On 03/13/2017 06:33 PM, Martin Sebor wrote:
The output of a floating point directive whose precision is specified by an asterisk with an argument that's in a range that includes both negative and positive values less than six may include between zero and six fractional digits plus a decimal point. For example,printf ("%.*e", p, x); results in the 14 bytes -1.797693e+308 when p == -1 and x == -DBL_MIN because a negative precision is ignored and the directive assumes the default 6, and in just the one byte 0 when p == 0 and x == 0.0. Current trunk doesn't handle this case correctly when the floating argument isn't known and uses the upper bound of the specified precision as the maximum number of fractional digits. As a result, it sets the range on the return value in this case as [5, 7] (plus 5 for the longest multibyte decimal decimal point) when the correct range is [5, 14] as explained above (plus 5 again). The attached patch corrects the handling of such precision ranges to avoid this unlikely bug. Martin gcc-79800.diff PR tree-optimization/79800 - wrong snprintf result range with precision in a narrow negative-positive range gcc/ChangeLog: PR tree-optimization/79800 * gimple-ssa-sprintf.c (format_floating: Add argument. Handle precision in negative-positive range. (format_floating): Call non-const overload with adjusted precision. gcc/testsuite/ChangeLog: PR tree-optimization/79800 * gcc.dg/tree-ssa/builtin-sprintf-warn-15.c: Add test cases. * gcc.dg/tree-ssa/pr79800.c: New test.
Thanks. Installed. jeff
