commit b13c65bb46544821a84ff2106d0710d77b0fb463
author Aart Bik <ajcbik@google.com> Tue Mar 21 20:14:07 2017 -0700
committer Aart Bik <ajcbik@google.com> Wed Mar 22 20:11:05 2017 +0000
tree 46ef54ce881e32c80901528cf21a7951cf219023
parent 2b864659d4399de6c17e93b8df8cdbf08c6a7ac9

Saves full XMM state along suspend check's slow path.

Rationale:
Break-out CL of ART Vectorizer. We need to save 128-bit
of data (default ABI of ART runtime only saves 64-bit)
Note that this is *only* done for xmm registers that
are live, so overhead is not too big.

Bug: 34083438
Test: test-art-host
Change-Id: Ic89988b0acb0c104634271d0c6c3e29b6596d59b

compiler/optimizing/code_generator_x86.cc

diff --git a/compiler/optimizing/code_generator_x86.cc b/compiler/optimizing/code_generator_x86.cc
index 0b50619..958c1a6 100644
--- a/compiler/optimizing/code_generator_x86.cc
+++ b/compiler/optimizing/code_generator_x86.cc
@@ -183,10 +183,13 @@
       : SlowPathCode(instruction), successor_(successor) {}
 
   void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
+    LocationSummary* locations = instruction_->GetLocations();
     CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen);
     __ Bind(GetEntryLabel());
+    SaveLiveRegisters(codegen, locations);  // only saves full width XMM for SIMD
     x86_codegen->InvokeRuntime(kQuickTestSuspend, instruction_, instruction_->GetDexPc(), this);
     CheckEntrypointTypes<kQuickTestSuspend, void, void>();
+    RestoreLiveRegisters(codegen, locations);  // only saves full width XMM for SIMD
     if (successor_ == nullptr) {
       __ jmp(GetReturnLabel());
     } else {
@@ -963,12 +966,20 @@
 }
 
 size_t CodeGeneratorX86::SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) {
-  __ movsd(Address(ESP, stack_index), XmmRegister(reg_id));
+  if (GetGraph()->HasSIMD()) {
+    __ movupd(Address(ESP, stack_index), XmmRegister(reg_id));
+  } else {
+    __ movsd(Address(ESP, stack_index), XmmRegister(reg_id));
+  }
   return GetFloatingPointSpillSlotSize();
 }
 
 size_t CodeGeneratorX86::RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) {
-  __ movsd(XmmRegister(reg_id), Address(ESP, stack_index));
+  if (GetGraph()->HasSIMD()) {
+    __ movupd(XmmRegister(reg_id), Address(ESP, stack_index));
+  } else {
+    __ movsd(XmmRegister(reg_id), Address(ESP, stack_index));
+  }
   return GetFloatingPointSpillSlotSize();
 }
 
@@ -5699,7 +5710,12 @@
 void LocationsBuilderX86::VisitSuspendCheck(HSuspendCheck* instruction) {
   LocationSummary* locations =
       new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnSlowPath);
-  locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty());  // No caller-save registers.
+  // In suspend check slow path, usually there are no caller-save registers at all.
+  // If SIMD instructions are present, however, we force spilling all live SIMD
+  // registers in full width (since the runtime only saves/restores lower part).
+  locations->SetCustomSlowPathCallerSaves(GetGraph()->HasSIMD()
+                                          ? RegisterSet::AllFpu()
+                                          : RegisterSet::Empty());
 }
 
 void InstructionCodeGeneratorX86::VisitSuspendCheck(HSuspendCheck* instruction) {