[PATCH] D88684: llvm-reduce: Don't replace intrinsic calls with undef

[Yuanfang Chen via Phabricator via llvm-commits]

ychen accepted this revision.
ychen added a comment.
This revision is now accepted and ready to land.

In D88684#2327662 <https://reviews.llvm.org/D88684#2327662>, @arsenm wrote:

> In D88684#2327661 <https://reviews.llvm.org/D88684#2327661>, @arsenm wrote:
>
>> In D88684#2318436 <https://reviews.llvm.org/D88684#2318436>, @ychen wrote:
>>
>>> In D88684#2318429 <https://reviews.llvm.org/D88684#2318429>, @dblaikie wrote:
>>>
>>>> In D88684#2318361 <https://reviews.llvm.org/D88684#2318361>, @arsenm wrote:
>>>>
>>>>> In D88684#2318108 <https://reviews.llvm.org/D88684#2318108>, @ychen wrote:
>>>>>
>>>>>> The test tieing to AMDGPU seems less ideal for general functionality. I've seen llvm.dbg.value caused the same problem, maybe we could use that?
>>>>>
>>>>> The debug intrinsics require a *lot* of metadata overhead simply to satsify the operations. The other metadata using intrinsics are experimental, so this seemed like the least fragile way to test it
>>>>
>>>> Given the functionality isn't implemented in terms of metadata-using-intrinsics in particular, a test using some non-metadata-using intrinsic would be fine.
>>>
>>> I think only the callee matters here. If anyone makes changes to llvm-reduce and not build AMDGPU, there is a chance (should be small I guess) to regress this.
>>
>> Developers should just always build all targets. We also have buildbots for this
>
> Also, building the target isn't actually necessary for the test. The intrinsic declaration works as is anyway

That's true. "REQUIRES: amdgpu-registered-target" is not needed then?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D88684/new/

https://reviews.llvm.org/D88684