From: Jeff Law <jeffreyalaw@gmail.com>
To: Philipp Tomsich <philipp.tomsich@vrull.eu>, gcc-patches@gcc.gnu.org
Cc: Kito Cheng <kito.cheng@gmail.com>,
Christoph Muellner <christoph.muellner@vrull.eu>,
Palmer Dabbelt <palmer@rivosinc.com>,
Andrew Waterman <andrew@sifive.com>,
Vineet Gupta <vineetg@rivosinc.com>
Subject: Re: [RFC PATCH v1 08/10] ifcvt: add if-conversion to conditional-zero instructions
Date: Mon, 13 Feb 2023 00:31:45 -0700 [thread overview]
Message-ID: <90ef6883-397b-8e6a-1ecb-3663fe854ac7@gmail.com> (raw)
In-Reply-To: <20230210224150.2801962-9-philipp.tomsich@vrull.eu>
On 2/10/23 15:41, Philipp Tomsich wrote:
> Some architectures, as it the case on RISC-V with the proposed
> ZiCondOps and the vendor-defined XVentanaCondOps, define a
> conditional-zero instruction that is equivalent to:
> - the positive form: rd = (rc != 0) ? rs : 0
> - the negated form: rd = (rc == 0) ? rs : 0
>
> While noce_try_store_flag_mask will somewhat work for this case, it
> will generate a number of atomic RTX that will misdirect the cost
> calculation and may be too long (i.e., 4 RTX and more) to successfully
> merge at combine-time.
>
> Instead, we add two new transforms that attempt to build up what we
> define as the canonical form of a conditional-zero expression:
>
> (set (match_operand 0 "register_operand" "=r")
> (and (neg (eq_or_ne (match_operand 1 "register_operand" "r")
> (const_int 0)))
> (match_operand 2 "register_operand" "r")))
>
> Architectures that provide a conditional-zero are thus expected to
> define an instruction matching this pattern in their backend.
>
> Based on this, we support the following cases:
> - noce_try_condzero:
> a ? a : b
> a ? b : 0 (and then/else swapped)
> !a ? b : 0 (and then/else swapped)
> - noce_try_condzero_arith:
> conditional-plus, conditional-minus, conditional-and,
> conditional-or, conditional-xor, conditional-shift,
> conditional-and
>
> Given that this is hooked into the CE passes, it is less powerful than
> a tree-pass (e.g., it can not transform cases where an extension, such
> as for uint16_t operations is in either the then or else-branch
> together with the arithmetic) but already covers a good array of cases
> and triggers across SPEC CPU 2017.
>
> Adding transformations in a tree pass should come in a future
> improvement.
>
> gcc/ChangeLog:
>
> * ifcvt.cc (noce_emit_insn): Add prototype.
> (noce_emit_condzero): Helper for noce_try_condzero and
> noce_try_condzero_arith transforms.
> (noce_try_condzero): New transform.
> (noce_try_condzero_arith): New transform for conditional
> arithmetic that can be built up by exploiting that the
> conditional-zero instruction will inject 0, which acts
> as the neutral element for operations.
> (noce_process_if_block): Call noce_try_condzero and
> noce_try_condzero_arith.
A few things we've noted while working with this patch internally.
1. The canonical conditional-zero-or-value requires operand 2 to be a
register. That will tend to inhibit if-conversion of something like a
shift-by-constant. This showed up somewhere in leela.
2. Internally we fixed #1 by forcing that argument into a register when
it was a constant and reload hasn't completed. This (naturally)
resulted in more if-conversions happening. That in turn caused
perlbench to fail. This was tracked down to a conditional-and sequence.
I think the call to noce_emit_condzero needs adjustment to pass arg0
instead of arg1 when OP == AND.
3. The canaonical conditional-zero-or-value assumes the target can do a
generic SEQ/SNE of two register values. As you know, on RISC-V we have
SEQZ/SNEZ. So we've added another fallback path to handle that case in
noce_emit_condzero. You subtract the two values, then you can do an
SEQZ/SNEZ on the result.
Formatting nits noted below.
> + if (!noce_emit_insn (gen_rtx_SET (tmp, cond))) {
> + end_sequence ();
> + return NULL_RTX;
> + }
Open-curly on new line, indented two places from the IF, similarly for
the close curly. That will probably require adjusting the indentation
of the statements.
> + if (code == NE && cond_arg1 == const0_rtx &&
> + REG_P (b) && rtx_equal_p (b, cond_arg0))
Don't end with &&, instead bring it down indented one position from the
open parenthesis.
