Avoid picture primitive copies via VecHelper #3362

kvark · 2018-11-27T20:41:58Z

This is a successor of #3360 that avoids the borrow checker dance via RAII
Addresses part of #3358

This change is

jrmuizel · 2018-11-27T20:46:28Z

webrender/src/util.rs

 // Matches the definition of SK_ScalarNearlyZero in Skia.
 const NEARLY_ZERO: f32 = 1.0 / 4096.0;

+/// A typesafe helper that separates new value construction from
+/// vector growing, which allows LLVM to elide the value copy.


"and ideally construct the element in place"

jrmuizel · 2018-11-27T20:48:38Z

webrender/src/util.rs

+
+impl<'a, T> Allocation<'a, T> {
+    #[inline(always)]
+    pub fn init(self, value: T) -> usize {


Maybe add:
// writing is safe because alloc() ensured enough capacity and Allocation holds a mutable borrow to prevent anyone else from breaking this invariant.

Are you sure this actually avoids the memcpy / memmove? At least for self arguments a bit back we couldn't work around it like this, see rust-lang/rust#42763.

Confirmed in playground. In fn foo() changing the push(xxx) to alloc().init(xxx) removes the memcpy.

I filed rust-lang/rust#56333 for a problem contributing to the first example not working well in the playground.

It looks like SmallVec also causes this to not work. I filed rust-lang/rust#56356 about that.

jrmuizel · 2018-11-27T20:50:50Z

webrender/src/frame_builder.rs

@@ -338,7 +338,7 @@ impl FrameBuilder {
        let mut profile_counters = FrameProfileCounters::new();
        profile_counters
            .total_primitives
-            .set(self.prim_store.prim_count);
+            .set(self.prim_store.prim_count());


It'd be nice if this only ran if profiling was on.

kvark · 2018-11-28T03:52:59Z

I'm not able to use @jrmuizel 's tool that find memcopies, since IR validation fails on webrender crate. However, I tried to reproduce this in playground, and in both cases (vec.push(x) and vec.alloc().init(x)) the compiler created the data on the stack first and then issued a single memcpy, so it doesn't appear to be fixing the copy.

Is my playground experiment incorrect? Otherwise, what assumptions are wrong?

bors-servo · 2018-11-28T03:57:56Z

☔ The latest upstream changes (presumably #3359) made this pull request unmergeable. Please resolve the merge conflicts.

jrmuizel · 2018-11-28T15:20:19Z

If you use an enum containing an array instead of an array and initialize with a small enum variant the copy is elided in your example

kvark · 2018-11-28T15:38:07Z

@jrmuizel indeed, I confirmed it with the playground experiment.
Therefore, we should proceed with this PR.

kvark · 2018-11-28T15:41:33Z

@bors-servo r=jrmuizel

bors-servo · 2018-11-28T15:41:34Z

📌 Commit d680e32 has been approved by jrmuizel

bors-servo · 2018-11-28T15:52:51Z

⌛ Testing commit d680e32 with merge be65949...

Avoid picture primitive copies via VecHelper This is a successor of #3360 that avoids the borrow checker dance via RAII Addresses part of #3358  --- This change is [<img src="https://reviewable.io/review_button.svg" height="34" align="absmiddle" alt="Reviewable"/>](https://reviewable.io/reviews/servo/webrender/3362)

bors-servo · 2018-11-28T16:45:42Z

☀️ Test successful - status-appveyor, status-taskcluster
Approved by: jrmuizel
Pushing be65949 to master...

jrmuizel · 2018-11-28T18:12:03Z

It looks like this may not have worked:

