Fix stack alignment corruption for MinGW32 build

Contributed by Peter Varga <pvarga@inf.u-szeged.hu>

BUG=
TEST=cctest/test-assembler-ia32/StackAlignmentForSSE2,cctest/test-assembler-x64/StackAlignmentForSSE2,cctest/test-platform/StackAlignment
R=jkummerow@chromium.org

Review URL: https://codereview.chromium.org/18300003

Patch from Peter Varga <pvarga@inf.u-szeged.hu>.

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@15502 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/platform-win32.cc

@@ -1479,6 +1479,10 @@ double OS::nan_value() {
 int OS::ActivationFrameAlignment() {
 #ifdef _WIN64
   return 16;  // Windows 64-bit ABI requires the stack to be 16-byte aligned.
+#elif defined(__MINGW32__)
+  // With gcc 4.4 the tree vectorization optimizer can generate code
+  // that requires 16 byte alignment such as movdqa on x86.
+  return 16;
 #else
   return 8;  // Floating-point math runs faster with 8-byte alignment.
 #endif

test/cctest/test-assembler-ia32.cc

@@ -473,6 +473,94 @@ TEST(AssemblerMultiByteNop) {
 }
 
 
+#ifdef __GNUC__
+#define ELEMENT_COUNT 4
+
+void DoSSE2(const v8::FunctionCallbackInfo<v8::Value>& args) {
+  Isolate* isolate = reinterpret_cast<Isolate*>(CcTest::isolate());
+  HandleScope scope(isolate);
+
+  CHECK(args[0]->IsArray());
+  v8::Local<v8::Array> vec = v8::Local<v8::Array>::Cast(args[0]);
+  CHECK_EQ(ELEMENT_COUNT, vec->Length());
+
+  v8::internal::byte buffer[256];
+  Assembler assm(isolate, buffer, sizeof buffer);
+
+  ASSERT(CpuFeatures::IsSupported(SSE2));
+  CpuFeatureScope fscope(&assm, SSE2);
+
+  // Remove return address from the stack for fix stack frame alignment.
+  __ pop(ecx);
+
+  // Store input vector on the stack.
+  for (int i = 0; i < ELEMENT_COUNT; ++i) {
+    __ push(Immediate(vec->Get(i)->Int32Value()));
+  }
+
+  // Read vector into a xmm register.
+  __ pxor(xmm0, xmm0);
+  __ movdqa(xmm0, Operand(esp, 0));
+  // Create mask and store it in the return register.
+  __ movmskps(eax, xmm0);
+
+  // Remove unused data from the stack.
+  __ add(esp, Immediate(ELEMENT_COUNT * sizeof(int32_t)));
+  // Restore return address.
+  __ push(ecx);
+
+  __ ret(0);
+
+  CodeDesc desc;
+  assm.GetCode(&desc);
+
+  Object* code = isolate->heap()->CreateCode(
+      desc,
+      Code::ComputeFlags(Code::STUB),
+      Handle<Code>())->ToObjectChecked();
+  CHECK(code->IsCode());
+
+  F0 f = FUNCTION_CAST<F0>(Code::cast(code)->entry());
+  int res = f();
+  args.GetReturnValue().Set(v8::Integer::New(res));
+}
+
+TEST(StackAlignmentForSSE2) {
+  CcTest::InitializeVM();
+  if (!CpuFeatures::IsSupported(SSE2)) return;
+
+  CHECK_EQ(0, OS::ActivationFrameAlignment() % 16);
+
+  v8::Isolate* isolate = v8::Isolate::GetCurrent();
+  v8::HandleScope handle_scope(isolate);
+  v8::Handle<v8::ObjectTemplate> global_template = v8::ObjectTemplate::New();
+  global_template->Set(v8_str("do_sse2"), v8::FunctionTemplate::New(DoSSE2));
+
+  LocalContext env(NULL, global_template);
+  CompileRun(
+      "function foo(vec) {"
+      "  return do_sse2(vec);"
+      "}");
+
+  v8::Local<v8::Object> global_object = env->Global();
+  v8::Local<v8::Function> foo =
+      v8::Local<v8::Function>::Cast(global_object->Get(v8_str("foo")));
+
+  int32_t vec[ELEMENT_COUNT] = { -1, 1, 1, 1 };
+  v8::Local<v8::Array> v8_vec = v8::Array::New(ELEMENT_COUNT);
+  for (int i = 0; i < ELEMENT_COUNT; i++) {
+      v8_vec->Set(i, v8_num(vec[i]));
+  }
+
+  v8::Local<v8::Value> args[] = { v8_vec };
+  v8::Local<v8::Value> result = foo->Call(global_object, 1, args);
+
+  // The mask should be 0b1000.
+  CHECK_EQ(8, result->Int32Value());
+}
+
+#undef ELEMENT_COUNT
+#endif  // __GNUC__
 
 
 #undef __

test/cctest/test-assembler-x64.cc

@@ -62,6 +62,7 @@ using v8::internal::rdx;
 using v8::internal::rsi;
 using v8::internal::rsp;
 using v8::internal::times_1;
+using v8::internal::xmm0;
 
 // Test the x64 assembler by compiling some simple functions into
 // a buffer and executing them.  These tests do not initialize the
@@ -429,6 +430,91 @@ TEST(AssemblerMultiByteNop) {
 }
 
