On Wed, Jun 21, 2017 at 04:40:01PM +0200, Jakub Jelinek wrote:
> So, I wrote following patch to do the subtraction in unsigned
> type. It passes bootstrap, but on both x86_64-linux and i686-linux
> regresses:
> +FAIL: gcc.dg/torture/pr66178.c -O* (test for excess errors)
> +FAIL: gcc.dg/tree-ssa/cmpexactdiv-2.c scan-tree-dump-not optimized
> "minus_expr"
> +FAIL: g++.dg/tree-ssa/pr21082.C -std=gnu++* (test for excess errors)
Another option is to do what the patch does only when sanitizing and accept
in that case less efficient code and rejection of weird corner case
testcases like the first one. We risk miscompilation of the pointer
differences, but I haven't managed to come up with a testcase where it would
show (I guess more likely is when we propagate constants into the pointers).
Jakub
