[llvm] [LV] Update interleaving count computation when scalar epilogue loop needs to run at least once (PR #79651)
Nilanjana Basu via llvm-commits
llvm-commits at lists.llvm.org
Fri Jan 26 14:19:57 PST 2024
https://github.com/nilanjana87 updated https://github.com/llvm/llvm-project/pull/79651
>From 44f51ab34c8210d7af5bddc75d40b60336b8a2be Mon Sep 17 00:00:00 2001
From: Nilanjana Basu <n_basu at apple.com>
Date: Thu, 25 Jan 2024 15:39:01 -0800
Subject: [PATCH 1/4] [LV][AArch64] Pre-commit tests for patch that changes
interleaving count computation for loops that require a scalar epilogue run
---
.../interleave_count_for_estimated_tc.ll | 48 +++++++
.../AArch64/interleave_count_for_known_tc.ll | 44 ++++++
...count_for_loops_needing_scalar_epilogue.ll | 132 ++++++++++++++++++
3 files changed, 224 insertions(+)
create mode 100644 llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_loops_needing_scalar_epilogue.ll
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_estimated_tc.ll b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_estimated_tc.ll
index 5552f9dd70c954..d557ad1ead2563 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_estimated_tc.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_estimated_tc.ll
@@ -125,6 +125,30 @@ for.end:
ret void
}
+; This has the same trip count as loop_with_profile_tc_64 but since the resulting interleaved group
+; in this case may access memory out-of-bounds, it requires a scalar epilogue iteration for
+; correctness, making at most 63 iterations available for interleaving.
+; TODO: When the auto-vectorizer chooses VF 16, it should choose IC 1 to leave a smaller scalar remainder
+; than IC 2
+; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 2)
+define void @loop_with_profile_tc_64_scalar_epilogue_reqd(ptr noalias %p, ptr noalias %q, i64 %n) {
+entry:
+ br label %for.body
+
+for.body:
+ %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
+ %gep.src = getelementptr inbounds [3 x i8], ptr %p, i64 %i, i64 0
+ %l = load i8, ptr %gep.src, align 1
+ %gep.dst = getelementptr inbounds i8, ptr %q, i64 %i
+ store i8 %l, ptr %gep.dst, align 1
+ %i.next = add nuw nsw i64 %i, 1
+ %cond = icmp eq i64 %i.next, %n
+ br i1 %cond, label %for.end, label %for.body, !prof !4
+
+for.end:
+ ret void
+}
+
; For a loop with a profile-guided estimated TC of 100, when the auto-vectorizer chooses VF 16,
; it should choose conservatively IC 2 so that the vector loop runs twice at least
; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 2)
@@ -173,6 +197,30 @@ for.end:
ret void
}
+; This has the same trip count as loop_with_profile_tc_128 but since the resulting interleaved group
+; in this case may access memory out-of-bounds, it requires a scalar epilogue iteration for
+; correctness, making at most 127 iterations available for interleaving.
+; TODO: When the auto-vectorizer chooses VF 16, it should choose IC 2 to leave a smaller scalar remainder
+; than IC 4
+; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 4)
+define void @loop_with_profile_tc_128_scalar_epilogue_reqd(ptr noalias %p, ptr noalias %q, i64 %n) {
+entry:
+ br label %for.body
+
+for.body:
+ %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
+ %gep.src = getelementptr inbounds [3 x i8], ptr %p, i64 %i, i64 0
+ %l = load i8, ptr %gep.src, align 1
+ %gep.dst = getelementptr inbounds i8, ptr %q, i64 %i
+ store i8 %l, ptr %gep.dst, align 1
+ %i.next = add nuw nsw i64 %i, 1
+ %cond = icmp eq i64 %i.next, %n
+ br i1 %cond, label %for.end, label %for.body, !prof !6
+
+for.end:
+ ret void
+}
+
; For a loop with a profile-guided estimated TC of 129, when the auto-vectorizer chooses VF 16,
; it should choose conservatively IC 4 so that the vector loop runs twice at least
; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 4)
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_known_tc.ll b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_known_tc.ll
index 0569bfb2ae4e02..4f90be1e3129fd 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_known_tc.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_known_tc.ll
@@ -29,6 +29,28 @@ for.end:
ret void
}
+; This has the same trip count as loop_with_tc_32 but since the resulting interleaved group
+; in this case may access memory out-of-bounds, it requires a scalar epilogue iteration for
+; correctness, making at most 31 iterations available for interleaving.
+; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 2)
+define void @loop_with_tc_32_scalar_epilogue_reqd(ptr noalias %p, ptr noalias %q) {
+entry:
+ br label %for.body
+
+for.body:
+ %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
+ %gep.src = getelementptr inbounds [3 x i8], ptr %p, i64 %i, i64 0
+ %l = load i8, ptr %gep.src, align 1
+ %gep.dst = getelementptr inbounds i8, ptr %q, i64 %i
+ store i8 %l, ptr %gep.dst, align 1
+ %i.next = add nuw nsw i64 %i, 1
+ %cond = icmp eq i64 %i.next, 32
+ br i1 %cond, label %for.end, label %for.body
+
+for.end:
+ ret void
+}
+
; For this loop with known TC of 33, when the auto-vectorizer chooses VF 16, it should choose
; IC 2 since there is a small remainder loop TC that needs to run after the vector loop.
; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 2)
@@ -197,6 +219,28 @@ for.end:
ret void
}
+; This has the same trip count as loop_with_tc_128 but since the resulting interleaved group
+; in this case may access memory out-of-bounds, it requires a scalar epilogue iteration for
+; correctness, making at most 31 iterations available for interleaving.
+; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 8)
+define void @loop_with_tc_128_scalar_epilogue_reqd(ptr noalias %p, ptr noalias %q) {
+entry:
+ br label %for.body
+
+for.body:
+ %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
+ %gep.src = getelementptr inbounds [3 x i8], ptr %p, i64 %i, i64 0
+ %l = load i8, ptr %gep.src, align 1
+ %gep.dst = getelementptr inbounds i8, ptr %q, i64 %i
+ store i8 %l, ptr %gep.dst, align 1
+ %i.next = add nuw nsw i64 %i, 1
+ %cond = icmp eq i64 %i.next, 128
+ br i1 %cond, label %for.end, label %for.body
+
+for.end:
+ ret void
+}
+
; For this loop with known TC of 129, when the auto-vectorizer chooses VF 16, it should choose
; IC 8 since there is a small remainder loop that needs to run after the vector loop
; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 8)
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_loops_needing_scalar_epilogue.ll b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_loops_needing_scalar_epilogue.ll
new file mode 100644
index 00000000000000..b1b9e6d1e20a95
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_loops_needing_scalar_epilogue.ll
@@ -0,0 +1,132 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
+; RUN: opt < %s -force-target-max-vector-interleave=8 -p loop-vectorize -S 2>&1 | FileCheck %s
+; RUN: opt < %s -force-target-max-vector-interleave=8 -p loop-vectorize -pass-remarks=loop-vectorize -disable-output -S 2>&1 | FileCheck %s -check-prefix=CHECK-REMARKS
+
+target triple = "aarch64-linux-gnu"
+
+%pair = type { i8, i8 }
+
+; For this loop with known TC of 128, when the auto-vectorizer chooses VF 16, it should choose
+; IC 8 since there is no remainder loop run needed after the vector loop runs
+; CHECK-REMARKS: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 8)
+define void @loop_with_tc_128(ptr noalias %p, ptr noalias %q) {
+; CHECK-LABEL: define void @loop_with_tc_128(
+; CHECK-SAME: ptr noalias [[P:%.*]], ptr noalias [[Q:%.*]]) {
+; CHECK-NEXT: entry:
+; CHECK-NEXT: br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
+entry:
+ br label %for.body
+
+for.body:
+ %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
+ %gep.src = getelementptr %pair, ptr %p, i64 %i, i32 0
+ %load.src = load i8, ptr %gep.src, align 1
+ %gep.dst = getelementptr %pair, ptr %p, i64 %i, i32 1
+ %load.dst = load i8, ptr %gep.dst, align 1
+ %add = add i8 %load.src, %load.dst
+ %qi = getelementptr i8, ptr %q, i64 %i
+ store i8 %add, ptr %qi, align 1
+ %i.next = add nuw nsw i64 %i, 1
+ %cond = icmp eq i64 %i.next, 128
+ br i1 %cond, label %for.end, label %for.body
+
+for.end:
+ ret void
+}
+
+; This function has the same trip count as loop_with_tc_128 but since the resulting interleaved group
+; in this case may access memory out-of-bounds, it requires a scalar epilogue iteration for
+; correctness, making at most 127 iterations available for interleaving.
