sw/source/core/text/itrform2.cxx | 11 ++++++++---
sw/source/core/text/portxt.cxx | 6 ++++++
vcl/source/gdi/CommonSalLayout.cxx | 9 ++++++---
vcl/source/gdi/impglyphitem.cxx | 9 ++++++++-
4 files changed, 28 insertions(+), 7 deletions(-)
New commits:
commit b0908a76d02e7babf23c4287f57f3d6e368e26e8
Author: Jonathan Clark <jonat...@libreoffice.org>
AuthorDate: Tue Sep 3 05:49:28 2024 -0600
Commit: Adolfo Jayme Barrientos <fit...@ubuntu.com>
CommitDate: Wed Sep 4 18:16:49 2024 +0200
tdf#92064 sw: Improve large paragraph layout performance
This change includes the following scalability improvements for
documents containing extremely large paragraphs:
- Reduces the size of layout contexts to account for LF control chars.
- Due to typical access patterns while laying out paragraphs, VCL was
making O(n^2) calls to vcl::ScriptRun::next(). VCL now uses an
existing global LRU cache for script runs, avoiding much of this
overhead.
Change-Id: Iee03938683c95776a817d4819fe9a43c65a7c3fc
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/172801
Tested-by: Jenkins
Reviewed-by: Jonathan Clark <jonat...@libreoffice.org>
Signed-off-by: Xisco Fauli <xiscofa...@libreoffice.org>
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/172815
Reviewed-by: Ilmari Lauhakangas <ilmari.lauhakan...@libreoffice.org>
Reviewed-by: Adolfo Jayme Barrientos <fit...@ubuntu.com>
diff --git a/sw/source/core/text/itrform2.cxx b/sw/source/core/text/itrform2.cxx
index 3673b46ba247..788d64fc538f 100644
--- a/sw/source/core/text/itrform2.cxx
+++ b/sw/source/core/text/itrform2.cxx
@@ -1460,6 +1460,7 @@ SwTextPortion *SwTextFormatter::NewTextPortion(
SwTextFormatInfo &rInf )
case CH_TXTATR_BREAKWORD:
case CH_TXTATR_INWORD:
case CH_TXTATR_TAB:
+ case CH_TXTATR_NEWLINE:
case CH_TXT_ATR_INPUTFIELDSTART:
case CH_TXT_ATR_INPUTFIELDEND:
case CH_TXT_ATR_FORMELEMENT:
diff --git a/vcl/source/gdi/CommonSalLayout.cxx
b/vcl/source/gdi/CommonSalLayout.cxx
index 0e4e3fe9dc48..115870dc9cf6 100644
--- a/vcl/source/gdi/CommonSalLayout.cxx
+++ b/vcl/source/gdi/CommonSalLayout.cxx
@@ -384,7 +384,7 @@ bool
GenericSalLayout::LayoutText(vcl::text::ImplLayoutArgs& rArgs, const SalLay
const int nLength = rArgs.mrStr.getLength();
const sal_Unicode *pStr = rArgs.mrStr.getStr();
- std::optional<vcl::text::TextLayoutCache> oNewScriptRun;
+ std::shared_ptr<const vcl::text::TextLayoutCache> pNewScriptRun;
vcl::text::TextLayoutCache const* pTextLayout;
if (rArgs.m_pTextLayoutCache)
{
@@ -392,8 +392,11 @@ bool
GenericSalLayout::LayoutText(vcl::text::ImplLayoutArgs& rArgs, const SalLay
}
else
{
- oNewScriptRun.emplace(pStr, rArgs.mnEndCharPos);
- pTextLayout = &*oNewScriptRun;
+ // tdf#92064, tdf#162663:
+ // Also use the global LRU cache for full string script runs.
+ // This obviates O(n^2) calls to vcl::ScriptRun::next() when laying
out large paragraphs.
+ pNewScriptRun = vcl::text::TextLayoutCache::Create(rArgs.mrStr);
+ pTextLayout = pNewScriptRun.get();
}
// nBaseOffset is used to align vertical text to the center of rotated
commit ef40759390de4eba93d0a1e9369fc8ba5c1ea534
Author: Jonathan Clark <jonat...@libreoffice.org>
AuthorDate: Fri Jul 12 13:40:34 2024 -0600
Commit: Adolfo Jayme Barrientos <fit...@ubuntu.com>
CommitDate: Wed Sep 4 18:16:38 2024 +0200
tdf#92064 sw: Improve Tibetan layout performance
This change includes the following scalability improvements for
documents containing extremely long paragraphs:
- Reduces the size of layout contexts to account for line breaks.
- Disables a misbehaving glyph cache performance optimization for long
strings.
