From: Bernd Edlinger <bernd.edlinger@hotmail.de>
To: "H.J. Lu" <hjl.tools@gmail.com>,
Richard Biener <richard.guenther@gmail.com>
Cc: GCC Patches <gcc-patches@gcc.gnu.org>,
Richard Sandiford <richard.sandiford@arm.com>,
Uros Bizjak <ubizjak@gmail.com>
Subject: Re: [PATCH] constructor: Elide expand_constructor when can move by pieces is true
Date: Fri, 21 May 2021 07:35:42 +0200 [thread overview]
Message-ID: <AM8PR10MB47085C3C75150670BF60D89DE4299@AM8PR10MB4708.EURPRD10.PROD.OUTLOOK.COM> (raw)
In-Reply-To: <CAMe9rOpEeTHmJwzdsVGPEm1XVzd=Ejw81_7yv0+yzDCXTrWyDw@mail.gmail.com>
On 5/20/21 4:03 PM, H.J. Lu wrote:
> On Thu, May 20, 2021 at 12:51 AM Richard Biener
> <richard.guenther@gmail.com> wrote:
>>
>> On Wed, May 19, 2021 at 3:22 PM H.J. Lu <hjl.tools@gmail.com> wrote:
>>>
>>> On Wed, May 19, 2021 at 2:33 AM Richard Biener
>>> <richard.guenther@gmail.com> wrote:
>>>>
>>>> On Tue, May 18, 2021 at 9:16 PM H.J. Lu <hjl.tools@gmail.com> wrote:
>>>>>
>>>>> When expanding a constant constructor, don't call expand_constructor if
>>>>> it is more efficient to load the data from the memory via move by pieces.
>>>>>
>>>>> gcc/
>>>>>
>>>>> PR middle-end/90773
>>>>> * expr.c (expand_expr_real_1): Don't call expand_constructor if
>>>>> it is more efficient to load the data from the memory.
>>>>>
>>>>> gcc/testsuite/
>>>>>
>>>>> PR middle-end/90773
>>>>> * gcc.target/i386/pr90773-24.c: New test.
>>>>> * gcc.target/i386/pr90773-25.c: Likewise.
>>>>> ---
>>>>> gcc/expr.c | 10 ++++++++++
>>>>> gcc/testsuite/gcc.target/i386/pr90773-24.c | 22 ++++++++++++++++++++++
>>>>> gcc/testsuite/gcc.target/i386/pr90773-25.c | 20 ++++++++++++++++++++
>>>>> 3 files changed, 52 insertions(+)
>>>>> create mode 100644 gcc/testsuite/gcc.target/i386/pr90773-24.c
>>>>> create mode 100644 gcc/testsuite/gcc.target/i386/pr90773-25.c
>>>>>
>>>>> diff --git a/gcc/expr.c b/gcc/expr.c
>>>>> index d09ee42e262..80e01ea1cbe 100644
>>>>> --- a/gcc/expr.c
>>>>> +++ b/gcc/expr.c
>>>>> @@ -10886,6 +10886,16 @@ expand_expr_real_1 (tree exp, rtx target, machine_mode tmode,
>>>>> unsigned HOST_WIDE_INT ix;
>>>>> tree field, value;
>>>>>
>>>>> + /* Check if it is more efficient to load the data from
>>>>> + the memory directly. FIXME: How many stores do we
>>>>> + need here if not moved by pieces? */
>>>>> + unsigned HOST_WIDE_INT bytes
>>>>> + = tree_to_uhwi (TYPE_SIZE_UNIT (type));
>>>>
>>>> that's prone to fail - it could be a VLA.
>>>
>>> What do you mean by fail? Is it ICE or missed optimization?
>>> Do you have a testcase?
>>>
>>>>
>>>>> + if ((bytes / UNITS_PER_WORD) > 2
>>>>> + && MOVE_MAX_PIECES > UNITS_PER_WORD
>>>>> + && can_move_by_pieces (bytes, TYPE_ALIGN (type)))
>>>>> + goto normal_inner_ref;
>>>>> +
>>>>
>>>> It looks like you're concerned about aggregate copies but this also handles
>>>> non-aggregates (which on GIMPLE might already be optimized of course).
>>>
>>> Here I check if we copy more than 2 words and we can move more than
>>> a word in a single instruction.
>>>
>>>> Also you say "if it's cheaper" but I see no cost considerations. How do
>>>> we generally handle immed const vs. load from constant pool costs?
>>>
>>> This trades 2 (update to 8) stores with one load plus one store. Is there
>>> a way to check which one is faster?
>>
>> I'm not sure - it depends on whether the target can do stores from immediates
>> at all or what restrictions apply, what the immediate value actually is
>> (zero or all-ones should be way cheaper than sth arbitrary) and how the
>> pressure on the load unit is. can_move_by_pieces (bytes, TYPE_ALIGN (type))
>> also does not guarantee it will actually move pieces larger than UNITS_PER_WORD,
>> that might depend on alignment. There's by_pieces_ninsns that might provide
>> some hint here.
