[Bug middle-end/64412] [regression] ICE in offload compiler: in extract_insn, at recog.c:2327

[iverbin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64412

--- Comment #18 from iverbin at gcc dot gnu.org ---
(In reply to Uroš Bizjak from comment #17)
> The x86 backend did survive many years just fine, so I think offload should
> be fixed to follow approach that generic middle-end takes. The testcase I
> posted expands without problems; in this case middle-end knows that symbol
> address has to be moved to a register. It looks like offload bypasses
> generic expansion functions (that would magically fix this issue).

It seems that the problem with offload is that -fPIC option is passed to the
offload compiler, but not passed to the host compiler. If I add -fPIC to the
host compiler as well, everything is ok.

Offload RTL for pic host, pic offload:
  (insn 8 6 9 4 (set (reg:DI 88)
          (mem/u/c:DI (const:DI (unspec:DI [
                          (symbol_ref:DI ("G")  <var_decl 0x7f2c78238c60 G>)
                      ] UNSPEC_GOTPCREL)) [0  S8 A8])) addr.c:6 -1
       (nil))

Offload RTL for nonpic host, pic offload:
  (insn 8 6 9 4 (set (reg:CC 17 flags)
          (compare:CC (mem/f/c:DI (plus:DI (reg/f:DI 82 virtual-stack-vars)
                      (const_int -8 [0xfffffffffffffff8])) [0 p+0 S8 A64])
              (symbol_ref:DI ("G")  <var_decl 0x7f5546a88c60 G>))) addr.c:6 -1
       (nil))

I don't know how -fPIC option affects IR before streaming out,
-fdump-tree-optimized are identical for pic/nonpic cases, but
.gnu.offload_lto_.decls sections are different. However debug_tree
(vnode->decl) for "G" in ipa_write_summaries are identical for pic/nonpic
cases.

So, the question is, how to figure out what is different in G's declaration in
IR, and how it can affect further expansion?