+; TODO: The entry block should branch into the vector loop, instead of the scalar epilogue.
+; TODO: When the auto-vectorizer chooses VF 16, it should choose IC 2, to have a smaller scalar remainder
+; than when using IC 4.
+; CHECK-REMARKS: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 8)
+define void @loop_with_tc_128_scalar_epilogue_reqd(ptr noalias %p, ptr noalias %q) {
+; CHECK-LABEL: define void @loop_with_tc_128_scalar_epilogue_reqd(
+; CHECK-SAME: ptr noalias [[P:%.*]], ptr noalias [[Q:%.*]]) {
+; CHECK-NEXT: entry:
+; CHECK-NEXT: br i1 true, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
+entry:
+ br label %for.body
+
+for.body:
+ %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
+ %gep.src = getelementptr inbounds [3 x i8], ptr %p, i64 %i, i64 0
+ %l = load i8, ptr %gep.src, align 1
+ %gep.dst = getelementptr inbounds i8, ptr %q, i64 %i
+ store i8 %l, ptr %gep.dst, align 1
+ %i.next = add nuw nsw i64 %i, 1
+ %cond = icmp eq i64 %i.next, 128
+ br i1 %cond, label %for.end, label %for.body
+
+for.end:
+ ret void
+}
+
+; For a loop with a profile-guided estimated TC of 128, when the auto-vectorizer chooses VF 16,
+; it should choose conservatively IC 4 so that the vector loop runs twice at least
+; CHECK-REMARKS: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 4)
+define void @loop_with_profile_tc_128(ptr noalias %p, ptr noalias %q, i64 %n) {
+; CHECK-LABEL: define void @loop_with_profile_tc_128(
+; CHECK-SAME: ptr noalias [[P:%.*]], ptr noalias [[Q:%.*]], i64 [[N:%.*]]) {
+; CHECK-NEXT: iter.check:
+; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], 8
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH:%.*]], label [[VECTOR_MAIN_LOOP_ITER_CHECK:%.*]], !prof [[PROF6:![0-9]+]]
+; CHECK: vector.main.loop.iter.check:
+; CHECK-NEXT: [[MIN_ITERS_CHECK1:%.*]] = icmp ult i64 [[N]], 64
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK1]], label [[VEC_EPILOG_PH:%.*]], label [[VECTOR_PH:%.*]], !prof [[PROF6]]
+;
+entry:
+ br label %for.body
+
+for.body:
+ %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
+ %gep.src = getelementptr %pair, ptr %p, i64 %i, i32 0
+ %load.src = load i8, ptr %gep.src, align 1
+ %gep.dst = getelementptr %pair, ptr %p, i64 %i, i32 1
+ %load.dst = load i8, ptr %gep.dst, align 1
+ %add = add i8 %load.src, %load.dst
+ %qi = getelementptr i8, ptr %q, i64 %i
+ store i8 %add, ptr %qi, align 1
+ %i.next = add nuw nsw i64 %i, 1
+ %cond = icmp eq i64 %i.next, %n
+ br i1 %cond, label %for.end, label %for.body, !prof !0
+
+for.end:
+ ret void
+}
+
+; This function has the same trip count as loop_with_profile_tc_128 but since the resulting interleaved group
+; in this case may access memory out-of-bounds, it requires a scalar epilogue iteration for
+; correctness, making at most 127 iterations available for interleaving.
+; TODO: When the auto-vectorizer chooses VF 16, it should choose IC 2, to have a smaller scalar remainder
+; than IC 4.