>>
>> I'm sure it works well for x86.
>>
>> I wonder if the existing code is in the appropriate place and we
>> shouldn't instead
>> handle this somewhere upthread where we ask to copy 'exp' into some other
>> memory location. For your testcase that's expand_assignment but I can
>> imagine passing array[0] by value to a function resulting in similar copying.
>> Testing that shows we get
>>
>> pushq array+56(%rip)
>> .cfi_def_cfa_offset 24
>> pushq array+48(%rip)
>> .cfi_def_cfa_offset 32
>> pushq array+40(%rip)
>> .cfi_def_cfa_offset 40
>> pushq array+32(%rip)
>> .cfi_def_cfa_offset 48
>> pushq array+24(%rip)
>> .cfi_def_cfa_offset 56
>> pushq array+16(%rip)
>> .cfi_def_cfa_offset 64
>> pushq array+8(%rip)
>> .cfi_def_cfa_offset 72
>> pushq array(%rip)
>> .cfi_def_cfa_offset 80
>> call bar
>>
>> for that. We do have the by-pieces infrastructure to generally do this kind of
>> copying but in both of these cases we do not seem to use it. I also wonder
>> if the by-pieces infrastructure can pick up constant initializers automagically
>> (we could native_encode the initializer part and feed the by-pieces
>> infrastructure with an array of bytes). There for example might be easy to
>> immediate-store byte parts and difficult ones where we could decide on a
>> case-by-case basis whether to load+store or immediate-store them.
>
> I opened:
>
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100704
>
>> For example if I change your testcase to have the array[] initializer
>> all-zero we currently emit
>>
>> pxor %xmm0, %xmm0
>> movups %xmm0, (%rdi)
>> movups %xmm0, 16(%rdi)
>> movups %xmm0, 32(%rdi)
>> movups %xmm0, 48(%rdi)
>> ret
>>
>> will your patch cause us to emit 4 loads? OTHO if I do
>>
>> const struct S array[] = {
>> { 0, 0, 0, 7241, 124764, 48, 16, 33, 10, 96, 2, 0, 0, 4 }
>> };
>>
>> we get
>>
>> movq $0, (%rdi)
>> movl $0, 8(%rdi)
>> movl $0, 12(%rdi)
>> movl $7241, 16(%rdi)
>> ...
>>
>> ideally we'd have sth like
>>
>> pxor %xmm0, %xmm0
>> movups %xmm0, (%rdi)
>> movaps array+16(%rip), %xmm0
>> movups %xmm0, 16(%rdi)
>> ...
>>
>> thus have the zeros written as immediates and the remaining pieces
>> with load+stores.
>>
>> The by-pieces infrastructure eventually get's to see
>>
>> (mem/u/c:BLK (symbol_ref:DI ("array") [flags 0x2] <var_decl
>> 0x7ffff7ff5b40 array>) [1 array+0 S64 A256])
>>
>> where the MEM_EXPR should provide a way to access the constant initializer.
>>
>> That said I do agree the current code is a bit premature optimization
>> - but maybe
>> it should be fend off in expand_constructor which has the cheap clear_storage
>> first and which already does check can_move_by_pieces with some heuristics,
>> but that seems to be guarded by
>>
>> || (tree_fits_uhwi_p (TYPE_SIZE_UNIT (type))
>> && (! can_move_by_pieces
>> (tree_to_uhwi (TYPE_SIZE_UNIT (type)),
>> TYPE_ALIGN (type)))
>> && ! mostly_zeros_p (exp))))
>>
>> which is odd (we _can_ move by pieces, but how does this apply to
>> TREE_CONSTANT CTORs and avoid_temp_mem?).
>>
>> That said, I wonder if we want to elide expand_constructor when the
>> CTOR is TREE_STATIC && TREE_CONSTANT and !mostly_zeros_p
>> and we can_move_by_pieces.
>>
>> So sth like
>>
>> diff --git a/gcc/expr.c b/gcc/expr.c
>> index 7139545d543..76b3bdf0c01 100644
>> --- a/gcc/expr.c
>> +++ b/gcc/expr.c
>> @@ -8504,6 +8504,12 @@ expand_constructor (tree exp, rtx target, enum
>> expand_modifier modifier,
>> && (! can_move_by_pieces
>> (tree_to_uhwi (TYPE_SIZE_UNIT (type)),
>> TYPE_ALIGN (type)))
>> + && ! mostly_zeros_p (exp))
>> + || (TREE_CONSTANT (exp)
>> + && tree_fits_uhwi_p (TYPE_SIZE_UNIT (type))
>> + && (can_move_by_pieces
>> + (tree_to_uhwi (TYPE_SIZE_UNIT (type)),
>> + TYPE_ALIGN (type)))
Just a minor nit: superfluous parentheses around can_move_by_pieces here.
Bernd.
>> && ! mostly_zeros_p (exp))))
>> || ((modifier == EXPAND_INITIALIZER || modifier == EXPAND_CONST_ADDRESS)
>> && TREE_CONSTANT (exp)))
>>
>> which handles your initializer and the all-zero one optimal?