 
+#ifdef __GNUC__
+#define ELEMENT_COUNT 4
+
+void DoSSE2(const v8::FunctionCallbackInfo<v8::Value>& args) {
+  CcTest::InitializeVM();
+  v8::HandleScope scope(CcTest::isolate());
+  v8::internal::byte buffer[1024];
+
+  CHECK(args[0]->IsArray());
+  v8::Local<v8::Array> vec = v8::Local<v8::Array>::Cast(args[0]);
+  CHECK_EQ(ELEMENT_COUNT, vec->Length());
+
+  Isolate* isolate = Isolate::Current();
+  Assembler assm(isolate, buffer, sizeof(buffer));
+
+  // Remove return address from the stack for fix stack frame alignment.
+  __ pop(rcx);
+
+  // Store input vector on the stack.
+  for (int i = 0; i < ELEMENT_COUNT; i++) {
+    __ movl(rax, Immediate(vec->Get(i)->Int32Value()));
+    __ shl(rax, Immediate(0x20));
+    __ or_(rax, Immediate(vec->Get(++i)->Int32Value()));
+    __ push(rax);
+  }
+
+  // Read vector into a xmm register.
+  __ xorps(xmm0, xmm0);
+  __ movdqa(xmm0, Operand(rsp, 0));
+  // Create mask and store it in the return register.
+  __ movmskps(rax, xmm0);
+
+  // Remove unused data from the stack.
+  __ addq(rsp, Immediate(ELEMENT_COUNT * sizeof(int32_t)));
+  // Restore return address.
+  __ push(rcx);
+
+  __ ret(0);
+
+  CodeDesc desc;
+  assm.GetCode(&desc);
+  Code* code = Code::cast(isolate->heap()->CreateCode(
+      desc,
+      Code::ComputeFlags(Code::STUB),
+      v8::internal::Handle<Code>())->ToObjectChecked());
+  CHECK(code->IsCode());
+
+  F0 f = FUNCTION_CAST<F0>(code->entry());
+  int res = f();
+  args.GetReturnValue().Set(v8::Integer::New(res));
+}
+
+TEST(StackAlignmentForSSE2) {
+  CHECK_EQ(0, OS::ActivationFrameAlignment() % 16);
+
+  v8::Isolate* isolate = v8::Isolate::GetCurrent();
+  v8::HandleScope handle_scope(isolate);
+  v8::Handle<v8::ObjectTemplate> global_template = v8::ObjectTemplate::New();
+  global_template->Set(v8_str("do_sse2"), v8::FunctionTemplate::New(DoSSE2));
+
+  LocalContext env(NULL, global_template);
+  CompileRun(
+      "function foo(vec) {"
+      "  return do_sse2(vec);"
+      "}");
+
+  v8::Local<v8::Object> global_object = env->Global();
+  v8::Local<v8::Function> foo =
+      v8::Local<v8::Function>::Cast(global_object->Get(v8_str("foo")));
+
+  int32_t vec[ELEMENT_COUNT] = { -1, 1, 1, 1 };
+  v8::Local<v8::Array> v8_vec = v8::Array::New(ELEMENT_COUNT);
+  for (int i = 0; i < ELEMENT_COUNT; i++) {
+    v8_vec->Set(i, v8_num(vec[i]));
+  }
+
+  v8::Local<v8::Value> args[] = { v8_vec };
+  v8::Local<v8::Value> result = foo->Call(global_object, 1, args);
+
+  // The mask should be 0b1000.
+  CHECK_EQ(8, result->Int32Value());
+}
+
+#undef ELEMENT_COUNT
+#endif  // __GNUC__
 
 
 #undef __

test/cctest/test-platform.cc

@@ -35,3 +35,65 @@ using namespace ::v8::internal;
 TEST(NumberOfCores) {
   CHECK_GT(OS::NumberOfCores(), 0);
 }
+
+
+#ifdef __GNUC__
+#define ASM __asm__ __volatile__
+
+#if defined(_M_X64) || defined(__x86_64__)
+#define GET_STACK_POINTER() \
+  static int sp_addr = 0; \
+  do { \
+    ASM("mov %%rsp, %0" : "=g" (sp_addr)); \
+  } while (0)
+#elif defined(_M_IX86) || defined(__i386__)
+#define GET_STACK_POINTER() \
+  static int sp_addr = 0; \
+  do { \
+    ASM("mov %%esp, %0" : "=g" (sp_addr)); \
+  } while (0)
+#elif defined(__ARMEL__)
+#define GET_STACK_POINTER() \
+  static int sp_addr = 0; \
+  do { \
+    ASM("str %%sp, %0" : "=g" (sp_addr)); \
+  } while (0)
+#elif defined(__MIPSEL__)
+#define GET_STACK_POINTER() \
+  static int sp_addr = 0; \
+  do { \
+    ASM("sw $sp, %0" : "=g" (sp_addr)); \
+  } while (0)
+#else
+#error Host architecture was not detected as supported by v8
+#endif
+
+void GetStackPointer(const v8::FunctionCallbackInfo<v8::Value>& args) {
+  GET_STACK_POINTER();
+  args.GetReturnValue().Set(v8_num(sp_addr));
+}
+
+TEST(StackAlignment) {
+  v8::Isolate* isolate = v8::Isolate::GetCurrent();
+  v8::HandleScope handle_scope(isolate);
+  v8::Handle<v8::ObjectTemplate> global_template = v8::ObjectTemplate::New();
+  global_template->Set(v8_str("get_stack_pointer"),
+                       v8::FunctionTemplate::New(GetStackPointer));
+
+  LocalContext env(NULL, global_template);
+  CompileRun(
+      "function foo() {"
+      "  return get_stack_pointer();"
+      "}");
+
+  v8::Local<v8::Object> global_object = env->Global();
+  v8::Local<v8::Function> foo =
+      v8::Local<v8::Function>::Cast(global_object->Get(v8_str("foo")));
+
+  v8::Local<v8::Value> result = foo->Call(global_object, 0, NULL);
+  CHECK_EQ(0, result->Int32Value() % OS::ActivationFrameAlignment());
+}
+
+#undef GET_STACK_POINTERS
+#undef ASM
+#endif  // __GNUC__