On 03/24/2017 01:27 PM, Jakub Jelinek wrote:
Hi!
I'm not aware of any target that would support sign extension of
something that can't be folded in the compiler into some type
larger than word/pointer. Zero extension is doable and assemble_variable
is able to deal with it by emitting low/high subregs of it (where the
low one will contain some expression that needs to be computed by assembler
or even have relocation and upper part will be all zeros), but for
sign extension we'd need assembler support that would for some
asm expression shift it arithmetically right.
So, this patch just rejects it in initializers (in C++ handles through
dynamic initialization) in that case. clang also rejects it, ICC silently
miscompiles (performs zero extension).
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2017-03-24 Jakub Jelinek <ja...@redhat.com>
PR middle-end/80163
* varasm.c (initializer_constant_valid_p_1): Disallow sign-extending
conversions to integer types wider than word and pointer.
* gcc.dg/pr80163.c: New test.
FWIW, the old PA 32 bit SOM model had relocations that would allow
supporting this case. We never exposed them as never seemed worth the
effort, given it's a relatively obscure issue on a dead platform, I
don't think it's worth trying to do anything to support it there.
OK for the trunk.
jeff
