Hi all,
The auto_inc_dec pass can handle 4 types of sequences, described in the comment
at the start of auto-inc-dec.c.
In two of those: FORM_PRE_ADD and FORM_POST_ADD the resulting sequence is a
move followed by a POST_INC or PRE_INC
memory operation.
In the FORM_POST_ADD case the pass transforms:
*a
...
b <- a + c
into
b <- a
...
*(b += c) post
However, the code in attempt_change that compares the costs of the before and
after sequences
has an oversight. When calculating the cost of the new sequence it doesn't take
into account the cost of the
b <- a move. This patch fixes the calculation by calling seq_cost on the result
of the emit_move_insn call
we do to emit that move.
With this patch I've seen less aggressive generation of POST_INC memory
instructions on arm, but the post_inc
tests we have in the arm testsuite still pass, so I don't think this kills the
usage of those instructions on
arm, just tames them somewhat.
No regressions on SPEC2000 on Cortex-A15.
SPECINT 2006 improves by a bit.
Bootstrapped and tested on arm, aarch64.
Ok for trunk?
Thanks,
Kyrill
2015-10-26 Kyrylo Tkachov <kyrylo.tkac...@arm.com>
* auto-inc-dec.c (insert_move_insn_before): Delete.
(attempt_change): Remember to cost the simple move in the
FORM_PRE_ADD and FORM_POST_ADD cases.
commit 569d1d9b789a38bf4305991d96cbb03d1665e311
Author: Kyrylo Tkachov <kyrylo.tkac...@arm.com>
Date: Fri Oct 16 13:46:51 2015 +0100
[auto-inc-dec.c] Account for cost of move operation in FORM_PRE_ADD and FORM_POST_ADD cases
diff --git a/gcc/auto-inc-dec.c b/gcc/auto-inc-dec.c
index 3b9a1f3..af3f8b3 100644
--- a/gcc/auto-inc-dec.c
+++ b/gcc/auto-inc-dec.c
@@ -438,24 +438,6 @@ move_dead_notes (rtx_insn *to_insn, rtx_insn *from_insn, rtx pattern)
}
}
-
-/* Create a mov insn DEST_REG <- SRC_REG and insert it before
- NEXT_INSN. */
-
-static rtx_insn *
-insert_move_insn_before (rtx_insn *next_insn, rtx dest_reg, rtx src_reg)
-{
- rtx_insn *insns;
-
- start_sequence ();
- emit_move_insn (dest_reg, src_reg);
- insns = get_insns ();
- end_sequence ();
- emit_insn_before (insns, next_insn);
- return insns;
-}
-
-
/* Change mem_insn.mem_loc so that uses NEW_ADDR which has an
increment of INC_REG. To have reached this point, the change is a
legitimate one from a dataflow point of view. The only questions
@@ -489,7 +471,23 @@ attempt_change (rtx new_addr, rtx inc_reg)
old_cost = (set_src_cost (mem, mode, speed)
+ set_rtx_cost (PATTERN (inc_insn.insn), speed));
- new_cost = set_src_cost (mem_tmp, mode, speed);
+
+ int new_mem_cost = set_src_cost (mem_tmp, mode, speed);
+ int new_mov_cost = 0;
+
+ /* In the FORM_PRE_ADD and FORM_POST_ADD cases we emit an extra move
+ whose cost we should account for. */
+ if (inc_insn.form == FORM_PRE_ADD
+ || inc_insn.form == FORM_POST_ADD)
+ {
+ start_sequence ();
+ emit_move_insn (inc_insn.reg_res, inc_insn.reg0);
+ mov_insn = get_insns ();
+ end_sequence ();
+ new_mov_cost = seq_cost (mov_insn, speed);
+ }
+
+ new_cost = new_mem_cost + new_mov_cost;
/* The first item of business is to see if this is profitable. */
if (old_cost < new_cost)
@@ -521,8 +519,8 @@ attempt_change (rtx new_addr, rtx inc_reg)
/* Replace the addition with a move. Do it at the location of
the addition since the operand of the addition may change
before the memory reference. */
- mov_insn = insert_move_insn_before (inc_insn.insn,
- inc_insn.reg_res, inc_insn.reg0);
+ gcc_assert (mov_insn);
+ emit_insn_before (mov_insn, inc_insn.insn);
move_dead_notes (mov_insn, inc_insn.insn, inc_insn.reg0);
regno = REGNO (inc_insn.reg_res);
@@ -547,8 +545,8 @@ attempt_change (rtx new_addr, rtx inc_reg)
break;
case FORM_POST_ADD:
- mov_insn = insert_move_insn_before (mem_insn.insn,
- inc_insn.reg_res, inc_insn.reg0);
+ gcc_assert (mov_insn);
+ emit_insn_before (mov_insn, mem_insn.insn);
move_dead_notes (mov_insn, inc_insn.insn, inc_insn.reg0);
/* Do not move anything to the mov insn because the instruction
