[x64] add Absps/d and Negps/d macro

src/x64/macro-assembler-x64.cc

Index: src/x64/macro-assembler-x64.cc
diff --git a/src/x64/macro-assembler-x64.cc b/src/x64/macro-assembler-x64.cc
index 407703ef8ffb303d4e46f4f35681edb31aa62a8d..28fa6f512125aae134be84fc749e8493dedb5846 100644
--- a/src/x64/macro-assembler-x64.cc
+++ b/src/x64/macro-assembler-x64.cc
@@ -2974,6 +2974,32 @@ void MacroAssembler::Ucomisd(XMMRegister src1, const Operand& src2) {
 }
 
 
+// ----------------------------------------------------------------------------
+
+void MacroAssembler::Absps(XMMRegister dst) {
+  static const struct V8_ALIGNED(16) {
+    uint32_t a;
+    uint32_t b;
+    uint32_t c;
+    uint32_t d;
+  } float_absolute_constant = {0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF};
+  Set(kScratchRegister, reinterpret_cast<intptr_t>(&float_absolute_constant));

[Benedikt Meurer, 2016/07/18 04:47:52]

Can we somehow put this somewhere near the roots array so we can use the
ExternalOperand mechanism to avoid this load?

Also can you add Abspd and Negpd versions as well?

+  Andps(dst, Operand(kScratchRegister, 0));
+}
+
+
+void MacroAssembler::Negps(XMMRegister dst) {
+  static const struct V8_ALIGNED(16) {
+    uint32_t a;
+    uint32_t b;
+    uint32_t c;
+    uint32_t d;
+  } float_negate_constant = {0x80000000, 0x80000000, 0x80000000, 0x80000000};
+  Set(kScratchRegister, reinterpret_cast<intptr_t>(&float_negate_constant));
+  Xorps(dst, Operand(kScratchRegister, 0));

[bbudge, 2016/07/18 14:08:09]

It seems like we could avoid the memory access with AVX:

vpxor kScratch, kScratch, kScratch
vsubps dst, kScratch, dst

Is there a reason we can't use 0 - dst?

A similar trick can be done with three instructions for SSE.

+}
+
+
 void MacroAssembler::Cmp(Register dst, Handle<Object> source) {
   AllowDeferredHandleDereference smi_check;
   if (source->IsSmi()) {