[llvm] LoopLoadElim: add pre-commit test for #96656 (PR #97598)
Ramkumar Ramachandra via llvm-commits
llvm-commits at lists.llvm.org
Wed Jul 3 09:09:54 PDT 2024
https://github.com/artagnon created https://github.com/llvm/llvm-project/pull/97598
After pr96656.ll were added to LAA and LoopVersioning, it was decided that the bug is in a caller of LoopVersioning, not in LAA or LoopVersioning itself. The caller has now been found to be LoopLoadElim. Hence, re-organize the added tests to avoid confusion, and add a new reduced-test for #96656 to LoopLoadElim, in preparation to fix the bug.
>From 93c9e035c7b45d253448a6b9301c58f5cba7a57b Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Wed, 3 Jul 2024 14:57:10 +0100
Subject: [PATCH] LoopLoadElim: add pre-commit test for #96656
After pr96656.ll were added to LAA and LoopVersioning, it was decided
that the bug is in a caller of LoopVersioning, not in LAA or
LoopVersioning itself. The caller has now been found to be
LoopLoadElim. Hence, re-organize the added tests to avoid confusion, and
add a new reduced-test for #96656 to LoopLoadElim, in preparation to fix
the bug.
---
.../Analysis/LoopAccessAnalysis/pr96656.ll | 49 -----------
.../LoopAccessAnalysis/symbolic-stride.ll | 48 ++++++++++
llvm/test/Transforms/LoopLoadElim/pr96656.ll | 87 +++++++++++++++++++
.../{pr96656.ll => single-iteration.ll} | 36 ++++----
4 files changed, 155 insertions(+), 65 deletions(-)
delete mode 100644 llvm/test/Analysis/LoopAccessAnalysis/pr96656.ll
create mode 100644 llvm/test/Transforms/LoopLoadElim/pr96656.ll
rename llvm/test/Transforms/LoopVersioning/{pr96656.ll => single-iteration.ll} (75%)
diff --git a/llvm/test/Analysis/LoopAccessAnalysis/pr96656.ll b/llvm/test/Analysis/LoopAccessAnalysis/pr96656.ll
deleted file mode 100644
index 5b9833553fa02..0000000000000
--- a/llvm/test/Analysis/LoopAccessAnalysis/pr96656.ll
+++ /dev/null
@@ -1,49 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --version 5
-; RUN: opt -passes='print<access-info>' -disable-output %s 2>&1 | FileCheck %s
-
-define void @false.equal.predicate(ptr %arg, ptr %arg1, i1 %arg2) {
-; CHECK-LABEL: 'false.equal.predicate'
-; CHECK-NEXT: loop.body:
-; CHECK-NEXT: Memory dependences are safe
-; CHECK-NEXT: Dependences:
-; CHECK-NEXT: Run-time memory checks:
-; CHECK-NEXT: Grouped accesses:
-; CHECK-EMPTY:
-; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
-; CHECK-NEXT: SCEV assumptions:
-; CHECK-NEXT: Equal predicate: %load == 1
-; CHECK-EMPTY:
-; CHECK-NEXT: Expressions re-written:
-; CHECK-NEXT: [PSE] %gep10 = getelementptr double, ptr %gep8, i64 %mul:
-; CHECK-NEXT: {(8 + %arg1),+,(8 * (sext i32 %load to i64))<nsw>}<%loop.body>
-; CHECK-NEXT: --> {(8 + %arg1),+,8}<%loop.body>
-;
-entry:
- %load = load i32, ptr %arg, align 4