+; CHECK-REMARKS: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 4)
+define void @loop_with_profile_tc_128_scalar_epilogue_reqd(ptr noalias %p, ptr noalias %q, i64 %n) {
+; CHECK-LABEL: define void @loop_with_profile_tc_128_scalar_epilogue_reqd(
+; CHECK-SAME: ptr noalias [[P:%.*]], ptr noalias [[Q:%.*]], i64 [[N:%.*]]) {
+; CHECK-NEXT: iter.check:
+; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ule i64 [[N]], 8
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH:%.*]], label [[VECTOR_MAIN_LOOP_ITER_CHECK:%.*]], !prof [[PROF6]]
+; CHECK: vector.main.loop.iter.check:
+; CHECK-NEXT: [[MIN_ITERS_CHECK1:%.*]] = icmp ule i64 [[N]], 64
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK1]], label [[VEC_EPILOG_PH:%.*]], label [[VECTOR_PH:%.*]], !prof [[PROF6]]
+;
+entry:
+ br label %for.body
+
+for.body:
+ %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
+ %gep.src = getelementptr inbounds [3 x i8], ptr %p, i64 %i, i64 0
+ %l = load i8, ptr %gep.src, align 1
+ %gep.dst = getelementptr inbounds i8, ptr %q, i64 %i
+ store i8 %l, ptr %gep.dst, align 1
+ %i.next = add nuw nsw i64 %i, 1
+ %cond = icmp eq i64 %i.next, %n
+ br i1 %cond, label %for.end, label %for.body, !prof !0
+
+for.end:
+ ret void
+}
+
+!0 = !{!"branch_weights", i32 1, i32 127}
>From f5cb46884d4eacd818df19676149d31fa616553c Mon Sep 17 00:00:00 2001
From: Nilanjana Basu <n_basu at apple.com>
Date: Fri, 26 Jan 2024 10:33:19 -0800
Subject: [PATCH 2/4] [LV] Update interleaving count computation when scalar
epilogue loop needs to run at least once
---
.../Transforms/Vectorize/LoopVectorize.cpp | 22 +++++++++++++------
.../interleave_count_for_estimated_tc.ll | 8 +++----
.../AArch64/interleave_count_for_known_tc.ll | 10 ++++++---
...count_for_loops_needing_scalar_epilogue.ll | 14 ++++++------
4 files changed, 33 insertions(+), 21 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index aa5d1bfa57d535..c2fad626f2ee50 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -5393,7 +5393,11 @@ LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF,
assert(EstimatedVF >= 1 && "Estimated VF shouldn't be less than 1");
unsigned KnownTC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
- if (KnownTC) {
+ if (KnownTC > 0) {
+ // At least one iteration must be scalar when this constraint holds. So the
+ // maximum available iterations for interleaving is one less.
+ unsigned availableTC = (requiresScalarEpilogue(VF.isVector())) ? KnownTC - 1 : KnownTC;
+
// If trip count is known we select between two prospective ICs, where
// 1) the aggressive IC is capped by the trip count divided by VF
// 2) the conservative IC is capped by the trip count divided by (VF * 2)
@@ -5403,27 +5407,31 @@ LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF,
// we run the vector loop at least twice.
unsigned InterleaveCountUB = bit_floor(
- std::max(1u, std::min(KnownTC / EstimatedVF, MaxInterleaveCount)));
+ std::max(1u, std::min(availableTC / EstimatedVF, MaxInterleaveCount)));
unsigned InterleaveCountLB = bit_floor(std::max(
- 1u, std::min(KnownTC / (EstimatedVF * 2), MaxInterleaveCount)));
+ 1u, std::min(availableTC / (EstimatedVF * 2), MaxInterleaveCount)));
MaxInterleaveCount = InterleaveCountLB;
if (InterleaveCountUB != InterleaveCountLB) {
- unsigned TailTripCountUB = (KnownTC % (EstimatedVF * InterleaveCountUB));
- unsigned TailTripCountLB = (KnownTC % (EstimatedVF * InterleaveCountLB));
+ unsigned TailTripCountUB = (availableTC % (EstimatedVF * InterleaveCountUB));
+ unsigned TailTripCountLB = (availableTC % (EstimatedVF * InterleaveCountLB));
// If both produce same scalar tail, maximize the IC to do the same work
// in fewer vector loop iterations
if (TailTripCountUB == TailTripCountLB)
MaxInterleaveCount = InterleaveCountUB;
}
- } else if (BestKnownTC) {
+ } else if (BestKnownTC > 0) {
+ // At least one iteration must be scalar when this constraint holds. So the
+ // maximum available iterations for interleaving is one less.
