[Bug rtl-optimization/36182] [4.3 Regression] Fix for PR 36090 causes libstdc++ regressions

[hp at gcc dot gnu dot org]

------- Comment #12 from hp at gcc dot gnu dot org  2008-05-09 19:18 -------
There has been some speculation that the back-ends where the bug
appear are *all* to blame by e.g. having a flawed
GO_IF_LEGITIMATE_ADDRESS.  A gdb session shows that the back-end
validation hook that is used for the (CONST (MINUS sym2 sym1))
expression is (naturally when you think of it)
LEGITIMATE_CONSTANT_P.  In cases similar to the failure for CRIS,
the flawed expression isn't used in an address-operand so
GO_IF_LEGITIMATE_ADDRESS won't hit.

But, consider the documentation, which recommends defining
LEGITIMATE_CONSTANT_P to 1, which indeed most targets do; very
few actually look inside a CONST, i386 being a notable
exception.

I think that kind of CONST should simply not be generated by gcc
for now, and certainly not on the 4.3 branch.

If this CONST *is* to be generated (on trunk), IMHO some new
constant-expression validation needs to take place in
CONSTANT_P, one where the target can refuse expressions that
aren't valid assembly-code (in this case, for CRIS, sym2-sym1
where sym2 and sym1 are in different sections). Alternatively,
LEGITIMATE_CONSTANT_P redefined in the documentation and most
back-ends.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36182