https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97767
--- Comment #2 from CVS Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Aldy Hernandez <al...@gcc.gnu.org>: https://gcc.gnu.org/g:3d3470e239e8040f642a8852721b4962b4ed36f2 commit r11-4864-g3d3470e239e8040f642a8852721b4962b4ed36f2 Author: Aldy Hernandez <al...@redhat.com> Date: Mon Nov 9 20:35:25 2020 +0100 Normalize VARYING for -fstrict-enums. The problem here is that the representation for VARYING in -fstrict-enums is different between value_range and irange. The helper function irange::normalize_min_max() will normalize to VARYING only if setting the range to the entire domain of the underlying type. That is, [0, 0xff..ff], not the domain as defined by -fstrict-enums. This causes problems because the multi-range version of varying_p() will return true if the range is the domain as defined by -fstrict-enums. Thus, normalize_min_max and varying_p have different concepts of varying for multi-ranges. (BTW, legacy ranges are different because they never look at the extremes of a range to determine varying-ness. They only look at the kind field.) One approach is to change all the code to limit ranges to the domain in the -fstrict-enums world, but this won't work because there are various instances of gimple where the values assigned or compared are beyond the limits of TYPE_{MIN,MAX}_VALUE. One example is the addition of 0xffffffff to represent subtraction. This patch fixes multi-range varying_p() and set_varying() to agree with the normalization code, using the extremes of the underlying type, to represent varying. gcc/ChangeLog: PR tree-optimization/97767 * value-range.cc (dump_bound_with_infinite_markers): Use wi::min_value and wi::max_value. (range_tests_strict_enum): New. (range_tests): Call range_tests_strict_enum. * value-range.h (irange::varying_p): Use wi::min_value and wi::max_value. (irange::set_varying): Same. (irange::normalize_min_max): Remove comment. gcc/testsuite/ChangeLog: * g++.dg/opt/pr97767.C: New test.