memcpy of 1592 in _ZN9webrender22display_list_flattener20DisplayListFlattener20pop_stacking_context17hc7e8ce6ec929e9eaE @
  write<webrender::picture::PicturePrimitive> /rustc/6f93e93af6f823948cc13d2938957757c6486d88/src/libcore/ptr.rs:739
  init<webrender::picture::PicturePrimitive> webrender/src/util.rs:33
  pop_stacking_context webrender/src/display_list_flattener.rs:1220
memcpy of 1592 in _ZN9webrender22display_list_flattener20DisplayListFlattener20pop_stacking_context17hc7e8ce6ec929e9eaE @
  write<webrender::picture::PicturePrimitive> /rustc/6f93e93af6f823948cc13d2938957757c6486d88/src/libcore/ptr.rs:739
  init<webrender::picture::PicturePrimitive> webrender/src/util.rs:33
  pop_stacking_context webrender/src/display_list_flattener.rs:1187
memcpy of 1592 in _ZN9webrender22display_list_flattener20DisplayListFlattener21push_stacking_context17h500d71065754d89aE @
  write<webrender::picture::PicturePrimitive> /rustc/6f93e93af6f823948cc13d2938957757c6486d88/src/libcore/ptr.rs:739
  init<webrender::picture::PicturePrimitive> webrender/src/util.rs:33
  cut_flat_item_sequence webrender/src/display_list_flattener.rs:2207
  push_stacking_context webrender/src/display_list_flattener.rs:994
memcpy of 1592 in _ZN9webrender22display_list_flattener20DisplayListFlattener20pop_stacking_context17hc7e8ce6ec929e9eaE @
  write<webrender::picture::PicturePrimitive> /rustc/6f93e93af6f823948cc13d2938957757c6486d88/src/libcore/ptr.rs:739
  init<webrender::picture::PicturePrimitive> webrender/src/util.rs:33
  pop_stacking_context webrender/src/display_list_flattener.rs:1158
memcpy of 1592 in _ZN9webrender22display_list_flattener20DisplayListFlattener12flatten_item17hfd2b2318f8c75547E @
  write<webrender::picture::PicturePrimitive> /rustc/6f93e93af6f823948cc13d2938957757c6486d88/src/libcore/ptr.rs:739
  init<webrender::picture::PicturePrimitive> webrender/src/util.rs:33
  pop_all_shadows webrender/src/display_list_flattener.rs:1564
  flatten_item webrender/src/display_list_flattener.rs:789
memcpy of 1592 in _ZN9webrender22display_list_flattener20DisplayListFlattener20pop_stacking_context17hc7e8ce6ec929e9eaE @
  write<webrender::picture::PicturePrimitive> /rustc/6f93e93af6f823948cc13d2938957757c6486d88/src/libcore/ptr.rs:739
  init<webrender::picture::PicturePrimitive> webrender/src/util.rs:33
  pop_stacking_context webrender/src/display_list_flattener.rs:1117

I wonder if llvm gets sad the enum variant is too big.

jrmuizel · 2018-11-29T04:09:47Z

It looks like SmallVec is contributing to the sadness:

#![crate_type = "lib"]

extern crate smallvec;

use smallvec::SmallVec;

#[derive(Default)]
pub struct L {
    a: SmallVec<[f64; 16]>,
    b: SmallVec<[f64; 16]>,
    c: SmallVec<[f64; 16]>
}

pub struct Allocation<T> {
    f: *mut T,
}
use std::ptr;

impl<T> Allocation<T> {
    pub fn init(self, value: T) {
        unsafe { ptr::write(self.f, value) };
    }
}

#[inline(never)]
pub fn bar(a: Allocation<L>) {
    a.init(L{a: SmallVec::new(), b: SmallVec::new(), c: SmallVec::new()});
}

compiles to

playground::bar:
	pushq	%rbx
	subq	$720, %rsp
	movq	%rdi, %rbx
	xorps	%xmm0, %xmm0
	movaps	%xmm0, 288(%rsp)
	movaps	%xmm0, (%rsp)
	movaps	%xmm0, 144(%rsp)
	leaq	432(%rsp), %rdi
	movq	%rsp, %rsi
	movl	$144, %edx
	callq	memcpy@PLT
	leaq	576(%rsp), %rdi
	leaq	144(%rsp), %rsi
	movl	$144, %edx
	callq	memcpy@PLT
	leaq	288(%rsp), %rsi
	movl	$432, %edx
	movq	%rbx, %rdi
	callq	memcpy@PLT
	addq	$720, %rsp
	popq	%rbx
	retq

…4c9170a70ef7 (WR PR #3362). r=kats servo/webrender#3362 Differential Revision: https://phabricator.services.mozilla.com/D13355 --HG-- extra : moz-landing-system : lando

…4c9170a70ef7 (WR PR #3362). r=kats servo/webrender#3362 Differential Revision: https://phabricator.services.mozilla.com/D13355

