Backup only the lower 128bits of the XMM callee save registers on Windows

Author: mur47x111

compiler/src/jdk.graal.compiler/src/jdk/graal/compiler/lir/alloc/SaveCalleeSaveRegisters.java

@@ -79,9 +79,11 @@ private static RegisterMap<AllocatableValue> saveAtEntry(LIR lir, LIRGeneratorTo
         List<Register> allocatables = lirGenRes.getRegisterConfig().getAllocatableRegisters();
         RegisterMap<AllocatableValue> saveMap = new RegisterMap<>(arch);
         for (Register register : calleeSaveRegisters) {
-            PlatformKind registerPlatformKind = arch.getLargestStorableKind(register.getRegisterCategory());
+            PlatformKind registerPlatformKind = lirGenRes.getRegisterConfig().getCalleeSaveRegisterStorageKind(arch, register);
             LIRKind lirKind = LIRKind.value(registerPlatformKind);
             RegisterValue registerValue = register.asValue(lirKind);
+            // Force a non-allocatable register to be saved to a stack slot
+            // so as to avoid unnecessary register pressure.
             AllocatableValue saveVariable = allocatables.contains(registerValue.getRegister()) ? lirGen.newVariable(lirKind) : lirGenRes.getFrameMapBuilder().allocateSpillSlot(lirKind);
             LIRInstruction save = lirGen.getSpillMoveFactory().createMove(saveVariable, registerValue);
             buffer.append(insertionIndex, save);

substratevm/src/com.oracle.svm.core.graal.amd64/src/com/oracle/svm/core/graal/amd64/SubstrateAMD64RegisterConfig.java

@@ -64,6 +64,7 @@
 import static jdk.vm.ci.amd64.AMD64.xmm7;
 import static jdk.vm.ci.amd64.AMD64.xmm8;
 import static jdk.vm.ci.amd64.AMD64.xmm9;
+import static jdk.vm.ci.amd64.AMD64Kind.V128_QWORD;
 
 import java.util.ArrayList;
 import java.util.Arrays;
@@ -87,6 +88,7 @@
 import jdk.graal.compiler.core.common.LIRKind;
 import jdk.vm.ci.amd64.AMD64;
 import jdk.vm.ci.amd64.AMD64Kind;
+import jdk.vm.ci.code.Architecture;
 import jdk.vm.ci.code.CallingConvention;
 import jdk.vm.ci.code.CallingConvention.Type;
 import jdk.vm.ci.code.Register;
@@ -239,6 +241,15 @@ public List<Register> getCalleeSaveRegisters() {
         return calleeSaveRegisters;
     }
 
+    @Override
+    public PlatformKind getCalleeSaveRegisterStorageKind(Architecture arch, Register calleeSaveRegister) {
+        if (Platform.includedIn(Platform.WINDOWS.class) && AMD64.XMM.equals(calleeSaveRegister.getRegisterCategory())) {
+            VMError.guarantee(calleeSaveRegister.encoding() >= xmm6.encoding() && calleeSaveRegister.encoding() <= xmm15.encoding(), "unexpected callee saved register");
+            return V128_QWORD;
+        }
+        return SubstrateRegisterConfig.super.getCalleeSaveRegisterStorageKind(arch, calleeSaveRegister);
+    }
+
     @Override
     public List<Register> getCallerSaveRegisters() {
         return getAllocatableRegisters();
@@ -410,7 +421,7 @@ public CallingConvention getCallingConvention(Type t, JavaType returnType, JavaT
             kinds = Arrays.copyOf(kinds, kinds.length + 1);
             locations = Arrays.copyOf(locations, locations.length + 1);
             kinds[kinds.length - 1] = JavaKind.Int;
-            locations[locations.length - 1] = AMD64.rax.asValue(LIRKind.value(AMD64Kind.DWORD));
+            locations[locations.length - 1] = rax.asValue(LIRKind.value(AMD64Kind.DWORD));
             if (type.customABI()) {
                 var extendsParametersAssignment = Arrays.copyOf(type.fixedParameterAssignment, type.fixedParameterAssignment.length + 1);
                 extendsParametersAssignment[extendsParametersAssignment.length - 1] = AssignedLocation.forRegister(rax, JavaKind.Long);