- br i1 %arg2, label %noloop.exit, label %loop.ph
-
-loop.ph: ; preds = %entry
- %sext7 = sext i32 %load to i64
- %gep8 = getelementptr i8, ptr %arg1, i64 8
- br label %loop.body
-
-loop.body: ; preds = %loop.body, %loop.ph
- %phi = phi i64 [ 0, %loop.ph ], [ %add, %loop.body ]
- %mul = mul i64 %phi, %sext7
- %gep10 = getelementptr double, ptr %gep8, i64 %mul
- %load11 = load double, ptr %gep10, align 8
- store double %load11, ptr %arg1, align 8
- %add = add i64 %phi, 1
- %icmp = icmp eq i64 %phi, 0
- br i1 %icmp, label %loop.exit, label %loop.body
-
-noloop.exit: ; preds = %entry
- %sext = sext i32 %load to i64
- %gep = getelementptr double, ptr %arg1, i64 %sext
- %load5 = load double, ptr %gep, align 8
- store double %load5, ptr %arg, align 8
- ret void
-
-loop.exit: ; preds = %loop.body
- ret void
-}
diff --git a/llvm/test/Analysis/LoopAccessAnalysis/symbolic-stride.ll b/llvm/test/Analysis/LoopAccessAnalysis/symbolic-stride.ll
index 7c1b11e22aef2..0871ae84c3711 100644
--- a/llvm/test/Analysis/LoopAccessAnalysis/symbolic-stride.ll
+++ b/llvm/test/Analysis/LoopAccessAnalysis/symbolic-stride.ll
@@ -223,6 +223,54 @@ exit:
ret void
}
+define double @single_iteration_unknown_stride(i32 %arg, ptr %arg1, i1 %arg2) {
+; CHECK-LABEL: 'single_iteration_unknown_stride'
+; CHECK-NEXT: loop.body:
+; CHECK-NEXT: Memory dependences are safe
+; CHECK-NEXT: Dependences:
+; CHECK-NEXT: Run-time memory checks:
+; CHECK-NEXT: Grouped accesses:
+; CHECK-EMPTY:
+; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
+; CHECK-NEXT: SCEV assumptions:
+; CHECK-NEXT: Equal predicate: %arg == 1
+; CHECK-EMPTY:
+; CHECK-NEXT: Expressions re-written:
+; CHECK-NEXT: [PSE] %gep10 = getelementptr double, ptr %gep8, i64 %mul:
+; CHECK-NEXT: {(8 + %arg1),+,(8 * (sext i32 %arg to i64))<nsw>}<%loop.body>
+; CHECK-NEXT: --> {(8 + %arg1),+,8}<%loop.body>
+;
+entry:
+ br i1 %arg2, label %noloop.exit, label %loop.ph
+
+loop.ph: ; preds = %entry
+ %sext7 = sext i32 %arg to i64
+ %gep8 = getelementptr i8, ptr %arg1, i64 8
+ br label %loop.body
+
+loop.body: ; preds = %loop.body, %loop.ph
+ %phi = phi i64 [ 0, %loop.ph ], [ %add, %loop.body ]
+ %mul = mul i64 %phi, %sext7
+ %gep10 = getelementptr double, ptr %gep8, i64 %mul
+ %load11 = load double, ptr %gep10, align 8
+ store double %load11, ptr %arg1, align 8
+ %add = add i64 %phi, 1
+ %icmp = icmp eq i64 %phi, 0
+ br i1 %icmp, label %loop.exit, label %loop.body
+
+noloop.exit: ; preds = %entry
+ %sext = sext i32 %arg to i64
+ %gep = getelementptr double, ptr %arg1, i64 %sext
+ %load5 = load double, ptr %gep, align 8
+ ret double %load5
+
+loop.exit: ; preds = %loop.body
+ %sext2 = sext i32 %arg to i64
+ %gep2 = getelementptr double, ptr %arg1, i64 %sext2
+ %load6 = load double, ptr %gep2, align 8
+ ret double %load6
+}
+
; A loop with two symbolic strides.
define void @two_strides(ptr noalias %A, ptr noalias %B, i64 %N, i64 %stride.1, i64 %stride.2) {
; CHECK-LABEL: 'two_strides'
diff --git a/llvm/test/Transforms/LoopLoadElim/pr96656.ll b/llvm/test/Transforms/LoopLoadElim/pr96656.ll
new file mode 100644
index 0000000000000..c6b336c0db8d3
--- /dev/null
+++ b/llvm/test/Transforms/LoopLoadElim/pr96656.ll
@@ -0,0 +1,87 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -passes=loop-load-elim -S %s | FileCheck %s
+
+define void @single_iteration_versioning(ptr %arg, ptr %arg1, i1 %arg2) {
+; CHECK-LABEL: define void @single_iteration_versioning(
+; CHECK-SAME: ptr [[ARG:%.*]], ptr [[ARG1:%.*]], i1 [[ARG2:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[LOAD:%.*]] = load i32, ptr [[ARG]], align 4
+; CHECK-NEXT: br i1 [[ARG2]], label %[[NOLOOP_EXIT:.*]], label %[[LOOP_LVER_CHECK:.*]]
+; CHECK: [[LOOP_LVER_CHECK]]:
+; CHECK-NEXT: [[SEXT7:%.*]] = sext i32 [[LOAD]] to i64
+; CHECK-NEXT: [[GEP8:%.*]] = getelementptr i8, ptr [[ARG1]], i64 8
+; CHECK-NEXT: [[GEP9:%.*]] = getelementptr i8, ptr [[ARG1]], i64 16
+; CHECK-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i32 [[LOAD]], 1
+; CHECK-NEXT: br i1 [[IDENT_CHECK]], label %[[LOOP_PH_LVER_ORIG:.*]], label %[[LOOP_PH:.*]]
+; CHECK: [[LOOP_PH_LVER_ORIG]]:
+; CHECK-NEXT: br label %[[LOOP_LVER_ORIG:.*]]
+; CHECK: [[LOOP_LVER_ORIG]]:
+; CHECK-NEXT: [[PHI:%.*]] = phi i64 [ 0, %[[LOOP_PH_LVER_ORIG]] ], [ [[ADD:%.*]], %[[LOOP_LVER_ORIG]] ]
+; CHECK-NEXT: [[MUL:%.*]] = mul i64 [[PHI]], [[SEXT7]]
+; CHECK-NEXT: [[GEP10:%.*]] = getelementptr double, ptr [[GEP8]], i64 [[MUL]]
+; CHECK-NEXT: [[LOAD11:%.*]] = load double, ptr [[GEP10]], align 8
+; CHECK-NEXT: [[GEP13_LVER_ORIG:%.*]] = getelementptr double, ptr [[GEP9]], i64 [[MUL]]
+; CHECK-NEXT: store double [[LOAD11]], ptr [[GEP13_LVER_ORIG]], align 8
+; CHECK-NEXT: [[ADD]] = add i64 [[PHI]], 1
+; CHECK-NEXT: [[ICMP_LVER_ORIG:%.*]] = icmp eq i64 [[PHI]], 1
+; CHECK-NEXT: br i1 [[ICMP_LVER_ORIG]], label %[[EXIT_LOOPEXIT_LOOPEXIT:.*]], label %[[LOOP_LVER_ORIG]]
+; CHECK: [[LOOP_PH]]:
+; CHECK-NEXT: [[LOAD_INITIAL:%.*]] = load double, ptr [[GEP8]], align 8
+; CHECK-NEXT: br label %[[LOOP:.*]]
+; CHECK: [[LOOP]]:
+; CHECK-NEXT: [[STORE_FORWARDED:%.*]] = phi double [ [[LOAD_INITIAL]], %[[LOOP_PH]] ], [ [[STORE_FORWARDED]], %[[LOOP]] ]
+; CHECK-NEXT: [[PHI1:%.*]] = phi i64 [ 0, %[[LOOP_PH]] ], [ [[ADD1:%.*]], %[[LOOP]] ]
+; CHECK-NEXT: [[MUL1:%.*]] = mul i64 [[PHI1]], [[SEXT7]]
