work in progress towards 64-bit ARM JIT support

This won't build, it's just a snapshot of what I have so far.

Conflicts:
	include/avian/codegen/architecture.h
	include/avian/codegen/registers.h
	src/codegen/compiler.cpp
	src/codegen/compiler/event.cpp
	src/codegen/compiler/site.cpp
	src/codegen/compiler/site.h
	src/codegen/registers.cpp
	src/codegen/target/arm/assembler.cpp
	src/codegen/target/arm/registers.h

include/avian/codegen/architecture.h

@@ -28,8 +28,6 @@ namespace codegen {
 
 class Assembler;
 
-class RegisterFile;
-
 class OperandMask {
  public:
   uint8_t typeMask;

src/codegen/compiler/site.h

@@ -123,7 +123,7 @@ class Site {
 
   virtual RegisterMask registerMask(Context*)
   {
-    return 0;
+    return RegisterMask(0);
   }
 
   virtual bool isVolatile(Context*)

src/codegen/target/arm/assembler.cpp

@@ -39,7 +39,7 @@ namespace isa {
 bool vfpSupported()
 {
 // TODO: Use at runtime detection
-#if defined(__ARM_PCS_VFP)
+#if (defined __ARM_PCS_VFP) || (defined ARCH_arm64)
   // armhf
   return true;
 #else
@@ -55,9 +55,9 @@ bool vfpSupported()
 const RegisterFile MyRegisterFileWithoutFloats(GPR_MASK, 0);
 const RegisterFile MyRegisterFileWithFloats(GPR_MASK, FPR_MASK);
 
-const unsigned FrameHeaderSize = 1;
+const unsigned FrameHeaderSize = TargetBytesPerWord / 4;
 
-const unsigned StackAlignmentInBytes = 8;
+const unsigned StackAlignmentInBytes = TargetBytesPerWord * 2;
 const unsigned StackAlignmentInWords = StackAlignmentInBytes
                                        / TargetBytesPerWord;
 
@@ -258,7 +258,7 @@ class MyArchitecture : public Architecture {
 
   virtual unsigned argumentRegisterCount()
   {
-    return 4;
+    return TargetBytesPerWord;
   }
 
   virtual Register argumentRegister(unsigned index)
@@ -434,11 +434,11 @@ class MyArchitecture : public Architecture {
       break;
 
     case lir::Float2Int:
-      // todo: Java requires different semantics than SSE for
+      // todo: Java requires different semantics than VFP for
       // converting floats to integers, we we need to either use
       // thunks or produce inline machine code which handles edge
       // cases properly.
-      if (false && vfpSupported() && bSize == 4) {
+      if (false && vfpSupported() && bSize <= TargetBytesPerWord) {
         aMask.typeMask = lir::Operand::RegisterPairMask;
         aMask.setLowHighRegisterMasks(FPR_MASK, FPR_MASK);
       } else {
@@ -447,7 +447,7 @@ class MyArchitecture : public Architecture {
       break;
 
     case lir::Int2Float:
-      if (vfpSupported() && aSize == 4) {
+      if (vfpSupported() && aSize <= TargetBytesPerWord) {
         aMask.typeMask = lir::Operand::RegisterPairMask;
         aMask.setLowHighRegisterMasks(GPR_MASK, GPR_MASK);
       } else {
@@ -544,7 +544,7 @@ class MyArchitecture : public Architecture {
     case lir::ShiftLeft:
     case lir::ShiftRight:
     case lir::UnsignedShiftRight:
-      if (bSize == 8)
+      if (bSize > TargetBytesPerWord)
         aMask.typeMask = bMask.typeMask = lir::Operand::RegisterPairMask;
       break;
 

src/codegen/target/arm/operations32.cpp

@@ -15,6 +15,8 @@
 #include "fixup.h"
 #include "multimethod.h"
 
+#if AVIAN_TARGET_ARCH == AVIAN_ARCH_ARM
+
 namespace avian {
 namespace codegen {
 namespace arm {
@@ -1554,3 +1556,5 @@ void storeLoadBarrier(Context* con)
 }  // namespace arm
 }  // namespace codegen
 }  // namespace avian
+
+#endif // AVIAN_TARGET_ARCH == AVIAN_ARCH_ARM

src/codegen/target/arm/registers.h

@@ -21,10 +21,29 @@ namespace arm {
 const uint64_t MASK_LO32 = 0xffffffff;
 const unsigned MASK_LO8 = 0xff;
 