+ unsigned availableTC = (requiresScalarEpilogue(VF.isVector())) ? (*BestKnownTC) - 1 : *BestKnownTC;
+
// If trip count is an estimated compile time constant, limit the
// IC to be capped by the trip count divided by VF * 2, such that the vector
// loop runs at least twice to make interleaving seem profitable when there
// is an epilogue loop present. Since exact Trip count is not known we
// choose to be conservative in our IC estimate.
MaxInterleaveCount = bit_floor(std::max(
- 1u, std::min(*BestKnownTC / (EstimatedVF * 2), MaxInterleaveCount)));
+ 1u, std::min(availableTC / (EstimatedVF * 2), MaxInterleaveCount)));
}
assert(MaxInterleaveCount > 0 &&
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_estimated_tc.ll b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_estimated_tc.ll
index d557ad1ead2563..2c046327f7c815 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_estimated_tc.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_estimated_tc.ll
@@ -128,9 +128,9 @@ for.end:
; This has the same trip count as loop_with_profile_tc_64 but since the resulting interleaved group
; in this case may access memory out-of-bounds, it requires a scalar epilogue iteration for
; correctness, making at most 63 iterations available for interleaving.
-; TODO: When the auto-vectorizer chooses VF 16, it should choose IC 1 to leave a smaller scalar remainder
+; When the auto-vectorizer chooses VF 16, it should choose IC 1 to leave a smaller scalar remainder
; than IC 2
-; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 2)
+; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 1)
define void @loop_with_profile_tc_64_scalar_epilogue_reqd(ptr noalias %p, ptr noalias %q, i64 %n) {
entry:
br label %for.body
@@ -200,9 +200,9 @@ for.end:
; This has the same trip count as loop_with_profile_tc_128 but since the resulting interleaved group
; in this case may access memory out-of-bounds, it requires a scalar epilogue iteration for
; correctness, making at most 127 iterations available for interleaving.
-; TODO: When the auto-vectorizer chooses VF 16, it should choose IC 2 to leave a smaller scalar remainder
+; When the auto-vectorizer chooses VF 16, it should choose IC 2 to leave a smaller scalar remainder
; than IC 4
-; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 4)
+; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 2)
define void @loop_with_profile_tc_128_scalar_epilogue_reqd(ptr noalias %p, ptr noalias %q, i64 %n) {
entry:
br label %for.body
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_known_tc.ll b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_known_tc.ll
index 4f90be1e3129fd..9209df400afe4a 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_known_tc.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_known_tc.ll
@@ -32,7 +32,9 @@ for.end:
; This has the same trip count as loop_with_tc_32 but since the resulting interleaved group
; in this case may access memory out-of-bounds, it requires a scalar epilogue iteration for
; correctness, making at most 31 iterations available for interleaving.
-; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 2)
+; When the auto-vectorizer chooses VF 16, it should choose IC 1 to leave a smaller scalar remainder
+; than IC 2
+; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 1)
define void @loop_with_tc_32_scalar_epilogue_reqd(ptr noalias %p, ptr noalias %q) {
entry:
br label %for.body
@@ -221,8 +223,10 @@ for.end:
; This has the same trip count as loop_with_tc_128 but since the resulting interleaved group
; in this case may access memory out-of-bounds, it requires a scalar epilogue iteration for
-; correctness, making at most 31 iterations available for interleaving.
-; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 8)
+; correctness, making at most 127 iterations available for interleaving.
+; When the auto-vectorizer chooses VF 16, it should choose IC 2 to leave a smaller scalar remainder
+; than IC 4
+; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 2)
define void @loop_with_tc_128_scalar_epilogue_reqd(ptr noalias %p, ptr noalias %q) {
entry:
br label %for.body
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_loops_needing_scalar_epilogue.ll b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_loops_needing_scalar_epilogue.ll
index b1b9e6d1e20a95..7a09327b4e1a37 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_loops_needing_scalar_epilogue.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count_for_loops_needing_scalar_epilogue.ll
@@ -37,15 +37,15 @@ for.end:
; This function has the same trip count as loop_with_tc_128 but since the resulting interleaved group
; in this case may access memory out-of-bounds, it requires a scalar epilogue iteration for
; correctness, making at most 127 iterations available for interleaving.
