------- Comment #5 from ubizjak at gmail dot com 2007-06-19 08:27 -------
(In reply to comment #4)
> No, this one is caused by dataflow.
Dataflow uncovered generic middle-end (RTL?) problem:
We have this comment in instantiate_virutal_regs():
/* Scan through all the insns, instantiating every virtual register still
present. */
for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
if (INSN_P (insn))
{
/* These patterns in the instruction stream can never be recognized.
Fortunately, they shouldn't contain virtual registers either. */
if (GET_CODE (PATTERN (insn)) == USE
|| GET_CODE (PATTERN (insn)) == CLOBBER
|| GET_CODE (PATTERN (insn)) == ADDR_VEC
|| GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC
|| GET_CODE (PATTERN (insn)) == ASM_INPUT)
continue;
instantiate_virtual_regs_in_insn (insn);
However, for reduced testcase we generate following creative sequence to put
*.LC0 into virtual-outgoing-args and 0 into virtual-outgoing-args+8 (these are
sections[0] values):
;; f (stderr, &""[0], &""[0], conf_name, 0, sections[0])
(insn 6 5 7 pr32374.c:13 (clobber (mem/s/c:BLK (plus:DI (reg/f:DI 54
virtual-stack-vars)
(const_int -16 [0xfffffffffffffff0])) [5 A128])) -1 (nil))
(insn 7 6 8 pr32374.c:13 (set (mem/s/c:DI (plus:DI (reg/f:DI 54
virtual-stack-vars)
(const_int -16 [0xfffffffffffffff0])) [4 S8 A128])
(symbol_ref/f:DI ("*.LC0") [flags 0x2] <string_cst 0xb7cf1498>)) -1
(nil))
(insn 8 7 9 pr32374.c:13 (set (mem/s/c:SI (plus:DI (reg/f:DI 54
virtual-stack-vars)
(const_int -8 [0xfffffffffffffff8])) [3 S4 A64])
(const_int 0 [0x0])) -1 (nil))
(insn 9 8 10 pr32374.c:13 (set (reg:DI 59)
(mem/s/c:DI (plus:DI (reg/f:DI 54 virtual-stack-vars)
(const_int -16 [0xfffffffffffffff0])) [5 S8 A128])) -1 (nil))
(insn 10 9 11 pr32374.c:13 (set (mem:DI (reg/f:DI 56 virtual-outgoing-args) [0
S8 A64])
(reg:DI 59)) -1 (nil))
(insn 11 10 12 pr32374.c:13 (set (reg:DI 60)
(mem/s/c:DI (plus:DI (reg/f:DI 54 virtual-stack-vars)
(const_int -8 [0xfffffffffffffff8])) [5 S8 A64])) -1 (nil))
(insn 12 11 13 pr32374.c:13 (set (mem:DI (plus:DI (reg/f:DI 56
virtual-outgoing-args)
(const_int 8 [0x8])) [0 S8 A64])
(reg:DI 60)) -1 (nil))
Putting a break on emit_insn_raw, we can backtrack where (invalid!) clobber is
generated:
#0 make_insn_raw (pattern=0xb7d0ad00) at
../../gcc-svn/trunk/gcc/emit-rtl.c:3312
#1 0x081861c2 in emit_insn (x=0xb7d0ad00) at
../../gcc-svn/trunk/gcc/emit-rtl.c:4347
#2 0x081a9821 in store_constructor (exp=0xb7d63dc8, target=0xb7d4f768,
cleared=0, size=16) at ../../gcc-svn/trunk/gcc/expr.c:5066
#3 0x081aee5b in expand_expr_real_1 (exp=0x11, target=0xb7d4f768,
tmode=VOIDmode, modifier=EXPAND_STACK_PARM, alt_rtl=0x0) at
../../gcc-svn/trunk/gcc/expr.c:7310
#4 0x081bf7cc in expand_expr_real (exp=0xb7d63dc8, target=0xb7d4f6f0,
tmode=VOIDmode, modifier=EXPAND_STACK_PARM, alt_rtl=0x0) at
../../gcc-svn/trunk/gcc/expr.c:6862
#5 0x081bf9b6 in expand_expr (exp=0x11, target=0xb7d0ad00, mode=VOIDmode,
modifier=EXPAND_STACK_PARM) at ../../gcc-svn/trunk/gcc/expr.h:504
#6 0x081bd4c4 in expand_expr_real_1 (exp=0xb7ca3058, target=0xb7d4f6f0,
tmode=dwarf2_read_address: Corrupted DWARF expression.
) at ../../gcc-svn/trunk/gcc/expr.c:7455
#7 0x081bf7cc in expand_expr_real (exp=0xb7ca3058, target=0xb7d4f6f0,
tmode=VOIDmode, modifier=EXPAND_STACK_PARM, alt_rtl=0x0) at
../../gcc-svn/trunk/gcc/expr.c:6862
#8 0x08115bd1 in store_one_arg (arg=0xbf96e8f0, argblock=0xb7c9c070, flags=0,
variable_size=0, reg_parm_stack_space=0) at ../../gcc-svn/trunk/gcc/expr.h:504
#9 0x0811ad60 in expand_call (exp=0xb7d5a240, target=0x0, ignore=1) at
../../gcc-svn/trunk/gcc/calls.c:2654
It can be tracked from expand_call() down to expr.c:7303, where we re-assign
target to virtual-stack-vars in this highly suspicious part of code (Irix 6?):
else
{
/* Handle calls that pass values in multiple non-contiguous
locations. The Irix 6 ABI has examples of this. */
if (target == 0 || ! safe_from_p (target, exp, 1)
|| GET_CODE (target) == PARALLEL
|| modifier == EXPAND_STACK_PARM)
target
= assign_temp (build_qualified_type (type,
(TYPE_QUALS (type)
| (TREE_READONLY (exp)
* TYPE_QUAL_CONST))),
0, TREE_ADDRESSABLE (exp), 1);
store_constructor (exp, target, 0, int_expr_size (exp));
return target;
}
Just before the call to store_constructor, we have:
p debug_rtx (target)
(mem/s/c:BLK (plus:DI (reg/f:DI 54 virtual-stack-vars)
(const_int -16 [0xfffffffffffffff0])) [5 A128])
and in store_constructor():
in store_constructor (exp=0xb7d83dc8, target=0xb7d6f768, cleared=0, size=16) at
../../gcc-svn/trunk/gcc/expr.c:5066
we proceed through:
if (! cleared)
emit_insn (gen_rtx_CLOBBER (VOIDmode, target));
to generate the clobber above.
--
ubizjak at gmail dot com changed:
What |Removed |Added
----------------------------------------------------------------------------
Component|rtl-optimization |middle-end
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32374
