encourage loads from memory directly into SSE registers where appropriate

src/compiler.cpp

@@ -3447,6 +3447,10 @@ maybeMove(Context* c, BinaryOperation type, unsigned srcSize,
       c->arch->planSource(type, dstSize, &srcTypeMask, &srcRegisterMask,
                           dstSize, &thunk);
 
+      if (src->type == ValueGeneral) {
+        srcRegisterMask &= c->arch->generalRegisterMask();
+      }
+
       assert(c, thunk == 0);
       assert(c, dstMask.typeMask & srcTypeMask & (1 << RegisterOperand));
 
@@ -6013,7 +6017,7 @@ class MyCompiler: public Compiler {
   virtual void push(unsigned footprint UNUSED) {
     assert(&c, footprint == 1);
 
-    Value* v = value(&c, ValueFloat);
+    Value* v = value(&c, ValueGeneral);
     Stack* s = local::stack(&c, v, c.stack);
 
     v->home = frameIndex(&c, s->index + c.localFootprint);

src/x86.cpp

@@ -946,6 +946,9 @@ void
 sseMoveRR(Context* c, unsigned aSize, Assembler::Register* a,
           unsigned bSize UNUSED, Assembler::Register* b)
 {
+  assert(c, aSize >= 4);
+  assert(c, aSize == bSize);
+
   if (floatReg(a) and floatReg(b)) {
     if (aSize == 4) {
       opcode(c, 0xf3);
@@ -1090,6 +1093,9 @@ void
 sseMoveMR(Context* c, unsigned aSize, Assembler::Memory* a,
           unsigned bSize UNUSED, Assembler::Register* b)
 {
+  assert(c, aSize >= 4);
+  assert(c, aSize == bSize);
+
   if (BytesPerWord == 4 and aSize == 8) {
     opcode(c, 0xf3);
     opcode(c, 0x0f, 0x7e);
@@ -1165,6 +1171,7 @@ void
 sseMoveRM(Context* c, unsigned aSize, Assembler::Register* a,
        UNUSED unsigned bSize, Assembler::Memory* b)
 {
+  assert(c, aSize >= 4);
   assert(c, aSize == bSize);
 
   if (BytesPerWord == 4 and aSize == 8) {
@@ -2942,8 +2949,7 @@ class MyArchitecture: public Assembler::Architecture {
 
     case Move:
       *aTypeMask = (1 << RegisterOperand) | (1 << MemoryOperand);
-      *aRegisterMask = GeneralRegisterMask
+      *aRegisterMask = ~static_cast<uint64_t>(0);
-        | (static_cast<uint64_t>(GeneralRegisterMask) << 32);
 
       if (BytesPerWord == 4) {
         if (aSize == 4 and bSize == 8) {