-; TODO: The entry block should branch into the vector loop, instead of the scalar epilogue.
-; TODO: When the auto-vectorizer chooses VF 16, it should choose IC 2, to have a smaller scalar remainder
+; The entry block should branch into the vector loop, instead of the scalar epilogue.
+; When the auto-vectorizer chooses VF 16, it should choose IC 2, to have a smaller scalar remainder
; than when using IC 4.
-; CHECK-REMARKS: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 8)
+; CHECK-REMARKS: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 2)
define void @loop_with_tc_128_scalar_epilogue_reqd(ptr noalias %p, ptr noalias %q) {
; CHECK-LABEL: define void @loop_with_tc_128_scalar_epilogue_reqd(
; CHECK-SAME: ptr noalias [[P:%.*]], ptr noalias [[Q:%.*]]) {
; CHECK-NEXT: entry:
-; CHECK-NEXT: br i1 true, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
+; CHECK-NEXT: br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
entry:
br label %for.body
@@ -99,9 +99,9 @@ for.end:
; This function has the same trip count as loop_with_profile_tc_128 but since the resulting interleaved group
; in this case may access memory out-of-bounds, it requires a scalar epilogue iteration for
; correctness, making at most 127 iterations available for interleaving.
-; TODO: When the auto-vectorizer chooses VF 16, it should choose IC 2, to have a smaller scalar remainder
+; When the auto-vectorizer chooses VF 16, it should choose IC 2, to have a smaller scalar remainder
; than IC 4.
-; CHECK-REMARKS: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 4)
+; CHECK-REMARKS: remark: <unknown>:0:0: vectorized loop (vectorization width: 16, interleaved count: 2)
define void @loop_with_profile_tc_128_scalar_epilogue_reqd(ptr noalias %p, ptr noalias %q, i64 %n) {
; CHECK-LABEL: define void @loop_with_profile_tc_128_scalar_epilogue_reqd(
; CHECK-SAME: ptr noalias [[P:%.*]], ptr noalias [[Q:%.*]], i64 [[N:%.*]]) {
@@ -109,7 +109,7 @@ define void @loop_with_profile_tc_128_scalar_epilogue_reqd(ptr noalias %p, ptr n
; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ule i64 [[N]], 8
; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH:%.*]], label [[VECTOR_MAIN_LOOP_ITER_CHECK:%.*]], !prof [[PROF6]]
; CHECK: vector.main.loop.iter.check:
-; CHECK-NEXT: [[MIN_ITERS_CHECK1:%.*]] = icmp ule i64 [[N]], 64
+; CHECK-NEXT: [[MIN_ITERS_CHECK1:%.*]] = icmp ule i64 [[N]], 32
; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK1]], label [[VEC_EPILOG_PH:%.*]], label [[VECTOR_PH:%.*]], !prof [[PROF6]]
;
entry:
>From d5ae1c4c7a1269a556b8646036621610d1db87d4 Mon Sep 17 00:00:00 2001
From: Nilanjana Basu <n_basu at apple.com>
Date: Fri, 26 Jan 2024 14:14:51 -0800
Subject: [PATCH 3/4] ran clang-format
---
llvm/lib/Transforms/Vectorize/LoopVectorize.cpp | 13 +++++++++----
1 file changed, 9 insertions(+), 4 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index c2fad626f2ee50..d8ef561f23a576 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -5396,7 +5396,8 @@ LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF,
if (KnownTC > 0) {
// At least one iteration must be scalar when this constraint holds. So the
// maximum available iterations for interleaving is one less.
