Changeset 204403 in webkit

Timestamp:

Aug 11, 2016 8:38:54 PM (8 years ago)

Author:

mark.lam@apple.com

Message:

OverridesHasInstance should not branch across register allocations.
​https://bugs.webkit.org/show_bug.cgi?id=160792
<rdar://problem/27361778>

Reviewed by Benjamin Poulain.

JSTests:

• stress/OverrideHasInstance-should-not-branch-across-register-allocations.js: Added.

Source/JavaScriptCore:

The OverrideHasInstance node has a branch test that is emitted conditionally.
It also has a bug where it allocated a register after this branch, which is not
allowed and would fail an assertion introduced in ​https://trac.webkit.org/r145931.
From the ChangeLog for r145931:

"This [assertion that register allocations are not branched around] protects
against the case where an allocation could have spilled register contents to free
up a register and that spill only occurs on one path of many through the code.
A subsequent fill of the spilled register may load garbage."

Because the branch isn't always emitted, this bug has gone unnoticed until now.
This patch fixes this issue by pre-allocating the registers before emitting the
branch in OverrideHasInstance.

Note: this issue is only present in DFGSpeculativeJIT64.cpp. The 32-bit version
is doing it right.

• dfg/DFGSpeculativeJIT64.cpp:

(JSC::DFG::SpeculativeJIT::compile):

Location:

trunk

Files:

• JSTests/ChangeLog (modified) (1 diff)
• JSTests/stress/OverrideHasInstance-should-not-branch-across-register-allocations.js (added)
• Source/JavaScriptCore/ChangeLog (modified) (1 diff)
• Source/JavaScriptCore/dfg/DFGSpeculativeJIT64.cpp (modified) (1 diff)

trunk/JSTests/ChangeLog

@@ +1 @@
+2016-08-11  Mark Lam  <mark.lam@apple.com>
+
+        OverridesHasInstance should not branch across register allocations.
+        https://bugs.webkit.org/show_bug.cgi?id=160792
+        <rdar://problem/27361778>
+
+        Reviewed by Benjamin Poulain.
+
+        * stress/OverrideHasInstance-should-not-branch-across-register-allocations.js: Added.
+
 2016-08-11  Mark Lam  <mark.lam@apple.com>
 

trunk/Source/JavaScriptCore/ChangeLog

@@ +1 @@
+2016-08-11  Mark Lam  <mark.lam@apple.com>
+
+        OverridesHasInstance should not branch across register allocations.
+        https://bugs.webkit.org/show_bug.cgi?id=160792
+        <rdar://problem/27361778>
+
+        Reviewed by Benjamin Poulain.
+
+        The OverrideHasInstance node has a branch test that is emitted conditionally.
+        It also has a bug where it allocated a register after this branch, which is not
+        allowed and would fail an assertion introduced in https://trac.webkit.org/r145931.
+        From the ChangeLog for r145931:
+
+        "This [assertion that register allocations are not branched around] protects
+        against the case where an allocation could have spilled register contents to free
+        up a register and that spill only occurs on one path of many through the code.
+        A subsequent fill of the spilled register may load garbage."
+
+        Because the branch isn't always emitted, this bug has gone unnoticed until now.
+        This patch fixes this issue by pre-allocating the registers before emitting the
+        branch in OverrideHasInstance.
+
+        Note: this issue is only present in DFGSpeculativeJIT64.cpp.  The 32-bit version
+        is doing it right.
+
+        * dfg/DFGSpeculativeJIT64.cpp:
+        (JSC::DFG::SpeculativeJIT::compile):
+
 2016-08-11  Benjamin Poulain  <bpoulain@apple.com>
 

trunk/Source/JavaScriptCore/dfg/DFGSpeculativeJIT64.cpp

@@ -4576 +4576 @@
 
         GPRReg resultGPR = result.gpr();
+        GPRReg baseGPR = base.gpr();
 
         // It would be great if constant folding handled automatically the case where we knew the hasInstance function
         // was a constant. Unfortunately, the folding rule for OverridesHasInstance is in the strength reduction phase
         // since it relies on OSR information. https://bugs.webkit.org/show_bug.cgi?id=154832
-        if (!hasInstanceValueNode->isCellConstant() || defaultHasInstanceFunction != hasInstanceValueNode->asCell())
-            notDefault = m_jit.branchPtr(MacroAssembler::NotEqual, hasInstanceValue.gpr(), TrustedImmPtr(defaultHasInstanceFunction));
+        if (!hasInstanceValueNode->isCellConstant() || defaultHasInstanceFunction != hasInstanceValueNode->asCell()) {
+            GPRReg hasInstanceValueGPR = hasInstanceValue.gpr();
+            notDefault = m_jit.branchPtr(MacroAssembler::NotEqual, hasInstanceValueGPR, TrustedImmPtr(defaultHasInstanceFunction));
+        }
 
         // Check that base 'ImplementsDefaultHasInstance'.
-        m_jit.test8(MacroAssembler::Zero, MacroAssembler::Address(base.gpr(), JSCell::typeInfoFlagsOffset()), MacroAssembler::TrustedImm32(ImplementsDefaultHasInstance), resultGPR);
+        m_jit.test8(MacroAssembler::Zero, MacroAssembler::Address(baseGPR, JSCell::typeInfoFlagsOffset()), MacroAssembler::TrustedImm32(ImplementsDefaultHasInstance), resultGPR);
         m_jit.or32(TrustedImm32(ValueFalse), resultGPR);
         MacroAssembler::Jump done = m_jit.jump();