[Bug tree-optimization/99728] code pessimization when using wrapper classes around SIMD types

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99728

--- Comment #19 from Richard Biener <rguenth at gcc dot gnu.org> ---
Created attachment 51150
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51150&action=edit
patch

This is what I bootstrapped & tested on x86_64-unknown-linux-gnu.  It causes

FAIL: g++.dg/warn/Wparentheses-22.C  -std=gnu++14 correct warning (test for
warnings, line 104)
FAIL: g++.dg/warn/Wparentheses-22.C  -std=gnu++14 correct warning (test for
warnings, line 33)
...
FAIL: g++.dg/warn/Wparentheses-23.C  -std=gnu++14  (test for warnings, line
117)
FAIL: g++.dg/warn/Wparentheses-23.C  -std=gnu++14  (test for warnings, line
118)
...

interfering with the intent to warn about code like

  if (a = b) // { dg-warning "assignment" "correct warning" }
    foo (0);

where a and b are aggregates convertible to bool.  We end up generating
something that acts as a valid "fix" for the diagnostic:

  if ((a = b), a)

similarly we'd beforehand not diagnose the case of an empty aggregate
assignment not wrapped in parentheses.

Adjusting the warning code to also look for the COMPOUND_EXPR case
wouldn't be valid if we'd not somehow mark those "fake" COMPOUND_EXPRs
in a special way.

Not sure what to do here.  Maybe simply ignore the lost diagnostic and
XFAIL the testcase or adjust it to have a second member in the aggregate
used.