[codegen] Save the full 128bit FP register for Push/PopCallerSaved

Note that Arm32 is already saving the full register in
https://source.chromium.org/chromium/chromium/src/+/main:v8/src/codegen/arm/macro-assembler-arm.cc;l=2250;drc=ec4fd32cf7f945923fa6bb332c061ecbdaaaa405

Change-Id: I1f5fe60ca350583fb4cb877ccad74f5e260c3665
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2912778
Commit-Queue: Santiago Aboy Solanes <solanes@chromium.org>
Reviewed-by: Zhi An Ng <zhin@chromium.org>
Cr-Commit-Position: refs/heads/master@{#74778}

src/codegen/arm64/macro-assembler-arm64.cc

@@ -45,17 +45,30 @@ CPURegList TurboAssembler::DefaultFPTmpList() {
   return CPURegList(fp_scratch1, fp_scratch2);
 }
 
+namespace {
+
+// For WebAssembly we care about the full floating point register. If we are not
+// running Wasm, we can get away with saving half of those registers.
+#if V8_ENABLE_WEBASSEMBLY
+constexpr int kStackSavedSavedFPSizeInBits = kQRegSizeInBits;
+#else
+constexpr int kStackSavedSavedFPSizeInBits = kDRegSizeInBits;
+#endif  // V8_ENABLE_WEBASSEMBLY
+
+}  // namespace
+
 int TurboAssembler::RequiredStackSizeForCallerSaved(SaveFPRegsMode fp_mode,
                                                     Register exclusion) const {
   auto list = kCallerSaved;
   list.Remove(exclusion);
   list.Align();
 
-  int bytes = list.Count() * kXRegSizeInBits / 8;
+  int bytes = list.TotalSizeInBytes();
 
   if (fp_mode == SaveFPRegsMode::kSave) {
-    DCHECK_EQ(kCallerSavedV.Count() % 2, 0);
-    bytes += kCallerSavedV.Count() * kDRegSizeInBits / 8;
+    auto fp_list = CPURegList::GetCallerSavedV(kStackSavedSavedFPSizeInBits);
+    DCHECK_EQ(fp_list.Count() % 2, 0);
+    bytes += fp_list.TotalSizeInBytes();
   }
   return bytes;
 }
@@ -68,12 +81,13 @@ int TurboAssembler::PushCallerSaved(SaveFPRegsMode fp_mode,
 
   PushCPURegList<kDontStoreLR>(list);
 
-  int bytes = list.Count() * kXRegSizeInBits / 8;
+  int bytes = list.TotalSizeInBytes();
 
   if (fp_mode == SaveFPRegsMode::kSave) {
-    DCHECK_EQ(kCallerSavedV.Count() % 2, 0);
-    PushCPURegList(kCallerSavedV);
-    bytes += kCallerSavedV.Count() * kDRegSizeInBits / 8;
+    auto fp_list = CPURegList::GetCallerSavedV(kStackSavedSavedFPSizeInBits);
+    DCHECK_EQ(fp_list.Count() % 2, 0);
+    PushCPURegList(fp_list);
+    bytes += fp_list.TotalSizeInBytes();
   }
   return bytes;
 }
@@ -81,9 +95,10 @@ int TurboAssembler::PushCallerSaved(SaveFPRegsMode fp_mode,
 int TurboAssembler::PopCallerSaved(SaveFPRegsMode fp_mode, Register exclusion) {
   int bytes = 0;
   if (fp_mode == SaveFPRegsMode::kSave) {
-    DCHECK_EQ(kCallerSavedV.Count() % 2, 0);
-    PopCPURegList(kCallerSavedV);
-    bytes += kCallerSavedV.Count() * kDRegSizeInBits / 8;
+    auto fp_list = CPURegList::GetCallerSavedV(kStackSavedSavedFPSizeInBits);
+    DCHECK_EQ(fp_list.Count() % 2, 0);
+    PopCPURegList(fp_list);
+    bytes += fp_list.TotalSizeInBytes();
   }
 
   auto list = kCallerSaved;
@@ -91,7 +106,7 @@ int TurboAssembler::PopCallerSaved(SaveFPRegsMode fp_mode, Register exclusion) {
   list.Align();
 
   PopCPURegList<kDontLoadLR>(list);
-  bytes += list.Count() * kXRegSizeInBits / 8;
+  bytes += list.TotalSizeInBytes();
 
   return bytes;
 }

src/codegen/ia32/macro-assembler-ia32.cc

@@ -296,7 +296,7 @@ int TurboAssembler::RequiredStackSizeForCallerSaved(SaveFPRegsMode fp_mode,
 
   if (fp_mode == SaveFPRegsMode::kSave) {
     // Count all XMM registers except XMM0.
-    bytes += kDoubleSize * (XMMRegister::kNumRegisters - 1);
+    bytes += kStackSavedSavedFPSize * (XMMRegister::kNumRegisters - 1);
   }
 
   return bytes;
@@ -318,11 +318,15 @@ int TurboAssembler::PushCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1,
 
