incorrect simplifications for fixed-point rtl expressions

incorrect simplifications for fixed-point rtl expressions

[Stefan M. Freudenberger]

Hello,

I am working with the current distribution of gcc 4.3.3.

The following two transformations are unsafe for (signed) fixed-point
numbers.  In both cases, a NEG operation is moved to the input of the MULT
(say, A).  However, if A is -1.0, then NEG A overflows, and the MULT yields
the wrong result.  Note that if the result of (-1.0 * B) is between -1.0 and
+1.0 than the NEG doesn't overflow (1st case); otherwise, it would be a user
error (undefined overflow behavior).

I didn't find this bug mentioned in the archives.  If it has been reported
before, please accept my apologies for sending a duplicate report.

Best regards,
Stefan Freudenberger

PS. I don't subscribe to gcc-bugs: please include me in replies.  Thank you.



$ LC_ALL=C TZ=UTC0 diff -Naur gcc-4.3.3-{ORIG,smf}/gcc/simplify-rtx.c
--- gcc-4.3.3-ORIG/gcc/simplify-rtx.c   2008-05-09 17:13:30.000000000 +0000
+++ gcc-4.3.3-smf/gcc/simplify-rtx.c    2009-03-27 09:52:38.712334115 +0000
@@ -542,7 +542,8 @@
       /* (neg (mult A B)) becomes (mult (neg A) B).
         This works even for floating-point values.  */
       if (GET_CODE (op) == MULT
-         && !HONOR_SIGN_DEPENDENT_ROUNDING (mode))
+         && !HONOR_SIGN_DEPENDENT_ROUNDING (mode)
+         && !SIGNED_FIXED_POINT_MODE_P (mode))
        {
          temp = simplify_gen_unary (NEG, mode, XEXP (op, 0), mode);
          return simplify_gen_binary (MULT, mode, temp, XEXP (op, 1));
@@ -1933,6 +1934,7 @@
       /* Canonicalize (minus (neg A) (mult B C)) to
         (minus (mult (neg B) C) A).  */
       if (!HONOR_SIGN_DEPENDENT_ROUNDING (mode)
+         && !SIGNED_FIXED_POINT_MODE_P (mode)
          && GET_CODE (op1) == MULT
          && GET_CODE (op0) == NEG)
        {
$