Changeset 50595 in webkit for trunk/JavaScriptCore/jit/JITArithmetic.cpp

Timestamp:

Nov 6, 2009, 1:35:03 AM (16 years ago)

Author:

[email protected]

Message:

Tidy up the shift methods on the macro-assembler interface.

Patch by Gavin Barraclough <​[email protected]> on 2009-11-06
Reviewed by Oliver Hunt.

Currently behaviour of shifts of a magnitude > 0x1f is undefined.
Instead defined that all shifts are masked to this range. This makes a lot of
practical sense, both since having undefined behaviour is not particularly
desirable, and because this behaviour is commonly required (particularly since
it is required bt ECMA-262 for shifts).

Update the ARM assemblers to provide this behaviour. Remove (now) redundant
masks from JITArithmetic, and remove rshiftPtr (this was used in case that
could be rewritten in a simpler form using rshift32, only optimized JSVALUE32
on x86-64, which uses JSVALUE64!)

• assembler/MacroAssembler.h:
• assembler/MacroAssemblerARM.h:

(JSC::MacroAssemblerARM::lshift32):
(JSC::MacroAssemblerARM::rshift32):

• assembler/MacroAssemblerARMv7.h:

(JSC::MacroAssemblerARMv7::lshift32):
(JSC::MacroAssemblerARMv7::rshift32):

• assembler/MacroAssemblerX86_64.h:
• jit/JITArithmetic.cpp:

(JSC::JIT::emit_op_lshift):
(JSC::JIT::emit_op_rshift):

File:

• trunk/JavaScriptCore/jit/JITArithmetic.cpp (modified) (2 diffs)

trunk/JavaScriptCore/jit/JITArithmetic.cpp

@@ -1144 +1144 @@
     emitFastArithImmToInt(regT0);
     emitFastArithImmToInt(regT2);
-#if !PLATFORM(X86)
-    // Mask with 0x1f as per ecma-262 11.7.2 step 7.
-    // On 32-bit x86 this is not necessary, since the shift anount is implicitly masked in the instruction.
-    and32(Imm32(0x1f), regT2);
-#endif
     lshift32(regT2, regT0);
-#if !USE(JSVALUE64)
+#if USE(JSVALUE32)
     addSlowCase(branchAdd32(Overflow, regT0, regT0));
     signExtend32ToPtr(regT0, regT0);
@@ -1226 +1221 @@
         }
         emitFastArithImmToInt(regT2);
-#if !PLATFORM(X86)
-        // Mask with 0x1f as per ecma-262 11.7.2 step 7.
-        // On 32-bit x86 this is not necessary, since the shift anount is implicitly masked in the instruction.
-        and32(Imm32(0x1f), regT2);
-#endif
-#if USE(JSVALUE64)
         rshift32(regT2, regT0);
-#else
-        rshiftPtr(regT2, regT0);
+#if USE(JSVALUE32)
+        signExtend32ToPtr(regT0, regT0);
 #endif
     }