[x64] fixed "vmovups reg, mem"

BUG=

Review-Url: https://codereview.chromium.org/2158923002
Cr-Commit-Position: refs/heads/master@{#37846}

src/x64/assembler-x64.cc

@@ -3898,6 +3898,7 @@ void Assembler::vps(byte op, XMMRegister dst, XMMRegister src1,
   emit_sse_operand(dst, src2);
 }
 
+
 void Assembler::vps(byte op, XMMRegister dst, XMMRegister src1,
                     const Operand& src2) {
   DCHECK(IsEnabled(AVX));
@@ -3907,14 +3908,6 @@ void Assembler::vps(byte op, XMMRegister dst, XMMRegister src1,
   emit_sse_operand(dst, src2);
 }
 
-void Assembler::vps(byte op, const Operand& dst, XMMRegister src1,
-                    XMMRegister src2) {
-  DCHECK(IsEnabled(AVX));
-  EnsureSpace ensure_space(this);
-  emit_vex_prefix(src2, src1, dst, kL128, kNone, k0F, kWIG);
-  emit(op);
-  emit_sse_operand(src2, dst);
-}
 
 void Assembler::vpd(byte op, XMMRegister dst, XMMRegister src1,
                     XMMRegister src2) {

src/x64/assembler-x64.h

@@ -1533,10 +1533,10 @@ class Assembler : public AssemblerBase {
   void vmovaps(XMMRegister dst, XMMRegister src) { vps(0x28, dst, xmm0, src); }
   void vmovups(XMMRegister dst, XMMRegister src) { vps(0x10, dst, xmm0, src); }
   void vmovups(XMMRegister dst, const Operand& src) {
-    vps(0x11, dst, xmm0, src);
+    vps(0x10, dst, xmm0, src);
   }
   void vmovups(const Operand& dst, XMMRegister src) {
-    vps(0x11, dst, xmm0, src);
+    vps(0x11, src, xmm0, dst);
   }
   void vmovapd(XMMRegister dst, XMMRegister src) { vpd(0x28, dst, xmm0, src); }
   void vmovmskpd(Register dst, XMMRegister src) {
@@ -1546,7 +1546,6 @@ class Assembler : public AssemblerBase {
 
   void vps(byte op, XMMRegister dst, XMMRegister src1, XMMRegister src2);
   void vps(byte op, XMMRegister dst, XMMRegister src1, const Operand& src2);
-  void vps(byte op, const Operand& dst, XMMRegister src1, XMMRegister src2);
   void vpd(byte op, XMMRegister dst, XMMRegister src1, XMMRegister src2);
   void vpd(byte op, XMMRegister dst, XMMRegister src1, const Operand& src2);
 

test/cctest/test-assembler-x64.cc

@@ -2316,4 +2316,40 @@ TEST(AssemblerX64PslldWithXmm15) {
   CHECK_EQ(V8_UINT64_C(0x22446688aaccef10), result);
 }
 
+typedef float (*F9)(float x, float y);
+TEST(AssemblerX64vmovups) {
+  CcTest::InitializeVM();
+  if (!CpuFeatures::IsSupported(AVX)) return;
+
+  Isolate* isolate = reinterpret_cast<Isolate*>(CcTest::isolate());
+  HandleScope scope(isolate);
+  v8::internal::byte buffer[256];
+  MacroAssembler assm(isolate, buffer, sizeof(buffer),
+                      v8::internal::CodeObjectRequired::kYes);
+  {
+    CpuFeatureScope avx_scope(&assm, AVX);
+    __ shufps(xmm0, xmm0, 0x0);  // brocast first argument
+    __ shufps(xmm1, xmm1, 0x0);  // brocast second argument
+    // copy xmm1 to xmm0 through the stack to test the "vmovups reg, mem".
+    __ subq(rsp, Immediate(kSimd128Size));
+    __ vmovups(Operand(rsp, 0), xmm1);
+    __ vmovups(xmm0, Operand(rsp, 0));
+    __ addq(rsp, Immediate(kSimd128Size));
+
+    __ ret(0);
+  }
+
+  CodeDesc desc;
+  assm.GetCode(&desc);
+  Handle<Code> code = isolate->factory()->NewCode(
+      desc, Code::ComputeFlags(Code::STUB), Handle<Code>());
+#ifdef OBJECT_PRINT
+  OFStream os(stdout);
+  code->Print(os);
+#endif
+
+  F9 f = FUNCTION_CAST<F9>(code->entry());
+  CHECK_EQ(-1.5, f(1.5, -1.5));
+}
+
 #undef __