Add MathPowStub to x64 platform, and fix error in stub on ia32 platform.

src/x64/code-stubs-x64.cc

 // Copyright 2011 the V8 project authors. All rights reserved.
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
 //
 //     * Redistributions of source code must retain the above copyright
 //       notice, this list of conditions and the following disclaimer.
 //     * Redistributions in binary form must reproduce the above
 //       copyright notice, this list of conditions and the following
 //       disclaimer in the documentation and/or other materials provided
@@ skipping 1999 matching lines @@
       }
       __ SmiNeg(rax, rax, &done);
       __ jmp(&slow);  // zero, if not handled above, and Smi::kMinValue.
 
       // Try floating point case.
       __ bind(&try_float);
     } else if (FLAG_debug_code) {
       __ AbortIfSmi(rax);
     }
 
-    __ movq(rdx, FieldOperand(rax, HeapObject::kMapOffset));
-    __ CompareRoot(rdx, Heap::kHeapNumberMapRootIndex);
+    __ CompareRoot(FieldOperand(rax, HeapObject::kMapOffset),
+                   Heap::kHeapNumberMapRootIndex);
     __ j(not_equal, &slow);
     // Operand is a float, negate its value by flipping sign bit.
     __ movq(rdx, FieldOperand(rax, HeapNumber::kValueOffset));
     __ movq(kScratchRegister, Immediate(0x01));
     __ shl(kScratchRegister, Immediate(63));
     __ xor_(rdx, kScratchRegister);  // Flip sign.
     // rdx is value to store.
     if (overwrite_ == UNARY_OVERWRITE) {
       __ movq(FieldOperand(rax, HeapNumber::kValueOffset), rdx);
     } else {
       __ AllocateHeapNumber(rcx, rbx, &slow);
       // rcx: allocated 'empty' number
       __ movq(FieldOperand(rcx, HeapNumber::kValueOffset), rdx);
       __ movq(rax, rcx);
     }
   } else if (op_ == Token::BIT_NOT) {
     if (include_smi_code_) {
       Label try_float;
       __ JumpIfNotSmi(rax, &try_float);
       __ SmiNot(rax, rax);
       __ jmp(&done);
       // Try floating point case.
       __ bind(&try_float);
     } else if (FLAG_debug_code) {
       __ AbortIfSmi(rax);
     }
 
     // Check if the operand is a heap number.
-    __ movq(rdx, FieldOperand(rax, HeapObject::kMapOffset));
-    __ CompareRoot(rdx, Heap::kHeapNumberMapRootIndex);
+    __ CompareRoot(FieldOperand(rax, HeapObject::kMapOffset),
+                   Heap::kHeapNumberMapRootIndex);
     __ j(not_equal, &slow);
 
     // Convert the heap number in rax to an untagged integer in rcx.
     IntegerConvert(masm, rax, rax);
 
     // Do the bitwise operation and smi tag the result.
     __ notl(rax);
     __ Integer32ToSmi(rax, rax);
   }
 
@@ skipping 12 matching lines @@
       break;
     case Token::BIT_NOT:
       __ InvokeBuiltin(Builtins::BIT_NOT, JUMP_FUNCTION);
       break;
     default:
       UNREACHABLE();
   }
 }
 
 
+void MathPowStub::Generate(MacroAssembler* masm) {
+  // Registers are used as follows:
+  // rdx = base
+  // rax = exponent
+  // rcx = temporary, result
+
+  CpuFeatures::Scope use_sse2(SSE2);

[Lasse Reichstein, 2011/02/28 13:21:41]

No need to have SSE2 scopes in X64.

