[llvm] [LV] Change loops' interleave count computation (PR #70141)
Nilanjana Basu via llvm-commits
llvm-commits at lists.llvm.org
Wed Oct 25 17:39:43 PDT 2023
https://github.com/nilanjana87 updated https://github.com/llvm/llvm-project/pull/70141
>From bdcbd522964ef728d94553a45aba4a21af078e47 Mon Sep 17 00:00:00 2001
From: Nilanjana Basu <n_basu at apple.com>
Date: Tue, 24 Oct 2023 15:48:46 -0700
Subject: [PATCH 1/3] [LV] Pre-committing tests for changing interleaving count
computation
---
.../LoopVectorize/AArch64/interleave_count.ll | 106 ++++++++++++++++++
1 file changed, 106 insertions(+)
create mode 100644 llvm/test/Transforms/LoopVectorize/AArch64/interleave_count.ll
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count.ll b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count.ll
new file mode 100644
index 000000000000000..74fd2f44fce0109
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count.ll
@@ -0,0 +1,106 @@
+; RUN: opt < %s -force-vector-width=64 -O3 -S -pass-remarks=loop-vectorize 2>&1 | FileCheck %s
+
+target triple = "aarch64-linux-gnu"
+
+%pair = type { i8, i8 }
+
+; For a loop with known trip count of 128, when we force VF 64, it should use
+; IC 2, since there is no remainder loop needed when the vector loop runs.
+; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 64, interleaved count: 2)
+define void @loop_with_tc_128(ptr %p, ptr %q) {
+entry:
+ br label %for.body
+
+for.body:
+ %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
+ %tmp0 = getelementptr %pair, ptr %p, i64 %i, i32 0
+ %tmp1 = load i8, ptr %tmp0, align 1
+ %tmp2 = getelementptr %pair, ptr %p, i64 %i, i32 1
+ %tmp3 = load i8, ptr %tmp2, align 1
+ %add = add i8 %tmp1, %tmp3
+ %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
+}
+
+; For a loop with known trip count of 129, when we force VF 64, it should use
+; IC 1, since there may be a remainder loop that needs to run after the vector loop.
+; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 64, interleaved count: 2)
+define void @loop_with_tc_129(ptr %p, ptr %q) {
+entry:
+ br label %for.body
+
+for.body:
+ %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
+ %tmp0 = getelementptr %pair, ptr %p, i64 %i, i32 0
+ %tmp1 = load i8, ptr %tmp0, align 1
+ %tmp2 = getelementptr %pair, ptr %p, i64 %i, i32 1
+ %tmp3 = load i8, ptr %tmp2, align 1
+ %add = add i8 %tmp1, %tmp3
+ %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, 129
+ br i1 %cond, label %for.end, label %for.body
+
+for.end:
+ ret void
+}
+
+; For a loop with unknown trip count but a profile showing an approx TC estimate of 128,
+; when we force VF 64, it should use IC 2, since chances are high that the remainder loop
+; won't need to run
+; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 64, interleaved count: 2)
+define void @loop_with_profile_tc_128(ptr %p, ptr %q, i64 %n) {
+entry:
+ br label %for.body
+
+for.body:
+ %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
+ %tmp0 = getelementptr %pair, ptr %p, i64 %i, i32 0
+ %tmp1 = load i8, ptr %tmp0, align 1
+ %tmp2 = getelementptr %pair, ptr %p, i64 %i, i32 1
+ %tmp3 = load i8, ptr %tmp2, align 1
+ %add = add i8 %tmp1, %tmp3
+ %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
+}
+
+; For a loop with unknown trip count but a profile showing an approx TC estimate of 129,
+; when we force VF 64, it should use IC 1, since chances are high that the remainder loop
+; will need to run
+; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 64, interleaved count: 2)
+define void @loop_with_profile_tc_129(ptr %p, ptr %q, i64 %n) {
+entry:
+ br label %for.body
+
+for.body:
+ %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
+ %tmp0 = getelementptr %pair, ptr %p, i64 %i, i32 0
