[Bug middle-end/78201] [7 Regression] ICE in tree_to_shwi, at tree.h:4037 (seen both on ARM32 an AArch64)

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78201

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 40015
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=40015&action=edit
gcc7-pr78201.patch

Untested fix.

The primary bug here is in the FEs/gimplifier, they shouldn't create a
TREE_STATIC artificial var with variable length size.  I believe we have PRs
for that though, and I think it is desirable to do the right thing even in this
target hook though.  There is no need to test decl != NULL again, this is all
in a block guarded with if (decl && DECL_P (decl)).  Furthermore, I think we
shouldn't put the odd-sized stuff into section anchors.  And, while
max_anchor_offset is HOST_WIDE_INT, the other uses cast those to unsigned
HOST_WIDE_INT anyway.  Plus using DECL_SIZE and comparing against
max_anchor_offset * BITS_PER_UNIT is too ugly, we already have DECL_SIZE_UNIT
in bytes.