+  Label allocate_return, call_runtime;
+
+  // Load input parameters.
+  __ movq(rdx, Operand(rsp, 2 * kPointerSize));
+  __ movq(rax, Operand(rsp, 1 * kPointerSize));
+
+  // Save 1 in xmm3 - we need this several times later on.
+  __ movl(rcx, Immediate(1));
+  __ cvtlsi2sd(xmm3, rcx);
+
+  Label exponent_nonsmi;
+  Label base_nonsmi;
+  // If the exponent is a heap number go to that specific case.
+  __ JumpIfNotSmi(rax, &exponent_nonsmi);
+  __ JumpIfNotSmi(rdx, &base_nonsmi);
+
+  // Optimized version when both exponent and base are smis.
+  Label powi;
+  __ SmiToInteger32(rdx, rdx);
+  __ cvtlsi2sd(xmm0, rdx);
+  __ jmp(&powi);
+  // exponent is smi and base is a heapnumber.
+  __ bind(&base_nonsmi);
+  __ CompareRoot(FieldOperand(rdx, HeapObject::kMapOffset),
+                 Heap::kHeapNumberMapRootIndex);
+  __ j(not_equal, &call_runtime);
+
+  __ movsd(xmm0, FieldOperand(rdx, HeapNumber::kValueOffset));
+
+  // Optimized version of pow if exponent is a smi.
+  // xmm0 contains the base.
+  __ bind(&powi);
+  __ SmiToInteger32(rax, rax);
+
+  // Save exponent in base as we need to check if exponent is negative later.
+  // We know that base and exponent are in different registers.
+  __ movq(rdx, rax);
+
+  // Get absolute value of exponent.
+  NearLabel no_neg;
+  __ cmpl(rax, Immediate(0));
+  __ j(greater_equal, &no_neg);
+  __ negl(rax);
+  __ bind(&no_neg);
+
+  // Load xmm1 with 1.
+  __ movsd(xmm1, xmm3);
+  NearLabel while_true;
+  NearLabel no_multiply;
+
+  __ bind(&while_true);
+  __ shrl(rax, Immediate(1));
+  __ j(not_carry, &no_multiply);
+  __ mulsd(xmm1, xmm0);
+  __ bind(&no_multiply);
+  __ mulsd(xmm0, xmm0);
+  __ j(not_zero, &while_true);
+
+  // base has the original value of the exponent - if the exponent  is
+  // negative return 1/result.
+  __ testl(rdx, rdx);
+  __ j(positive, &allocate_return);
+  // Special case if xmm1 has reached infinity.
+  __ movl(rcx, Immediate(0x7FB00000));

[Lasse Reichstein, 2011/02/28 13:21:41]

Could you detect infinity by adding the value to itself and see that it doesn't
change? (Need to filter out zero at the beginning then, ofcourse).
Or subtract it from itself and not get zero (but NaN)?

[William Hesse, 2011/02/28 14:36:11]

We now divide first, then compare to 0.


On 2011/02/28 13:21:41, Lasse Reichstein wrote:

+  __ movd(xmm0, rcx);
+  __ cvtss2sd(xmm0, xmm0);
+  __ ucomisd(xmm0, xmm1);
+  __ j(equal, &call_runtime);
+  __ divsd(xmm3, xmm1);
+  __ movsd(xmm1, xmm3);

[Lasse Reichstein, 2011/02/28 13:21:41]

It's worth noticing that if xmm1 is very small, the division might give Infinity
- but that's the correct result for the operation.

[William Hesse, 2011/02/28 14:36:11]

Done.

+  __ jmp(&allocate_return);
+
+  // exponent (or both) is a heapnumber - no matter what we should now work

[Lasse Reichstein, 2011/02/28 13:21:41]

Capitalize exponent.

+  // on doubles.
+  __ bind(&exponent_nonsmi);
+  __ CompareRoot(FieldOperand(rax, HeapObject::kMapOffset),
+                 Heap::kHeapNumberMapRootIndex);
+  __ j(not_equal, &call_runtime);
+  __ movsd(xmm1, FieldOperand(rax, HeapNumber::kValueOffset));
+  // Test if exponent is nan.
+  __ ucomisd(xmm1, xmm1);
+  __ j(parity_even, &call_runtime);
+
+  NearLabel base_not_smi;
+  NearLabel handle_special_cases;
+  __ JumpIfNotSmi(rdx, &base_not_smi);
+  __ SmiToInteger32(rdx, rdx);
+  __ cvtlsi2sd(xmm0, rdx);
+  __ jmp(&handle_special_cases);
+
+  __ bind(&base_not_smi);
+  __ CompareRoot(FieldOperand(rdx, HeapObject::kMapOffset),
+                 Heap::kHeapNumberMapRootIndex);
+  __ j(not_equal, &call_runtime);
+  __ movl(rcx, FieldOperand(rdx, HeapNumber::kExponentOffset));
+  __ andl(rcx, Immediate(HeapNumber::kExponentMask));
+  __ cmpl(rcx, Immediate(HeapNumber::kExponentMask));
+  // base is NaN or +/-Infinity
+  __ j(greater_equal, &call_runtime);

