[PATCH] D72467: [WIP] Remove "mask" operand from shufflevector.

[Eli Friedman via Phabricator via llvm-commits]

efriedma created this revision.
Herald added subscribers: dmgreen, hiraditya.
Herald added a reviewer: rengolin.
Herald added a project: LLVM.

  Instead, represent the mask as out-of-line data in the instruction. This
  should be more efficient in the places that currently use
  getShuffleVector(), and paves the way for further changes to add new
  shuffles for scalable vectors.
  
  This doesn't change the syntax in textual IR. And I don't currently plan
  to change the bitcode encoding in this patch, although we'll probably
  need to do something once we extend shufflevector for scalable types.
  
  The biggest part of the patch that still doesn't work is the bitcode
  writer/reader bits.  It turns out to be a bit more complicated than I
  anticipated.  For writing, my attempt at conjuring the constant out of
  thin air doesn't really work out because it needs to be found earlier;
  not sure how I'll fix this.  For reading, we need to change the way we
  construct constant expressions, I think; currently, we can sometimes
  construct an expression before its operands, which doesn't really work
  out here.
  
  There are a few other smaller cleanups I'd want to do here, but the new
  APIs are in essentially the form I expect them to take in the final
  patch.
  
  I expect that once this is finished, we can then replace the raw "mask"
  with something more appropriate for scalable vectors.  Not sure exactly
  what this looks like at the moment, but there are a few different ways
  we could handle it.  Maybe we could try to describe specific shuffles.
  Or maybe we could define it in terms of a function to convert a fixed-length
  array into an appropriate scalable vector, using a "step", or something
  like that.
  
  The big question here, of course, is "do we want to do this?", and if we
  do, "do we want to do this now?". For the immediate needs of SVE, we
  don't need fully general shuffles; we have intriniscs for the native
  instructions, and we can add additional intrinsics or intrinsic
  overloads for the various "extended" patterns we expect to need in the
  immediate future that don't involve native SVE vectors, like
  concatenating two "<vscale x 2 x i1>" vectors.  The biggest risk if we
  don't prioritize fixing shufflevector is that we end up duplicating work
  when we add another target supporting vscale.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D72467

Files:
  llvm/include/llvm/Analysis/ConstantFolding.h
  llvm/include/llvm/Analysis/InstructionSimplify.h
  llvm/include/llvm/IR/Constants.h
  llvm/include/llvm/IR/Instructions.h
  llvm/include/llvm/IR/PatternMatch.h
  llvm/lib/Analysis/ConstantFolding.cpp
  llvm/lib/Analysis/InstructionSimplify.cpp
  llvm/lib/Analysis/TargetTransformInfo.cpp
  llvm/lib/Analysis/ValueTracking.cpp
  llvm/lib/Analysis/VectorUtils.cpp
  llvm/lib/Bitcode/Writer/BitcodeWriter.cpp
  llvm/lib/CodeGen/CodeGenPrepare.cpp
  llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
  llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
  llvm/lib/IR/AsmWriter.cpp
  llvm/lib/IR/ConstantFold.cpp
  llvm/lib/IR/ConstantFold.h
  llvm/lib/IR/Constants.cpp
  llvm/lib/IR/ConstantsContext.h
  llvm/lib/IR/Instruction.cpp
  llvm/lib/IR/Instructions.cpp
  llvm/lib/IR/Verifier.cpp
  llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
  llvm/lib/Target/ARM/ARMISelLowering.cpp
  llvm/lib/Target/ARM/MVETailPredication.cpp
  llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
  llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
  llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
  llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
  llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
  llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp
  llvm/lib/Transforms/Scalar/ConstantProp.cpp
  llvm/lib/Transforms/Scalar/GVN.cpp
  llvm/lib/Transforms/Scalar/GVNSink.cpp
  llvm/lib/Transforms/Scalar/NewGVN.cpp
  llvm/lib/Transforms/Scalar/RewriteStatepointsForGC.cpp
  llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
  llvm/unittests/IR/PatternMatch.cpp

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D72467.237124.patch
Type: text/x-patch
Size: 70834 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200109/a2fff011/attachment-0001.bin>