fix x86->powerpc boot image cross build

This fixes the remaining cross-endian translation issues needed to
build powerpc boot images on x86.

src/arm.cpp

@@ -1783,6 +1783,10 @@ class MyArchitecture: public Assembler::Architecture {
     return 0;
   }
 
+  virtual int scratch() {
+    return 5;
+  }
+
   virtual int stack() {
     return StackRegister;
   }

src/assembler.h

@@ -329,6 +329,7 @@ class Assembler {
     virtual uint32_t generalRegisterMask() = 0;
     virtual uint32_t floatRegisterMask() = 0;
 
+    virtual int scratch() = 0;
     virtual int stack() = 0;
     virtual int thread() = 0;
     virtual int returnLow() = 0;

src/bootimage.cpp

@@ -672,7 +672,7 @@ makeCodeImage(Thread* t, Zone* zone, BootImage* image, uint8_t* code,
     expect(t, value >= code);
 
     addresses->listener->resolve
-      (targetVW(static_cast<target_intptr_t>(value - code)), 0);
+      (static_cast<target_intptr_t>(value - code), 0);
   }
 
   for (; methods; methods = pairSecond(t, methods)) {
@@ -987,7 +987,7 @@ copy(Thread* t, object typeMaps, object p, uint8_t* dst)
         if (field->type == Type_object) {
           unsigned offset = field->targetOffset / TargetBytesPerWord;
           reinterpret_cast<uint32_t*>(mask)[offset / 32]
-            |= static_cast<uint32_t>(1) << (offset % 32);
+            |= targetV4(static_cast<uint32_t>(1) << (offset % 32));
         }
       }
 
@@ -1027,14 +1027,15 @@ copy(Thread* t, object typeMaps, object p, uint8_t* dst)
         switch (field->type) {
         case Type_object:
           reinterpret_cast<uint32_t*>(objectMask)[i / 32]
-            |= static_cast<uint32_t>(1) << (i % 32);
+            |= targetV4(static_cast<uint32_t>(1) << (i % 32));
           break;
 
         case Type_float:
         case Type_double:
           reinterpret_cast<target_uintptr_t*>(poolMask)
             [i / TargetBitsPerWord]
-            |= static_cast<target_uintptr_t>(1) << (i % TargetBitsPerWord);
+            |= targetVW
+            (static_cast<target_uintptr_t>(1) << (i % TargetBitsPerWord));
           break;
 
         default:
@@ -1083,7 +1084,7 @@ copy(Thread* t, object typeMaps, object referer, unsigned refererOffset,
       if (field->type == Type_object) {
         unsigned offset = field->targetOffset / TargetBytesPerWord;
         reinterpret_cast<uint32_t*>(dst + (TargetBytesPerWord * 2))
-          [offset / 32] |= static_cast<uint32_t>(1) << (offset % 32);
+          [offset / 32] |= targetV4(static_cast<uint32_t>(1) << (offset % 32));
       }
     }
 
@@ -1092,7 +1093,7 @@ copy(Thread* t, object typeMaps, object referer, unsigned refererOffset,
     {
       unsigned offset = map->targetFixedSizeInWords;
       reinterpret_cast<uint32_t*>(dst + (TargetBytesPerWord * 2))
-        [offset / 32] |= static_cast<uint32_t>(1) << (offset % 32);
+        [offset / 32] |= targetV4(static_cast<uint32_t>(1) << (offset % 32));
     }
   } else {
     copy(t, typeMaps, p, dst);

src/compile.cpp

@@ -8347,8 +8347,14 @@ class MyProcessor: public Processor {
 #define THUNK(s) thunkTable[s##Index] = voidPointer(s);
 #include "thunks.cpp"
 #undef THUNK
+    // Set the dummyIndex entry to a constant which should require the
+    // maximum number of bytes to represent in assembly code
+    // (i.e. can't be represented by a smaller number of bytes and
+    // implicitly sign- or zero-extended).  We'll use this property
+    // later to determine the maximum size of a thunk in the thunk
+    // table.
     thunkTable[dummyIndex] = reinterpret_cast<void*>
-      (~static_cast<uintptr_t>(0));
+      (static_cast<uintptr_t>(UINT64_C(0x5555555555555555)));
   }
 
   virtual Thread*
@@ -9396,10 +9402,7 @@ compileCall(MyThread* t, Context* c, ThunkIndex index, bool call = true)
 
   if (processor(t)->bootImage) {
     Assembler::Memory table(t->arch->thread(), TargetThreadThunkTable);
-    // use Architecture::virtualCallTarget register here as a scratch
+    Assembler::Register scratch(t->arch->scratch());
-    // register; any register that isn't used to pass arguments would
-    // be acceptable:
-    Assembler::Register scratch(t->arch->virtualCallTarget());
     a->apply(Move, TargetBytesPerWord, MemoryOperand, &table,
              TargetBytesPerWord, RegisterOperand, &scratch);
     Assembler::Memory proc(scratch.low, index * TargetBytesPerWord);

src/compiler.cpp

@@ -6953,14 +6953,14 @@ class MyCompiler: public Compiler {
         (c.machineCode + pad(c.machineCodeSize, TargetBytesPerWord) + i);
 
       if (n->promise->resolved()) {
-        *target = n->promise->value();
+        *target = targetVW(n->promise->value());
       } else {
         class Listener: public Promise::Listener {
          public:
           Listener(target_intptr_t* target): target(target){ }
 
           virtual bool resolve(int64_t value, void** location) {
-            *target = value;
+            *target = targetVW(value);
             if (location) *location = target;
             return true;
           }

src/powerpc.cpp

@@ -789,14 +789,14 @@ updateImmediate(System* s, void* dst, int32_t src, unsigned size, bool address)
   switch (size) {
   case 4: {
     int32_t* p = static_cast<int32_t*>(dst);
-    int r = (p[1] >> 21) & 31;
+    int r = (targetV4(p[1]) >> 21) & 31;
 
     if (address) {
-      p[0] = lis(r, ha16(src));
+      p[0] = targetV4(lis(r, ha16(src)));
-      p[1] |= (src & 0xFFFF);
+      p[1] |= targetV4(src & 0xFFFF);
     } else {
-      p[0] = lis(r, src >> 16);
+      p[0] = targetV4(lis(r, src >> 16));
-      p[1] = ori(r, r, src);
+      p[1] = targetV4(ori(r, r, src));
     }
   } break;
 
@@ -2076,6 +2076,10 @@ class MyArchitecture: public Assembler::Architecture {
     return 0;
   }
 
+  virtual int scratch() {
+    return 31;
+  }
+
   virtual int stack() {
     return StackRegister;
   }

src/x86.cpp

@@ -2763,6 +2763,10 @@ class MyArchitecture: public Assembler::Architecture {
     return useSSE(&c) ? FloatRegisterMask : 0;
   }
 
+  virtual int scratch() {
+    return rax;
+  }
+
   virtual int stack() {
     return rsp;
   }