Implement Math.pow using FPU instructions and inline it in crankshaft (ia32).

src/ia32/code-stubs-ia32.cc

Index: src/ia32/code-stubs-ia32.cc
diff --git a/src/ia32/code-stubs-ia32.cc b/src/ia32/code-stubs-ia32.cc
index 68eebd3a040f239d2ba721117aed017a6b734e0b..e4783165edc9a5aa0d81cf92f72248dc145bb7e3 100644
--- a/src/ia32/code-stubs-ia32.cc
+++ b/src/ia32/code-stubs-ia32.cc
@@ -2938,157 +2938,258 @@ void FloatingPointHelper::CheckFloatOperandsAreInt32(MacroAssembler* masm,
 
 
 void MathPowStub::Generate(MacroAssembler* masm) {
-  // Registers are used as follows:
-  // edx = base
-  // eax = exponent
-  // ecx = temporary, result
-
   CpuFeatures::Scope use_sse2(SSE2);
-  Label allocate_return, call_runtime;
-
-  // Load input parameters.
-  __ mov(edx, Operand(esp, 2 * kPointerSize));
-  __ mov(eax, Operand(esp, 1 * kPointerSize));
-
+  Factory* factory = masm->isolate()->factory();
+  Label double_int_runtime, generic_runtime, done;
+  Label base_is_smi, unpack_exponent, exponent_not_smi, int_exponent;
   // Save 1 in xmm3 - we need this several times later on.
   __ mov(ecx, Immediate(1));
   __ cvtsi2sd(xmm3, ecx);
 
-  Label exponent_nonsmi;
-  Label base_nonsmi;
-  // If the exponent is a heap number go to that specific case.
-  __ JumpIfNotSmi(eax, &exponent_nonsmi);
-  __ JumpIfNotSmi(edx, &base_nonsmi);
+  switch (exponent_type_) {
+    case ON_STACK:
+      // Load input parameters from stack
+      __ mov(edx, Operand(esp, 2 * kPointerSize));
+      __ mov(eax, Operand(esp, 1 * kPointerSize));
+      // edx: base (smi or heap number)
+      // eax: exponent (smi or heap number)
+      __ JumpIfSmi(edx, &base_is_smi, Label::kNear);
+      __ cmp(FieldOperand(edx, HeapObject::kMapOffset),
+             factory->heap_number_map());
+      __ j(not_equal, &generic_runtime);
+
+      // Check base for NaN or +/-Infinity
+      __ mov(ecx, FieldOperand(edx, HeapNumber::kExponentOffset));
+      __ and_(ecx, HeapNumber::kExponentMask);
+      __ cmp(ecx, Immediate(HeapNumber::kExponentMask));
+      __ j(greater_equal, &generic_runtime);
+      __ movdbl(xmm1, FieldOperand(edx, HeapNumber::kValueOffset));
+
+      __ jmp(&unpack_exponent, Label::kNear);
+      __ bind(&base_is_smi);
+      __ SmiUntag(edx);
+      __ cvtsi2sd(xmm1, edx);
+      __ bind(&unpack_exponent);
+      // Fall through is intended.
+    case TAGGED:
+      // xmm1: base as double
+      // eax: exponent (smi or heap number)
+      __ JumpIfNotSmi(eax, &exponent_not_smi, Label::kNear);
+      __ SmiUntag(eax);
+      __ jmp(&int_exponent);
+      __ bind(&exponent_not_smi);
+      if (exponent_type_ == ON_STACK) {
+        // Heap number check not necessary in optimized code as we will have
+        // already deoptimized if eax was neither smi nor heap number.
+        __ cmp(FieldOperand(eax, HeapObject::kMapOffset),
+               factory->heap_number_map());
+        __ j(not_equal, &generic_runtime);
+      }
+      __ movdbl(xmm2, FieldOperand(eax, HeapNumber::kValueOffset));
+      break;
+    case INTEGER:
+      // xmm1: base as double
+      // eax: exponent as untagged integer
+    case DOUBLE:
+      // xmm1: base as double
+      // xmm2: exponent as double
+      break;
+    default:
+      UNREACHABLE();

[ulan, 2011/12/01 18:11:26]

Since you listed all the cases above, the default case can be removed. This will
enable static checks by compiler instead of run-time checks by UNREACHABLE().

