fix valist_size

Author: tgymnich

src/device/intrinsics/output.jl

@@ -50,6 +50,19 @@ end
     end
 end
 
+function valist_size(dl, param_types)
+    size = 0
+    for pty in param_types
+        ps = sizeof(dl, pty)
+        if size % ps == 0
+            size += ps
+        else
+            size += (size % ps) + ps
+        end
+    end
+
+    return size
+end
 
 """
     @mtlprintf("%Fmt", args...)
@@ -99,9 +112,8 @@ end
             log_type = LLVM.ConstantInt(T_int32, __METAL_OS_LOG_TYPE_DEBUG__)
 
             # compute argsize
-            argtypes = LLVM.StructType(param_types)
-            dl = datalayout(mod) 
-            arg_size = LLVM.ConstantInt(T_int64, sizeof(dl, argtypes))
+            dl = datalayout(mod)
+            arg_size = LLVM.ConstantInt(T_int64, valist_size(dl, param_types))
 
             alloc = alloca!(builder, T_pint8)
             buffer = bitcast!(builder, alloc, T_pint8)