[ARM] Add NEON instructions for implementing SIMD.

test/cctest/test-macro-assembler-arm.cc

Index: test/cctest/test-macro-assembler-arm.cc
diff --git a/test/cctest/test-macro-assembler-arm.cc b/test/cctest/test-macro-assembler-arm.cc
index 6b69296e11bdcae4311b153d0e4deb57e44408f5..1c1c5db96b892ca7571927b1d98149931878c5d5 100644
--- a/test/cctest/test-macro-assembler-arm.cc
+++ b/test/cctest/test-macro-assembler-arm.cc
@@ -379,4 +379,116 @@ TEST(ReplaceLane) {
   }
 }
 
+#define CHECK_EQ_32X4(field, v0, v1, v2, v3) \
+  CHECK_EQ(v0, t.field[0]);                  \
+  CHECK_EQ(v1, t.field[1]);                  \
+  CHECK_EQ(v2, t.field[2]);                  \
+  CHECK_EQ(v3, t.field[3]);
+
+TEST(Swizzle) {
+  if (!CpuFeatures::IsSupported(NEON)) return;
+
+  // Allocate an executable page of memory.
+  size_t actual_size;
+  byte* buffer = static_cast<byte*>(v8::base::OS::Allocate(
+      Assembler::kMinimalBufferSize, &actual_size, true));
+  CHECK(buffer);
+  Isolate* isolate = CcTest::i_isolate();
+  HandleScope handles(isolate);
+  MacroAssembler assembler(isolate, buffer, static_cast<int>(actual_size),
+                           v8::internal::CodeObjectRequired::kYes);
+  MacroAssembler* masm = &assembler;  // Create a pointer for the __ macro.
+
+  typedef struct {
+    int32_t _32x4_3210[4];  // identity
+    int32_t _32x4_1032[4];  // high / low swap
+    int32_t _32x4_0000[4];  // vdup's
+    int32_t _32x4_1111[4];
+    int32_t _32x4_2222[4];
+    int32_t _32x4_3333[4];
+    int32_t _32x4_2103[4];           // rotate left
+    int32_t _32x4_0321[4];           // rotate right
+    int32_t _32x4_1132[4];           // irregular
+    int32_t _32x4_1132_in_place[4];  // irregular, in-place
+  } T;
+  T t;
+
+  __ stm(db_w, sp, r4.bit() | r5.bit() | r6.bit() | r7.bit() | lr.bit());
+
+  const Register kScratch = r5;
+
+  // Make test vector [0, 1, 2, 3]
+  __ veor(q1, q1, q1);  // Zero
+  for (int i = 0; i < 4; i++) {
+    __ mov(r4, Operand(i));
+    __ ReplaceLane(q1, q1, r4, NeonS32, i);
+  }
+  __ Swizzle(q0, q1, kScratch, Neon32, 0x3210);
+  __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, _32x4_3210))));
+  __ vst1(Neon8, NeonListOperand(q0), NeonMemOperand(r4));
+
+  __ Swizzle(q0, q1, kScratch, Neon32, 0x1032);
+  __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, _32x4_1032))));
+  __ vst1(Neon8, NeonListOperand(q0), NeonMemOperand(r4));
+
+  __ Swizzle(q0, q1, kScratch, Neon32, 0x0000);
+  __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, _32x4_0000))));
+  __ vst1(Neon8, NeonListOperand(q0), NeonMemOperand(r4));
+
+  __ Swizzle(q0, q1, kScratch, Neon32, 0x1111);
+  __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, _32x4_1111))));
+  __ vst1(Neon8, NeonListOperand(q0), NeonMemOperand(r4));
+
+  __ Swizzle(q0, q1, kScratch, Neon32, 0x2222);
+  __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, _32x4_2222))));
+  __ vst1(Neon8, NeonListOperand(q0), NeonMemOperand(r4));
+
+  __ Swizzle(q0, q1, kScratch, Neon32, 0x3333);
+  __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, _32x4_3333))));
+  __ vst1(Neon8, NeonListOperand(q0), NeonMemOperand(r4));
+
+  __ Swizzle(q0, q1, kScratch, Neon32, 0x2103);
+  __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, _32x4_2103))));
+  __ vst1(Neon8, NeonListOperand(q0), NeonMemOperand(r4));
+
+  __ Swizzle(q0, q1, kScratch, Neon32, 0x0321);
+  __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, _32x4_0321))));
+  __ vst1(Neon8, NeonListOperand(q0), NeonMemOperand(r4));
+
+  __ Swizzle(q0, q1, kScratch, Neon32, 0x1132);
+  __ add(r4, r0, Operand(static_cast<int32_t>(offsetof(T, _32x4_1132))));
+  __ vst1(Neon8, NeonListOperand(q0), NeonMemOperand(r4));
+
+  __ vmov(q0, q1);
+  __ Swizzle(q0, q0, kScratch, Neon32, 0x1132);
+  __ add(r4, r0,
+         Operand(static_cast<int32_t>(offsetof(T, _32x4_1132_in_place))));
+  __ vst1(Neon8, NeonListOperand(q0), NeonMemOperand(r4));
+
+  __ ldm(ia_w, sp, r4.bit() | r5.bit() | r6.bit() | r7.bit() | pc.bit());
+  __ bx(lr);

[Rodolph Perfetta (ARM), 2016/12/16 20:50:11]

nit: unnecessary.

[bbudge, 2016/12/17 01:08:32]

Done.

+
+  CodeDesc desc;
+  masm->GetCode(&desc);
+  Handle<Code> code = isolate->factory()->NewCode(
+      desc, Code::ComputeFlags(Code::STUB), Handle<Code>());
+#ifdef DEBUG
+  OFStream os(stdout);
+  code->Print(os);
+#endif
+  F3 f = FUNCTION_CAST<F3>(code->entry());
+  Object* dummy = CALL_GENERATED_CODE(isolate, f, &t, 0, 0, 0, 0);
+  USE(dummy);
+  CHECK_EQ_32X4(_32x4_3210, 0, 1, 2, 3);
+  CHECK_EQ_32X4(_32x4_1032, 2, 3, 0, 1);
+  CHECK_EQ_32X4(_32x4_0000, 0, 0, 0, 0);
+  CHECK_EQ_32X4(_32x4_1111, 1, 1, 1, 1);
+  CHECK_EQ_32X4(_32x4_2222, 2, 2, 2, 2);
+  CHECK_EQ_32X4(_32x4_3333, 3, 3, 3, 3);
+  CHECK_EQ_32X4(_32x4_2103, 3, 0, 1, 2);
+  CHECK_EQ_32X4(_32x4_0321, 1, 2, 3, 0);
+  CHECK_EQ_32X4(_32x4_1132, 2, 3, 1, 1);
+  CHECK_EQ_32X4(_32x4_1132_in_place, 2, 3, 1, 1);
+}
+
 #undef __