[llvm] [LV] Support EVL for partial reduction and VPExpressionRecipe. (PR #205741)

[Florian Hahn via llvm-commits]

================
@@ -0,0 +1,91 @@
+; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -mtriple riscv64 -passes=loop-vectorize -force-vector-width=4 -scalable-vectorization=on -mattr=+v,+experimental-zvdot4a8i -disable-output 2>&1 -vplan-print-after=optimizeEVLMasks < %s | FileCheck %s
+
+define i32 @vdota4(ptr %a, ptr %b) #0 {
+; CHECK-LABEL: VPlan for loop in 'vdota4'
+; CHECK:  VPlan 'Initial VPlan for VF={vscale x 4},UF={1}' {
+; CHECK-NEXT:  Live-in vp<[[VP0:%[0-9]+]]> = VF * UF
+; CHECK-NEXT:  Live-in vp<[[VP1:%[0-9]+]]> = vector-trip-count
+; CHECK-NEXT:  Live-in ir<1024> = original trip-count
+; CHECK-EMPTY:
+; CHECK-NEXT:  ir-bb<entry>:
+; CHECK-NEXT:  Successor(s): scalar.ph, vector.ph
+; CHECK-EMPTY:
+; CHECK-NEXT:  vector.ph:
+; CHECK-NEXT:    EMIT vp<[[VP2:%[0-9]+]]> = reduction-start-vector ir<0>, ir<0>, ir<4>
+; CHECK-NEXT:  Successor(s): vector loop
+; CHECK-EMPTY:
+; CHECK-NEXT:  <x1> vector loop: {
+; CHECK-NEXT:  vp<[[VP3:%[0-9]+]]> = CANONICAL-IV
+; CHECK-EMPTY:
+; CHECK-NEXT:    vector.body:
+; CHECK-NEXT:      CURRENT-ITERATION-PHI vp<[[VP5:%[0-9]+]]> = phi ir<0>, vp<%current.iteration.next>
+; CHECK-NEXT:      WIDEN-REDUCTION-PHI ir<%accum> = phi (add) vp<[[VP2]]>, vp<[[VP10:%[0-9]+]]> (VF scaled by 1/4)
+; CHECK-NEXT:      EMIT-SCALAR vp<%avl> = phi [ ir<1024>, vector.ph ], [ vp<%avl.next>, vector.body ]
+; CHECK-NEXT:      EMIT-SCALAR vp<%evl> = EXPLICIT-VECTOR-LENGTH vp<%avl>
+; CHECK-NEXT:      EMIT-SCALAR vp<[[VP6:%[0-9]+]]> = zext vp<%evl> to i64
+; CHECK-NEXT:      vp<[[VP7:%[0-9]+]]> = SCALAR-STEPS vp<[[VP5]]>, ir<1>, vp<[[VP6]]>
+; CHECK-NEXT:      CLONE ir<%gep.a> = getelementptr ir<%a>, vp<[[VP7]]>
+; CHECK-NEXT:      vp<[[VP8:%[0-9]+]]> = vector-pointer i8, ir<%gep.a>, ir<1>
+; CHECK-NEXT:      WIDEN ir<%load.a> = vp.load vp<[[VP8]]>, vp<%evl>
+; CHECK-NEXT:      CLONE ir<%gep.b> = getelementptr ir<%b>, vp<[[VP7]]>
+; CHECK-NEXT:      vp<[[VP9:%[0-9]+]]> = vector-pointer i8, ir<%gep.b>, ir<1>
+; CHECK-NEXT:      WIDEN ir<%load.b> = vp.load vp<[[VP9]]>, vp<%evl>
+; CHECK-NEXT:      EXPRESSION vp<[[VP10]]> = ir<%accum> + partial.reduce.add (mul (ir<%load.b> zext to i32), (ir<%load.a> sext to i32), vp<%evl>)
+; CHECK-NEXT:      EMIT-SCALAR vp<[[VP11:%[0-9]+]]> = zext vp<%evl> to i64
+; CHECK-NEXT:      EMIT vp<%current.iteration.next> = add nuw vp<[[VP11]]>, vp<[[VP5]]>
+; CHECK-NEXT:      EMIT vp<%avl.next> = sub nuw vp<%avl>, vp<[[VP11]]>
+; CHECK-NEXT:      EMIT vp<%index.next> = add nuw vp<[[VP3]]>, vp<[[VP0]]>
+; CHECK-NEXT:      EMIT branch-on-count vp<%index.next>, vp<[[VP1]]>
+; CHECK-NEXT:    No successors
+; CHECK-NEXT:  }
+; CHECK-NEXT:  Successor(s): middle.block
+; CHECK-EMPTY:
+; CHECK-NEXT:  middle.block:
+; CHECK-NEXT:    EMIT vp<[[VP13:%[0-9]+]]> = compute-reduction-result (add) vp<[[VP10]]>
+; CHECK-NEXT:  Successor(s): ir-bb<for.exit>
+; CHECK-EMPTY:
+; CHECK-NEXT:  ir-bb<for.exit>:
+; CHECK-NEXT:    IR   %add.lcssa = phi i32 [ %add, %for.body ] (extra operand: vp<[[VP13]]> from middle.block)
+; CHECK-NEXT:  No successors
+; CHECK-EMPTY:
+; CHECK-NEXT:  scalar.ph:
+; CHECK-NEXT:  Successor(s): ir-bb<for.body>
----------------
fhahn wrote:

could filter-out-after=scalar.ph

https://github.com/llvm/llvm-project/pull/205741