Speedup Function::getIntrinsicID() with caching of result

[Michael Ilseman]

On Feb 28, 2013, at 2:21 AM, Duncan Sands <baldrick at free.fr> wrote:

> Hi Jean-Luc,
> 
> On 28/02/13 02:16, Jean-Luc Duprat wrote:
>> The attached patch caches the result of Function::getIntrinsicID() in a DenseMap attached to the LLVMContext.  This reduces the time actually spent doing string to ID conversion and shows a 10% improvement in compile time for a particularly bad case that involves ARM Neon intrinsics (these have many overloads).
>> 
>> This changes passes the regression tests and the nightly tests suite.
> 
> if you delete an intrinsic function declaration from the module, don't you get a
> stale pointer to freed memory in the cache?  To avoid all such issues, maybe it
> is simpler to have the cache map the name (a StringRef) to the intrinsic ID,
> rather than mapping the function pointer to the intrinsic ID.  You probably lose
> some of the speedup then though.

Can he instead have the destructor remove itself from the cache? This would preserve the speedups of key-ing off of the Function pointer directly.

Also, since it's the Function* that's the key, I believe the situation you described wouldn't provide an error so much as it would under-utilize the map by having extra meaningless entries. However, maybe there would be an error if anyone ever iterated over the cache. Replacing the Module's destructor clear() call with a Function's destructor erase() call might be a good way to go.

> 
> Ciao, Duncan.
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits