From: Jiufu Guo <guojiufu@linux.ibm.com>
To: Segher Boessenkool <segher@kernel.crashing.org>
Cc: Richard Biener <rguenther@suse.de>,
gcc-patches@gcc.gnu.org, dje.gcc@gmail.com, jlaw@tachyum.com,
wschmidt@linux.ibm.com
Subject: Re: [PATCH] Check if loading const from mem is faster
Date: Tue, 08 Mar 2022 19:25:26 +0800 [thread overview]
Message-ID: <7pcziwy9cp.fsf@linux.ibm.com> (raw)
In-Reply-To: <7pilsvqref.fsf@linux.ibm.com> (Jiufu Guo's message of "Thu, 03 Mar 2022 18:09:12 +0800")
Jiufu Guo <guojiufu@linux.ibm.com> writes:
Hi!
> Hi Sehger,
>
> Segher Boessenkool <segher@kernel.crashing.org> writes:
>
>> On Tue, Mar 01, 2022 at 10:28:57PM +0800, Jiufu Guo wrote:
>>> Segher Boessenkool <segher@kernel.crashing.org> writes:
>>> > No. insn_cost is only for correct, existing instructions, not for
>>> > made-up nonsense. I created insn_cost precisely to get away from that
>>> > aspect of rtx_cost (and some other issues, like, it is incredibly hard
>>> > and cumbersome to write a correct rtx_cost).
>>>
>>> Thanks! The implementations of hook insn_cost are align with this
>>> design, they are checking insn's attributes and COSTS_N_INSNS.
>>>
>>> One question on the speciall case:
>>> For instruction: "r119:DI=0x100803004101001"
>>> Would we treat it as valid instruction?
>>
>> Currently we do, alternative 6 in *movdi_internal64: we allow any r<-n.
>> This is costed as 5 insns (cost=20).
>>
>> It generally is better to split things into patterns close to the
>> eventual machine isntructions as early as possible: all the more generic
>> optimisations can take advantage of that then.
> Get it!
>>
>>> A patch, which is attached the end of this mail, accepts
>>> "r119:DI=0x100803004101001" as input of insn_cost.
>>> In this patch,
>>> - A tmp instruction is generated via make_insn_raw.
>>> - A few calls to rtx_cost (in cse_insn) is replaced by insn_cost.
>>> - In hook of insn_cost, checking the special 'constant' instruction.
>>> Are these make sense?
>>
>> I'll review that patch inline.
I drafted a new patch that replace rtx_cost with insn_cost for cse.cc.
Different from the previous partial patch, this patch replaces all usage
of rtx_cost. It may be better/aggressive than previous one.
With this patch, bootstrap pass.
From regtest, only output of fusion-p10-ldcmpi.c is changed, and the
change seems as expected.
BR,
Jiufu
diff --git a/gcc/cse.cc b/gcc/cse.cc
index a18b599d324..e623ad298db 100644
--- a/gcc/cse.cc
+++ b/gcc/cse.cc
@@ -262,6 +262,9 @@ static struct qty_table_elem *qty_table;
static rtx_insn *this_insn;
static bool optimize_this_for_speed_p;
+/* Used for insn_cost. */
+static rtx_insn *estimate_insn;
+
/* Index by register number, gives the number of the next (or
previous) register in the chain of registers sharing the same
value.
@@ -445,7 +448,7 @@ struct table_elt
/* Compute cost of X, as stored in the `cost' field of a table_elt. Fixed
hard registers and pointers into the frame are the cheapest with a cost
of 0. Next come pseudos with a cost of one and other hard registers with
- a cost of 2. Aside from these special cases, call `rtx_cost'. */
+ a cost of 2. Aside from these special cases, call `insn_cost'. */
#define CHEAP_REGNO(N) \
(REGNO_PTR_FRAME_P (N) \
@@ -698,18 +701,33 @@ preferable (int cost_a, int regcost_a, int cost_b, int regcost_b)
from COST macro to keep it simple. */
static int
-notreg_cost (rtx x, machine_mode mode, enum rtx_code outer, int opno)
+notreg_cost (rtx x, machine_mode mode, enum rtx_code /*outer*/, int /*opno*/)
{
scalar_int_mode int_mode, inner_mode;
- return ((GET_CODE (x) == SUBREG
- && REG_P (SUBREG_REG (x))
- && is_int_mode (mode, &int_mode)
- && is_int_mode (GET_MODE (SUBREG_REG (x)), &inner_mode)
- && GET_MODE_SIZE (int_mode) < GET_MODE_SIZE (inner_mode)
- && subreg_lowpart_p (x)
- && TRULY_NOOP_TRUNCATION_MODES_P (int_mode, inner_mode))
- ? 0
- : rtx_cost (x, mode, outer, opno, optimize_this_for_speed_p) * 2);
+ if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))
+ && is_int_mode (mode, &int_mode)
+ && is_int_mode (GET_MODE (SUBREG_REG (x)), &inner_mode)
+ && GET_MODE_SIZE (int_mode) < GET_MODE_SIZE (inner_mode)
+ && subreg_lowpart_p (x)
+ && TRULY_NOOP_TRUNCATION_MODES_P (int_mode, inner_mode))
+ return 0;
+
+ if (estimate_insn == NULL)
+ {
+ estimate_insn = make_insn_raw (
+ gen_rtx_SET (gen_rtx_REG (mode, LAST_VIRTUAL_REGISTER + 1), x));
+ SET_PREV_INSN (estimate_insn) = NULL_RTX;
+ SET_NEXT_INSN (estimate_insn) = NULL_RTX;
+ INSN_LOCATION (estimate_insn) = 0;
+ }
+ else
+ {
+ /* Update for new context. */
+ INSN_CODE (estimate_insn) = -1;
+ PUT_MODE (SET_DEST (PATTERN (estimate_insn)), mode);
+ SET_SRC (PATTERN (estimate_insn)) = x;
+ }
+ return insn_cost (estimate_insn, optimize_this_for_speed_p);
}
\f
@@ -6667,6 +6685,7 @@ cse_main (rtx_insn *f ATTRIBUTE_UNUSED, int nregs)
init_recog ();
init_alias_analysis ();
+ estimate_insn = NULL;
reg_eqv_table = XNEWVEC (struct reg_eqv_elem, nregs);
next prev parent reply other threads:[~2022-03-08 11:25 UTC|newest]
Thread overview: 34+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-02-22 6:53 Jiufu Guo
2022-02-22 7:26 ` Richard Biener
2022-02-23 11:32 ` guojiufu
2022-02-23 13:02 ` Richard Biener
2022-02-23 21:14 ` Segher Boessenkool
2022-02-24 5:56 ` Jiufu Guo
2022-02-24 6:33 ` Jiufu Guo
2022-02-24 8:50 ` Richard Biener
2022-02-25 4:35 ` Jiufu Guo
2022-02-25 8:45 ` Richard Biener
2022-02-25 13:32 ` Jiufu Guo
2022-02-25 13:57 ` Richard Biener
2022-02-28 9:15 ` Jiufu Guo
2022-02-28 17:03 ` Segher Boessenkool
2022-03-01 2:59 ` Jiufu Guo
2022-03-01 7:47 ` Richard Biener
2022-03-01 13:47 ` Jiufu Guo
2022-03-02 19:15 ` Jeff Law
2022-03-03 10:08 ` Jiufu Guo
2022-02-23 21:27 ` Segher Boessenkool
2022-02-24 7:48 ` Jiufu Guo
2022-02-28 16:45 ` Segher Boessenkool
2022-03-01 14:28 ` Jiufu Guo
2022-03-02 20:24 ` Segher Boessenkool
2022-03-03 10:09 ` Jiufu Guo
2022-03-08 11:25 ` Jiufu Guo [this message]
2022-03-08 11:50 ` Richard Biener
2022-03-09 4:37 ` Jiufu Guo
2022-03-09 7:41 ` Richard Biener
2022-03-10 2:09 ` Jiufu Guo
2022-03-10 7:09 ` Richard Biener
2022-03-11 6:33 ` Jiufu Guo
2022-02-22 17:30 ` Segher Boessenkool
2022-02-23 3:31 ` guojiufu
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=7pcziwy9cp.fsf@linux.ibm.com \
--to=guojiufu@linux.ibm.com \
--cc=dje.gcc@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=jlaw@tachyum.com \
--cc=rguenther@suse.de \
--cc=segher@kernel.crashing.org \
--cc=wschmidt@linux.ibm.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).