The following patch avoids simplifying x-0.0 to x when -fsignaling-nans is
specified,
which resolves PR rtl-optimization 61494. Indeed, running the test program
attached
to that PR now reports no failures. Alas reducing that validation program
to a portable
test for the GCC testsuite is a challenge that I plan to leave to a
volunteer. The fix
itself is a trivial one-liner, so it's curious that the compiler has been
broken for so long.
Hopefully folks agree that correctness is more important than testability
(but both
are desirable).
The following patch has been tested on x86_64-pc-linux-gnu with a "make
bootstrap"
and "make -k check" with no new regressions.
Ok for mainline?
2020-08-02 Roger Sayle <ro...@nextmovesoftware.com>
gcc/ChangeLog
PR rtl-optimization/61494
* simplify-rtx.c (simplify_binary_operation_1) [MINUS]: Don't
simplify x - 0.0 with -fsignaling-nans.
Thanks in advance,
Roger
--
Roger Sayle
NextMove Software
Cambridge, UK
diff --git a/gcc/simplify-rtx.c b/gcc/simplify-rtx.c
index d221168..a9d8e8c 100644
--- a/gcc/simplify-rtx.c
+++ b/gcc/simplify-rtx.c
@@ -2678,11 +2678,12 @@ simplify_binary_operation_1 (enum rtx_code code,
machine_mode mode,
&& !contains_symbolic_reference_p (op1))
return simplify_gen_unary (NOT, mode, op1, mode);
- /* Subtracting 0 has no effect unless the mode has signed zeros
- and supports rounding towards -infinity. In such a case,
- 0 - 0 is -0. */
+ /* Subtracting 0 has no effect unless the mode has signalling NaNs,
+ or has signed zeros and supports rounding towards -infinity.
+ In such a case, 0 - 0 is -0. */
if (!(HONOR_SIGNED_ZEROS (mode)
&& HONOR_SIGN_DEPENDENT_ROUNDING (mode))
+ && !HONOR_SNANS (mode)
&& trueop1 == CONST0_RTX (mode))
return op0;
