Use SSE2 instructions when available on ia32 platform.

src/ia32/codegen-ia32.cc

 // Copyright 2006-2009 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 750 matching lines @@
   // be either smi or heap number objects (fp values). Requirements:
   // operand_1 on TOS+1 , operand_2 on TOS+2; Returns operands as
   // floating point numbers on FPU stack.
   static void LoadFloatOperands(MacroAssembler* masm, Register scratch);
   // Test if operands are smi or number objects (fp). Requirements:
   // operand_1 in eax, operand_2 in edx; falls through on float
   // operands, jumps to the non_float label otherwise.
   static void CheckFloatOperands(MacroAssembler* masm,
                                  Label* non_float,
                                  Register scratch);
+  // Test if operands are numbers (smi or HeapNumber objects), and load
+  // them into xmm0 and xmm1 if they are.  Jump to label not_numbers if
+  // either operand is not a number.  Operands are in edx and eax.
+  // Leaves operands unchanged.
+  static void LoadSse2Operands(MacroAssembler* masm, Label* not_numbers);
   // Allocate a heap number in new space with undefined value.
   // Returns tagged pointer in eax, or jumps to need_gc if new space is full.
   static void AllocateHeapNumber(MacroAssembler* masm,
                                  Label* need_gc,
                                  Register scratch1,
                                  Register scratch2,
                                  Register result);
 };
 
 
@@ skipping 5911 matching lines @@
   __ mov(edx, Operand(esp, 2 * kPointerSize));  // get x
 
   // Floating point case.
   switch (op_) {
     case Token::ADD:
     case Token::SUB:
     case Token::MUL:
     case Token::DIV: {
       // eax: y
       // edx: x
-      FloatingPointHelper::CheckFloatOperands(masm, &call_runtime, ebx);
-      // Fast-case: Both operands are numbers.
-      // Allocate a heap number, if needed.
-      Label skip_allocation;
-      switch (mode_) {
-        case OVERWRITE_LEFT:
-          __ mov(eax, Operand(edx));
-          // Fall through!
-        case OVERWRITE_RIGHT:
-          // If the argument in eax is already an object, we skip the
-          // allocation of a heap number.
-          __ test(eax, Immediate(kSmiTagMask));
-          __ j(not_zero, &skip_allocation, not_taken);
-          // Fall through!
-        case NO_OVERWRITE:
-          FloatingPointHelper::AllocateHeapNumber(masm,
-                                                  &call_runtime,
-                                                  ecx,
-                                                  edx,
-                                                  eax);
-          __ bind(&skip_allocation);
-          break;
-        default: UNREACHABLE();
+
+      if (CpuFeatures::IsSupported(CpuFeatures::SSE2)) {
+        CpuFeatures::Scope use_sse2(CpuFeatures::SSE2);
+        FloatingPointHelper::LoadSse2Operands(masm, &call_runtime);
+
+        switch (op_) {
+          case Token::ADD: __ addsd(xmm0, xmm1); break;
+          case Token::SUB: __ subsd(xmm0, xmm1); break;
+          case Token::MUL: __ mulsd(xmm0, xmm1); break;
+          case Token::DIV: __ divsd(xmm0, xmm1); break;
+          default: UNREACHABLE();
+        }
+        // Allocate a heap number, if needed.
+        Label skip_allocation;
+        switch (mode_) {
+          case OVERWRITE_LEFT:
+            __ mov(eax, Operand(edx));
+            // Fall through!
+          case OVERWRITE_RIGHT:
+            // If the argument in eax is already an object, we skip the
+            // allocation of a heap number.
+            __ test(eax, Immediate(kSmiTagMask));
+            __ j(not_zero, &skip_allocation, not_taken);
+            // Fall through!
+          case NO_OVERWRITE:
+            FloatingPointHelper::AllocateHeapNumber(masm,
+                                                    &call_runtime,
+                                                    ecx,
+                                                    edx,
+                                                    eax);
+            __ bind(&skip_allocation);
+            break;
+          default: UNREACHABLE();
+        }
+        __ movdbl(FieldOperand(eax, HeapNumber::kValueOffset), xmm0);
+        __ ret(2 * kPointerSize);
+
+      } else {  // SSE2 not available, use FPU.
+        FloatingPointHelper::CheckFloatOperands(masm, &call_runtime, ebx);

[Mark Mentovai, 2009/09/09 15:09:55]

Shouldn't this be done for both SSE2 and x87, or should there be a
CheckSse2Operands?

[William Hesse, 2009/09/10 07:27:37]

The checking and the loading are combined into one function in LoadSse2Operands,
so that we can be more efficient.  The Smi tag is only checked once.  This is
documented in the declaration of LoadSse2Operands().

+        // Allocate a heap number, if needed.

[Mark Mentovai, 2009/09/09 15:09:55]

Is there a reason that this is done before the x87 operations in this branch,
but after the SSE2 operations in the branch above?

