[dynarmic] pre-LZCNT count leading zeroes (integer)

src/dynarmic/src/dynarmic/backend/x64/emit_x64_data_processing.cpp

@@ -1568,14 +1568,14 @@ void EmitX64::EmitCountLeadingZeros32(EmitContext& ctx, IR::Inst* inst) {
     } else {
         const Xbyak::Reg32 source = ctx.reg_alloc.UseScratchGpr(args[0]).cvt32();
         const Xbyak::Reg32 result = ctx.reg_alloc.ScratchGpr().cvt32();
+        const Xbyak::Reg32 temp = ctx.reg_alloc.ScratchGpr().cvt32();
 
         // The result of a bsr of zero is undefined, but zf is set after it.
         code.bsr(result, source);
-        code.mov(source, 0xFFFFFFFF);
-        code.cmovz(result, source);
-        code.neg(result);
-        code.add(result, 31);
-
+        code.mov(temp, 32);
+        code.xor_(result, 31);
+        code.test(source, source);
+        code.cmove(result, temp);
         ctx.reg_alloc.DefineValue(inst, result);
     }
 }
@@ -1592,14 +1592,14 @@ void EmitX64::EmitCountLeadingZeros64(EmitContext& ctx, IR::Inst* inst) {
     } else {
         const Xbyak::Reg64 source = ctx.reg_alloc.UseScratchGpr(args[0]).cvt64();
         const Xbyak::Reg64 result = ctx.reg_alloc.ScratchGpr().cvt64();
+        const Xbyak::Reg64 temp = ctx.reg_alloc.ScratchGpr().cvt64();
 
         // The result of a bsr of zero is undefined, but zf is set after it.
         code.bsr(result, source);
-        code.mov(source.cvt32(), 0xFFFFFFFF);
-        code.cmovz(result.cvt32(), source.cvt32());
-        code.neg(result.cvt32());
-        code.add(result.cvt32(), 63);
-
+        code.mov(temp.cvt32(), 64);
+        code.xor_(result.cvt32(), 63);
+        code.test(source, source);
+        code.cmove(result.cvt32(), temp.cvt32());
         ctx.reg_alloc.DefineValue(inst, result);
     }
 }

src/dynarmic/tests/A64/a64.cpp

@@ -2345,3 +2345,43 @@ TEST_CASE("A64: RBIT{16b}", "[a64]") {
     REQUIRE(jit.GetVector(2)[0] == 0xcafedead);
     REQUIRE(jit.GetVector(2)[1] == 0xbabebeef);
 }
+
+TEST_CASE("A64: CLZ{X}", "[a64]") {
+    A64TestEnv env;
+    A64::UserConfig conf{};
+    conf.callbacks = &env;
+    A64::Jit jit{conf};
+    env.code_mem.emplace_back(0xdac01060); // clz x0, x3
+    env.code_mem.emplace_back(0xdac01081); // clz x1, x4
+    env.code_mem.emplace_back(0xdac010a2); // clz x2, x5
+    env.code_mem.emplace_back(0x14000000); // b .
+    jit.SetRegister(3, 0xfffffffffffffff0);
+    jit.SetRegister(4, 0x0fffffff0ffffff0);
+    jit.SetRegister(5, 0x07fffffeffeffef0);
+    jit.SetPC(0); // at _start
+    env.ticks_left = 4;
+    jit.Run();
+    REQUIRE(jit.GetRegister(0) == 0);
+    REQUIRE(jit.GetRegister(1) == 4);
+    REQUIRE(jit.GetRegister(2) == 5);
+}
+
+TEST_CASE("A64: CLZ{W}", "[a64]") {
+    A64TestEnv env;
+    A64::UserConfig conf{};
+    conf.callbacks = &env;
+    A64::Jit jit{conf};
+    env.code_mem.emplace_back(0x5ac01060); // clz w0, w3
+    env.code_mem.emplace_back(0x5ac01081); // clz w1, w4
+    env.code_mem.emplace_back(0x5ac010a2); // clz w2, w5
+    env.code_mem.emplace_back(0x14000000); // b .
+    jit.SetRegister(3, 0xffff1110);
+    jit.SetRegister(4, 0x0fff1110);
+    jit.SetRegister(5, 0x07fffffe);
+    jit.SetPC(0); // at _start
+    env.ticks_left = 4;
+    jit.Run();
+    REQUIRE(jit.GetRegister(0) == 0);
+    REQUIRE(jit.GetRegister(1) == 4);
+    REQUIRE(jit.GetRegister(2) == 5);
+}