+ %tmp1 = load i8, ptr %tmp0, align 1
+ %tmp2 = getelementptr %pair, ptr %p, i64 %i, i32 1
+ %tmp3 = load i8, ptr %tmp2, align 1
+ %add = add i8 %tmp1, %tmp3
+ %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 !1
+
+for.end:
+ ret void
+}
+
+!0 = !{!"branch_weights", i32 1, i32 127}
+!1 = !{!"branch_weights", i32 1, i32 128}
>From 2ef770ddff79ebf0e971418a022869c89a61ebf9 Mon Sep 17 00:00:00 2001
From: Nilanjana Basu <n_basu at apple.com>
Date: Sun, 22 Oct 2023 23:19:19 -0700
Subject: [PATCH 2/3] [LV] Change loops' interleave count computation
A set of microbenchmarks in llvm-test-suite (https://github.com/llvm/llvm-test-suite/pull/26), when tested on a AArch64 platform, demonstrates that loop interleaving is beneficial in two cases:
1) when TC > 2 * VW * IC, such that the interleaved vectorized portion of the loop runs at least twice
2) when TC is an exact multiple of VW * IC, such that there is no epilogue loop to run
where, TC = trip count, VW = vectorization width, IC = interleaving count
We change the interleave count computation based on this information but we leave it the same when the flag InterleaveSmallLoopScalarReductionTrue is set to true, since it handles a special case (https://reviews.llvm.org/D81416).
---
.../Transforms/Vectorize/LoopVectorize.cpp | 21 +++++++++++++------
.../LoopVectorize/AArch64/interleave_count.ll | 4 ++--
2 files changed, 17 insertions(+), 8 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index cc17d91d4f43727..5786e8e38cb4974 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -5745,8 +5745,12 @@ LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF,
}
// If trip count is known or estimated compile time constant, limit the
- // interleave count to be less than the trip count divided by VF, provided it
- // is at least 1.
+ // interleave count to be less than the trip count divided by VF * 2,
+ // provided VF is at least 1 and the trip count is not an exact multiple of
+ // VF, such that the vector loop runs at least twice to make interleaving seem
+ // profitable when there is an epilogue loop present. When
+ // InterleaveSmallLoopScalarReduction is true or trip count is an exact
+ // multiple of VF, we allow interleaving even when the vector loop runs once.
//
// For scalable vectors we can't know if interleaving is beneficial. It may
// not be beneficial for small loops if none of the lanes in the second vector
@@ -5755,10 +5759,15 @@ LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF,
// the InterleaveCount as if vscale is '1', although if some information about
// the vector is known (e.g. min vector size), we can make a better decision.
if (BestKnownTC) {
- MaxInterleaveCount =
- std::min(*BestKnownTC / VF.getKnownMinValue(), MaxInterleaveCount);
- // Make sure MaxInterleaveCount is greater than 0.
- MaxInterleaveCount = std::max(1u, MaxInterleaveCount);
+ if (InterleaveSmallLoopScalarReduction ||
+ (*BestKnownTC % VF.getKnownMinValue() == 0))
+ MaxInterleaveCount =
+ std::min(*BestKnownTC / VF.getKnownMinValue(), MaxInterleaveCount);
+ else
+ MaxInterleaveCount = std::min(*BestKnownTC / (VF.getKnownMinValue() * 2),
+ MaxInterleaveCount);
+ // Make sure MaxInterleaveCount is greater than 0 & a power of 2.
+ MaxInterleaveCount = llvm::bit_floor(std::max(1u, MaxInterleaveCount));
}
assert(MaxInterleaveCount > 0 &&
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count.ll b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count.ll
index 74fd2f44fce0109..501fc334bc8bf48 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/interleave_count.ll
@@ -30,7 +30,7 @@ for.end:
; For a loop with known trip count of 129, when we force VF 64, it should use
; IC 1, since there may be a remainder loop that needs to run after the vector loop.
-; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 64, interleaved count: 2)
+; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 64, interleaved count: 1)
define void @loop_with_tc_129(ptr %p, ptr %q) {
entry:
br label %for.body
@@ -80,7 +80,7 @@ for.end:
; For a loop with unknown trip count but a profile showing an approx TC estimate of 129,
; when we force VF 64, it should use IC 1, since chances are high that the remainder loop
; will need to run
-; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 64, interleaved count: 2)
+; CHECK: remark: <unknown>:0:0: vectorized loop (vectorization width: 64, interleaved count: 1)
define void @loop_with_profile_tc_129(ptr %p, ptr %q, i64 %n) {
entry:
br label %for.body
>From 566e26b5b9e6baa57353383697dbfef681ba18a3 Mon Sep 17 00:00:00 2001
From: Nilanjana Basu <n_basu at apple.com>
Date: Wed, 25 Oct 2023 17:33:16 -0700
Subject: [PATCH 3/3] [LV] Adding tests for loop unrolling.
Since this patch is the first part of a two-part patch for changing loop interleaving count computation, followed by removing loop interleaving threshold, these tests remain unchanged for the current (first) patch but will be modified for the second patch. The original PR for these tests is in https://github.com/llvm/llvm-project/pull/70272.
---
.../LoopVectorize/X86/unroll-small-loops.ll | 194 +++++++++++++++++-
1 file changed, 190 insertions(+), 4 deletions(-)
diff --git a/llvm/test/Transforms/LoopVectorize/X86/unroll-small-loops.ll b/llvm/test/Transforms/LoopVectorize/X86/unroll-small-loops.ll
index 290be569bc12524..6b71cfef06be149 100644
--- a/llvm/test/Transforms/LoopVectorize/X86/unroll-small-loops.ll
+++ b/llvm/test/Transforms/LoopVectorize/X86/unroll-small-loops.ll
@@ -8,20 +8,20 @@ target triple = "x86_64-apple-macosx10.8.0"
; We don't unroll this loop because it has a small constant trip count.
;
-; CHECK-VECTOR-LABEL: @foo(
+; CHECK-VECTOR-LABEL: @foo_trip_count_8(
; CHECK-VECTOR: load <4 x i32>
; CHECK-VECTOR-NOT: load <4 x i32>
; CHECK-VECTOR: store <4 x i32>
; CHECK-VECTOR-NOT: store <4 x i32>
; CHECK-VECTOR: ret
;
-; CHECK-SCALAR-LABEL: @foo(
+; CHECK-SCALAR-LABEL: @foo_trip_count_8(
; CHECK-SCALAR: load i32, ptr
; CHECK-SCALAR-NOT: load i32, ptr
; CHECK-SCALAR: store i32
; CHECK-SCALAR-NOT: store i32
; CHECK-SCALAR: ret
-define i32 @foo(ptr nocapture %A) nounwind uwtable ssp {
+define i32 @foo_trip_count_8(ptr nocapture %A) nounwind uwtable ssp {
br label %1
; <label>:1 ; preds = %1, %0
@@ -32,7 +32,193 @@ define i32 @foo(ptr nocapture %A) nounwind uwtable ssp {
store i32 %4, ptr %2, align 4
%indvars.iv.next = add i64 %indvars.iv, 1
%lftr.wideiv = trunc i64 %indvars.iv.next to i32
- %exitcond = icmp eq i32 %lftr.wideiv, 100
+ %exitcond = icmp eq i32 %lftr.wideiv, 8
+ br i1 %exitcond, label %5, label %1
+
+; <label>:5 ; preds = %1
+ ret i32 undef
+}
+
