Skip to contentSkip to conversation

isomorphism with unlikely

Cocci Wolfram Sang 3 participants 8 messages May 20, 2013 – May 26, 2013

lore
Share & tools
lore
#0 lore raw

Hi,

using 1.0-rc16 from debian, I use the following semantic patch:

@@
expression r, p, n;
@@

	r = platform_get_resource(p, IORESOURCE_MEM, n);
	...

(
- if (!r) { ... }
|
- if (unlikely(!r)) { ... }
|
- if (unlikely(r == NULL)) { ... }
)

and running against the latest kernel tree (~3.10-rc1):

spatch --sp-file unlikely_iso.cocci ~/Kernel/linux/drivers/net/ethernet/renesas/sh_eth.c

gives me a match:

- if (unlikely(res == NULL)) {
...

Note that I have two alternations with 'unlikely' because using '!r' alone will not give me a match. I guess the isomorphism is skipped for some reason when used as an argument of unlikely. If so, is it possible to enforce it somehow? Or would it be even better to make 'unlikely' an isomorphism itself, so that the semantic patch doesn't need to know anything about 'unlikely'?

Thanks,

Wolfram

#1 lore raw
Quoted 26 lines · Wolfram Sang

Hi,

using 1.0-rc16 from debian, I use the following semantic patch:

@@
expression r, p, n;
@@

	r = platform_get_resource(p, IORESOURCE_MEM, n);
	...

(

(quoted diff) +0 -1
-	if (!r) { ... }

|

(quoted diff) +0 -1
-	if (unlikely(!r)) { ... }

|

(quoted diff) +0 -1
-	if (unlikely(r == NULL)) { ... }

)

and running against the latest kernel tree (~3.10-rc1):

spatch --sp-file unlikely_iso.cocci ~/Kernel/linux/drivers/net/ethernet/renesas/sh_eth.c

gives me a match:

(quoted diff) +0 -1
-	if (unlikely(res == NULL)) {
...
Quoted 7 lines

Note that I have two alternations with 'unlikely' because using '!r' alone will not give me a match. I guess the isomorphism is skipped for some reason when used as an argument of unlikely. If so, is it possible to enforce it somehow? Or would it be even better to make 'unlikely' an isomorphism itself, so that the semantic patch doesn't need to know anything about 'unlikely'?

There is already an isomorphism for unlikely. But it is in one direction only, so 'unlikely(E)' in your cocci script will also match 'E', but 'E' wont match 'unlikely(E)'.

So in your case the script could be reduced to:

Review

Quoted 6 lines

@@
expression r, p, n;
@@

	r = platform_get_resource(p, IORESOURCE_MEM, n);
	...
(quoted diff) +0 -1
-	if (unlikely(r == NULL)) { ... }

Btw. if you use 'statement S;' and 'if(...) S' instead of '{ ... }' you'll also be able to match if there is only a single statement without brackets after the if.

- Lars

#2 lore raw
Quoted 4 lines

There is already an isomorphism for unlikely. But it is in one direction only, so 'unlikely(E)' in your cocci script will also match 'E', but 'E' wont match 'unlikely(E)'.

Okay, but I would then still need 'unlikely(!r)' and 'unlikely(r == NULL)' because the !r isomorphisms won't apply. I'd like to have both :)

Quoted 3 lines

Btw. if you use 'statement S;' and 'if(...) S' instead of '{ ... }' you'll also be able to match if there is only a single statement without brackets after the if.

True, thanks! Slowly getting back to it...

#3 lore raw
Quoted 7 lines · Wolfram Sang
Quoted 4 lines

There is already an isomorphism for unlikely. But it is in one direction only, so 'unlikely(E)' in your cocci script will also match 'E', but 'E' wont match 'unlikely(E)'.

Okay, but I would then still need 'unlikely(!r)' and 'unlikely(r == NULL)' because the !r isomorphisms won't apply. I'd like to have both :)

No, just unlikley(r == NULL) also matches !r (and unlikely(!r)). This should also work for unlikely(!r) matching unlikely(r == NULL) but for some reason it doesn't. I just did a quick test and it looks as if even 'if(!r) ..." doesn't match 'if(r == NULL) ...' either.