- unsigned availableTC = (requiresScalarEpilogue(VF.isVector())) ? KnownTC - 1 : KnownTC;
+ unsigned availableTC =
+ requiresScalarEpilogue(VF.isVector()) ? KnownTC - 1 : KnownTC;
// If trip count is known we select between two prospective ICs, where
// 1) the aggressive IC is capped by the trip count divided by VF
@@ -5413,8 +5414,10 @@ LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF,
MaxInterleaveCount = InterleaveCountLB;
if (InterleaveCountUB != InterleaveCountLB) {
- unsigned TailTripCountUB = (availableTC % (EstimatedVF * InterleaveCountUB));
- unsigned TailTripCountLB = (availableTC % (EstimatedVF * InterleaveCountLB));
+ unsigned TailTripCountUB =
+ (availableTC % (EstimatedVF * InterleaveCountUB));
+ unsigned TailTripCountLB =
+ (availableTC % (EstimatedVF * InterleaveCountLB));
// If both produce same scalar tail, maximize the IC to do the same work
// in fewer vector loop iterations
if (TailTripCountUB == TailTripCountLB)
@@ -5423,7 +5426,9 @@ LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF,
} else if (BestKnownTC > 0) {
// At least one iteration must be scalar when this constraint holds. So the
// maximum available iterations for interleaving is one less.
- unsigned availableTC = (requiresScalarEpilogue(VF.isVector())) ? (*BestKnownTC) - 1 : *BestKnownTC;
+ unsigned availableTC = requiresScalarEpilogue(VF.isVector())
+ ? (*BestKnownTC) - 1
+ : *BestKnownTC;
// If trip count is an estimated compile time constant, limit the
// IC to be capped by the trip count divided by VF * 2, such that the vector
>From e1595d3f5c3ae24e44f29f0dae3ee151c0e751cc Mon Sep 17 00:00:00 2001
From: Nilanjana Basu <n_basu at apple.com>
Date: Fri, 26 Jan 2024 14:18:24 -0800
Subject: [PATCH 4/4] nit: updating variable name
---
llvm/lib/Transforms/Vectorize/LoopVectorize.cpp | 14 +++++++-------
1 file changed, 7 insertions(+), 7 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index d8ef561f23a576..af197e006b4dd0 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -5396,7 +5396,7 @@ LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF,
if (KnownTC > 0) {
// At least one iteration must be scalar when this constraint holds. So the
// maximum available iterations for interleaving is one less.
- unsigned availableTC =
+ unsigned AvailableTC =
requiresScalarEpilogue(VF.isVector()) ? KnownTC - 1 : KnownTC;
// If trip count is known we select between two prospective ICs, where
@@ -5408,16 +5408,16 @@ LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF,
// we run the vector loop at least twice.
unsigned InterleaveCountUB = bit_floor(
- std::max(1u, std::min(availableTC / EstimatedVF, MaxInterleaveCount)));
+ std::max(1u, std::min(AvailableTC / EstimatedVF, MaxInterleaveCount)));
unsigned InterleaveCountLB = bit_floor(std::max(
- 1u, std::min(availableTC / (EstimatedVF * 2), MaxInterleaveCount)));
+ 1u, std::min(AvailableTC / (EstimatedVF * 2), MaxInterleaveCount)));
MaxInterleaveCount = InterleaveCountLB;
if (InterleaveCountUB != InterleaveCountLB) {
unsigned TailTripCountUB =
- (availableTC % (EstimatedVF * InterleaveCountUB));
+ (AvailableTC % (EstimatedVF * InterleaveCountUB));
unsigned TailTripCountLB =
- (availableTC % (EstimatedVF * InterleaveCountLB));
+ (AvailableTC % (EstimatedVF * InterleaveCountLB));
// If both produce same scalar tail, maximize the IC to do the same work
// in fewer vector loop iterations
if (TailTripCountUB == TailTripCountLB)
@@ -5426,7 +5426,7 @@ LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF,
} else if (BestKnownTC > 0) {
// At least one iteration must be scalar when this constraint holds. So the
// maximum available iterations for interleaving is one less.
- unsigned availableTC = requiresScalarEpilogue(VF.isVector())
+ unsigned AvailableTC = requiresScalarEpilogue(VF.isVector())
? (*BestKnownTC) - 1
: *BestKnownTC;
@@ -5436,7 +5436,7 @@ LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF,
// is an epilogue loop present. Since exact Trip count is not known we
// choose to be conservative in our IC estimate.
MaxInterleaveCount = bit_floor(std::max(
- 1u, std::min(availableTC / (EstimatedVF * 2), MaxInterleaveCount)));
+ 1u, std::min(AvailableTC / (EstimatedVF * 2), MaxInterleaveCount)));
}
assert(MaxInterleaveCount > 0 &&
More information about the llvm-commits
mailing list