Atsushi Nemoto <firstname.lastname@example.org> writes:
> Is this a get_user's problem or gcc's?
The latter. gcc is putting the empty asm:
__asm__ ("":"=r" (__gu_val));
into the delay slot of the call.
Part of the problem is that gcc estimates the length of an asm to be the
number of instruction separators + 1. This means that it estimates the
asm above to be one instruction long, which is perhaps a little silly
for an empty string.
But the real problem is that gcc should never trust this estimate anyway,
since each "instruction" could obviously be a multi-instruction macro.
gcc should certainly never put asms into delay slots.
FWIW, I don't think the bug is specific to 3.3 or 3.4. It could
probably trigger for other gcc versions too. It is highly dependent
on scheduling though.
The attached 3.4.x patch fixes the problem there, but if you want to work
around it for old versions, just avoid using empty asms if you can,
or make them volatile if you can't.
Of course, the problem isn't confined to empty asms. If you have an asm
with a single, multi-instruction macro, gcc might try putting that in a
delay slot too. You should at least get an assembler warning in that case.
RCS file: /cvs/gcc/gcc/gcc/config/mips/mips.md,v
retrieving revision 126.96.36.199
diff -u -p -F^\([(a-zA-Z0-9_]\|#define\) -r188.8.131.52 mips.md
--- config/mips/mips.md 25 Jun 2004 07:35:30 -0000 184.108.40.206
+++ config/mips/mips.md 1 Sep 2004 08:26:02 -0000
@@ -251,7 +251,7 @@ (define_attr "single_insn" "no,yes"
;; Can the instruction be put into a delay slot?
(define_attr "can_delay" "no,yes"
- (if_then_else (and (eq_attr "type" "!branch,call,jump")
+ (if_then_else (and (eq_attr "type" "!branch,call,jump,multi")
(and (eq_attr "hazard" "none")
(eq_attr "single_insn" "yes")))