+#ifdef ARCH_arm64
+constexpr Register ThreadRegister(19);
+constexpr Register StackRegister(31);
+constexpr Register LinkRegister(30);
+constexpr Register LinkRegister(29);
+constexpr Register ProgramCounter(0xFE); // i.e. unaddressable
+
+const int N_GPRS = 32;
+const int N_FPRS = 32;
+const uint64_t GPR_MASK = 0xffffffff;
+const uint64_t FPR_MASK = 0xffffffff00000000;
+#else
+constexpr Register ThreadRegister(8);
+constexpr Register StackRegister(13);
+constexpr Register LinkRegister(14);
+constexpr Register FrameRegister(0xFE); // i.e. there is none
+constexpr Register ProgramCounter(15);
+
 const int N_GPRS = 16;
 const int N_FPRS = 16;
 const RegisterMask GPR_MASK = 0xffff;
 const RegisterMask FPR_MASK = 0xffff0000;
+#endif
 
 inline bool isFpr(lir::RegisterPair* reg)
 {
@@ -48,18 +67,6 @@ inline int fpr32(lir::RegisterPair* reg)
   return fpr64(reg) << 1;
 }
 
-#ifdef ARCH_arm64
-constexpr Register ThreadRegister(19);
-constexpr Register StackRegister(31);
-constexpr Register LinkRegister(30);
-constexpr Register ProgramCounter(0xFE); // i.e. unaddressable
-#else
-constexpr Register ThreadRegister(8);
-constexpr Register StackRegister(13);
-constexpr Register LinkRegister(14);
-constexpr Register ProgramCounter(15);
-#endif
-
 }  // namespace arm
 }  // namespace codegen
 }  // namespace avian

src/compile-arm.S

@@ -16,20 +16,122 @@
 #define BYTES_PER_WORD 4
 
 #define LOCAL(x) .L##x
-   
+
 #ifdef __APPLE__
 #  define GLOBAL(x) _##x
 #else
-#  define GLOBAL(x) x   
+#  define GLOBAL(x) x
 #endif
 
