The recently added test gcc.dg/torture/builtin-sprintf.c
to verify that the sprintf result computed by GCC matches
libc's for Infinity and NaN has been failing on AIX which
formats NaN as either QNaN or SNaN, contrary to C/POSIX
requirements. The attached tweak adjusts the result
computed by GCC to include the AIX format. If there are
no objections I'll commit the tweak later this week and
backport it to GCC 8 the next.
Martin
PR tree-optimization/86571 - AIX NaNQ and NaNS output format conflicts with __builtin_sprintf
gcc/ChangeLog:
PR tree-optimization/86571
* gimple-ssa-sprintf.c (format_floating): Extend upper bound of
NaN output to 4.
Index: gcc/gimple-ssa-sprintf.c
===================================================================
--- gcc/gimple-ssa-sprintf.c (revision 263268)
+++ gcc/gimple-ssa-sprintf.c (working copy)
@@ -2014,8 +2014,15 @@ format_floating (const directive &dir, tree arg, v
res.range.likely = res.range.min;
res.range.max = res.range.min;
- /* The inlikely maximum is "[-/+]infinity" or "[-/+]nan". */
- res.range.unlikely = sign + (real_isinf (rvp) ? 8 : 3);
+ /* The unlikely maximum is "[-/+]infinity" or "[-/+][qs]nan".
+ For NaN, the C/POSIX standards specify two formats:
+ "[-/+]nan"
+ and
+ "[-/+]nan(n-char-sequence)"
+ No known printf implementation outputs the latter format but AIX
+ outputs QNaN and SNaN for quiet and signalling NaN, respectively,
+ so the unlikely maximum reflects that. */
+ res.range.unlikely = sign + (real_isinf (rvp) ? 8 : 4);
/* The range for infinity and NaN is known unless either width
or precision is unknown. Width has the same effect regardless
