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

[Luke Lau via llvm-commits]

================
@@ -0,0 +1,858 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --filter-out-after "^scalar.ph" --version 6
+; RUN: opt -passes=loop-vectorize -mtriple riscv64 -mattr=+v,+experimental-zvdot4a8i -S < %s | FileCheck %s --check-prefixes=CHECK
+; RUN: opt -passes=loop-vectorize -mtriple riscv64 -mattr=+v,+experimental-zvdot4a8i -scalable-vectorization=off -tail-folding-policy=dont-fold-tail -S < %s | FileCheck %s --check-prefixes=CHECK-FIXED
+
+define i32 @pred_reduction(ptr %src, ptr %cond, i64 %N) #0 {
+; CHECK-LABEL: define i32 @pred_reduction(
+; CHECK-SAME: ptr [[SRC:%.*]], ptr [[COND:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    br label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[CURRENT_ITERATION_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[PARTIAL_REDUCE4:%.*]] = phi <vscale x 4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP6:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[AVL:%.*]] = phi i64 [ [[N]], %[[VECTOR_PH]] ], [ [[AVL_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[TMP0:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 4, i1 true)
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds nuw i8, ptr [[COND]], i64 [[INDEX]]
+; CHECK-NEXT:    [[VP_OP_LOAD:%.*]] = call <vscale x 4 x i8> @llvm.vp.load.nxv4i8.p0(ptr align 1 [[TMP4]], <vscale x 4 x i1> splat (i1 true), i32 [[TMP0]])
+; CHECK-NEXT:    [[TMP2:%.*]] = icmp ne <vscale x 4 x i8> [[VP_OP_LOAD]], zeroinitializer
+; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[SRC]], i64 [[INDEX]]
+; CHECK-NEXT:    [[VP_OP_LOAD1:%.*]] = call <vscale x 4 x i8> @llvm.vp.load.nxv4i8.p0(ptr align 1 [[TMP8]], <vscale x 4 x i1> [[TMP2]], i32 [[TMP0]])
+; CHECK-NEXT:    [[PARTIAL_REDUCE:%.*]] = zext <vscale x 4 x i8> [[VP_OP_LOAD1]] to <vscale x 4 x i32>
+; CHECK-NEXT:    [[BIN_RDX:%.*]] = add <vscale x 4 x i32> [[PARTIAL_REDUCE4]], [[PARTIAL_REDUCE]]
+; CHECK-NEXT:    [[PREDPHI:%.*]] = select <vscale x 4 x i1> [[TMP2]], <vscale x 4 x i32> [[BIN_RDX]], <vscale x 4 x i32> [[PARTIAL_REDUCE4]]
+; CHECK-NEXT:    [[TMP6]] = call <vscale x 4 x i32> @llvm.vp.merge.nxv4i32(<vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> [[PREDPHI]], <vscale x 4 x i32> [[PARTIAL_REDUCE4]], i32 [[TMP0]])
----------------
lukel97 wrote:

I was expecting to see a vp.merge with a non-all-ones mask, but I guess that's handled by the select above? Do you know why we end up with the select, and not with the mask taken from `VPReductionEVLRecipe::getCondOp`? 

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