As the testcases shows, the -Wconversion warning behaves quite differently
when -fsanitize=undefined vs. when not sanitizing, but in the end it is
not something specific to sanitizing, if a user uses
  return static_cast<uc>(static_cast<uc>((d++, a) << 1U) | b) | c;
instead of
  return static_cast<uc>(static_cast<uc>(a << 1U) | b) | c;
and thus there is some COMPOUND_EXPR involved, cp_build_binary_op behaves
significantly different, e.g. shorten_binary_op will have different result
(uc for the case without COMPOUND_EXPR, int with it), but it isn't limited
to that.

> How about improving get_narrower to handle COMPOUND_EXPR?  I'd think that
> would do the trick without affecting evaluation order.

Not completely, had to change also warnings_for_convert_and_check, but with
that it works.  The float-cast-overflow* changes are needed because now with
-O1+ we emit lots of -Woverflow warnings on the testcase, but we do emit
those warnings on the testcase even when compiling just with -O1 and without
-fsanitize=float-cast-overflow, so it seems to me a change in the right
direction, having -fsanitize= or explicit comma expressions smaller effect
on the warnings that are emitted.

2020-03-23  Jakub Jelinek  <jakub@redhat.com>

	PR c++/91993
	* tree.c (get_narrower): Handle COMPOUND_EXPR by recursing on
	ultimate rhs and if returned something different, reconstructing
	the COMPOUND_EXPRs.

	* c-warn.c (warnings_for_convert_and_check): For expr and/or
	result being COMPOUND_EXPRs, skip to ultimate rhs.

	* g++.dg/warn/Wconversion-pr91993.C: New test.
	* g++.dg/ubsan/pr91993.C: New test.
	* c-c++-common/ubsan/float-cast-overflow-1.c: Add -Wno-overflow
	to dg-options.
	* c-c++-common/ubsan/float-cast-overflow-2.c: Likewise.
	* c-c++-common/ubsan/float-cast-overflow-4.c: Likewise.