------- Comment #14 from danglin at gcc dot gnu dot org 2005-12-19 04:45
-------
I did a checkout as of Sep 12, 2005 00:00:00 UTC and get the same assembly
code as in the initial report using the following compilation options:
"-Os -w -S -mschedule=7100LC". The build was done on hppa2.0w-hp-hpux11.11.
Here's some rtl from 54.barriers and 55.dbr showing the end of the initial
loop in the function testB:
(jump_insn:TI 18 15 186 (set (pc)
(if_then_else (ne (reg/v/f:SI 6 %r6 [orig:98 p___4463 ] [98])
(reg/f:SI 19 %r19 [111]))
(label_ref:SI 11)
(pc))) 25 {*pa.md:1700} (insn_list:REG_DEP_TRUE 17
(insn_list:REG_DE
P_ANTI 13 (insn_list:REG_DEP_TRUE 15 (nil))))
(expr_list:REG_DEAD (reg/f:SI 19 %r19 [111])
(expr_list:REG_BR_PROB (const_int 8750 [0x222e])
(nil))))
(note 186 18 20 NOTE_INSN_LOOP_END)
(note 20 186 45 [bb 2] NOTE_INSN_BASIC_BLOCK)
(note 45 20 23 NOTE_INSN_DELETED)
(insn:TI 23 45 21 (set (reg:SI 28 %r28 [114])
(mem/s/j/c:SI (plus:SI (reg/v/f:SI 6 %r6 [orig:98 p___4463 ] [98])
(const_int -8 [0xfffffff8])) [0+0 S4 A32])) 37 {*pa.md:2291}
(ni
l)
(expr_list:REG_EQUAL (mem/s/j/c:SI (plus:SI (reg/v/f:SI 6 %r6 [orig:98
p___4
463 ] [98])
(const_int -8 [0xfffffff8])) [0+0 S4 A32])
(nil)))
(insn 21 23 196 (set (reg:SI 1 %r1)
(high:SI (symbol_ref:SI ("sB") [flags 0x80] <var_decl 7afc1bb0 sB>)))
48
{*pa.md:2737} (nil)
(expr_list:REG_EQUIV (high:SI (symbol_ref:SI ("sB") [flags 0x80] <var_decl
7
afc1bb0 sB>))
(nil)))
----
(insn 253 15 186 (sequence [
(jump_insn:TI 18 15 23 (set (pc)
(if_then_else (ne (reg/v/f:SI 6 %r6 [orig:98 p___4463 ]
[98]
)
(reg/f:SI 19 %r19 [111]))
(label_ref:SI 11)
(pc))) 25 {*pa.md:1700} (insn_list:REG_DEP_TRUE 17
(insn
_list:REG_DEP_ANTI 13 (insn_list:REG_DEP_TRUE 15 (nil))))
(expr_list:REG_BR_PRED (const_int 6 [0x6])
(expr_list:REG_DEAD (reg/f:SI 19 %r19 [111])
(expr_list:REG_BR_PROB (const_int 8750 [0x222e])
(nil)))))
(insn:TI 23 18 186 (set (reg:SI 28 %r28 [114])
(mem/s/j/c:SI (plus:SI (reg/v/f:SI 6 %r6 [orig:98 p___4463
]
[98])
(const_int -8 [0xfffffff8])) [0+0 S4 A32])) 37
{*pa.
md:2291} (nil)
(expr_list:REG_EQUAL (mem/s/j/c:SI (plus:SI (reg/v/f:SI 6 %r6
[o
rig:98 p___4463 ] [98])
(const_int -8 [0xfffffff8])) [0+0 S4 A32])
(nil)))
]) -1 (nil)
(nil))
(note 186 253 20 NOTE_INSN_LOOP_END)
(note 20 186 45 [bb 2] NOTE_INSN_BASIC_BLOCK)
(note 45 20 21 NOTE_INSN_DELETED)
(insn 21 45 196 (set (reg:SI 1 %r1)
(high:SI (symbol_ref:SI ("sB") [flags 0x80] <var_decl 7afc1bb0 sB>)))
48
{*pa.md:2737} (nil)
(expr_list:REG_EQUIV (high:SI (symbol_ref:SI ("sB") [flags 0x80] <var_decl
7
afc1bb0 sB>))
(nil)))
As can be seen, the memory load that causes the segmentation fault was outside
the loop in the 54.barriers pass.
>From an optimisation standpoint, I'm a bit surprised that the memory load
was selected for placement into the delay slot rather than the following set.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23954