+; CHECK-VECTOR-LABEL: @foo_trip_count_16(
+; CHECK-VECTOR: load <4 x i32>
+; CHECK-VECTOR-NOT: load <4 x i32>
+; CHECK-VECTOR: store <4 x i32>
+; CHECK-VECTOR-NOT: store <4 x i32>
+; CHECK-VECTOR: ret
+;
+; CHECK-SCALAR-LABEL: @foo_trip_count_16(
+; CHECK-SCALAR: load i32, ptr
+; CHECK-SCALAR-NOT: load i32, ptr
+; CHECK-SCALAR: store i32
+; CHECK-SCALAR-NOT: store i32
+; CHECK-SCALAR: ret
+define i32 @foo_trip_count_16(ptr nocapture %A) nounwind uwtable ssp {
+ br label %1
+
+; <label>:1 ; preds = %1, %0
+ %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+ %2 = getelementptr inbounds i32, ptr %A, i64 %indvars.iv
+ %3 = load i32, ptr %2, align 4
+ %4 = add nsw i32 %3, 6
+ store i32 %4, ptr %2, align 4
+ %indvars.iv.next = add i64 %indvars.iv, 1
+ %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+ %exitcond = icmp eq i32 %lftr.wideiv, 16
+ br i1 %exitcond, label %5, label %1
+
+; <label>:5 ; preds = %1
+ ret i32 undef
+}
+
+; CHECK-VECTOR-LABEL: @foo_trip_count_17(
+; CHECK-VECTOR: load <4 x i32>
+; CHECK-VECTOR-NOT: load <4 x i32>
+; CHECK-VECTOR: store <4 x i32>
+; CHECK-VECTOR-NOT: store <4 x i32>
+; CHECK-VECTOR: ret
+;
+; CHECK-SCALAR-LABEL: @foo_trip_count_17(
+; CHECK-SCALAR: load i32, ptr
+; CHECK-SCALAR-NOT: load i32, ptr
+; CHECK-SCALAR: store i32
+; CHECK-SCALAR-NOT: store i32
+; CHECK-SCALAR: ret
+define i32 @foo_trip_count_17(ptr nocapture %A) nounwind uwtable ssp {
+ br label %1
+
+; <label>:1 ; preds = %1, %0
+ %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+ %2 = getelementptr inbounds i32, ptr %A, i64 %indvars.iv
+ %3 = load i32, ptr %2, align 4
+ %4 = add nsw i32 %3, 6
+ store i32 %4, ptr %2, align 4
+ %indvars.iv.next = add i64 %indvars.iv, 1
+ %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+ %exitcond = icmp eq i32 %lftr.wideiv, 17
+ br i1 %exitcond, label %5, label %1
+
+; <label>:5 ; preds = %1
+ ret i32 undef
+}
+
+; CHECK-VECTOR-LABEL: @foo_trip_count_24(
+; CHECK-VECTOR: load <4 x i32>
+; CHECK-VECTOR-NOT: load <4 x i32>
+; CHECK-VECTOR: store <4 x i32>
+; CHECK-VECTOR-NOT: store <4 x i32>
+; CHECK-VECTOR: ret
+;
+; CHECK-SCALAR-LABEL: @foo_trip_count_24(
+; CHECK-SCALAR: load i32, ptr
+; CHECK-SCALAR-NOT: load i32, ptr
+; CHECK-SCALAR: store i32
+; CHECK-SCALAR-NOT: store i32
+; CHECK-SCALAR: ret
+define i32 @foo_trip_count_24(ptr nocapture %A) nounwind uwtable ssp {
+ br label %1
+
+; <label>:1 ; preds = %1, %0
+ %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+ %2 = getelementptr inbounds i32, ptr %A, i64 %indvars.iv
+ %3 = load i32, ptr %2, align 4
+ %4 = add nsw i32 %3, 6
+ store i32 %4, ptr %2, align 4
+ %indvars.iv.next = add i64 %indvars.iv, 1
+ %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+ %exitcond = icmp eq i32 %lftr.wideiv, 24
+ br i1 %exitcond, label %5, label %1
+
+; <label>:5 ; preds = %1
+ ret i32 undef
+}
+
