[llvm] [LV][EVL] Support reduction idioms using tail folding with EVL. (PR #90184)

[Florian Hahn via llvm-commits]

================
@@ -0,0 +1,2253 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -passes=loop-vectorize \
+; RUN: -force-tail-folding-style=data-with-evl \
+; RUN: -prefer-predicate-over-epilogue=predicate-dont-vectorize \
+; RUN: -mtriple=riscv64 -mattr=+v,+f -S < %s| FileCheck %s --check-prefix=IF-EVL
+
+; RUN: opt -passes=loop-vectorize \
+; RUN: -force-tail-folding-style=none \
+; RUN: -prefer-predicate-over-epilogue=predicate-else-scalar-epilogue \
+; RUN: -mtriple=riscv64 -mattr=+v,+f -S < %s| FileCheck %s --check-prefix=NO-VP
+
+define i32 @add(ptr %a, i64 %n, i32 %start) {
+; IF-EVL-LABEL: @add(
+; IF-EVL-NEXT:  entry:
+; IF-EVL-NEXT:    [[TMP0:%.*]] = sub i64 -1, [[N:%.*]]
+; IF-EVL-NEXT:    [[TMP1:%.*]] = call i64 @llvm.vscale.i64()
+; IF-EVL-NEXT:    [[TMP2:%.*]] = mul i64 [[TMP1]], 4
+; IF-EVL-NEXT:    [[TMP3:%.*]] = icmp ult i64 [[TMP0]], [[TMP2]]
+; IF-EVL-NEXT:    br i1 [[TMP3]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
+; IF-EVL:       vector.ph:
+; IF-EVL-NEXT:    [[TMP4:%.*]] = call i64 @llvm.vscale.i64()
+; IF-EVL-NEXT:    [[TMP5:%.*]] = mul i64 [[TMP4]], 4
+; IF-EVL-NEXT:    [[TMP6:%.*]] = call i64 @llvm.vscale.i64()
+; IF-EVL-NEXT:    [[TMP7:%.*]] = mul i64 [[TMP6]], 4
+; IF-EVL-NEXT:    [[TMP8:%.*]] = sub i64 [[TMP7]], 1
+; IF-EVL-NEXT:    [[N_RND_UP:%.*]] = add i64 [[N]], [[TMP8]]
+; IF-EVL-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP5]]
+; IF-EVL-NEXT:    [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
+; IF-EVL-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i64 [[N]], 1
+; IF-EVL-NEXT:    [[TMP9:%.*]] = call i64 @llvm.vscale.i64()
+; IF-EVL-NEXT:    [[TMP10:%.*]] = mul i64 [[TMP9]], 4
+; IF-EVL-NEXT:    [[TMP11:%.*]] = insertelement <vscale x 4 x i32> zeroinitializer, i32 [[START:%.*]], i32 0
+; IF-EVL-NEXT:    [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TRIP_COUNT_MINUS_1]], i64 0
+; IF-EVL-NEXT:    [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT1]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; IF-EVL-NEXT:    br label [[VECTOR_BODY:%.*]]
+; IF-EVL:       vector.body:
+; IF-EVL-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
+; IF-EVL-NEXT:    [[EVL_BASED_IV:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_EVL_NEXT:%.*]], [[VECTOR_BODY]] ]
+; IF-EVL-NEXT:    [[VEC_PHI:%.*]] = phi <vscale x 4 x i32> [ [[TMP11]], [[VECTOR_PH]] ], [ [[TMP20:%.*]], [[VECTOR_BODY]] ]
+; IF-EVL-NEXT:    [[TMP12:%.*]] = sub i64 [[N]], [[EVL_BASED_IV]]
+; IF-EVL-NEXT:    [[TMP13:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[TMP12]], i32 4, i1 true)
+; IF-EVL-NEXT:    [[TMP14:%.*]] = add i64 [[EVL_BASED_IV]], 0
+; IF-EVL-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[EVL_BASED_IV]], i64 0
+; IF-EVL-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
+; IF-EVL-NEXT:    [[TMP15:%.*]] = call <vscale x 4 x i64> @llvm.experimental.stepvector.nxv4i64()
+; IF-EVL-NEXT:    [[TMP16:%.*]] = add <vscale x 4 x i64> zeroinitializer, [[TMP15]]
+; IF-EVL-NEXT:    [[VEC_IV:%.*]] = add <vscale x 4 x i64> [[BROADCAST_SPLAT]], [[TMP16]]
+; IF-EVL-NEXT:    [[TMP17:%.*]] = icmp ule <vscale x 4 x i64> [[VEC_IV]], [[BROADCAST_SPLAT2]]
+; IF-EVL-NEXT:    [[TMP18:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[TMP14]]
+; IF-EVL-NEXT:    [[TMP19:%.*]] = getelementptr inbounds i32, ptr [[TMP18]], i32 0
+; IF-EVL-NEXT:    [[VP_OP_LOAD:%.*]] = call <vscale x 4 x i32> @llvm.vp.load.nxv4i32.p0(ptr align 4 [[TMP19]], <vscale x 4 x i1> shufflevector (<vscale x 4 x i1> insertelement (<vscale x 4 x i1> poison, i1 true, i64 0), <vscale x 4 x i1> poison, <vscale x 4 x i32> zeroinitializer), i32 [[TMP13]])
+; IF-EVL-NEXT:    [[TMP20]] = add <vscale x 4 x i32> [[VP_OP_LOAD]], [[VEC_PHI]]
----------------
fhahn wrote:

IIUC it should be safe for in-loop reductions, right? If that's the case, how about restricting codegen to in-loop reductions for now? this still allows brining up most of the code without miscompiles or the need for a temporary flag.

I am curious if at the moment it would be feasible to catch this issue at runtime, i.e. is there HW that implements EVL in a way that performs the last 2 iterations < VF?

https://github.com/llvm/llvm-project/pull/90184