1330187 - DisplayList building shows signs of poor CPU usage

Status: NEW

(Core :: Layout, defect, P3)

Description

[Bas Schouten (:bas.schouten)]

Profiling indicates that DisplayList building is doing a lot of work during which the CPU is not retiring intructions but waiting for the Front-End and Back-End portions of the pipeline. During Scrolling on Google Docs, where display list building and processing takes a lot of the time, on my P50 the full 16 GB/s bandwidth of the DRAM bus is completely filled up, which is causing a lot of back, and we're seeing a large amount of instruction cache misses as well as a high ITLB overhead. This is likely because of accessing a lot of functions which are far apart in memory, as well as accessing a lot of objects in memory in suboptimal order. These seem to be the areas we should focus on when trying to improve display list building and processing performance.

Comment 1

[Bas Schouten (:bas.schouten)]

Here are the top functions in a profile of scrolling on Google Docs, the one with high CPI rates in particular are concerning, I will file individual bugs for things which seem like they might be actionable.

Function / Memory Object / Allocation Stack	Clockticks	Instructions Retired	CPI Rate	Front-End Bound	Bad Speculation	Back-End Bound	Retiring	Module	Function (Full)	Source File	Start Address
memcpy	2,525,600,000	198,800,000	12.704	2.7%	0.0%	92.3%	5.0%	msvcrt.dll	memcpy	[Unknown]	0x10188620
PLDHashTable::SearchTable<0>	1,514,800,000	221,200,000	6.848	9.9%	0.0%	80.3%	11.4%	xul.dll	PLDHashTable::SearchTable<0>(void const *,unsigned int)	pldhashtable.cpp	0x1014e060
MergeCompressedGlyph1Bit<unsigned char>	1,442,000,000	2,755,200,000	0.523	6.4%	8.8%	45.0%	39.9%	dwrite.dll	MergeCompressedGlyph1Bit<unsigned char>(struct GlyphBitmapLayout const &,int,int,unsigned int,unsigned int,unsigned int,unsigned int *)	[Unknown]	0x100b3939
MergeUncompressedGlyph1Bit	1,414,000,000	2,035,600,000	0.695	4.9%	11.4%	50.4%	33.3%	dwrite.dll	MergeUncompressedGlyph1Bit(struct GlyphBitmapLayout const &,int,int,unsigned int,unsigned int,unsigned int,unsigned int *)	[Unknown]	0x1008fdd0
nsRuleNode::GetStyleDisplay<1>	1,030,400,000	126,000,000	8.178	25.7%	0.0%	69.9%	12.3%	xul.dll	nsRuleNode::GetStyleDisplay<1>(class nsStyleContext *)	nsstylestructlist.h	0x106da2db
MergeScanLine	988,400,000	1,542,800,000	0.641	8.1%	0.0%	65.1%	46.5%	dwrite.dll	MergeScanLine(unsigned int * const,unsigned int,unsigned int,unsigned int,unsigned int,unsigned int,unsigned short)	[Unknown]	0x100b4881
nsStyleContext::DoGetStyleDisplay<1>	977,200,000	95,200,000	10.265	20.0%	0.0%	76.5%	3.5%	xul.dll	nsStyleContext::DoGetStyleDisplay<1>(void)	nsstylestructlist.h	0x106da2c3
nsIFrame::BuildDisplayListForChild	954,800,000	257,600,000	3.707	19.3%	4.8%	67.5%	8.4%	xul.dll	nsIFrame::BuildDisplayListForChild(class nsDisplayListBuilder *,class nsIFrame *,struct nsRect const &,class nsDisplayListSet const &,unsigned int)	nsframe.cpp	0x1120608d