+  }
 
-  // Optimized version when both exponent and base are smis.
-  Label powi;
-  __ SmiUntag(edx);
-  __ cvtsi2sd(xmm0, edx);
-  __ jmp(&powi);
-  // exponent is smi and base is a heapnumber.
-  __ bind(&base_nonsmi);
-  Factory* factory = masm->isolate()->factory();
-  __ cmp(FieldOperand(edx, HeapObject::kMapOffset),
-         factory->heap_number_map());
-  __ j(not_equal, &call_runtime);
+  if (exponent_type_ != ON_STACK) {
+    // Check base in xmm1 for NaN or +/-Infinity
+    if (CpuFeatures::IsSupported(SSE4_1)) {
+      __ extractps(ecx, xmm1, 4 * kBitsPerByte);
+    } else {
+      __ movsd(xmm4, xmm1);
+      __ psrlq(xmm4, 4 * kBitsPerByte);
+      __ movd(ecx, xmm4);
+    }
+    __ and_(ecx, HeapNumber::kExponentMask);
+    __ cmp(ecx, Immediate(HeapNumber::kExponentMask));
+    __ j(greater_equal, &generic_runtime);
+  }
 
-  __ movdbl(xmm0, FieldOperand(edx, HeapNumber::kValueOffset));
+  if (exponent_type_ != INTEGER) {
+    Label not_minus_half, fast_power;
+    // xmm1: base as double that is not +/- Infinity or NaN
+    // xmm2: exponent as double
+    // Detect integer exponents stored as double.
+    __ cvttsd2si(eax, Operand(xmm2));
+    __ cmp(eax, Immediate(0x80000000));  // Skip to runtime if possibly NaN.
+    __ j(equal, &generic_runtime);
+    __ cvtsi2sd(xmm4, eax);
+    __ ucomisd(xmm2, xmm4);
+    __ j(equal, &int_exponent);
+
+    // Detect square root case.
+    // Test for -0.5.
+    // Load xmm4 with -0.5.
+    __ mov(ecx, Immediate(0xBF000000));
+    __ movd(xmm4, ecx);
+    __ cvtss2sd(xmm4, xmm4);
+    // xmm3 now has -0.5.
+    __ ucomisd(xmm4, xmm2);
+    __ j(not_equal, &not_minus_half, Label::kNear);
+
+    // Calculates reciprocal of square root.eax
+    // sqrtsd returns -0 when input is -0.  ECMA spec requires +0.
+    __ xorps(xmm2, xmm2);
+    __ addsd(xmm2, xmm1);
+    __ sqrtsd(xmm2, xmm2);
+    __ divsd(xmm3, xmm2);
+    __ jmp(&done);
+    __ extractps(ecx, xmm1, 4 * kBitsPerByte);
+    // Test for 0.5.
+    __ bind(&not_minus_half);
+    // Load xmm2 with 0.5.
+    // Since xmm3 is 1 and xmm4 is -0.5 this is simply xmm4 + xmm3.
+    __ addsd(xmm4, xmm3);
+    // xmm2 now has 0.5.
+    __ ucomisd(xmm4, xmm2);
+    __ j(not_equal, &fast_power, Label::kNear);
+    // Calculates square root.
+    // sqrtsd returns -0 when input is -0.  ECMA spec requires +0.
+    __ xorps(xmm4, xmm4);
+    __ addsd(xmm4, xmm1);
+    __ sqrtsd(xmm3, xmm4);
+    __ jmp(&done);
+
+    // Using FPU instructions to calculate power.
+    Label fast_power_failed;
+    __ bind(&fast_power);
+    __ fnclex();  // Clear flags to catch exceptions later.
+    // Transfer (B)ase and (E)xponent onto the FPU register stack.
+    __ sub(esp, Immediate(kDoubleSize));
+    __ movdbl(Operand(esp, 0), xmm2);
+    __ fld_d(Operand(esp, 0));  // E
+    __ movdbl(Operand(esp, 0), xmm1);
+    __ fld_d(Operand(esp, 0));  // B, E
+
+    // Exponent is in st(1) and base is in st(0)
+    // B ^ E = (2^(E * log2(B)) - 1) + 1 = (2^X - 1) + 1 for X = E * log2(B)
+    // FYL2X calculates st(1) * log2(st(0))
+    __ fyl2x();    // X
+    __ fld(0);     // X, X
+    __ frndint();  // rnd(X), X
+    __ fsub(1);    // rnd(X), X-rnd(X)
+    __ fxch(1);    // X - rnd(X), rnd(X)
+    // F2XM1 calculates 2^st(0) - 1 for -1 < st(0) < 1
+    __ f2xm1();    // 2^(X-rnd(X)) - 1, rnd(X)
+    __ fld1();     // 1, 2^(X-rnd(X)) - 1, rnd(X)
+    __ faddp(1);   // 1, 2^(X-rnd(X)), rnd(X)
+    // FSCALE calculates st(0) * 2^st(1)
+    __ fscale();   // 2^X, rnd(X)
+    __ fstp(1);
+    // Bail out to runtime in case of exceptions in the status word.
+    __ fnstsw_ax();
+    __ test_b(eax, 0x5F);  // We check for all but precision exception.
+    __ j(not_zero, &fast_power_failed, Label::kNear);
+    __ fstp_d(Operand(esp, 0));
+    __ movdbl(xmm3, Operand(esp, 0));
+    __ add(esp, Immediate(kDoubleSize));
+    __ jmp(&done);
 
