Merge pull request #665 from undingen/mov_opt

assembler: Use 32bit moves if the immediate fits inside 32bit

src/asm_writing/assembler.cpp

@@ -145,18 +145,20 @@ void Assembler::emitSIB(uint8_t scalebits, uint8_t index, uint8_t base) {
     emitByte((scalebits << 6) | (index << 3) | base);
 }
 
-void Assembler::mov(Immediate val, Register dest) {
-    int rex = REX_W;
+void Assembler::mov(Immediate val, Register dest, bool force_64bit_load) {
+    force_64bit_load = force_64bit_load || !val.fitsInto32Bit();
 
+    int rex = force_64bit_load ? REX_W : 0;
     int dest_idx = dest.regnum;
     if (dest_idx >= 8) {
         rex |= REX_B;
         dest_idx -= 8;
     }
 
-    emitRex(rex);
+    if (rex)
+        emitRex(rex);
     emitByte(0xb8 + dest_idx);
-    emitInt(val.val, 8);
+    emitInt(val.val, force_64bit_load ? 8 : 4);
 }
 
 void Assembler::movq(Immediate src, Indirect dest) {
@@ -975,7 +977,9 @@ void Assembler::leave() {
 }
 
 uint8_t* Assembler::emitCall(void* ptr, Register scratch) {
-    mov(Immediate(ptr), scratch);
+    // emit a 64bit movabs because some caller expect a fixed number of bytes.
+    // until they are fixed use the largest encoding.
+    mov(Immediate(ptr), scratch, true /* force_64bit_load */);
     callq(scratch);
     return addr;
 }

src/asm_writing/assembler.h

@@ -114,8 +114,8 @@ public:
     void nop() { emitByte(0x90); }
     void trap() { emitByte(0xcc); }
 
-    // some things (such as objdump) call this "movabs" if the immediate is 64-bit
-    void mov(Immediate imm, Register dest);
+    // emits a movabs if the immediate is a 64bit value or force_64bit_load = true otherwise it emits a 32bit mov
+    void mov(Immediate imm, Register dest, bool force_64bit_load = false);
     // not sure if we should use the 'q' suffix here, but this is the most ambiguous one;
     // this does a 64-bit store of a 32-bit value.
     void movq(Immediate imm, Indirect dest);

src/asm_writing/types.h

@@ -158,6 +158,11 @@ struct Immediate {
 
     explicit Immediate(uint64_t val) : val(val) {}
     explicit Immediate(void* val) : val((uint64_t)val) {}
+
+    bool fitsInto32Bit() const {
+        uint32_t val_32bit = (uint32_t)val;
+        return val_32bit == val;
+    }
 };
 
 struct JumpDestination {