volatile loads). The change that introduced this is no longer necessary as the PRE set filling was re-organized completely, avoiding the original issue (PRE indeed should not insert throwing stmts - well, at least if they throw internally, but we can't distinguish the case we have two matching loads, one throwing internally and one externally). I have a patch. Richard. > -- > Eric Botcazou
