Bug 100899 – JSC: C++ llint 64-bit backend needs to zero extend results of int32 operations

• Status: RESOLVED
• Resolution: FIXED
• Component: JavaScriptCore

[Mark Lam]

The BaseIndex address calculation should use a different type/sized index depending on the type of the instruction that is using it.  This is an x86_64-ism that wasn't previously captured.

[Mark Lam]

Created attachment 171736 [details]
Fix.

[Filip Pizlo]

Comment on attachment 171736 [details]
Fix.

Mark and I found a bug while looking at this.

[Mark Lam]

Also, the index register used in BaseIndex addressing is expected to be of size intptr_t.

[Mark Lam]

Created attachment 171780 [details]
The real fix.

[Filip Pizlo]

Comment on attachment 171780 [details]
The real fix.

Woohoo!

[Mark Lam]

Landed in r133131: <http://trac.webkit.org/changeset/133131>.

The svn commit message for it was erroneous.  It should have said:
=== BEGIN ===
C++ llint 64-bit backend needs to zero extend results of int32 operations.
https://bugs.webkit.org/show_bug.cgi?id=100899.

Reviewed by Filip Pizlo.

llint asm instructions ending in "i" for a 64-bit machine expects the
high 32-bit of registers to be zero'ed out when a 32-bit instruction
writes into a register. Fixed the C++ llint to honor this.

Fixed the index register used in BaseIndex addressing to be of size
intptr_t as expected.

Updated CLoopRegister to handle different endiannesss configurations.

* llint/LowLevelInterpreter.cpp:
  (JSC::CLoopRegister::clearHighWord):
  - new method to clear the high 32-bit of a 64-bit register.
     It's a no-op for the 32-bit build. 
  (CLoopRegister):
  - CLoopRegister now takes care of packing and byte endianness order.
  (JSC::CLoop::execute): - Added an assert.
* offlineasm/cloop.rb:
   - Add calls to clearHighWord() wherever needed.
=== END ===