Hi!
IEEE754 says that x + (-x) and x - x result in +0 in all rounding modes
but rounding towards negative infinity, in which case the result is -0
for all finite x. x + x and x - (-x) if it is zero retain sign of x.
Now, range_arithmetic implements the normal rounds to even rounding,
and as the addition or subtraction in those cases is exact, we don't do any
further rounding etc. and e.g. on the testcase below distilled from glibc
compute a range [+0, +INF], which is fine for -fno-rounding-math or
if we'd have a guarantee that those statements aren't executed with rounding
towards negative infinity.
I believe it is only +- which has this problematic behavior and I think
it is best to deal with it in frange_arithmetic; if we know -frounding-math
is on, it is x + (-x) or x - x and we are asked to round to negative
infinity (i.e. want low bound rather than high bound), change +0 result to
-0.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk and
after a while for 13.3? I'm afraid rushing this so late into 13.2...
2023-07-24 Jakub Jelinek <ja...@redhat.com>
PR tree-optimization/110755
* range-op-float.cc (frange_arithmetic): Change +0 result to -0
for PLUS_EXPR or MINUS_EXPR if -frounding-math, inf is negative and
it is exact op1 + (-op1) or op1 - op1.
* gcc.dg/pr110755.c: New test.
--- gcc/range-op-float.cc.jj 2023-07-23 19:32:20.832434105 +0200
+++ gcc/range-op-float.cc 2023-07-24 09:41:26.231030258 +0200
@@ -324,6 +324,24 @@ frange_arithmetic (enum tree_code code,
bool inexact = real_arithmetic (&value, code, &op1, &op2);
real_convert (&result, mode, &value);
+ /* When rounding towards negative infinity, x + (-x) and
+ x - x is -0 rather than +0 real_arithmetic computes.
+ So, when we are looking for lower bound (inf is negative),
+ use -0 rather than +0. */
+ if (flag_rounding_math
+ && (code == PLUS_EXPR || code == MINUS_EXPR)
+ && !inexact
+ && real_iszero (&result)
+ && !real_isneg (&result)
+ && real_isneg (&inf))
+ {
+ REAL_VALUE_TYPE op2a = op2;
+ if (code == PLUS_EXPR)
+ op2a.sign ^= 1;
+ if (real_isneg (&op1) == real_isneg (&op2a) && real_equal (&op1, &op2a))
+ result.sign = 1;
+ }
+
// Be extra careful if there may be discrepancies between the
// compile and runtime results.
bool round = false;
--- gcc/testsuite/gcc.dg/pr110755.c.jj 2023-07-21 10:34:05.037251433 +0200
+++ gcc/testsuite/gcc.dg/pr110755.c 2023-07-21 10:35:10.986326816 +0200
@@ -0,0 +1,29 @@
+/* PR tree-optimization/110755 */
+/* { dg-do run } */
+/* { dg-require-effective-target fenv } */
+/* { dg-require-effective-target hard_float } */
+/* { dg-options "-O2 -frounding-math" } */
+
+#include <fenv.h>
+
+__attribute__((noipa)) float
+foo (float x)
+{
+ if (x > 0.0)
+ {
+ x += 0x1p+23;
+ x -= 0x1p+23;
+ x = __builtin_fabsf (x);
+ }
+ return x;
+}
+
+int
+main ()
+{
+#ifdef FE_DOWNWARD
+ fesetround (FE_DOWNWARD);
+ if (__builtin_signbit (foo (0.5)))
+ __builtin_abort ();
+#endif
+}
Jakub
