I'd be very happy to see this turned on as a warning; I'd be very     reluctant to see it turned on as a error.  During a lot of code     cleanups, the intermediate state of "deleted, but dead args still     passed in" is a very useful intermediate step.  It greatly     simplifies review to be able to separate the meat of the semantic     change from the "just deleting a bunch of dead args" patch.  I'd be     very very reluctant to create a situation where the patch with the     semantic change but without the argument cleanup won't pass all the     build bots.  

Yep - I recall Chandler expressing similar consternation when he & I have discussed -Wunreachable-code so I was a bit surprised to see the suggestion.

But that said, in this case it doesn't seem too costly to comment out the name of the parameter in the first patch - it is an extra change, but doesn't require touching all the callers, etc. Then do the removal in a follow up patch still.
 
    
    Philip
    
    (1 inline comment below as well)
    
    
On 07/18/2015 02:48 AM, Chandler       Carruth wrote:
    
           LLVM and Clang both have lots of objects that are         passed through many different API boundaries. Things like         AliasAnalysis in LLVM or Sema in Clang get threaded all over the         place.         

        
        
Over times, refactoring can often cause the parameters (or           local variables, or member variables, etc) to become dead. If           we notice this, we can often un-thread the interface through           our APIs, sometimes even reducing coupling, etc.
        

        
        
But currently -Wunused-parameter is hard disabled. I've not           checked to see if there are any other disabling of -Wunused-*           variants, but I'd like to move the project toward firmly           enabling them and being clean with them.
        

        
        
The technique I'd like to use is leaving out names of           unused parameters, using a /*CommentedName*/ if there is a           useful name, otherwise just omitting the name.
                Sorry, could you rephrase?  Not sure what you're getting at here.
                    

        
        
All of this will require a reasonable amount of cleanup           across the projects which I'm happy to do prior to flipping           defaults around. Thoughts? Any concerns or objections?
        

        
        
(I'll ask the same question and ofter to either enact the           cleanups or toggle the warning(s) back off for each of the           less intertwined subprojects like LLD, LLDB, Polly, etc.)
        

        
        
-Chandler
             
      
      
      
_______________________________________________
LLVM Developers mailing list
LLVMdev@cs.uiuc.edu         http://llvm.cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
         
  ">

(original) (raw)



On Mon, Jul 20, 2015 at 4:18 PM, Philip Reames <listmail@philipreames.com> wrote:
I'd be very happy to see this turned on as a warning; I'd be very reluctant to see it turned on as a error. During a lot of code cleanups, the intermediate state of "deleted, but dead args still passed in" is a very useful intermediate step. It greatly simplifies review to be able to separate the meat of the semantic change from the "just deleting a bunch of dead args" patch. I'd be very very reluctant to create a situation where the patch with the semantic change but without the argument cleanup won't pass all the build bots.

Yep - I recall Chandler expressing similar consternation when he & I have discussed -Wunreachable-code so I was a bit surprised to see the suggestion.

But that said, in this case it doesn't seem too costly to comment out the name of the parameter in the first patch - it is an extra change, but doesn't require touching all the callers, etc. Then do the removal in a follow up patch still.

Philip

(1 inline comment below as well)

On 07/18/2015 02:48 AM, Chandler Carruth wrote:
LLVM and Clang both have lots of objects that are passed through many different API boundaries. Things like AliasAnalysis in LLVM or Sema in Clang get threaded all over the place.

Over times, refactoring can often cause the parameters (or local variables, or member variables, etc) to become dead. If we notice this, we can often un-thread the interface through our APIs, sometimes even reducing coupling, etc.

But currently -Wunused-parameter is hard disabled. I've not checked to see if there are any other disabling of -Wunused-\* variants, but I'd like to move the project toward firmly enabling them and being clean with them.

The technique I'd like to use is leaving out names of unused parameters, using a /\*CommentedName\*/ if there is a useful name, otherwise just omitting the name.
 Sorry, could you rephrase? Not sure what you're getting at here.

All of this will require a reasonable amount of cleanup across the projects which I'm happy to do prior to flipping defaults around. Thoughts? Any concerns or objections?

(I'll ask the same question and ofter to either enact the cleanups or toggle the warning(s) back off for each of the less intertwined subprojects like LLD, LLDB, Polly, etc.)

-Chandler


\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_  
LLVM Developers mailing list  
LLVMdev@cs.uiuc.edu http://llvm.cs.uiuc.edu  
http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev


_______________________________________________

LLVM Developers mailing list

LLVMdev@cs.uiuc.edu http://llvm.cs.uiuc.edu

http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev