From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jason Merrill To: Jan Hubicka Cc: Richard Kenner , gcc-patches@gcc.gnu.org Subject: Re: spec2000 regression Date: Wed, 15 Aug 2001 06:36:00 -0000 Message-id: References: <10108151302.AA17923@vlsi1.ultra.nyu.edu> <20010815150251.Y19872@atrey.karlin.mff.cuni.cz> X-SW-Source: 2001-08/msg00909.html >>>>> "Jan" == Jan Hubicka writes: >> I'd prefer to fix things so they don't set DECL_RTL first. MEM_ALIAS_SET >> is not a test to see if the alias set has been set. The alias set might >> have been zero in which case we *do* want to use it. > That what I was trying to do first, but there are number of places this > is happening and fixing each of it still keeps open door for possible > new bugs - it is natural to save the RTL to proper place first. I disagree. Even without this bug, it's bad style. > If I understand the purpose of original patch properly, it was hooting for > case variable gets unique class, but this class is recomputed several times > so it get several unique classes. > For class 0 this does not matter. On the contrary, it is needed for the case where a variable gets a class different from that of its type, so recomputing it would give the wrong answer. How does this work for you? 2001-08-15 Jason Merrill * varasm.c (make_decl_rtl): Don't SET_DECL_RTL until we're done modifying the rtl.