[committed] Fix multiple recent sh3/sh3eb regressions
Commit Message
So my tester started showing even more regressions on the sh3/sh4 runs
recently (beyond the one recently reported in BZ triggered by some DCE
related changes). Bisection kept showing inconsistent results. I was
starting to think memory management error, but valgrind didn't flag
anything.
After a bit of head-banging I was able to track it down to predicate
tests called from the SH specific combiner passes. And once I started
getting inside the actual code for the predicate function it became
pretty obvious. The predicate routines are supposed to return a bool,
fine and they dutifully set the low bit in %eax properly.
The *caller* was looking at the full register. Uh-oh. Naturally we
became dependent on what happened to be in the upper 31 bits of a
register. That's why the bug would come and go so willy-nilly. This was
ultimately chased down to an incorrect prototype in sh_treg_combine.cc
for predicates functions defined via define_predicate.
Removing the bogus prototypes and instead including the generated
tm-preds.h fixes this problem. I also checked the other ports for
similar problems (specifically looking for a extern int.*_operand, then
for each of the hits looking to see if the predicate was defined via
define_predicate). No other ports had similar braindamage.
This fixes the most recent regressions in my tester for sh3/sh3eb and I
strongly suspect sh4. It does not fix 107704, but I think Richi and I
both agree that's a visitation order issue and we were just getting
lucky before.
Committing to the trunk.
Jeff
commit e214cab68cb34e77622b91113f7698cf137bbdd6
Author: Jeff Law <jlaw@ventanamicro.com>
Date: Wed Nov 16 18:47:59 2022 -0700
Fix multiple recent sh3/sh3eb regressions
So my tester started showing even more regressions on the sh3/sh4 runs
recently (beyond the one recently reported in BZ triggered by some DCE
related changes). Bisection kept showing inconsistent results. I was
starting to think memory management error, but valgrind didn't flag anything.
After a bit of head-banging I was able to track it down to predicate
tests called from the SH specific combiner passes. And once I started
getting inside the actual code for the predicate function it became
pretty obvious. The predicate routines are supposed to return a bool,
fine and they dutifully set the low bit in %eax properly.
The *caller* was looking at the full register. Uh-oh. Naturally we
became dependent on what happened to be in the upper 31 bits of a register.
That's why the bug would come and go so willy-nilly. This was ultimately
chased down to an incorrect prototype in sh_treg_combine.cc for predicate
functions defined via define_predicate.
Removing the bogus prototypes and instead including the generated
tm-preds.h fixes this problem. I also checked the other ports for
similar problems (specifically looking for a extern int.*_operand, then
for each of the hits looking to see if the predicate was defined via
define_predicate). No other ports had similar braindamage.
This fixes the most recent regressions in my tester for sh3/sh3eb
and I strongly suspect sh4. It does not fix 107704, but I think
Richi and I both agree that's a visitation order issue and we were
just getting lucky before.
gcc/
* config/sh/sh_treg_combine.cc: Include tm-preds.h.
(t_reg_operand): Remove bogus prototype.
(negt_reg_operand): Likewise.
@@ -37,6 +37,7 @@ along with GCC; see the file COPYING3. If not see
#include "cfgrtl.h"
#include "tree-pass.h"
#include "expr.h"
+#include "tm-preds.h"
/*
This pass tries to optimize for example this:
@@ -426,10 +427,6 @@ is_conditional_insn (rtx_insn* i)
return GET_CODE (p) == SET && GET_CODE (XEXP (p, 1)) == IF_THEN_ELSE;
}
-// FIXME: Remove dependency on SH predicate function somehow.
-extern int t_reg_operand (rtx, machine_mode);
-extern int negt_reg_operand (rtx, machine_mode);
-
// - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
// RTL pass class