- Lars

#4 lore raw
Quoted 6 lines · Lars-Peter Clausen
Quoted 5 lines · Wolfram Sang
Quoted 4 lines

There is already an isomorphism for unlikely. But it is in one direction only, so 'unlikely(E)' in your cocci script will also match 'E', but 'E' wont match 'unlikely(E)'.

Quoted 7 lines
Quoted 2 lines

Okay, but I would then still need 'unlikely(!r)' and 'unlikely(r == NULL)' because the !r isomorphisms won't apply. I'd like to have both :)

No, just unlikley(r == NULL) also matches !r (and unlikely(!r)). This should also work for unlikely(!r) matching unlikely(r == NULL) but for some reason it doesn't. I just did a quick test and it looks as if even 'if(!r) ..." doesn't match 'if(r == NULL) ...' either.

It should work if r has pointer type. Otherwise, if r is not a pointer, then !r should become r == 0.

julia

#5 lore raw
Quoted 11 lines · Julia Lawall
Quoted 9 lines · Lars-Peter Clausen
Quoted 7 lines · Wolfram Sang
Quoted 4 lines

There is already an isomorphism for unlikely. But it is in one direction only, so 'unlikely(E)' in your cocci script will also match 'E', but 'E' wont match 'unlikely(E)'.

Okay, but I would then still need 'unlikely(!r)' and 'unlikely(r == NULL)' because the !r isomorphisms won't apply. I'd like to have both :)

Quoted 7 lines
Quoted 4 lines

No, just unlikley(r == NULL) also matches !r (and unlikely(!r)). This should also work for unlikely(!r) matching unlikely(r == NULL) but for some reason it doesn't. I just did a quick test and it looks as if even 'if(!r) ..." doesn't match 'if(r == NULL) ...' either.

It should work if r has pointer type. Otherwise, if r is not a pointer, then !r should become r == 0.

But for some reason it does not. With --verbose-match I can see that it generates the correct isomorphisms (both for the case that r is a pointer or that r is an integer), but for some reason they don't seem to match.

The testcase is simple:

cocci file:

@@
expression *r;
statement S;
@@
*if(!r) S

c file:

int foo(struct resource *r)

{
	if (r == NULL)
		return 0;
	return 1;
}

- Lars

#6 lore raw
Quoted 34 lines · Lars-Peter Clausen
Quoted 11 lines · Julia Lawall
Quoted 9 lines · Lars-Peter Clausen
Quoted 7 lines · Wolfram Sang
Quoted 4 lines

There is already an isomorphism for unlikely. But it is in one direction only, so 'unlikely(E)' in your cocci script will also match 'E', but 'E' wont match 'unlikely(E)'.

Okay, but I would then still need 'unlikely(!r)' and 'unlikely(r == NULL)' because the !r isomorphisms won't apply. I'd like to have both :)

Quoted 7 lines
Quoted 4 lines

No, just unlikley(r == NULL) also matches !r (and unlikely(!r)). This should also work for unlikely(!r) matching unlikely(r == NULL) but for some reason it doesn't. I just did a quick test and it looks as if even 'if(!r) ..." doesn't match 'if(r == NULL) ...' either.

It should work if r has pointer type. Otherwise, if r is not a pointer, then !r should become r == 0.

But for some reason it does not. With --verbose-match I can see that it generates the correct isomorphisms (both for the case that r is a pointer or that r is an integer), but for some reason they don't seem to match.

The testcase is simple:

cocci file:

@@
expression *r;
statement S;
@@
*if(!r) S

Quoted 8 lines

c file:

int foo(struct resource *r)

{
	if (r == NULL)
		return 0;
	return 1;
}

Thanks for the report. I have found the problem, and will try to fix it shortly.

julia

#7 lore raw
In reply to Lars-Peter Clausen

A patch is attached. I will make a new release including this patch shortly.

julia

Quoted 36 lines · Lars-Peter Clausen
Quoted 11 lines · Julia Lawall
Quoted 9 lines · Lars-Peter Clausen
Quoted 7 lines · Wolfram Sang
Quoted 4 lines

There is already an isomorphism for unlikely. But it is in one direction only, so 'unlikely(E)' in your cocci script will also match 'E', but 'E' wont match 'unlikely(E)'.

