Changeset 209725 in webkit for trunk/Source/JavaScriptCore/dfg/DFGInPlaceAbstractState.cpp

Timestamp:

Dec 12, 2016, 1:46:45 PM (9 years ago)

Author:

[email protected]

Message:

REGRESSION(r209653): speedometer crashes making virtual slow path tailcalls
​https://bugs.webkit.org/show_bug.cgi?id=165748

Reviewed by Filip Pizlo.

JSTests:

New regression test.

• stress/regress-165748.js: Added.

(sum1):
(sum2):
(sum3):
(sum4):
(sum5):
(sum6):
(tailCaller):
(test):

Source/JavaScriptCore:

The virtual slow path for tailcalls always passes arguments on the stack.
The fix here is to link to the stack argument entrypoint instead of a register
argument entrypoint.

While fixing this bug, I found that we weren't clearing the code origin when
shuffling the call frame for a register argument tailcall.

Also rolling back in r209653, r209654, r209663, and r209673.

• jit/CallFrameShuffler.cpp:

(JSC::CallFrameShuffler::prepareAny):

• jit/ThunkGenerators.cpp:

(JSC::virtualThunkFor):

Source/WTF:

Rolling back in r209653, r209654, r209663, and r209673.

• wtf/Platform.h:

File:

• trunk/Source/JavaScriptCore/dfg/DFGInPlaceAbstractState.cpp (modified) (1 diff)

trunk/Source/JavaScriptCore/dfg/DFGInPlaceAbstractState.cpp

@@ -107 +107 @@
             format = m_graph.m_argumentFormats[i];
         else {
-            Node* node = m_graph.m_arguments[i];
+            Node* node = m_graph.m_argumentsOnStack[i];
             if (!node)
                 format = FlushedJSValue;
             else {
-                ASSERT(node->op() == SetArgument);
+                ASSERT(node->op() == SetArgument || node->op() == SetLocal);
                 format = node->variableAccessData()->flushFormat();
             }