------- Comment #35 from jakub at gcc dot gnu dot org 2007-12-11 21:28 -------
df_get_entry_block_def_set has this:
/* Once the prologue has been generated, all of these registers
should just show up in the first regular block. */
if (HAVE_prologue && epilogue_completed)
{
/* Defs for the callee saved registers are inserted so that the
pushes have some defining location. */
for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
if ((call_used_regs[i] == 0) && (df_regs_ever_live_p (i)))
bitmap_set_bit (entry_block_defs, i);
}
else
{
...
r = targetm.calls.struct_value_rtx (current_function_decl, true);
if (r && REG_P (r))
bitmap_set_bit (entry_block_defs, REGNO (r));
}
So, apparently struct_value_rtx is already considered, but only before
prologues/epilogues are inserted (or the arch doesn't HAVE_prologue).
On PA call_used_regs[28] != 0, so the comment "Once the prologue ..."
doesn't make sense for it, because the if part won't ever set %r28 in
entry_block_defs, even when it is needed. I've tried:
--- df-scan.c.jj 2007-10-26 13:45:44.000000000 +0200
+++ df-scan.c 2007-12-11 22:24:21.000000000 +0100
@@ -3530,12 +3530,12 @@ df_get_entry_block_def_set (bitmap entry
bitmap_set_bit (entry_block_defs, STATIC_CHAIN_REGNUM);
#endif
#endif
-
- r = targetm.calls.struct_value_rtx (current_function_decl, true);
- if (r && REG_P (r))
- bitmap_set_bit (entry_block_defs, REGNO (r));
}
+ r = targetm.calls.struct_value_rtx (current_function_decl, true);
+ if (r && REG_P (r))
+ bitmap_set_bit (entry_block_defs, REGNO (r));
+
if ((!reload_completed) || frame_pointer_needed)
{
/* Any reference to any pseudo before reload is a potential
patch and it indeed probably cures this testcase. The diff in assembly is:
...
comib,>= 0,%r24,L$0025
stw %r4,-72(%r30)
comib,= 1,%r24,L$0026
- ldil LR'L$C0002,%r28
+ ldil LR'L$C0003,%r19
stw %r7,-52(%r30)
ldw 4(%r26),%r20
ldo -112(%r30),%r25
ldo -120(%r30),%r26
stw %r20,-116(%r30)
ldw 4(%r6),%r19
ldw 0(%r3),%r20
stw %r19,-108(%r30)
stw %r20,-120(%r30)
ldw 0(%r6),%r19
.CALL ARGW0=GR,ARGW1=GR,ARGW2=GR,ARGW3=GR
bl _Z4foo1I1DIiEPFbiiEET_S4_S4_iRKiT0_,%r2
stw %r19,-112(%r30)
...
L$0026:
- ldil LR'L$C0003,%r19
+ ldil LR'L$C0002,%r28
ldw 4(%r26),%r20
ldo RR'L$C0002(%r28),%r9
ldil LR'L$C0001,%r28
...
which looks promissing (though of course without access to PA I can't be sure
it works).
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32636
