[Bug middle-end/103031] New: [12 Regression] Missing static initializer folding with -frounding-math

[jsm28 at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103031

            Bug ID: 103031
           Summary: [12 Regression] Missing static initializer folding
                    with -frounding-math
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jsm28 at gcc dot gnu.org
                CC: rguenth at gcc dot gnu.org
  Target Milestone: ---

One of the recent fixes to avoid inappropriate constant folding with
-frounding-math caused the following test to produce an error when building
with -frounding-math.  Static initializers in C are evaluated at translation
time, in the default rounding mode and with exceptions discarded;
-frounding-math should not affect this.  Indeed, fold-const.c has
START_FOLD_INIT and END_FOLD_INIT that are supposed to be used for static
initializer folding to implement this, but apparently aren't being effective in
this case.

Test case:

double x = -1ULL;

Errors:

t.c:1:12: error: initializer element is not computable at load time
    1 | double x = -1ULL;
      |            ^

Note: if the initializer is cast to double, things work fine; it's only the
case of implicit conversion that has the problem, so maybe something is handled
differently for folding an implicit conversion of a static initializer.

This breaks building glibc tests for some platforms:
https://sourceware.org/pipermail/libc-alpha/2021-October/132475.html