+#ifdef __aarch64__
+
+.globl GLOBAL(vmInvoke)
+.align 2
+GLOBAL(vmInvoke):
+  // arguments:
+  //   x0 : thread
+  //   x1 : function
+  //   x2 : arguments
+  //   w3 : argumentFootprint
+  //   w4 : frameSize (not used)
+  //   w5 : returnType
+
+  // allocate frame
+  stp   x29, x30, [sp,#-96]!
+
+  // save callee-saved register values
+  stp   x19, x20, [sp,#16]
+  stp   x21, x22, [sp,#32]
+  stp   x23, x24, [sp,#48]
+  stp   x25, x26, [sp,#64]
+  stp   x27, x28, [sp,#80]
+
+  // save return type
+  str   w5, [sp,#-16]!
+
+  mov   x5, sp
+  str   x5, [x0,#TARGET_THREAD_SCRATCH]
+
+  // copy arguments into place
+  sub   sp, sp, w3
+  mov   x5, #0
+  b     LOCAL(vmInvoke_argumentTest)
+
+LOCAL(vmInvoke_argumentLoop):
+  ldr   x5, [x2, x4]
+  str   x5, [sp, x4]
+  add   x4, x4, #BYTES_PER_WORD
+
+LOCAL(vmInvoke_argumentTest):
+  cmp   x4, x3
+  blt   LOCAL(vmInvoke_argumentLoop)
+
+  // we use x19 to hold the thread pointer, by convention
+  mov   x19, x0
+
+  // load and call function address
+  blr   x1
+
+.globl GLOBAL(vmInvoke_returnAddress)
+.align 2
+GLOBAL(vmInvoke_returnAddress):
+  // restore stack pointer
+  ldr   x5, [x19, #TARGET_THREAD_SCRATCH]
+  mov   sp, x5
+
+  // clear MyThread::stack to avoid confusing another thread calling
+  // java.lang.Thread.getStackTrace on this one.  See
+  // MyProcess::getStackTrace in compile.cpp for details on how we get
+  // a reliable stack trace from a thread that might be interrupted at
+  // any point in its execution.
+  mov  x5, #0
+  str  x5, [x19, #TARGET_THREAD_STACK]
+
+.globl GLOBAL(vmInvoke_safeStack)
+.align 2
+GLOBAL(vmInvoke_safeStack):
+
+#ifdef AVIAN_CONTINUATIONS
+#error todo
+#endif // AVIAN_CONTINUATIONS
+
+  mov   x5, #0
+  str   x5, [x19, #TARGET_THREAD_STACK]
+
+  // restore return type
+  ldr   w5, [sp], #4
+
+  // restore callee-saved register values
+  ldp   x19, x20, [sp,#16]
+  ldp   x21, x22, [sp,#32]
+  ldp   x23, x24, [sp,#48]
+  ldp   x25, x26, [sp,#64]
+  ldp   x27, x28, [sp,#80]
+  ldp   x29, x30, [sp],#96
+
+LOCAL(vmInvoke_return):
+  br    x30
+
+.globl GLOBAL(vmJumpAndInvoke)
+.align 2
+GLOBAL(vmJumpAndInvoke):
+#ifdef AVIAN_CONTINUATIONS
+#error todo
+#else // not AVIAN_CONTINUATIONS
+   // vmJumpAndInvoke should only be called when continuations are
+   // enabled, so we force a crash if we reach here:
+   brk 0
+#endif // not AVIAN_CONTINUATIONS
+
+#elif defined __arm__
+
 #define CONTINUATION_NEXT 4
 #define CONTINUATION_ADDRESS 16
 #define CONTINUATION_RETURN_ADDRESS_OFFSET 20
 #define CONTINUATION_FRAME_POINTER_OFFSET 24
 #define CONTINUATION_LENGTH 28
 #define CONTINUATION_BODY 32
-   
+
 .globl GLOBAL(vmInvoke)
 .align 2
 GLOBAL(vmInvoke):
@@ -56,7 +158,7 @@ GLOBAL(vmInvoke):
   eor   r4, sp, r3
   tst   r4, #4
   subne sp, sp, #4
-   
+
   // copy arguments into place
   sub   sp, r3
   mov   r4, #0
@@ -87,7 +189,7 @@ LOCAL(vmInvoke_argumentTest):
 GLOBAL(vmInvoke_returnAddress):
   // restore stack pointer
   ldr   sp, [r8, #TARGET_THREAD_SCRATCH]
-   
+
    // clear MyThread::stack to avoid confusing another thread calling
    // java.lang.Thread.getStackTrace on this one.  See
    // MyProcess::getStackTrace in compile.cpp for details on how we get
@@ -201,7 +303,7 @@ GLOBAL(vmJumpAndInvoke):
    // which is not true in this case
    sub  r2,r2,r6
    sub  r2,r2,#84
-   
+
    mov  r8,r0
 
    // copy arguments into place
@@ -220,7 +322,7 @@ LOCAL(vmJumpAndInvoke_argumentTest):
    // the arguments have been copied, so we can set the real stack
    // pointer now
    mov  sp,r2
-   
+
    // set return address to vmInvoke_returnAddress
 #ifdef __APPLE__
    movw r11, :lower16:(GLOBAL(vmInvoke_returnAddress)-(LOCAL(vmJumpAndInvoke_getAddress)+8))
@@ -246,10 +348,12 @@ LOCAL(vmInvoke_getAddress_word):
 LOCAL(vmJumpAndInvoke_getAddress_word):
    .word _GLOBAL_OFFSET_TABLE_-(LOCAL(vmJumpAndInvoke_getAddress)+8)
 #endif // not __APPLE__
-   
+
 #else // not AVIAN_CONTINUATIONS
    // vmJumpAndInvoke should only be called when continuations are
    // enabled, so we force a crash if we reach here:
    mov  r1,#0
    ldr  r1,[r1]
 #endif // not AVIAN_CONTINUATIONS
+
+#endif // __arm__