[Lasse Reichstein, 2011/02/28 13:21:41]

If the value isn't -Infinity, won't the result be NaN/Infinity too. We shouldn't
need to go to runtime. 
It's only if it's -Infinity that we need to do something (and I have no idea
what :).

[William Hesse, 2011/02/28 14:36:11]

The value b^infinity depends on whether b is >, =, or < 1, and >, =, or < 0.  It
is complicated.

On 2011/02/28 13:21:41, Lasse Reichstein wrote:

+  __ movsd(xmm0, FieldOperand(rdx, HeapNumber::kValueOffset));
+
+  // base is in xmm0 and exponent is in xmm1.
+  __ bind(&handle_special_cases);
+  NearLabel not_minus_half;
+  // Test for -0.5.
+  // Load xmm2 with -0.5.
+  __ movl(rcx, Immediate(0xBF000000));

[Lasse Reichstein, 2011/02/28 13:21:41]

Is this smaller than loading the exact representation of 1/2 as 64-bit?

[William Hesse, 2011/02/28 14:36:11]

Changed.

On 2011/02/28 13:21:41, Lasse Reichstein wrote:

+  __ movd(xmm2, rcx);
+  __ cvtss2sd(xmm2, xmm2);
+  // xmm2 now has -0.5.
+  __ ucomisd(xmm2, xmm1);
+  __ j(not_equal, &not_minus_half);
+
+  // Calculates reciprocal of square root.
+  // sqrtsd returns -0 when input is -0.  ECMA spec requires +0.
+  __ xorpd(xmm1, xmm1);
+  __ addsd(xmm1, xmm0);
+  __ sqrtsd(xmm1, xmm1);
+  __ divsd(xmm3, xmm1);
+  __ movsd(xmm1, xmm3);
+  __ jmp(&allocate_return);
+
+  // Test for 0.5.
+  __ bind(&not_minus_half);
+  // Load xmm2 with 0.5.
+  // Since xmm3 is 1 and xmm2 is -0.5 this is simply xmm2 + xmm3.
+  __ addsd(xmm2, xmm3);
+  // xmm2 now has 0.5.
+  __ ucomisd(xmm2, xmm1);
+  __ j(not_equal, &call_runtime);
+  // Calculates square root.
+  // sqrtsd returns -0 when input is -0.  ECMA spec requires +0.
+  __ xorpd(xmm1, xmm1);
+  __ addsd(xmm1, xmm0);
+  __ sqrtsd(xmm1, xmm1);
+
+  __ bind(&allocate_return);
+  __ AllocateHeapNumber(rcx, rax, &call_runtime);
+  __ movsd(FieldOperand(rcx, HeapNumber::kValueOffset), xmm1);
+  __ movq(rax, rcx);
+  __ ret(2 * kPointerSize);
+
+  __ bind(&call_runtime);
+  __ TailCallRuntime(Runtime::kMath_pow_cfunction, 2, 1);
+}
+
+
 void ArgumentsAccessStub::GenerateReadElement(MacroAssembler* masm) {
   // The key is in rdx and the parameter count is in rax.
 
   // The displacement is used for skipping the frame pointer on the
   // stack. It is the offset of the last parameter (if any) relative
   // to the frame pointer.
   static const int kDisplacement = 1 * kPointerSize;
 
   // Check that the key is a smi.
   Label slow;
@@ skipping 2841 matching lines @@
           FieldOperand(elements, PixelArray::kExternalPointerOffset));
   __ movb(Operand(external_pointer, untagged_key, times_1, 0), untagged_value);
   __ ret(0);  // Return value in eax.
 }
 
 #undef __
 
 } }  // namespace v8::internal
 
 #endif  // V8_TARGET_ARCH_X64