[ia32] Introduce FMA3 instructions on scalar data elements.

test/cctest/test-disasm-ia32.cc

Index: test/cctest/test-disasm-ia32.cc
diff --git a/test/cctest/test-disasm-ia32.cc b/test/cctest/test-disasm-ia32.cc
index fb0c56283f6c388da054a0ad8bcbd280b14bc55c..a2eaa15ed341fd6513ca75ec1e0cd7d130222656 100644
--- a/test/cctest/test-disasm-ia32.cc
+++ b/test/cctest/test-disasm-ia32.cc
@@ -51,7 +51,7 @@ TEST(DisasmIa320) {
   CcTest::InitializeVM();
   Isolate* isolate = CcTest::i_isolate();
   HandleScope scope(isolate);
-  v8::internal::byte buffer[2048];
+  v8::internal::byte buffer[4096];
   Assembler assm(isolate, buffer, sizeof buffer);
   DummyStaticFunction(NULL);  // just bloody use it (DELETE; debugging)
 
@@ -401,6 +401,14 @@ TEST(DisasmIa320) {
     __ xorps(xmm0, Operand(ebx, ecx, times_4, 10000));
 
     // Arithmetic operation
+    __ addss(xmm1, xmm0);
+    __ addss(xmm1, Operand(ebx, ecx, times_4, 10000));
+    __ mulss(xmm1, xmm0);
+    __ mulss(xmm1, Operand(ebx, ecx, times_4, 10000));
+    __ subss(xmm1, xmm0);
+    __ subss(xmm1, Operand(ebx, ecx, times_4, 10000));
+    __ divss(xmm1, xmm0);
+    __ divss(xmm1, Operand(ebx, ecx, times_4, 10000));
     __ addps(xmm1, xmm0);
     __ addps(xmm1, Operand(ebx, ecx, times_4, 10000));
     __ subps(xmm1, xmm0);
@@ -409,6 +417,9 @@ TEST(DisasmIa320) {
     __ mulps(xmm1, Operand(ebx, ecx, times_4, 10000));
     __ divps(xmm1, xmm0);
     __ divps(xmm1, Operand(ebx, ecx, times_4, 10000));
+
+    __ ucomiss(xmm0, xmm1);
+    __ ucomiss(xmm0, Operand(ebx, ecx, times_4, 10000));
   }
   {
     __ cvttss2si(edx, Operand(ebx, ecx, times_4, 10000));
@@ -486,6 +497,68 @@ TEST(DisasmIa320) {
     }
   }
 
+  // FMA3 instruction
+  {
+    if (CpuFeatures::IsSupported(FMA3)) {
+      CpuFeatureScope scope(&assm, FMA3);
+      __ vfmadd132sd(xmm0, xmm1, xmm2);
+      __ vfmadd132sd(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfmadd213sd(xmm0, xmm1, xmm2);
+      __ vfmadd213sd(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfmadd231sd(xmm0, xmm1, xmm2);
+      __ vfmadd231sd(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+
+      __ vfmsub132sd(xmm0, xmm1, xmm2);
+      __ vfmsub132sd(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfmsub213sd(xmm0, xmm1, xmm2);
+      __ vfmsub213sd(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfmsub231sd(xmm0, xmm1, xmm2);
+      __ vfmsub231sd(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+
+      __ vfnmadd132sd(xmm0, xmm1, xmm2);
+      __ vfnmadd132sd(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfnmadd213sd(xmm0, xmm1, xmm2);
+      __ vfnmadd213sd(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfnmadd231sd(xmm0, xmm1, xmm2);
+      __ vfnmadd231sd(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+
+      __ vfnmsub132sd(xmm0, xmm1, xmm2);
+      __ vfnmsub132sd(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfnmsub213sd(xmm0, xmm1, xmm2);
+      __ vfnmsub213sd(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfnmsub231sd(xmm0, xmm1, xmm2);
+      __ vfnmsub231sd(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+
+      __ vfmadd132ss(xmm0, xmm1, xmm2);
+      __ vfmadd132ss(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfmadd213ss(xmm0, xmm1, xmm2);
+      __ vfmadd213ss(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfmadd231ss(xmm0, xmm1, xmm2);
+      __ vfmadd231ss(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+
+      __ vfmsub132ss(xmm0, xmm1, xmm2);
+      __ vfmsub132ss(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfmsub213ss(xmm0, xmm1, xmm2);
+      __ vfmsub213ss(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfmsub231ss(xmm0, xmm1, xmm2);
+      __ vfmsub231ss(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+
+      __ vfnmadd132ss(xmm0, xmm1, xmm2);
+      __ vfnmadd132ss(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfnmadd213ss(xmm0, xmm1, xmm2);
+      __ vfnmadd213ss(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfnmadd231ss(xmm0, xmm1, xmm2);
+      __ vfnmadd231ss(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+
+      __ vfnmsub132ss(xmm0, xmm1, xmm2);
+      __ vfnmsub132ss(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfnmsub213ss(xmm0, xmm1, xmm2);
+      __ vfnmsub213ss(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+      __ vfnmsub231ss(xmm0, xmm1, xmm2);
+      __ vfnmsub231ss(xmm0, xmm1, Operand(ebx, ecx, times_4, 10000));
+    }
+  }
+
   // xchg.
   {
     __ xchg(eax, eax);