Title: [195939] trunk/Source/_javascript_Core
- Revision
- 195939
- Author
- [email protected]
- Date
- 2016-01-31 16:07:06 -0800 (Sun, 31 Jan 2016)
Log Message
B3->Air lowering should use MoveFloat more
https://bugs.webkit.org/show_bug.cgi?id=153714
Reviewed by Sam Weinig.
This is a very minor and benign bug. It just means that we will use the more canonical
MoveFloat instruction when moving floats, rather than using MoveDouble.
* b3/B3LowerToAir.cpp:
(JSC::B3::Air::LowerToAir::relaxedMoveForType):
Modified Paths
Diff
Modified: trunk/Source/_javascript_Core/ChangeLog (195938 => 195939)
--- trunk/Source/_javascript_Core/ChangeLog 2016-01-31 23:05:10 UTC (rev 195938)
+++ trunk/Source/_javascript_Core/ChangeLog 2016-02-01 00:07:06 UTC (rev 195939)
@@ -1,3 +1,16 @@
+2016-01-30 Filip Pizlo <[email protected]>
+
+ B3->Air lowering should use MoveFloat more
+ https://bugs.webkit.org/show_bug.cgi?id=153714
+
+ Reviewed by Sam Weinig.
+
+ This is a very minor and benign bug. It just means that we will use the more canonical
+ MoveFloat instruction when moving floats, rather than using MoveDouble.
+
+ * b3/B3LowerToAir.cpp:
+ (JSC::B3::Air::LowerToAir::relaxedMoveForType):
+
2016-01-31 Yusuke Suzuki <[email protected]>
Should not predict OtherObj for ToThis with primitive types under strict mode
Modified: trunk/Source/_javascript_Core/b3/B3LowerToAir.cpp (195938 => 195939)
--- trunk/Source/_javascript_Core/b3/B3LowerToAir.cpp 2016-01-31 23:05:10 UTC (rev 195938)
+++ trunk/Source/_javascript_Core/b3/B3LowerToAir.cpp 2016-02-01 00:07:06 UTC (rev 195939)
@@ -855,8 +855,25 @@
switch (type) {
case Int32:
case Int64:
+ // For Int32, we could return Move or Move32. It's a trade-off.
+ //
+ // Move32: Using Move32 guarantees that we use the narrower move, but in cases where the
+ // register allocator can't prove that the variables involved are 32-bit, this will
+ // disable coalescing.
+ //
+ // Move: Using Move guarantees that the register allocator can coalesce normally, but in
+ // cases where it can't prove that the variables are 32-bit and it doesn't coalesce,
+ // this will force us to use a full 64-bit Move instead of the slightly cheaper
+ // 32-bit Move32.
+ //
+ // Coalescing is a lot more profitable than turning Move into Move32. So, it's better to
+ // use Move here because in cases where the register allocator cannot prove that
+ // everything is 32-bit, we still get coalescing.
return Move;
case Float:
+ // MoveFloat is always coalescable and we never convert MoveDouble to MoveFloat, so we
+ // should use MoveFloat when we know that the temporaries involved are 32-bit.
+ return MoveFloat;
case Double:
return MoveDouble;
case Void:
_______________________________________________
webkit-changes mailing list
[email protected]
https://lists.webkit.org/mailman/listinfo/webkit-changes
