Various optimizations in fold_inf_compare can cause "invalid" floating-point exceptions to be present when they should be absent, or vice versa.
Folding x > +Inf to 0 should not be about HONOR_SNANS - ordered comparisons of both quiet and signaling NaNs should raise invalid. So if -ftrapping-math and HONOR_NANS, this is invalid (given a strict interpretation of -ftrapping-math, anyway).
x <= +Inf is not the same as x == x, because again that loses an exception (equality comparisons don't raise exceptions except for signaling NaNs).
x == +Inf is not the same as x > DBL_MAX, and a similar issue applies with the x != +Inf case - that transformation causes a spurious exception (in particular, I've seen this in __gcc_qadd for powerpc-linux-gnu soft-float, causing glibc testsuite failures, which is how I found this bug), not just a missing exception. (The existing HONOR_NANS code for NE_EXPR is only about the result, not the exceptions raised.)
Date: Tue Jan 9 13:25:38 2018
New Revision: 256380
Fix folding of Inf/NaN comparisons for -ftrapping-math (PR tree-optimization/64811).
The folding of comparisons against Inf (to constants or comparisons
with the maximum finite value) has various cases where it introduces
or loses "invalid" exceptions for comparisons with NaNs.
Folding x > +Inf to 0 should not be about HONOR_SNANS - ordered
comparisons of both quiet and signaling NaNs should raise invalid.
x <= +Inf is not the same as x == x, because again that loses an
exception (equality comparisons don't raise exceptions except for
x == +Inf is not the same as x > DBL_MAX, and a similar issue applies
with the x != +Inf case - that transformation causes a spurious
This patch fixes the conditionals on the folding to avoid such
introducing or losing exceptions.
Bootstrapped with no regressions on x86_64-pc-linux-gnu (where the
cases involving spurious exceptions wouldn't have failed anyway before
GCC 8 because of unordered comparisons wrongly always having formerly
been used by the back end). Also tested for powerpc-linux-gnu
soft-float that this fixes many glibc math/ test failures that arose
in that configuration because this folding affected the IBM long
double support in libgcc (no such failures appeared for hard-float
because of the bug of powerpc hard-float always using unordered
comparisons) - some failures remain, but I believe them to be
* match.pd: When optimizing comparisons with Inf, avoid
introducing or losing exceptions from comparisons with NaN.
* gcc.dg/torture/inf-compare-1.c, gcc.dg/torture/inf-compare-2.c,
* gcc.c-torture/execute/ieee/fp-cmp-7.x: New file.
Fixed for GCC 8.