Is there a way to share this duplicated code between the two branches?

[William Hesse, 2009/09/10 07:27:37]

This code has to go after the check, but before the store of the result, so it
is enclosed by code that depends on whether we use SSE2.  It also needs to have
the values still in the eax and edx registers.  Most of the common code has been
factored out into AllocateHeapNumber() already.

So either we need to put it in a separate function, or we need to end the
use_sse2 scope, run the common code, and the n run a test and potentially enter
a new sse2 scope again.  This doesn't seem worth it.

+        Label skip_allocation;
+        switch (mode_) {
+          case OVERWRITE_LEFT:
+            __ mov(eax, Operand(edx));
+            // Fall through!
+          case OVERWRITE_RIGHT:
+            // If the argument in eax is already an object, we skip the
+            // allocation of a heap number.
+            __ test(eax, Immediate(kSmiTagMask));
+            __ j(not_zero, &skip_allocation, not_taken);
+            // Fall through!
+          case NO_OVERWRITE:
+            FloatingPointHelper::AllocateHeapNumber(masm,
+                                                    &call_runtime,
+                                                    ecx,
+                                                    edx,
+                                                    eax);
+            __ bind(&skip_allocation);
+            break;
+          default: UNREACHABLE();
+        }
+        FloatingPointHelper::LoadFloatOperands(masm, ecx);
+
+        switch (op_) {
+          case Token::ADD: __ faddp(1); break;
+          case Token::SUB: __ fsubp(1); break;
+          case Token::MUL: __ fmulp(1); break;
+          case Token::DIV: __ fdivp(1); break;
+          default: UNREACHABLE();
+        }
+        __ fstp_d(FieldOperand(eax, HeapNumber::kValueOffset));
+        __ ret(2 * kPointerSize);
       }
-      FloatingPointHelper::LoadFloatOperands(masm, ecx);
-
-      switch (op_) {
-        case Token::ADD: __ faddp(1); break;
-        case Token::SUB: __ fsubp(1); break;
-        case Token::MUL: __ fmulp(1); break;
-        case Token::DIV: __ fdivp(1); break;
-        default: UNREACHABLE();
-      }
-      __ fstp_d(FieldOperand(eax, HeapNumber::kValueOffset));
-      __ ret(2 * kPointerSize);
     }
     case Token::MOD: {
       // For MOD we go directly to runtime in the non-smi case.
       break;
     }
     case Token::BIT_OR:
     case Token::BIT_AND:
     case Token::BIT_XOR:
     case Token::SAR:
     case Token::SHL:
@@ skipping 227 matching lines @@
   __ bind(&load_smi);
   __ sar(number, kSmiTagSize);
   __ push(number);
   __ fild_s(Operand(esp, 0));
   __ pop(number);
 
   __ bind(&done);
 }
 
 
+void FloatingPointHelper::LoadSse2Operands(MacroAssembler* masm,
+                                           Label* not_numbers) {
+  Label load_smi_edx, load_eax, load_smi_eax, load_float_eax, done;
+  // Load operand in edx into xmm0, or branch to not_numbers.
+  __ test(edx, Immediate(kSmiTagMask));
+  __ j(zero, &load_smi_edx, not_taken);  // Argument in edx is a smi.
+  __ cmp(FieldOperand(edx, HeapObject::kMapOffset), Factory::heap_number_map());
+  __ j(not_equal, not_numbers);  // Argument in edx is not a number.
+  __ movdbl(xmm0, FieldOperand(edx, HeapNumber::kValueOffset));
+  __ bind(&load_eax);
+  // Load operand in eax into xmm1, or branch to not_numbers.
+  __ test(eax, Immediate(kSmiTagMask));
+  __ j(zero, &load_smi_eax, not_taken);  // Argument in eax is a smi.
+  __ cmp(FieldOperand(eax, HeapObject::kMapOffset), Factory::heap_number_map());
+  __ j(equal, &load_float_eax);
+  __ jmp(not_numbers);  // Argument in eax is not a number.
+  __ bind(&load_smi_edx);
+  __ sar(edx, 1);  // Untag smi before converting to float.
+  __ cvtsi2sd(xmm0, Operand(edx));
+  __ shl(edx, 1);  // Retag smi for heap number overwriting test.
+  __ jmp(&load_eax);
+  __ bind(&load_smi_eax);
+  __ sar(eax, 1);  // Untag smi before converting to float.
+  __ cvtsi2sd(xmm1, Operand(eax));
+  __ shl(eax, 1);  // Retag smi for heap number overwriting test.
+  __ jmp(&done);
+  __ bind(&load_float_eax);
+  __ movdbl(xmm1, FieldOperand(eax, HeapNumber::kValueOffset));
+  __ bind(&done);
+}
+
+
 void FloatingPointHelper::LoadFloatOperands(MacroAssembler* masm,
                                             Register scratch) {
   Label load_smi_1, load_smi_2, done_load_1, done;
   __ mov(scratch, Operand(esp, 2 * kPointerSize));
   __ test(scratch, Immediate(kSmiTagMask));
   __ j(zero, &load_smi_1, not_taken);
   __ fld_d(FieldOperand(scratch, HeapNumber::kValueOffset));
   __ bind(&done_load_1);
 
