[Bug target/45233] FAIL: gcc.c-torture/compile/pr44707.c

[iains at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45233

--- Comment #2 from Iain Sandoe <iains at gcc dot gnu.org> 2011-11-06 12:36:23 UTC ---
Trying to simplify this, the following is enough to trigger the behavior:

extern int w;

void
foo (void)
{
  int e2 = w;
  __asm__ volatile ("/* %0 */" : : "ro" (e2));
}

==

(1) the code DTRT when the constraint is "r" or "o" ... but not when it is
"ro".

(2) essentially, up to combine, we have:


(insn 5 2 6 2 (set:SI (reg:SI 122)
        (plus:SI (reg:SI 31 r31)
            (high:SI (const:SI (unspec:SI [
                            (symbol_ref:SI ("&L_w$non_lazy_ptr") [flags 0x400] 
<var_decl 0x420a3a20 w>)
                        ] UNSPEC_MACHOPIC_OFFSET))))) ../tests/pr44707-x.c:7 96
{addsi3_high}
     (nil))

(insn 6 5 7 2 (set (reg/f:SI 121)
        (mem/u/c:SI (lo_sum:SI (reg:SI 122)
                (const:SI (unspec:SI [
                            (symbol_ref:SI ("&L_w$non_lazy_ptr") [flags 0x400] 
<var_decl 0x420a3a20 w>)
                        ] UNSPEC_MACHOPIC_OFFSET))) [0 S4 A8]))
../tests/pr44707-x.c:7 348 {movsi_low}
     (expr_list:REG_DEAD (reg:SI 122)
        (expr_list:REG_EQUAL (symbol_ref:SI ("w") [flags 0x240]  <var_decl
0x420a3a20 w>)
            (nil))))


(3) when "ro" is given, combine decides to delete insn 5 & 6  ...

(4) ira decides to re-insert the load of a pointer to w .. but it doesn't
(re-)insert the machopic indirect stuff...

(note 5 2 6 2 NOTE_INSN_DELETED)

(note 6 5 12 2 NOTE_INSN_DELETED)

(insn 12 6 13 2 (set (reg:SI 2 r2)
        (plus:SI (reg:SI 31 r31)
            (high:SI (const:SI (unspec:SI [
                            (symbol_ref:SI ("w") [flags 0x240]  <var_decl
0x420a3a20 w>)
                        ] UNSPEC_MACHOPIC_OFFSET))))) ../tests/pr44707-x.c:7 96
{addsi3_high}
     (nil))

(insn 13 12 7 2 (set (reg:SI 2 r2)
        (lo_sum:SI (reg:SI 2 r2)
            (const:SI (unspec:SI [
                        (symbol_ref:SI ("w") [flags 0x240]  <var_decl
0x420a3a20 w>)
                    ] UNSPEC_MACHOPIC_OFFSET)))) ../tests/pr44707-x.c:7 14
{macho_low_si}
     (nil))


which leads to the assembler fail...

====

So ... I don't know whether the problem lies in combine (incorrectly
eliminating the address load) -- or in ira (not re-introducing the indirect
reference).

Would welcome input from one of you gurus ... 
this is an irritating long-term fail (and, I guess, it might even have
implications for other targets...)