http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45233
--- Comment #8 from Iain Sandoe <iains at gcc dot gnu.org> 2011-11-09 10:03:31
UTC ---
well, I was trying to be too complicated - we should just avoid trying to do
the substitution unless we can see the var in the TU. When this is done:
Index: gcc/config/rs6000/rs6000.c
===================================================================
--- gcc/config/rs6000/rs6000.c (revision 181150)
+++ gcc/config/rs6000/rs6000.c (working copy)
@@ -6169,6 +6169,7 @@ rs6000_legitimize_reload_address (rtx x, enum mach
#if TARGET_MACHO
&& DEFAULT_ABI == ABI_DARWIN
&& (flag_pic || MACHO_DYNAMIC_NO_PIC_P)
+ && machopic_symbol_defined_p (x)
#else
&& DEFAULT_ABI == ABI_V4
&& !flag_pic
.. the test-case generates beautiful (i.e. what one would have generated by
hand) asm.
reg-strapped on trunk (although there's a lot of noise on trunk right now
because of recent changes)
- doing 4.6 branch now.
will post to patches if the 4.6 regstrap is also OK.
====
As an aside (non-Darwin comment - since powerpc-eabisim does exactly the same).
It's not obvious why the perfectly sensible RTL pre-combine is thrown away and
the process started over (I'm guessing it's because the asm insn has a zero
computed cost). Effectively (to my untutored eye), we end up rebuilding what
was there before combine from the reload ...