-  // Optimized version of pow if exponent is a smi.
-  // xmm0 contains the base.
-  __ bind(&powi);
-  __ SmiUntag(eax);
+    __ bind(&fast_power_failed);
+    __ fninit();
+    __ add(esp, Immediate(kDoubleSize));
+    __ jmp(&generic_runtime);
+  }
 
+  // Calculate power with integer exponent.
+  __ bind(&int_exponent);
+  // xmm1: base as double that is not +/- Infinity or NaN
+  // eax: exponent as untagged integer
   // Save exponent in base as we need to check if exponent is negative later.
   // We know that base and exponent are in different registers.
-  __ mov(edx, eax);
+  __ mov(ecx, eax);      // Back up exponent.
+  __ movsd(xmm4, xmm1);  // Back up base.
+  __ movsd(xmm2, xmm3);  // Load xmm2 with 1.
 
   // Get absolute value of exponent.
-  Label no_neg;
+  Label no_neg, while_true, no_multiply;
   __ cmp(eax, 0);
   __ j(greater_equal, &no_neg, Label::kNear);
   __ neg(eax);
   __ bind(&no_neg);
 
-  // Load xmm1 with 1.
-  __ movsd(xmm1, xmm3);
-  Label while_true;
-  Label no_multiply;
-
   __ bind(&while_true);
   __ shr(eax, 1);
   __ j(not_carry, &no_multiply, Label::kNear);
-  __ mulsd(xmm1, xmm0);
+  __ mulsd(xmm3, xmm1);
   __ bind(&no_multiply);
-  __ mulsd(xmm0, xmm0);
+  __ mulsd(xmm1, xmm1);
   __ j(not_zero, &while_true);
 
   // base has the original value of the exponent - if the exponent  is
   // negative return 1/result.
-  __ test(edx, edx);
-  __ j(positive, &allocate_return);
-  // Special case if xmm1 has reached infinity.
-  __ mov(ecx, Immediate(0x7FB00000));
-  __ movd(xmm0, ecx);
-  __ cvtss2sd(xmm0, xmm0);
-  __ ucomisd(xmm0, xmm1);
-  __ j(equal, &call_runtime);
-  __ divsd(xmm3, xmm1);
-  __ movsd(xmm1, xmm3);
-  __ jmp(&allocate_return);
-
-  // exponent (or both) is a heapnumber - no matter what we should now work
-  // on doubles.
-  __ bind(&exponent_nonsmi);
-  __ cmp(FieldOperand(eax, HeapObject::kMapOffset),
-         factory->heap_number_map());
-  __ j(not_equal, &call_runtime);
-  __ movdbl(xmm1, FieldOperand(eax, HeapNumber::kValueOffset));
-  // Test if exponent is nan.
-  __ ucomisd(xmm1, xmm1);
-  __ j(parity_even, &call_runtime);
+  __ test(ecx, ecx);
+  __ j(positive, &done);
+  __ divsd(xmm2, xmm3);
+  __ movsd(xmm3, xmm2);
+  // Test whether result is zero.  Bail out to check for subnormal result.
+  // Due to subnormals, x^-y == (1/x)^y does not hold in all cases.
+  __ xorps(xmm2, xmm2);
+  __ ucomisd(xmm2, xmm3);
+  __ j(equal, &double_int_runtime);
+
+  // Returning or bailing out.
+  if (exponent_type_ == ON_STACK) {
+    // We expect the result as heap number in eax.
+    __ bind(&done);
+    // xmm1: result
+    __ AllocateHeapNumber(eax, ecx, edx, &generic_runtime);
+    __ movdbl(FieldOperand(eax, HeapNumber::kValueOffset), xmm3);
+    __ ret(2 * kPointerSize);
 