nsTArray_Impl<struct nsNameSpaceEntry,struct nsTArrayInfallibleAllocator>::IndexOf<class nsIAtom *,class nsDefaultComparator<struct nsNameSpaceEntry,class nsIAtom *> >	764,400,000	126,000,000	6.067	10.5%	7.5%	77.4%	4.5%	xul.dll	nsTArray_Impl<struct nsNameSpaceEntry,struct nsTArrayInfallibleAllocator>::IndexOf<class nsIAtom *,class nsDefaultComparator<struct nsNameSpaceEntry,class nsIAtom *> >(class nsIAtom * const &,unsigned int,class nsDefaultComparator<struct nsNameSpaceEntry,class nsIAtom *> const &)	nstarray.h	0x10f85505
fastzero_I	618,800,000	84,000,000	7.367	1.9%	0.0%	96.3%	1.9%	msvcrt.dll	fastzero_I	[Unknown]	0x1018a63e
nsIFrame::GetOverflowRect	590,800,000	53,200,000	11.105	17.5%	5.8%	74.7%	1.9%	xul.dll	nsIFrame::GetOverflowRect(enum nsOverflowType)	nsframe.cpp	0x1121472a
CachedBitmapReader<unsigned char>::MoveNextX	588,000,000	1,366,400,000	0.430	15.6%	0.0%	39.4%	45.0%	dwrite.dll	CachedBitmapReader<unsigned char>::MoveNextX(void)	[Unknown]	0x100b4523
nsIFrame::GetOverflowAreasProperty	546,000,000	25,200,000	21.667	19.0%	0.0%	78.9%	2.1%	xul.dll	nsIFrame::GetOverflowAreasProperty(void)	nsframe.cpp	0x112145df
mozilla::PresShell::MarkFramesInSubtreeApproximatelyVisible	506,800,000	131,600,000	3.851	9.1%	6.8%	79.6%	4.5%	xul.dll	mozilla::PresShell::MarkFramesInSubtreeApproximatelyVisible(class nsIFrame *,struct nsRect const &,class mozilla::Maybe<class nsClassHashtable<class nsUint64HashKey,class mozilla::gfx::IntRegionTyped<struct mozilla::CSSPixel> > > &,bool)	presshell.cpp	0x111b7233
mozilla::EffectCompositor::GetAnimationElementAndPseudoForFrame	434,000,000	19,600,000	22.143	42.4%	0.0%	55.0%	2.6%	xul.dll	mozilla::EffectCompositor::GetAnimationElementAndPseudoForFrame(class nsIFrame const *)	effectcompositor.cpp	0x107bd135
CachedBitmapReader<unsigned char>::MoveNextY	434,000,000	1,131,200,000	0.384	10.6%	2.6%	31.1%	55.6%	dwrite.dll	CachedBitmapReader<unsigned char>::MoveNextY(unsigned int)	[Unknown]	0x100b44a9
GlyphRunFiller::FillSubrect	420,000,000	67,200,000	6.250	11.0%	0.0%	83.6%	5.5%	d2d1.dll	GlyphRunFiller::FillSubrect(struct GlyphFillData const &,enum DWRITE_RASTER_TYPE)	[Unknown]	0x1037e14d
nsIFrame::Extend3DContext	389,200,000	42,000,000	9.267	35.5%	3.0%	61.6%	0.0%	xul.dll	nsIFrame::Extend3DContext(void)	nsframe.cpp	0x1120f18a
mozilla::ContainerState::ProcessDisplayItems	364,000,000	182,000,000	2.000	41.1%	22.1%	17.9%	19.0%	xul.dll	mozilla::ContainerState::ProcessDisplayItems(class nsDisplayList *)	framelayerbuilder.cpp	0x113264bf
nsIFrame::Combines3DTransformWithAncestors	347,200,000	11,200,000	31.000	36.4%	0.0%	63.6%	9.9%	xul.dll	nsIFrame::Combines3DTransformWithAncestors(void)	nsframe.cpp	0x11208bc3
RtlLeaveCriticalSection	316,400,000	232,400,000	1.361	32.7%	7.3%	45.5%	14.5%	ntdll.dll	RtlLeaveCriticalSection	[Unknown]	0x4b2bfea0
arena_bin_nonfull_run_get	302,400,000	11,200,000	27.000	7.6%	7.6%	84.8%	0.0%	mozglue.dll	arena_bin_nonfull_run_get	jemalloc.c	0x1000399a
mozilla::FrameLayerBuilder::GetDisplayItemData	302,400,000	5,600,000	54.000	11.4%	0.0%	88.6%	0.0%	xul.dll	mozilla::FrameLayerBuilder::GetDisplayItemData(class nsIFrame *,unsigned int)	framelayerbuilder.cpp	0x113192c6

Comment 2

[Bas Schouten (:bas.schouten)]

Attachment: scrollbar scrolling.csv

More readable and better profile data.