mirror/bitcoin
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge #20447: depends: Patch qt_intersect_spans to avoid non-deterministic behavior in LLVM 8

Browse Source 
8f7d1b39ef Fix QPainter non-determinism on macOS (Andrew Chow)

Pull request description:

  Aplies a patch to Qt that fixes the non-determinism by modifying Qt. The source of the non-determinism is how LLVM 8 optimizes qt_intersect_spans when compiling. The particular optimization that seems to be causing the problems is that a temp variable is being added for spans->y. For some reason, when it does this, it chooses different instructions to use when making that variable. We bypass this problem by patching qt_intersect_spans to always make and use this local variable.

  Potential alternative to #20436 and #20440

ACKs for top commit:
  hebasto:
    re-ACK 8f7d1b39ef ~for merging into the 0.21 branch, but [not into the master](https://github.com/bitcoin/bitcoin/pull/20454) branch.~
  fanquake:
    ACK 8f7d1b39ef

Tree-SHA512: b0d00a77643554021736524fb64611462ef2ec849a220543c12d99edb0f52f2e8128d2cc61fa82176b7e13b294574774a92d6b649badf8b7630c6d6a7e70ce10
This commit is contained in:
fanquake committed 2020-11-24 21:12:02 +08:00
2 parents 9402159b9b 8f7d1b39ef
commit 31c9987976
2 files changed
+65 -1
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

depends/patches/qt/fix_qpainter_non_determinism.patch
+63
View File
@@ -0,0 +1,63 @@
commit 2a8f7dc6ddfc414a66491522501c1574a1343ee1
Author: Andrew Chow <achow101-github@achow101.com>
Date: Sat Nov 21 01:11:04 2020 -0500
build: Fix determinism issue when building with Clang 8
When building Qt with LLVM/Clang 8 under -O3 (the default), we run into
a determinism issue in `qt_interset_spans`. The issue has been fixed for
LLVM/Clang 9, see
https://github.com/llvm/llvm-project/commit/db101864bdc938deb1d63fe4f7da761bd38e5cae
and https://reviews.llvm.org/D64601, however this fix was not backported
to 8.x. Once LLVM/Clang 9 is used, this patch can be dropped.
The particular issue appears to be an optimization done by -O3 which
adds a temporary variable for `spans->y` in `qt_intersect_spans`. When
it does this, sometimes it chooses to use a 32-bit movs instruction
(movswl), and other times it chooses a 64-bit movs instruction (movswq).
By patching `qt_intersect_spans` to always make a temporary variable for
`spans->y`, we are able to sidestep this problem.
diff --git a/qtbase/src/gui/painting/qpaintengine_raster.cpp b/qtbase/src/gui/painting/qpaintengine_raster.cpp
index 92ab6e8375..f018009e0b 100644
--- a/qtbase/src/gui/painting/qpaintengine_raster.cpp
+++ b/qtbase/src/gui/painting/qpaintengine_raster.cpp
@@ -3971,22 +3971,23 @@ static const QSpan *qt_intersect_spans(const QClipData *clip, int *currentClip,
const QSpan *clipEnd = clip->m_spans + clip->count;
while (available && spans < end ) {
+ const short spans_y = spans->y;
if (clipSpans >= clipEnd) {
spans = end;
break;
}
- if (clipSpans->y > spans->y) {
+ if (clipSpans->y > spans_y) {
++spans;
continue;
}
- if (spans->y != clipSpans->y) {
- if (spans->y < clip->count && clip->m_clipLines[spans->y].spans)
- clipSpans = clip->m_clipLines[spans->y].spans;
+ if (spans_y != clipSpans->y) {
+ if (spans_y < clip->count && clip->m_clipLines[spans_y].spans)
+ clipSpans = clip->m_clipLines[spans_y].spans;
else
++clipSpans;
continue;
}
- Q_ASSERT(spans->y == clipSpans->y);
+ Q_ASSERT(spans_y == clipSpans->y);
int sx1 = spans->x;
int sx2 = sx1 + spans->len;
@@ -4005,7 +4006,7 @@ static const QSpan *qt_intersect_spans(const QClipData *clip, int *currentClip,
if (len) {
out->x = qMax(sx1, cx1);
out->len = qMin(sx2, cx2) - out->x;
- out->y = spans->y;
+ out->y = spans_y;
out->coverage = qt_div_255(spans->coverage * clipSpans->coverage);
++out;
--available;