Okay, but I would then still need 'unlikely(!r)' and 'unlikely(r == NULL)' because the !r isomorphisms won't apply. I'd like to have both :)

Quoted 7 lines
Quoted 4 lines

No, just unlikley(r == NULL) also matches !r (and unlikely(!r)). This should also work for unlikely(!r) matching unlikely(r == NULL) but for some reason it doesn't. I just did a quick test and it looks as if even 'if(!r) ..." doesn't match 'if(r == NULL) ...' either.

It should work if r has pointer type. Otherwise, if r is not a pointer, then !r should become r == 0.

But for some reason it does not. With --verbose-match I can see that it generates the correct isomorphisms (both for the case that r is a pointer or that r is an integer), but for some reason they don't seem to match.

The testcase is simple:

cocci file:

@@
expression *r;
statement S;
@@
*if(!r) S

c file:

Quoted 9 lines

int foo(struct resource *r)

{
	if (r == NULL)
		return 0;
	return 1;
}

- Lars

-------------- next part --------------

parsing_cocci/ast0_cocci.ml +1 -0
index a9f2651..901f438 100644
@@ -554,6 +554,7 @@ let get_test_pos x      = x.true_if_test
 let set_test_pos x      = {x with true_if_test = true}
 let get_test_exp x      = x.true_if_test_exp
 let set_test_exp x      = {x with true_if_test_exp = true}
+let clear_test_exp x      = {x with true_if_test_exp = false}
 let get_iso x           = x.iso_info
 let set_iso x i = if !Flag.track_iso_usage then {x with iso_info = i} else x
 let set_mcode_data data (_,ar,info,mc,pos,adj) = (data,ar,info,mc,pos,adj)
parsing_cocci/ast0_cocci.mli +1 -0
index 72712b4..ba50c7b 100644
@@ -511,6 +511,7 @@ val get_arg_exp : expression -> bool
 val set_test_pos : expression -> expression
 val get_test_pos : 'a wrap -> bool
 val set_test_exp : expression -> expression
+val clear_test_exp : expression -> expression
 val get_test_exp : 'a wrap -> bool
 val set_iso : 'a wrap -> (string*anything) list -> 'a wrap
 val get_iso : 'a wrap -> (string*anything) list
parsing_cocci/iso_pattern.ml +6 -5
index d934e53..8efdc6a 100644
@@ -1649,7 +1649,7 @@ let instantiate bindings mv_bindings =
       Ast0.MetaExpr(name,constraints,x,form,pure) ->
 	(rebuild_mcode None).VT0.rebuilder_rec_expression
 	  (match lookup name bindings mv_bindings with
-	    Common.Left(Ast0.ExprTag(exp)) -> exp
+	    Common.Left(Ast0.ExprTag(exp)) -> Ast0.clear_test_exp exp
 	  | Common.Left(_) -> failwith "not possible 1"
 	  | Common.Right(new_mv) ->
 	      let new_types =
@@ -1678,10 +1678,11 @@ let instantiate bindings mv_bindings =
 			  Type_cocci.Array(renamer ty)
 		      | t -> t in
 		    Some(List.map renamer types) in
-	      Ast0.rewrap e
-		(Ast0.MetaExpr
-		   (Ast0.set_mcode_data new_mv name,constraints,
-		    new_types,form,pure)))
+	      Ast0.clear_test_exp 
+		(Ast0.rewrap e
+		   (Ast0.MetaExpr
+		      (Ast0.set_mcode_data new_mv name,constraints,
+		       new_types,form,pure))))
     | Ast0.MetaErr(namea,_,pure) -> failwith "metaerr not supported"
     | Ast0.MetaExprList(namea,lenname,pure) ->
 	failwith "metaexprlist not supported"