diff --git a/gen/asm-x86-64.h b/gen/asm-x86-64.h
index 16f999b2..074715c9 100644
--- a/gen/asm-x86-64.h
+++ b/gen/asm-x86-64.h
@@ -630,8 +630,7 @@ namespace AsmParserx8664
     static AsmOpEnt opData[] =
     {
         { "adc",    Op_UpdSrcF },
-
-        { "add",    Op_DstSrcNT }, //Op_UpdSrcF },
+        { "add",    Op_UpdSrcF },
         { "addpd",  Op_DstSrcSSE },
         { "addps",  Op_DstSrcSSE },
         { "addq",    Op_DstSrcSSE },
diff --git a/gen/asmstmt.cpp b/gen/asmstmt.cpp
index 95e6b5b0..44a4cac1 100644
--- a/gen/asmstmt.cpp
+++ b/gen/asmstmt.cpp
@@ -396,8 +396,9 @@ assert(0);
             // Add input operand with same value, with original as "matching output".
             std::ostringstream ss;
             ss << '*' << (n + asmblock->outputcount);
-            input_constraints.push_front(ss.str());
-            input_values.push_front(output_values[n]);
+            // Must be at the back; unused operands before used ones screw up numbering.
+            input_constraints.push_back(ss.str());
+            input_values.push_back(output_values[n]);
         }
         llvmOutConstraints += *i;
         llvmOutConstraints += ",";
@@ -440,7 +441,7 @@ AsmBlockStatement::AsmBlockStatement(Loc loc, Statements* s)
 }
 
 // rewrite argument indices to the block scope indices
-static void remap_outargs(std::string& insnt, size_t nargs, size_t& idx)
+static void remap_outargs(std::string& insnt, size_t nargs, size_t idx)
 {
     static const std::string digits[10] =
     {
@@ -465,7 +466,7 @@ static void remap_outargs(std::string& insnt, size_t nargs, size_t& idx)
 }
 
 // rewrite argument indices to the block scope indices
-static void remap_inargs(std::string& insnt, size_t nargs, size_t& idx)
+static void remap_inargs(std::string& insnt, size_t nargs, size_t idx)
 {
     static const std::string digits[10] =
     {
@@ -644,6 +645,7 @@ void AsmBlockStatement::toIR(IRState* p)
             out_c += a->out_c;
         }
         remap_outargs(a->code, onn+a->in.size(), asmIdx);
+        asmIdx += onn;
     }
 
     Logger::println("do inputs");
@@ -662,6 +664,7 @@ void AsmBlockStatement::toIR(IRState* p)
             in_c += a->in_c;
         }
         remap_inargs(a->code, inn+a->out.size(), asmIdx);
+        asmIdx += inn;
         if (!code.empty())
             code += "\n\t";
         code += a->code;
@@ -707,8 +710,12 @@ void AsmBlockStatement::toIR(IRState* p)
     if (Logger::enabled()) {
         Logger::cout() << "Arguments:" << '\n';
         Logger::indent();
-        for (std::vector<LLValue*>::iterator b = args.begin(), i = b, e = args.end(); i != e; ++i)
-            Logger::cout() << '$' << (i - b) << " ==> " << **i;
+        for (std::vector<LLValue*>::iterator b = args.begin(), i = b, e = args.end(); i != e; ++i) {
+            std::ostream& cout = Logger::cout();
+            cout << '$' << (i - b) << " ==> " << **i;
+            if (llvm::isa<LLConstant>(*i))
+                cout << '\n';
+        }
         Logger::undent();
     }
 
diff --git a/tests/mini/asm9.d b/tests/mini/asm9.d
new file mode 100644
index 00000000..bd4a459b
--- /dev/null
+++ b/tests/mini/asm9.d
@@ -0,0 +1,28 @@
+module asm9;
+
+version(X86)            version = DoSome;
+else version(X86_64)    version = DoSome;
+
+T add(T, T t)(T a) {
+    asm {
+        add a, t;
+    }
+    return a;
+}
+
+void main() {
+    version (DoSome) {
+        assert(add!(ubyte, 20)(10) == 30);
+        assert(add!(ushort, 20_000)(10_000) == 30_000);
+        assert(add!(uint, 2_000_000)(1_000_000) == 3_000_000);
+    }
+    version(X86_64) {
+        // 64-bit immediates aren't allowed on "ADD", nor are
+        // unsigned 32-bit ones, so make the template parameter
+        // fit in a 32-bit signed int.
+        // These values were chosen so that the lower 32-bits overflow
+        // and we can see the upper half of the 64-bit input increment.
+        auto result = add!(long, 2_000_000_000)(21_000_000_000);
+        assert(result == 23_000_000_000);
+    }
+}