[Bug tree-optimization/128356] [13/14/15/16/17 Regression] wrong code with -funswitch-loops: switch with more than 255 case ranges on one edge

[lzy20000419 at outlook dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128356

--- Comment #2 from liu zhenyu <lzy20000419 at outlook dot com> ---
PR126536 is the same constructor with exactly 255 pairs: invert () needs 256
pairs and overflows the buffer (ICE /heap corruption). The trunk fix (r17-3206,
"Allow range invert to fail") only handles that overflow. Here the first case
is 0 (the type minimum) so the inverse fits and invert () succeeds, and the
wrong range is used silently. A possible fix is to give up (set both ranges to
varying) when the union had to be merged, for example when the number of case
ranges on the edge exceeds what the range can hold.

Disclaimer: This test case was generated by a fuzzer, and the report was
written with LLM assistance.