+; CHECK-VECTOR-LABEL: @foo_trip_count_25(
+; CHECK-VECTOR: load <4 x i32>
+; CHECK-VECTOR-NOT: load <4 x i32>
+; CHECK-VECTOR: store <4 x i32>
+; CHECK-VECTOR-NOT: store <4 x i32>
+; CHECK-VECTOR: ret
+;
+; CHECK-SCALAR-LABEL: @foo_trip_count_25(
+; CHECK-SCALAR: load i32, ptr
+; CHECK-SCALAR-NOT: load i32, ptr
+; CHECK-SCALAR: store i32
+; CHECK-SCALAR-NOT: store i32
+; CHECK-SCALAR: ret
+define i32 @foo_trip_count_25(ptr nocapture %A) nounwind uwtable ssp {
+ br label %1
+
+; <label>:1 ; preds = %1, %0
+ %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+ %2 = getelementptr inbounds i32, ptr %A, i64 %indvars.iv
+ %3 = load i32, ptr %2, align 4
+ %4 = add nsw i32 %3, 6
+ store i32 %4, ptr %2, align 4
+ %indvars.iv.next = add i64 %indvars.iv, 1
+ %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+ %exitcond = icmp eq i32 %lftr.wideiv, 25
+ br i1 %exitcond, label %5, label %1
+
+; <label>:5 ; preds = %1
+ ret i32 undef
+}
+
+; CHECK-VECTOR-LABEL: @foo_trip_count_33(
+; CHECK-VECTOR: load <4 x i32>
+; CHECK-VECTOR-NOT: load <4 x i32>
+; CHECK-VECTOR: store <4 x i32>
+; CHECK-VECTOR-NOT: store <4 x i32>
+; CHECK-VECTOR: ret
+;
+; CHECK-SCALAR-LABEL: @foo_trip_count_33(
+; CHECK-SCALAR: load i32, ptr
+; CHECK-SCALAR-NOT: load i32, ptr
+; CHECK-SCALAR: store i32
+; CHECK-SCALAR-NOT: store i32
+; CHECK-SCALAR: ret
+define i32 @foo_trip_count_33(ptr nocapture %A) nounwind uwtable ssp {
+ br label %1
+
+; <label>:1 ; preds = %1, %0
+ %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+ %2 = getelementptr inbounds i32, ptr %A, i64 %indvars.iv
+ %3 = load i32, ptr %2, align 4
+ %4 = add nsw i32 %3, 6
+ store i32 %4, ptr %2, align 4
+ %indvars.iv.next = add i64 %indvars.iv, 1
+ %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+ %exitcond = icmp eq i32 %lftr.wideiv, 33
+ br i1 %exitcond, label %5, label %1
+
+; <label>:5 ; preds = %1
+ ret i32 undef
+}
+
+; CHECK-VECTOR-LABEL: @foo_trip_count_101(
+; CHECK-VECTOR: load <4 x i32>
+; CHECK-VECTOR-NOT: load <4 x i32>
+; CHECK-VECTOR: store <4 x i32>
+; CHECK-VECTOR-NOT: store <4 x i32>
+; CHECK-VECTOR: ret
+;
+; CHECK-SCALAR-LABEL: @foo_trip_count_101(
+; CHECK-SCALAR: load i32, ptr
+; CHECK-SCALAR-NOT: load i32, ptr
+; CHECK-SCALAR: store i32
+; CHECK-SCALAR-NOT: store i32
+; CHECK-SCALAR: ret
+define i32 @foo_trip_count_101(ptr nocapture %A) nounwind uwtable ssp {
+ br label %1
+
+; <label>:1 ; preds = %1, %0
+ %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+ %2 = getelementptr inbounds i32, ptr %A, i64 %indvars.iv
+ %3 = load i32, ptr %2, align 4
+ %4 = add nsw i32 %3, 6
+ store i32 %4, ptr %2, align 4
+ %indvars.iv.next = add i64 %indvars.iv, 1
+ %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+ %exitcond = icmp eq i32 %lftr.wideiv, 101
br i1 %exitcond, label %5, label %1
; <label>:5 ; preds = %1
More information about the llvm-commits
mailing list