-  Label base_not_smi;
-  Label handle_special_cases;
-  __ JumpIfNotSmi(edx, &base_not_smi, Label::kNear);
-  __ SmiUntag(edx);
-  __ cvtsi2sd(xmm0, edx);
-  __ jmp(&handle_special_cases, Label::kNear);
-
-  __ bind(&base_not_smi);
-  __ cmp(FieldOperand(edx, HeapObject::kMapOffset),
-         factory->heap_number_map());
-  __ j(not_equal, &call_runtime);
-  __ mov(ecx, FieldOperand(edx, HeapNumber::kExponentOffset));
-  __ and_(ecx, HeapNumber::kExponentMask);
-  __ cmp(ecx, Immediate(HeapNumber::kExponentMask));
-  // base is NaN or +/-Infinity
-  __ j(greater_equal, &call_runtime);
-  __ movdbl(xmm0, FieldOperand(edx, HeapNumber::kValueOffset));
+    // The arguments are still on the stack.
+    __ bind(&generic_runtime);
+    __ bind(&double_int_runtime);
+    __ TailCallRuntime(Runtime::kMath_pow_cfunction, 2, 1);
+  } else {
+    __ jmp(&done);
 
-  // base is in xmm0 and exponent is in xmm1.
-  __ bind(&handle_special_cases);
-  Label not_minus_half;
-  // Test for -0.5.
-  // Load xmm2 with -0.5.
-  __ mov(ecx, Immediate(0xBF000000));
-  __ movd(xmm2, ecx);
-  __ cvtss2sd(xmm2, xmm2);
-  // xmm2 now has -0.5.
-  __ ucomisd(xmm2, xmm1);
-  __ j(not_equal, &not_minus_half, Label::kNear);
-
-  // Calculates reciprocal of square root.
-  // sqrtsd returns -0 when input is -0.  ECMA spec requires +0.
-  __ xorps(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.
-  __ xorps(xmm1, xmm1);
-  __ addsd(xmm1, xmm0);
-  __ sqrtsd(xmm1, xmm1);
-
-  __ bind(&allocate_return);
-  __ AllocateHeapNumber(ecx, eax, edx, &call_runtime);
-  __ movdbl(FieldOperand(ecx, HeapNumber::kValueOffset), xmm1);
-  __ mov(eax, ecx);
-  __ ret(2 * kPointerSize);
+    Label return_from_runtime;
+    StubRuntimeCallHelper callhelper;
+    __ bind(&generic_runtime);
+    // xmm1: base
+    // xmm2: exponent
+    {
+      AllowExternalCallThatCantCauseGC scope(masm);
+      __ PrepareCallCFunction(4, eax);
+      __ movdbl(Operand(esp, 0 * kDoubleSize), xmm1);
+      __ movdbl(Operand(esp, 1 * kDoubleSize), xmm2);
+      __ CallCFunction(
+          ExternalReference::power_double_double_function(masm->isolate()), 4);
+    }
+    __ jmp(&return_from_runtime, Label::kNear);
 
-  __ bind(&call_runtime);
-  __ TailCallRuntime(Runtime::kMath_pow_cfunction, 2, 1);
+    __ bind(&double_int_runtime);
+    // xmm4: base
+    // ecx: exponent
+    {
+      __ PrepareCallCFunction(4, eax);
+      __ movdbl(Operand(esp, 0 * kDoubleSize), xmm4);
+      __ mov(Operand(esp, 1 * kDoubleSize), ecx);
+      AllowExternalCallThatCantCauseGC scope(masm);
+      __ CallCFunction(
+          ExternalReference::power_double_int_function(masm->isolate()), 4);
+    }
+
+    __ bind(&return_from_runtime);
+    // Return value is in st(0) on ia32.
+    // Store it into the (fixed) result register.
+    __ sub(esp, Immediate(kDoubleSize));
+    __ fstp_d(Operand(esp, 0));
+    __ movdbl(xmm3, Operand(esp, 0));
+    __ add(esp, Immediate(kDoubleSize));
+
+    // We expect the result in xmm3.
+    __ bind(&done);
+    __ ret(0);
+  }
 }