   if (fp_mode == SaveFPRegsMode::kSave) {
     // Save all XMM registers except XMM0.
-    int delta = kDoubleSize * (XMMRegister::kNumRegisters - 1);
+    const int delta = kStackSavedSavedFPSize * (XMMRegister::kNumRegisters - 1);
     AllocateStackSpace(delta);
     for (int i = XMMRegister::kNumRegisters - 1; i > 0; i--) {
       XMMRegister reg = XMMRegister::from_code(i);
-      movsd(Operand(esp, (i - 1) * kDoubleSize), reg);
+#if V8_ENABLE_WEBASSEMBLY
+      movdqu(Operand(esp, (i - 1) * kStackSavedSavedFPSize), reg);
+#else
+      movsd(Operand(esp, (i - 1) * kStackSavedSavedFPSize), reg);
+#endif  // V8_ENABLE_WEBASSEMBLY
     }
     bytes += delta;
   }
@@ -335,10 +339,14 @@ int TurboAssembler::PopCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1,
   int bytes = 0;
   if (fp_mode == SaveFPRegsMode::kSave) {
     // Restore all XMM registers except XMM0.
-    int delta = kDoubleSize * (XMMRegister::kNumRegisters - 1);
+    const int delta = kStackSavedSavedFPSize * (XMMRegister::kNumRegisters - 1);
     for (int i = XMMRegister::kNumRegisters - 1; i > 0; i--) {
       XMMRegister reg = XMMRegister::from_code(i);
-      movsd(reg, Operand(esp, (i - 1) * kDoubleSize));
+#if V8_ENABLE_WEBASSEMBLY
+      movdqu(reg, Operand(esp, (i - 1) * kStackSavedSavedFPSize));
+#else
+      movsd(reg, Operand(esp, (i - 1) * kStackSavedSavedFPSize));
+#endif  // V8_ENABLE_WEBASSEMBLY
     }
     add(esp, Immediate(delta));
     bytes += delta;

src/codegen/shared-ia32-x64/macro-assembler-shared-ia32-x64.h

@@ -21,6 +21,14 @@ namespace v8 {
 namespace internal {
 class Assembler;
 
+// For WebAssembly we care about the full floating point register. If we are not
+// running Wasm, we can get away with saving half of those registers.
+#if V8_ENABLE_WEBASSEMBLY
+constexpr int kStackSavedSavedFPSize = 2 * kDoubleSize;
+#else
+constexpr int kStackSavedSavedFPSize = kDoubleSize;
+#endif  // V8_ENABLE_WEBASSEMBLY
+
 class V8_EXPORT_PRIVATE SharedTurboAssembler : public TurboAssemblerBase {
  public:
   using TurboAssemblerBase::TurboAssemblerBase;

src/codegen/x64/macro-assembler-x64.cc

@@ -690,7 +690,7 @@ int TurboAssembler::RequiredStackSizeForCallerSaved(SaveFPRegsMode fp_mode,
 
   // R12 to r15 are callee save on all platforms.
   if (fp_mode == SaveFPRegsMode::kSave) {
-    bytes += kDoubleSize * XMMRegister::kNumRegisters;
+    bytes += kStackSavedSavedFPSize * XMMRegister::kNumRegisters;
   }
 
   return bytes;
@@ -712,11 +712,15 @@ int TurboAssembler::PushCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1,
 
   // R12 to r15 are callee save on all platforms.
   if (fp_mode == SaveFPRegsMode::kSave) {
-    int delta = kDoubleSize * XMMRegister::kNumRegisters;
+    const int delta = kStackSavedSavedFPSize * XMMRegister::kNumRegisters;
     AllocateStackSpace(delta);
     for (int i = 0; i < XMMRegister::kNumRegisters; i++) {
       XMMRegister reg = XMMRegister::from_code(i);
-      Movsd(Operand(rsp, i * kDoubleSize), reg);
+#if V8_ENABLE_WEBASSEMBLY
+      Movdqu(Operand(rsp, i * kStackSavedSavedFPSize), reg);
+#else
+      Movsd(Operand(rsp, i * kStackSavedSavedFPSize), reg);
+#endif  // V8_ENABLE_WEBASSEMBLY
     }
     bytes += delta;
   }
@@ -730,10 +734,14 @@ int TurboAssembler::PopCallerSaved(SaveFPRegsMode fp_mode, Register exclusion1,
   if (fp_mode == SaveFPRegsMode::kSave) {
     for (int i = 0; i < XMMRegister::kNumRegisters; i++) {
       XMMRegister reg = XMMRegister::from_code(i);
-      Movsd(reg, Operand(rsp, i * kDoubleSize));
+#if V8_ENABLE_WEBASSEMBLY
+      Movdqu(reg, Operand(rsp, i * kStackSavedSavedFPSize));
+#else
+      Movsd(reg, Operand(rsp, i * kStackSavedSavedFPSize));
+#endif  // V8_ENABLE_WEBASSEMBLY
     }
-    int delta = kDoubleSize * XMMRegister::kNumRegisters;
-    addq(rsp, Immediate(kDoubleSize * XMMRegister::kNumRegisters));
+    const int delta = kStackSavedSavedFPSize * XMMRegister::kNumRegisters;
+    addq(rsp, Immediate(delta));
     bytes += delta;
   }