[PATCH] D46827: [VPlan] Add VPInstruction to VPRecipe transformation.

[Satish K Guggilla via Phabricator via llvm-commits]

sguggill added a comment.



> Sure, we can start with just creating the basic recipes.  Should we only support widening for a start and not relying on the cost model at all? I suppose that would make things easier for outer loop code gen.

Hi Florian,

For the initial outer loop vector code generation, we were planning to only support widening and not rely on the cost model at all. Where the current code generation relies on results of cost modeling, we were planning to put changes in place to return a conservative result for the VPlanNativePath. Given the current constraints for the supported outer loops in the VPlanNativePath, this is what we had in mind specifically:

- generate gather/scatter for loads/stores until we have proper divergence analysis in place. We may be able to infer unit-stride information using current SCEV based analysis. However, for the initial implementation we can generate gathers/scatters by returning CM_GatherScatter in getWideningDecision for outer loops in VPlanNativePath.

- widen non-induction PHIs. The initial implementation will run some part of legality checks to recognize inductions as you pointed out earlier. PHIs that are not in the outer loop header will be widened. Due to back edges from inner loops we may need to generate place-holder PHIs to begin with and fixup such PHIs at the end of code generation. This is due to the fact that vector values corresponding to all the phi-operands may not be available when widening the PHI instruction. This will be similar to the way reductions/recurrences are currently handled.

We will need to either guarantee that the phi operand order matches the parent BBlock successor order or have a mapping between scalar BBlocks and the corresponding vector BBlock.

- treat all branch instructions as uniform. During vector CG, the branch operands need to be replaced with the vector equivalents. We will need an approach similar to handling of PHIs i.e. fixup branches and control flow at the end of vector CG.

- all other instructions will simply be widened using vector values or generating broadcasts(for loop invariants)

Initially, we will need at the least - widenMemory, widen, and widenPhi recipes. Branch instruction can be part of a widenRecipe and CG needs to handle it appropriately for the VPlanNativePath. If that is not acceptable, we can think about having a new recipe to handle branch instructions.

Please let me know if this matches with what you had in mind. We will try to send you a patch by the end of this week so that you can see what we have.

Satish


https://reviews.llvm.org/D46827