This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[PATCH, m68k] Fix floating-point comparisons with zero


Hi,

It appears that m68k floating-point comparisons may have been
subtly broken since 2007 when the following patch was applied:

  http://gcc.gnu.org/ml/gcc-patches/2007-09/msg01570.html

The bug shows with our internal 4.6-based branch, but appears to be
latent on mainline (generated code is perturbed by r171236, an
apparently-unrelated change -- if that commit alone is reverted, the bug
reappears on current trunk). The discussion below was written wrt. the
previously-mentioned 4.6 branch, so details may be slightly different on
mainline, but the outcome is the same.

The trouble is the code introduced in m68k.c:notice_update_cc, intended
to reverse the sense of comparisons for one particular
*cmp<mode>_68881/*cmp<mode>_cf alternative, inadvertently also inverts
the sense for tst<mode>_68881/tst<mode>_cf (FP comparisons against
zero).

For the test case gcc.c-torture/execute/ieee/20010114-2.c, the
badness only triggers when -freorder-blocks is in effect (-O2 and
above). Two comparisons (of the same variable) against zero are munged
into one in final.c after basic-block reordering makes them
consecutive, and we get:

	ftst.s %d0	| x
	fjgt .L13	|
(final.c omits a redundant "ftst.s %d0" here)
	fjngt .L2	|

when we ought to have got:

	ftst.s %d0	| x
	fjgt .L13	|
	fjnlt .L2	|

The bug only triggers due to the omission of the test, else it would
happen (far) more often: if the ftst is actually emitted, the flags get
reset before any harm can be done:

(define_insn "tst<mode>_68881"
  [(set (cc0)
	(compare (match_operand:FP 0 "general_operand" "f<FP:dreg>m")
		 (match_operand:FP 1 "const0_operand" "H")))]
  "TARGET_68881"
{
  cc_status.flags = CC_IN_68881;  /* unsets CC_REVERSED & other flags */
  ...

but, when final.c decides the second ftst is redundant, nothing unsets
CC_REVERSED, so we actually (incorrectly) treat the comparison as
having been done with reversed operands.

We can fix this by never setting CC_REVERSED for ftst instructions in
the first place. The problem is that the test (in notice_update_cc),

  if (cc_status.value2 && GET_CODE (cc_status.value2) == COMPARE
      && GET_MODE_CLASS (GET_MODE (XEXP (cc_status.value2, 0))) == MODE_FLOAT)
    {
      cc_status.flags = CC_IN_68881;
      if (!FP_REG_P (XEXP (cc_status.value2, 0)))
	cc_status.flags |= CC_REVERSED;
    }

is not strict enough, and triggers for ftst instructions (even elided
ones), as well as for the intended third alternative of fcmp.

So, the fix is to tighten up the inner condition to not trigger for
ftst. The attached patch does that, and we get the following test result
delta (as well as some noise in libstdc++ results, but I think those
are environment-related):

FAIL -> PASS: gcc.c-torture/execute/ieee/20010114-2.c execution,  -O2 
FAIL -> PASS: gcc.c-torture/execute/ieee/20010114-2.c execution,  -O3 -fomit-frame-pointer 
FAIL -> PASS: gcc.c-torture/execute/ieee/20010114-2.c execution,  -O3 -g 
FAIL -> PASS: gcc.c-torture/execute/ieee/20030331-1.c execution,  -O2 
FAIL -> PASS: gcc.c-torture/execute/ieee/20030331-1.c execution,  -O3 -fomit-frame-pointer 
FAIL -> PASS: gcc.c-torture/execute/ieee/20030331-1.c execution,  -O3 -g 

These results are against mainline *with r171236 reverted*, cross to
ColdFire Linux -- otherwise results are the same before/after. I believe
the patch is still needed regardless though.

OK to apply?

Thanks,

Julian

ChangeLog

    gcc/
    * config/m68k/m68k.c (notice_update_cc): Tighten condition for
    setting CC_REVERSED for FP comparisons.
Index: gcc/config/m68k/m68k.c
===================================================================
--- gcc/config/m68k/m68k.c	(revision 180197)
+++ gcc/config/m68k/m68k.c	(working copy)
@@ -4206,7 +4206,8 @@ notice_update_cc (rtx exp, rtx insn)
       && GET_MODE_CLASS (GET_MODE (XEXP (cc_status.value2, 0))) == MODE_FLOAT)
     {
       cc_status.flags = CC_IN_68881;
-      if (!FP_REG_P (XEXP (cc_status.value2, 0)))
+      if (!FP_REG_P (XEXP (cc_status.value2, 0))
+	  && FP_REG_P (XEXP (cc_status.value2, 1)))
 	cc_status.flags |= CC_REVERSED;
     }
 }

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]