Change-Id: Ie9a3365076c0d112a7a655988d672a9f4609b42b
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/170420
Tested-by: Jenkins
Reviewed-by: Jonathan Clark <jonat...@libreoffice.org>
Signed-off-by: Xisco Fauli <xiscofa...@libreoffice.org>
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/172858
Reviewed-by: Adolfo Jayme Barrientos <fit...@ubuntu.com>
Reviewed-by: Ilmari Lauhakangas <ilmari.lauhakan...@libreoffice.org>
diff --git a/sw/source/core/text/itrform2.cxx b/sw/source/core/text/itrform2.cxx
index 71b78bc0914d..3673b46ba247 100644
--- a/sw/source/core/text/itrform2.cxx
+++ b/sw/source/core/text/itrform2.cxx
@@ -1418,8 +1418,11 @@ SwTextPortion *SwTextFormatter::NewTextPortion(
SwTextFormatInfo &rInf )
if (!nCharWidthGuess)
nCharWidthGuess = 1;
auto nExpect = rInf.GetIdx() + TextFrameIndex(rInf.GetLineWidth() /
nCharWidthGuess);
- if (nExpect > rInf.GetIdx() && nNextChg > nExpect)
- nNextChg = nExpect;
+ if (nExpect > rInf.GetIdx())
+ {
+ nNextChg = std::min(nNextChg, nExpect);
+ nNextContext = std::min(nNextContext, nExpect);
+ }
// we keep an invariant during method calls:
// there are no portion ending characters like hard spaces
@@ -1444,7 +1447,8 @@ SwTextPortion *SwTextFormatter::NewTextPortion(
SwTextFormatInfo &rInf )
// for the first text portion in a paragraph, or for any successive
// portions that are outside of the bounds of the previous context.
if (!rInf.GetLayoutContext().has_value()
- || rInf.GetLayoutContext()->m_nEnd <= rInf.GetIdx().get())
+ || rInf.GetLayoutContext()->m_nBegin < rInf.GetLineStart().get()
+ || rInf.GetLayoutContext()->m_nEnd < nNextChg.get())
{
// The layout context must terminate at special characters
sal_Int32 nEnd = rInf.GetIdx().get();
diff --git a/sw/source/core/text/portxt.cxx b/sw/source/core/text/portxt.cxx
index ee39f156c24a..ec0303c0665c 100644
--- a/sw/source/core/text/portxt.cxx
+++ b/sw/source/core/text/portxt.cxx
@@ -445,6 +445,12 @@ bool SwTextPortion::Format_( SwTextFormatInfo &rInf )
SetLen( pGuess->BreakPos() - rInf.GetIdx() );
+ // Clamp layout context to the end of the line
+ if(auto stClampedContext = GetLayoutContext();
stClampedContext.has_value()) {
+ stClampedContext->m_nEnd = pGuess->BreakPos().get();
+ SetLayoutContext(stClampedContext);
+ }
+
OSL_ENSURE( pGuess->BreakStart() >= pGuess->FieldDiff(),
"Trouble with expanded field portions during line break" );
TextFrameIndex const nRealStart = pGuess->BreakStart() -
pGuess->FieldDiff();
diff --git a/vcl/source/gdi/impglyphitem.cxx b/vcl/source/gdi/impglyphitem.cxx
index 1b40291736b7..946230a51762 100644
--- a/vcl/source/gdi/impglyphitem.cxx
+++ b/vcl/source/gdi/impglyphitem.cxx
@@ -364,6 +364,13 @@ const SalLayoutGlyphs*
SalLayoutGlyphsCache::GetLayoutGlyphs(
// part being underlined. Doing this for any two segments allows
this optimization
// even when the prefix of the string would use a different font.
// TODO: Can those font differences be ignored?
+
+ // Shaping performance seems to scale poorly with respect to
string length. Certain
+ // writing systems involve extremely long strings (for example,
Tibetan: tdf#92064).
+ // In such cases, this optimization would be a net loss, and must
be disabled.
+ constexpr sal_Int32 nOptLengthThreshold = 20000;
+ bool bEnableOptimization = (text.getLength() <
nOptLengthThreshold);
+
// Writer layouts tests enable SAL_NON_APPLICATION_FONT_USE=abort
in order
// to make PrintFontManager::Substitute() abort if font fallback
happens. When
// laying out the entire string the chance this happens increases
(e.g. testAbi11870
@@ -374,7 +381,7 @@ const SalLayoutGlyphs*
SalLayoutGlyphsCache::GetLayoutGlyphs(
const char* pEnv = getenv("SAL_NON_APPLICATION_FONT_USE");
return pEnv && strcmp(pEnv, "abort") == 0;
}();
- if (mLastSubstringKey.has_value() && !bAbortOnFontSubstitute)
+ if (bEnableOptimization && mLastSubstringKey.has_value() &&
!bAbortOnFontSubstitute)
{
sal_Int32 pos = nIndex;
if (mLastSubstringKey->len < pos && text[pos - 1] == nbSpace)