@jrmuizel

Reduce data copies during internation Excessive copying in `fn intern()` is something we noticed yesterday with @jrmuizel . This PR attempts to improve them in two ways (each in a separate commit): 1. reduce the `Update` enum size by moving the data out 2. adopt entry-like API to accommodate the common pattern of `if index == v.len() { v.push(xxx) } else { v[index] = xxx; }` without panics between element construction and actual assignment. It builds upon the `VecHelper` work of #3362.  --- This change is [<img src="https://reviewable.io/review_button.svg" height="34" align="absmiddle" alt="Reviewable"/>](https://reviewable.io/reviews/servo/webrender/3366)

jrmuizel · 2019-02-02T18:34:14Z

The smallvec example now generates better code with Nightly:

playground::bar: # @playground::bar
# %bb.0:
	subq	$440, %rsp              # imm = 0x1B8
	xorps	%xmm0, %xmm0
	movaps	%xmm0, (%rsp)
	movaps	%xmm0, 144(%rsp)
	movaps	%xmm0, 288(%rsp)
	movq	%rsp, %rsi
	movl	$432, %edx              # imm = 0x1B0
	callq	*memcpy@GOTPCREL(%rip)
	addq	$440, %rsp              # imm = 0x1B8
	retq

However it looks like there's still an extra memcpy

jrmuizel · 2019-02-02T20:03:34Z

I filed a follow up at rust-lang/rust#58082

…4c9170a70ef7 (WR PR #3362). r=kats servo/webrender#3362 Differential Revision: https://phabricator.services.mozilla.com/D13355 UltraBlame original commit: 54d56b9a77fb9d41eff3ce1662ba5fff538365d2

kvark requested a review from jrmuizel November 27, 2018 20:41

kvark mentioned this pull request Nov 27, 2018

[WIP] Avoid copies of large picture primitives during flattening #3360

Closed

3 tasks

jrmuizel approved these changes Nov 27, 2018

View reviewed changes

kvark force-pushed the picture-copies2 branch from 83b830f to 18455ee Compare November 28, 2018 02:42

kvark changed the title ~~Avoid picture primitive copies via VecHelper~~ [WIP] Avoid picture primitive copies via VecHelper Nov 28, 2018

Avoid picture primitive copies via VecHelper

d680e32

kvark force-pushed the picture-copies2 branch from 18455ee to d680e32 Compare November 28, 2018 15:41

kvark changed the title ~~[WIP] Avoid picture primitive copies via VecHelper~~ Avoid picture primitive copies via VecHelper Nov 28, 2018

bors-servo merged commit d680e32 into servo:master Nov 28, 2018

kvark deleted the picture-copies2 branch November 28, 2018 17:31

kvark mentioned this pull request Nov 28, 2018

Reduce data copies during internation #3366

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Avoid picture primitive copies via VecHelper #3362

Avoid picture primitive copies via VecHelper #3362

kvark commented Nov 27, 2018 •

edited by larsbergstrom

Loading

jrmuizel Nov 27, 2018

jrmuizel Nov 27, 2018

emilio Nov 28, 2018

kvark Nov 28, 2018

jrmuizel Nov 29, 2018

jrmuizel Nov 29, 2018

jrmuizel Nov 27, 2018

kvark commented Nov 28, 2018

bors-servo commented Nov 28, 2018

jrmuizel commented Nov 28, 2018 via email

kvark commented Nov 28, 2018

kvark commented Nov 28, 2018

bors-servo commented Nov 28, 2018

bors-servo commented Nov 28, 2018

bors-servo commented Nov 28, 2018

jrmuizel commented Nov 28, 2018

jrmuizel commented Nov 29, 2018

jrmuizel commented Feb 2, 2019

jrmuizel commented Feb 2, 2019

Avoid picture primitive copies via VecHelper #3362

Avoid picture primitive copies via VecHelper #3362

Conversation

kvark commented Nov 27, 2018 • edited by larsbergstrom Loading

jrmuizel Nov 27, 2018

Choose a reason for hiding this comment

jrmuizel Nov 27, 2018

Choose a reason for hiding this comment

emilio Nov 28, 2018

Choose a reason for hiding this comment

kvark Nov 28, 2018

Choose a reason for hiding this comment

jrmuizel Nov 29, 2018

Choose a reason for hiding this comment

jrmuizel Nov 29, 2018

Choose a reason for hiding this comment

jrmuizel Nov 27, 2018

Choose a reason for hiding this comment

kvark commented Nov 28, 2018

bors-servo commented Nov 28, 2018

jrmuizel commented Nov 28, 2018 via email

kvark commented Nov 28, 2018

kvark commented Nov 28, 2018

bors-servo commented Nov 28, 2018

bors-servo commented Nov 28, 2018

bors-servo commented Nov 28, 2018

jrmuizel commented Nov 28, 2018

jrmuizel commented Nov 29, 2018

jrmuizel commented Feb 2, 2019

jrmuizel commented Feb 2, 2019

kvark commented Nov 27, 2018 •

edited by larsbergstrom

Loading