[PATCH] D86995: [VectorCombine][SVE] Do not fold bitcast shuffle for scalable type.

[Huihui Zhang via Phabricator via llvm-commits]

huihuiz created this revision.
huihuiz added reviewers: efriedma, sdesmalen, paulwalker-arm, spatel.
huihuiz added a project: LLVM.
Herald added subscribers: psnobl, hiraditya, tschuett.
huihuiz requested review of this revision.

First, shuffle cost for scalable type is not known for scalable type;
Second, we cannot reason if the narrowed shuffle mask for scalable type
is a splat or not.

E.g., Bitcast splat vector from type <vscale x 4 x i32> to <vscale x 8 x i16>
will involve narrowing shuffle mask <vscale x 4 x i32> zeroinitializer to
<vscale x 8 x i32> with element sequence of <0, 1, 0, 1, ...>, which cannot be
reasoned if it's a valid splat or not.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D86995

Files:
  llvm/lib/Transforms/Vectorize/VectorCombine.cpp
  llvm/test/Transforms/VectorCombine/AArch64/lit.local.cfg
  llvm/test/Transforms/VectorCombine/AArch64/vscale-bitcast-shuffle.ll


Index: llvm/test/Transforms/VectorCombine/AArch64/vscale-bitcast-shuffle.ll
===================================================================
--- /dev/null
+++ llvm/test/Transforms/VectorCombine/AArch64/vscale-bitcast-shuffle.ll
@@ -0,0 +1,21 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -vector-combine -S -mtriple=aarch64-- | FileCheck %s --check-prefixes=CHECK
+
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+target triple = "aarch64-unknown-linux-gnu"
+
+; This test checks we are not crashing with TTI when trying to get shuffle cost.
+; This test also check that shuffle mask <vscale x 4 x i32> zeroinitializer is
+; not narrowed into <0, 1, 0, 1, ...>, which we cannot reason if it's a valid
+; splat or not.
+
+define <vscale x 8 x i16> @bitcast_shuffle(<vscale x 4 x i32> %a) {
+; CHECK-LABEL: @bitcast_shuffle(
+; CHECK-NEXT:    [[I:%.*]] = shufflevector <vscale x 4 x i32> [[A:%.*]], <vscale x 4 x i32> undef, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[R:%.*]] = bitcast <vscale x 4 x i32> [[I]] to <vscale x 8 x i16>
+; CHECK-NEXT:    ret <vscale x 8 x i16> [[R]]
+;
+  %i = shufflevector <vscale x 4 x i32> %a, <vscale x 4 x i32> undef, <vscale x 4 x i32> zeroinitializer
+  %r = bitcast <vscale x 4 x i32> %i to <vscale x 8 x i16>
+  ret <vscale x 8 x i16> %r
+}
Index: llvm/test/Transforms/VectorCombine/AArch64/lit.local.cfg
===================================================================
--- /dev/null
+++ llvm/test/Transforms/VectorCombine/AArch64/lit.local.cfg
@@ -0,0 +1,2 @@
+if not 'AArch64' in config.root.targets:
+    config.unsupported = True
Index: llvm/lib/Transforms/Vectorize/VectorCombine.cpp
===================================================================
--- llvm/lib/Transforms/Vectorize/VectorCombine.cpp
+++ llvm/lib/Transforms/Vectorize/VectorCombine.cpp
@@ -431,6 +431,12 @@
   if (!DestTy || I.getOperand(0)->getType() != SrcTy)
     return false;
 
+  // Do not fold bitcast shuffle for scalable type. First, shuffle cost for
+  // scalable type is unknown; Second, we cannot reason if the narrowed shuffle
+  // mask for scalable type is a splat or not.
+  if (isa<ScalableVectorType>(SrcTy) || isa<ScalableVectorType>(DestTy))
+    return false;
+
   // The new shuffle must not cost more than the old shuffle. The bitcast is
   // moved ahead of the shuffle, so assume that it has the same cost as before.
   if (TTI.getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, DestTy) >


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D86995.289336.patch
Type: text/x-patch
Size: 2529 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200902/bb300a49/attachment.bin>