[cfe-commits] [PATCH] Driver reengineering - first patch

[Jean-Daniel Dupas]

Le 18 janv. 2012 à 12:26, James Molloy a écrit :

> Hi,
> 
> The attached patch is the first of the clang driver revamp. This patch has
> no intended functionality change unless you use the new hidden flag
> "-ccc-dynamic-driver".
> 
> This patch addresses two points brought up during the driver BoF last
> November, namely
>  * "There should be a 'driver database' of known platforms"
>    * "This should fall back for the unusual case: load from some sort of
> config file (in JSON format)"
>  * "We should be able to ship a compiler that can always act like it's
> targetting Ubuntu".
> 
> Firstly, this first patch has some limitations:
>  * It has only been fully implemented for Linux, and only tested (to
> exhaustion) on Ubuntu Natty and RHEL5.
>  * It only deals with what should be the "uncommon" case - loading JSON at
> runtime.
>  * It doesn't yet copy the target DB into the correct place in the build or
> install dirs.
> 
> Secondly, this is what it does do:
>  * llvm.diff adds SMLoc tracking, raw MemoryBuffer input and '// ...' style
> comment support to the JSON parser. This makes developing the target DB
> files a lot easier, and the comment support is invaluable (and means we can
> add the copyright header).
>  * Adds an implementation of the Target Database for linux, and minix (as
> an example of a trivial platform). The database parser is completely
> implemented, and a schema is implemented and described that should be
> sufficient for everything we need, and also simple enough to grasp easily.
> This has been at the expense of some terseness in the DB definition.
> 
> The input to the target database is a JSON file. This JSON file can include
> other files. The file consists of a list of lists. Each of those lists is a
> "group" of objects. When queried, the targetdb will iterate over these
> groups. For each group, all objects are iterated in order. If an object
> "matches" the query criteria, its contents are added to the query result,
> and no more objects are iterated over in that group.
> 
> See lib/Driver/TargetDB_Linux.json for an example of how it works. A query
> to the target database will result in a structure containing everything you
> should need to compile/link/assemble on the target. The format of this
> structure is derived from a tablegen file (the patch includes a new tablegen
> backend), so there is no boilerplate if you want to add a new struct member.
> 
> The patch has been tested by taking an indicative subset of all possible
> Clang flags (especially interesting ones such as -static, -shared, -pie, -S
> etc) and brute force running through the entire set of combinations of these
> flags testing the targetdb-based driver's generated command lines against
> the current one.
> 
> The new patch adds a new command flag "-ccc-target-triple" to be
> distinguished from -ccc-host-triple. (When this patch and its followups are
> accepted and turned on by default, this will change into "-target"). This
> way we can actually (finally!) test "compile targetting X as if you were on
> Y", which should let us increase our driver test coverage (I also added
> -ccc-linux-distro which allows you to pretend to be a specific distro, as
> this is not encoded in the triple).
> 
> Speaking of tests, the submitted patch does not have any yet. I intend to
> submit them in a later patch, as I'd like this to be at least concept
> approved before I spend an infinite amount of time writing a decent test
> set.
> 
> Heckling... commence! :-)


>From dicsussions at http://llvm.org/bugs/show_bug.cgi?id=4127 it was suggested to not use target triple as an argument, but instead having something like 

'--target=configdb-entry-alias' which would define the corresponding triple. 

Using this flag, you will be able to use --target=linux_distro directly instead of having 2 arguments for this.

What do you think about this ? 

-- Jean-Daniel