> + {
> + orig_b = b;
> + b = const0_rtx;
> + }
> +
> + /* We may encounter the form "(b == 0) ? b : a", which can be
> + simplied to "(b == 0) ? 0 : a". */
> + if (code == EQ && cond_arg1 == const0_rtx &&
> + REG_P (b) && rtx_equal_p (b, cond_arg0))
Likewise.
> + {
> + target = noce_emit_condzero(if_info, reversep ? a : b, reversep);
Missing whitespace before open-paren of function call.
> +
> + if (op != PLUS && op != MINUS && op != IOR && op != XOR &&
> + op != ASHIFT && op != ASHIFTRT && op != LSHIFTRT && op != AND)
Bring the "&&" down to the next line here too.
> +
> + if (!rtx_equal_p (if_info->x, arg0))
> + return FALSE;
?!? What's this for?
> +
> + target = noce_emit_condzero(if_info, arg1, op != AND ? true : false);
Missing space before open-paren.
> +
> + emit_insn_before_setloc(seq, if_info->jump,
> + INSN_LOCATION(if_info->insn_a));
Here too.
I think the ChangeLog entry for the testsuite needs its filenames
updated :-)
While this was submitted before stage1 closed, I think we're probably
too late for it to go forward until we re-open stage1.
jeff
next prev parent reply other threads:[~2023-02-13 7:31 UTC|newest]
Thread overview: 31+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-02-10 22:41 [RFC PATCH v1 00/10] RISC-V: Support the Zicond (conditional-operations) extension Philipp Tomsich
2023-02-10 22:41 ` [RFC PATCH v1 01/10] docs: Document a canonical RTL for a conditional-zero insns Philipp Tomsich
2023-02-10 23:18 ` Andrew Pinski
2023-02-10 22:41 ` [RFC PATCH v1 02/10] RISC-V: Recognize Zicond (conditional operations) extension Philipp Tomsich
2023-04-20 17:44 ` Jeff Law
2023-02-10 22:41 ` [RFC PATCH v1 03/10] RISC-V: Generate czero.eqz/nez on noce_try_store_flag_mask if-conversion Philipp Tomsich
2023-04-20 17:53 ` Jeff Law
2023-02-10 22:41 ` [RFC PATCH v1 04/10] RISC-V: Support immediates in Zicond Philipp Tomsich
2023-04-20 18:00 ` Jeff Law
2023-02-10 22:41 ` [RFC PATCH v1 05/10] RISC-V: Support noce_try_store_flag_mask as czero.eqz/czero.nez Philipp Tomsich
2023-04-21 19:31 ` Jeff Law
2023-02-10 22:41 ` [RFC PATCH v1 06/10] RISC-V: Recognize sign-extract + and cases for czero.eqz/nez Philipp Tomsich
2023-04-21 19:40 ` Jeff Law
2023-02-10 22:41 ` [RFC PATCH v1 07/10] RISC-V: Recognize bexti in negated if-conversion Philipp Tomsich
2023-04-21 19:56 ` Jeff Law
2023-02-10 22:41 ` [RFC PATCH v1 08/10] ifcvt: add if-conversion to conditional-zero instructions Philipp Tomsich
2023-02-10 23:07 ` Andrew Pinski
2023-02-13 17:32 ` Richard Sandiford
2023-02-13 18:43 ` Jeff Law
2023-02-13 18:53 ` Andrew Pinski
2023-02-13 7:31 ` Jeff Law [this message]
2023-02-28 16:42 ` Maciej W. Rozycki
2023-03-11 15:50 ` Jeff Law
2023-02-10 22:41 ` [RFC PATCH v1 09/10] RISC-V: Recognize xventanacondops extension Philipp Tomsich
2023-04-21 19:57 ` Jeff Law
2023-04-25 9:53 ` Kito Cheng
2023-04-25 10:15 ` Philipp Tomsich
2023-04-25 10:43 ` Kito Cheng
2023-04-26 2:28 ` Jeff Law
2023-02-10 22:41 ` [RFC PATCH v1 10/10] RISC-V: Support XVentanaCondOps extension Philipp Tomsich
2023-04-21 19:58 ` Jeff Law
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=90ef6883-397b-8e6a-1ecb-3663fe854ac7@gmail.com \
--to=jeffreyalaw@gmail.com \
--cc=andrew@sifive.com \
--cc=christoph.muellner@vrull.eu \
--cc=gcc-patches@gcc.gnu.org \
--cc=kito.cheng@gmail.com \
--cc=palmer@rivosinc.com \
--cc=philipp.tomsich@vrull.eu \
--cc=vineetg@rivosinc.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).