[PATCH] avoid using upper bound of width and precision in -Wformat-overlow (PR 79692)

[Jeff Law]

On 02/25/2017 03:46 PM, Martin Sebor wrote:
> In an arithmetic directive with the width or precision specified
> by an argument to the asterisk (e.g., "%*x") and where the argument
> range is unknown, for the purposes of the return value optimization
> the pass must assume it's potentially as large as INT_MAX.  But for
> the purposes of issuing a warning, that assumption leads to false
> positives because the value of the argument can and in reality
> usually is much smaller.
>
> The attached patch adjusts the checker to use the lower bound in
> this case to avoid these false positives.  It does that for both
> integer and floating directives (for the latter it uses the lesser
> of 3 and the lower bound in this case).
>
> In addition, the patch corrects the handling of the pound flag ('#')
> in "%#.*g" directives which makes it keep as many trailing zeros after
> the radix character as specified by the precision.  (Without the '#',
> "%.*g" trims trailing zeros.)
>
> Martin
>
>
>
> gcc-79692.diff
>
>
> PR middle-end/79692 - [7 Regression] -Wformat-overflow false positive
>
> gcc/ChangeLog:
>
> 	PR c/79692
> 	* gimple-ssa-sprintf.c
> 	(directive::known_width_and_precision): New function.
> 	(format_integer): Use it.
> 	(get_mpfr_format_length): Consider the full range of precision
> 	when computing %g output with the # flag.  Set the likely byte
> 	count to 3 rather than 1 when precision is indeterminate.
> 	(format_floating): Correct the lower bound of precision.
>
> gcc/testsuite/ChangeLog:
>
> 	PR c/79692
> 	* gcc.dg/tree-ssa/builtin-sprintf-2.c: Add test cases.
> 	* gcc.dg/tree-ssa/builtin-sprintf-warn-10.c: Correct %#g.
> 	* gcc.dg/tree-ssa/builtin-sprintf-warn-15.c: New test.

So in some cases you use

+  /* For an indeterminate precision the lower bound must be assumed
+     to be zero.  */
+  if (prec[0] < 0 && prec[1] >= 0)

Note prec[1] >= 0

In other cases you have:

+	/* The minimum output with unknown precision is a single byte
+	   (e.g., "0") but the more likely output is 3 bytes ("0.0").  */
+	if (dir.prec[0] < 0 && dir.prec[1] > 0)

Note  dir.prec[1] > 0

Shouldn't one of those be changed to be consistent with the other?

Similarly in known_width_and_precision.  Please review the patch to 
ensure that we're as consistent as possible for these tests.


OK after doing that review and making any necessary trivial adjustments.

Jeff