>>
>
> It works. Here is the updated patch.
>
> Thanks.
>
next prev parent reply other threads:[~2021-05-21 5:35 UTC|newest]
Thread overview: 52+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-05-18 19:16 [PATCH v4 00/12] Allow TImode/OImode/XImode in op_by_pieces operations H.J. Lu
2021-05-18 19:16 ` [PATCH v4 01/12] Add TARGET_READ_MEMSET_VALUE/TARGET_GEN_MEMSET_VALUE H.J. Lu
2021-05-19 9:25 ` Richard Biener
2021-05-19 12:55 ` H.J. Lu
2021-05-20 20:49 ` [PATCH] Add 3 target hooks for memset H.J. Lu
2021-05-21 5:42 ` Bernd Edlinger
2021-05-21 11:53 ` H.J. Lu
2021-05-25 14:34 ` Richard Biener
2021-05-25 15:11 ` H.J. Lu
2021-05-26 8:28 ` Richard Biener
2021-05-31 12:09 ` [PATCH] Add integer_extract and vec_const_duplicate optabs H.J. Lu
2021-05-31 12:46 ` Richard Biener
2021-05-31 13:12 ` H.J. Lu
2021-05-31 13:25 ` Richard Biener
2021-05-31 13:32 ` H.J. Lu
2021-05-31 13:36 ` H.J. Lu
2021-05-31 20:22 ` [PATCH v2] Add vec_const_duplicate optab and TARGET_GEN_MEMSET_SCRATCH_RTX H.J. Lu
2021-06-01 5:50 ` Richard Sandiford
2021-06-01 5:54 ` Jeff Law
2021-06-01 13:05 ` H.J. Lu
2021-06-01 13:25 ` Richard Biener
2021-06-01 13:29 ` H.J. Lu
2021-06-01 14:21 ` Jeff Law
2021-06-01 23:07 ` H.J. Lu
2021-06-02 1:21 ` Hongtao Liu
2021-06-02 1:54 ` H.J. Lu
2021-06-02 7:02 ` Richard Biener
2021-06-02 13:50 ` H.J. Lu
2021-05-18 19:16 ` [PATCH v4 02/12] x86: Add TARGET_READ_MEMSET_VALUE/TARGET_GEN_MEMSET_VALUE H.J. Lu
2021-05-18 19:16 ` [PATCH v4 03/12] x86: Avoid stack realignment when copying data H.J. Lu
2021-05-18 19:16 ` [PATCH v4 04/12] Remove MAX_BITSIZE_MODE_ANY_INT H.J. Lu
2021-05-25 14:37 ` Richard Biener
2021-05-18 19:16 ` [PATCH v4 05/12] x86: Update piecewise move and store H.J. Lu
2021-05-18 19:16 ` [PATCH v4 06/12] x86: Add AVX2 tests for PR middle-end/90773 H.J. Lu
2021-05-18 19:16 ` [PATCH v4 07/12] x86: Add tests for piecewise move and store H.J. Lu
2021-05-18 19:16 ` [PATCH v4 08/12] x86: Also pass -mno-avx to pr72839.c H.J. Lu
2021-05-18 19:16 ` [PATCH v4 09/12] x86: Also pass -mno-avx to cold-attribute-1.c H.J. Lu
2021-05-18 19:16 ` [PATCH v4 10/12] x86: Also pass -mno-avx to sw-1.c for ia32 H.J. Lu
2021-05-18 19:16 ` [PATCH v4 11/12] x86: Update gcc.target/i386/incoming-11.c H.J. Lu
2021-05-18 19:16 ` [PATCH v4 12/12] constructor: Check if it is faster to load constant from memory H.J. Lu
2021-05-19 9:33 ` Richard Biener
2021-05-19 13:22 ` H.J. Lu
2021-05-19 13:27 ` Bernd Edlinger
2021-05-19 19:04 ` H.J. Lu
2021-05-20 6:57 ` Richard Biener
2021-05-20 7:51 ` Richard Biener
2021-05-20 14:03 ` [PATCH] constructor: Elide expand_constructor when can move by pieces is true H.J. Lu
2021-05-21 5:35 ` Bernd Edlinger [this message]
2021-05-21 6:57 ` Richard Biener
2021-05-21 7:30 ` Bernd Edlinger
2021-05-21 13:13 ` H.J. Lu
2021-05-21 13:09 ` [PATCH] Elide expand_constructor if move by pieces is preferred H.J. Lu
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=AM8PR10MB47085C3C75150670BF60D89DE4299@AM8PR10MB4708.EURPRD10.PROD.OUTLOOK.COM \
--to=bernd.edlinger@hotmail.de \
--cc=gcc-patches@gcc.gnu.org \
--cc=hjl.tools@gmail.com \
--cc=richard.guenther@gmail.com \
--cc=richard.sandiford@arm.com \
--cc=ubizjak@gmail.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).