https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109008
--- Comment #38 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I've also repeated the testing with the above generator generated testcase with
" - " replaced with " / " and " + " with " * ", with #c32 + #c33 patches this
resulted
in
1 15 50
1 17 50
1 18 50
1 23 50
1 29 50
1 32 50
1 50 14
1 50 18
1 50 33
1 50 35
2 11 50
2 13 50
2 50 12
2 50 8
2 7 50
3 50 10
3 9 50
4 50 6
4 50 9
4 6 50
5 10 50
5 8 50
6 50 7
9 5 50
10 4 50
10 50 4
12 50 5
28 3 50
45 50 3
171 50 2
194 2 50
366 1 50
411 50 1
8744 50 50
33627 2 2
45607 1 2
52776 1 1
70436 2 1
85436 -1 -1
so again, seems at least for the 297936 tests with finite ranges it is
conservatively correct and while it still has various 50+ulps cases, in most
cases it is within an ulp or two from the minimal range.
If we find some important issue where it is too conservative, we can deal with
it incrementally.
