[LLVMdev] [RFC] Stripping unusable intrinsics
beanz at apple.com
Tue Dec 23 11:50:29 PST 2014
I had patches earlier on this thread that stripped the uses of intrinsics from LLVM using ifdefs.
These files have intrinsics as switch cases. I’ve listed which target the intrinsics belong to as well for reference.
lib/Analysis/BasicAliasAnalysis.cpp - ARM
lib/Analysis/ConstantFolding.cpp - X86
lib/Analysis/ValueTracking.cpp - X86
lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp - X86
lib/Transforms/InstCombine/InstCombineCalls.cpp - PowerPC, X86, ARM, AArch64, and R600
lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp - X86
lib/Transforms/Instrumentation/MemorySanitizer.cpp - X86
lib/Transforms/Scalar/LoopStrengthReduce.cpp - X86
In InstCombineCalls some of the case bodies are shared by ARM and AArch64 intrinsics. In those cases there were also some if statements that matched on the intrinsic enums too. As an example see the case at InstCombineCalls.cpp line 922-1028.
Other than that, lib/IR/AutoUpgrade.cpp has a bunch of string matched X86 Intrinsics being mapped to enums.
Clang also has a bunch of places that use target intrinsics that were not as easy to ifdef, so I didn’t try.
As Bob said, I don’t think this is particularly a difficult undertaking, but it will be a big change. I think a lot of the switch statements could be replaced by per-target hooks.
Personally I’d love to be in a world where intrinsics lived in the targets, but we won’t get there quickly. I think a stop-gap solution for our more immediate needs could be accomplished with Alex’s suggestion of a clang attribute. I’ve sent patches to clang (http://reviews.llvm.org/D6763 <http://reviews.llvm.org/D6763>). One of the allures I see to that approach is that it would allow us to get what we need today without significant modifications to LLVM, and it would not get in the way of working towards the “Right™” solution.
> On Dec 23, 2014, at 10:54 AM, Bob Wilson <bob.wilson at apple.com> wrote:
>> On Dec 23, 2014, at 10:41 AM, Chris Lattner <clattner at apple.com <mailto:clattner at apple.com>> wrote:
>> On Dec 23, 2014, at 10:28 AM, Chris Bieneman <beanz at apple.com <mailto:beanz at apple.com>> wrote:
>>>>> It should be straight-forward to have something like LLVMInitializeX86Target/RegisterTargetMachine install the intrinsics into a registry.
>>>> I tried doing that a few years ago. It’s not nearly as easy as it sounds because we’ve got hardcoded references to various target intrinsics scattered throughout the code.
>>> I was just writing to say exactly this. There are a number of ways we could work toward something like this. I’m completely in favor of a world where Intrinsics are properties of the targets and don’t leach out, however today they do in a lot of places.
>> What are the specific problems here? Anything that does an equality comparison with the IntrinsicID can be changed to do strcmp with the name. That would handle the one-off cases like InstCombiner::SimplifyDemandedUseBits in InstCombine.
>> The other cases in InstCombine could be handled similarly, but may be better handled by adding a intrinsic behavior query APIs to the intrinsic registry, or would be better handled (eventually) by adding new attributes to the intrinsics themselves.
> I don’t think there’s anything fundamentally difficult, but it’s a big change. For example:
> $ git grep Intrinsic:: | wc
> 3364 12286 281078
> The vast majority of those 3,364 lines have hardcoded references to specific intrinsics. Many of them are used in contexts where you can’t easily insert a strcmp (e.g., case values in large switch statements, or worse, the m_Intrinsic PatternMatch templates).
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-dev