Difference between gcc and Microsoft preprocessor

# define M3(x, y, z) x + y + z
# define M2(x, y) M3(x, y)
# define P(x, y) {x, y}
# define M(x, y) M2(x, P(x, y))
M(a, b)

Let us roll this out manually, step by step:

M(a, b)
--> M2(a, P(a, b))
--> M2(a, {a, b})

The standard says:

The individual arguments within the list are separated by comma preprocessing tokens, but comma preprocessing tokens between matching inner parentheses do not separate

only parentheses are mentioned, so ...

--> M3(a, {a, b})
--> a + {a + b}

Important:

M3(a, {a, b})

Here, according to the previous quote from the standard, three "arguments" are passed to M3 (using single-quotes to describe tokens/arguments):

M3('a', '{a', 'b}')

which are expanded to

'a' + '{a' + 'b}'

And this is what cpp (4.6.1) gives verbatim:

# 1 "cpp.cpp"
# 1 "<built-in>"
# 1 "<command-line>"
# 1 "cpp.cpp"




a + {a + b}

cpp (or gcc and g++) are correct, MSVC isn't.

As a nobleman make sure a bug report exists.

The only logic that explains such a behavior looks like this.

CL way:

 M(a,b) 
 M2(a,P(a,b)) 
 M3(a,P(a,b))
 M3(a,{a,b}) -> M3 gets 2 arguments ( 'a' and '{a,b}') instead of 3.
    |  \ /
  arg1  |
      arg2 

Gcc way:

M(a,b) 
M2(a,P(a,b)) 
M3(a,P(a,b))
M3(a,{a,b}) -> Gcc probably thinks there are 3 arguments here ('a', '{a', 'b}').
   |  | |
 arg1 | |
   arg2 |
     arg3

I think gcc gets it right, what Microsoft does is incorrect.

When macro substitution is done for the line

M2(a, P(a, b))

the standard (section 6.10.3.1) requires that before replacing the second parameter ("y") in the macro's replacement list ("M3(x, y)") with its argument ("P(a, b)"), macro replacement is to be performed for that argument. This means "P(a, b)" is processed to "{a, b}" before it is inserted, resulting in

M3(a, {a, b})

which is then further replaced to

a + {a + b}