   __ mov(scratch, Operand(esp, 1 * kPointerSize));
@@ skipping 342 matching lines @@
 
   // Push arguments below the return address.
   __ pop(ecx);
   __ push(eax);
   __ push(edx);
   __ push(ecx);
 
   // Inlined floating point compare.
   // Call builtin if operands are not floating point or smi.
   Label check_for_symbols;
-  FloatingPointHelper::CheckFloatOperands(masm, &check_for_symbols, ebx);
-  FloatingPointHelper::LoadFloatOperands(masm, ecx);
-  __ FCmp();
+  Label unordered;
+  if (CpuFeatures::IsSupported(CpuFeatures::SSE2)) {
+    CpuFeatures::Scope use_sse2(CpuFeatures::SSE2);
+    CpuFeatures::Scope use_cmov(CpuFeatures::CMOV);
 
-  // Jump to builtin for NaN.
-  __ j(parity_even, &call_builtin, not_taken);
+    FloatingPointHelper::LoadSse2Operands(masm, &check_for_symbols);
+    __ comisd(xmm0, xmm1);
 
-  // TODO(1243847): Use cmov below once CpuFeatures are properly hooked up.
-  Label below_lbl, above_lbl;
-  // use edx, eax to convert unsigned to signed comparison
-  __ j(below, &below_lbl, not_taken);
-  __ j(above, &above_lbl, not_taken);
+    // Jump to builtin for NaN.
+    __ j(parity_even, &unordered, not_taken);
+    __ mov(eax, 0);  // equal
+    __ mov(ecx, Immediate(Smi::FromInt(1)));
+    __ cmov(above, eax, Operand(ecx));
+    __ mov(ecx, Immediate(Smi::FromInt(-1)));
+    __ cmov(below, eax, Operand(ecx));
+    __ ret(2 * kPointerSize);
+  } else {
+    FloatingPointHelper::CheckFloatOperands(masm, &check_for_symbols, ebx);
+    FloatingPointHelper::LoadFloatOperands(masm, ecx);
+    __ FCmp();
 
-  __ xor_(eax, Operand(eax));  // equal
-  __ ret(2 * kPointerSize);
+    // Jump to builtin for NaN.
+    __ j(parity_even, &unordered, not_taken);
 
-  __ bind(&below_lbl);
-  __ mov(eax, -1);
-  __ ret(2 * kPointerSize);
+    Label below_lbl, above_lbl;
+    // Return a result of -1, 0, or 1, to indicate result of comparison.
+    __ j(below, &below_lbl, not_taken);
+    __ j(above, &above_lbl, not_taken);
 
-  __ bind(&above_lbl);
-  __ mov(eax, 1);
+    __ xor_(eax, Operand(eax));  // equal
+    // Both arguments were pushed in case a runtime call was needed.
+    __ ret(2 * kPointerSize);
+
+    __ bind(&below_lbl);
+    __ mov(eax, Immediate(Smi::FromInt(-1)));
+    __ ret(2 * kPointerSize);
+
+    __ bind(&above_lbl);
+    __ mov(eax, Immediate(Smi::FromInt(1)));
+    __ ret(2 * kPointerSize);  // eax, edx were pushed
+  }
+  // If one of the numbers was NaN, then the result is always false.
+  // The cc is never not-equal.
+  __ bind(&unordered);
+  ASSERT(cc_ != not_equal);
+  if (cc_ == less || cc_ == less_equal) {
+    __ mov(eax, Immediate(Smi::FromInt(1)));
+  } else {
+    __ mov(eax, Immediate(Smi::FromInt(-1)));
+  }
   __ ret(2 * kPointerSize);  // eax, edx were pushed
 
   // Fast negative check for symbol-to-symbol equality.
   __ bind(&check_for_symbols);
   if (cc_ == equal) {
     BranchIfNonSymbol(masm, &call_builtin, eax, ecx);
     BranchIfNonSymbol(masm, &call_builtin, edx, ecx);
 
     // We've already checked for object identity, so if both operands
     // are symbols they aren't equal. Register eax already holds a
@@ skipping 498 matching lines @@
 
 int CompareStub::MinorKey() {
   // Encode the two parameters in a unique 16 bit value.
   ASSERT(static_cast<unsigned>(cc_) < (1 << 15));
   return (static_cast<unsigned>(cc_) << 1) | (strict_ ? 1 : 0);
 }
 
 #undef __
 
 } }  // namespace v8::internal