------- Comment #2 from kkojima at gcc dot gnu dot org 2010-07-30 00:50 -------
With -g, .ira for the above test case has the insns:
(insn 143 141 229 14 (set (reg/f:SI 1 r1 [238])
(const_int 2046 [0x7fe])) yyy.c:25 176 {movsi_ie}
(expr_list:REG_EQUAL (const_int 2046 [0x7fe])
(nil)))
(debug_insn 229 143 144 14 (var_location:SI D#2 (reg/v/f:SI 8 r8 [orig:210 buf
] [210])) -1
(nil))
(insn 144 229 145 14 (set (reg/f:SI 1 r1 [238])
(plus:SI (reg/f:SI 1 r1 [238])
(reg/v/f:SI 8 r8 [orig:210 buf ] [210]))) yyy.c:25 35
{*addsi3_compact}
(expr_list:REG_EQUAL (plus:SI (reg/v/f:SI 8 r8 [orig:210 buf ] [210])
(const_int 2046 [0x7fe]))
(nil)))
(insn 145 144 146 14 (set (reg:QI 2 r2 [240])
(const_int 0 [0])) yyy.c:25 186 {movqi_i}
(expr_list:REG_EQUIV (const_int 0 [0])
(nil)))
(insn 146 145 38 14 (set (mem:QI (reg/f:SI 1 r1 [238]) [0 MEM[(char *)buf_1(D)
+ 2046B]+0 S1 A8])
(reg:QI 2 r2 [240])) yyy.c:25 186 {movqi_i}
(expr_list:REG_EQUAL (const_int 0 [0])
(nil)))
When debug_insn 229 is absent, reload_combine_recognize_pattern
converts them to the insns like
(insn 143 141 145 14 (set (reg:SI 0 r0)
(const_int 2046 [0x7fe])) xxx.c:25 176 {movsi_ie}
(expr_list:REG_EQUAL (const_int 2046 [0x7fe])
(nil)))
(insn 145 143 146 14 (set (reg:QI 2 r2 [240])
(const_int 0 [0])) xxx.c:25 186 {movqi_i}
(expr_list:REG_EQUIV (const_int 0 [0])
(nil)))
(insn 146 145 38 14 (set (mem:QI (plus:SI (reg:SI 0 r0)
(reg/v/f:SI 8 r8 [orig:210 buf ] [210])) [0 MEM[(char
*)buf_1(D) + 2046B]+0 S1 A8])
(reg:QI 2 r2 [240])) xxx.c:25 186 {movqi_i}
(expr_list:REG_EQUAL (const_int 0 [0])
(nil)))
but this isn't done with debug_insn 229.
It seems that reload_combine_recognize_pattern only looks
prev_nonnote_insn of insn 145 as the candidate of
(set (REGX) (CONST_INT)) insn and the debug_insn confuses
it. I've confirmed that the patch
--- ORIG/trunk/gcc/postreload.c 2010-07-28 09:38:05.000000000 +0900
+++ trunk/gcc/postreload.c 2010-07-30 08:42:50.000000000 +0900
@@ -1113,12 +1113,20 @@ reload_combine_recognize_pattern (rtx in
&& last_label_ruid < reg_state[regno].use_ruid)
{
rtx base = XEXP (src, 1);
- rtx prev = prev_nonnote_insn (insn);
- rtx prev_set = prev ? single_set (prev) : NULL_RTX;
+ rtx prev;
+ rtx prev_set;
rtx index_reg = NULL_RTX;
rtx reg_sum = NULL_RTX;
int i;
+ /* Find the previous non-note/dubug insn. */
+ prev = insn;
+ do {
+ prev = prev_nonnote_insn (prev);
+ } while (prev && DEBUG_INSN_P (prev));
+
+ prev_set = prev ? single_set (prev) : NULL_RTX;
+
/* Now we need to set INDEX_REG to an index register (denoted as
REGZ in the illustration above) and REG_SUM to the expression
register+register that we want to use to substitute uses of REG
fixes the problem.
Perhaps {next,prev}_nonnote_nondebug_insn should be added
to emit-rtl.c and use it here. Thoughts?
--
kkojima at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
Summary|-g changes the generated |[4.6 Regression] -g changes
|code for gcc/sched-vis.c on |the generated code for
|SH |gcc/sched-vis.c on SH
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45137