+; CHECK-NEXT: [[GEP11:%.*]] = getelementptr double, ptr [[GEP8]], i64 [[MUL1]]
+; CHECK-NEXT: [[LOAD12:%.*]] = load double, ptr [[GEP11]], align 8
+; CHECK-NEXT: [[GEP13:%.*]] = getelementptr double, ptr [[GEP9]], i64 [[MUL1]]
+; CHECK-NEXT: store double [[STORE_FORWARDED]], ptr [[GEP13]], align 8
+; CHECK-NEXT: [[ADD1]] = add i64 [[PHI1]], 1
+; CHECK-NEXT: [[ICMP:%.*]] = icmp eq i64 [[PHI1]], 1
+; CHECK-NEXT: br i1 [[ICMP]], label %[[EXIT_LOOPEXIT_LOOPEXIT1:.*]], label %[[LOOP]]
+; CHECK: [[NOLOOP_EXIT]]:
+; CHECK-NEXT: [[SEXT2:%.*]] = sext i32 [[LOAD]] to i64
+; CHECK-NEXT: [[GEP2:%.*]] = getelementptr double, ptr [[ARG1]], i64 [[SEXT2]]
+; CHECK-NEXT: [[LOAD6:%.*]] = load double, ptr [[GEP2]], align 8
+; CHECK-NEXT: store double [[LOAD6]], ptr [[ARG]], align 8
+; CHECK-NEXT: br label %[[EXIT:.*]]
+; CHECK: [[EXIT_LOOPEXIT_LOOPEXIT]]:
+; CHECK-NEXT: br label %[[EXIT_LOOPEXIT:.*]]
+; CHECK: [[EXIT_LOOPEXIT_LOOPEXIT1]]:
+; CHECK-NEXT: br label %[[EXIT_LOOPEXIT]]
+; CHECK: [[EXIT_LOOPEXIT]]:
+; CHECK-NEXT: br label %[[EXIT]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ %load = load i32, ptr %arg, align 4
+ br i1 %arg2, label %noloop.exit, label %loop.ph
+
+loop.ph: ; preds = %entry
+ %sext7 = sext i32 %load to i64
+ %gep8 = getelementptr i8, ptr %arg1, i64 8
+ %gep9 = getelementptr i8, ptr %arg1, i64 16
+ br label %loop
+
+loop: ; preds = %loop, %loop.ph
+ %phi = phi i64 [ 0, %loop.ph ], [ %add, %loop ]
+ %mul = mul i64 %phi, %sext7
+ %gep11 = getelementptr double, ptr %gep8, i64 %mul
+ %load12 = load double, ptr %gep11, align 8
+ %gep13 = getelementptr double, ptr %gep9, i64 %mul
+ store double %load12, ptr %gep13, align 8
+ %add = add i64 %phi, 1
+ %icmp = icmp eq i64 %phi, 1
+ br i1 %icmp, label %exit, label %loop
+
+noloop.exit: ; preds = %loop.ph
+ %sext = sext i32 %load to i64
+ %gep = getelementptr double, ptr %arg1, i64 %sext
+ %load5 = load double, ptr %gep, align 8
+ store double %load5, ptr %arg, align 8
+ br label %exit
+
+exit: ; preds = %loop.body
+ ret void
+}
diff --git a/llvm/test/Transforms/LoopVersioning/pr96656.ll b/llvm/test/Transforms/LoopVersioning/single-iteration.ll
similarity index 75%
rename from llvm/test/Transforms/LoopVersioning/pr96656.ll
rename to llvm/test/Transforms/LoopVersioning/single-iteration.ll
index 0264fe40a9430..7be34f0678585 100644
--- a/llvm/test/Transforms/LoopVersioning/pr96656.ll
+++ b/llvm/test/Transforms/LoopVersioning/single-iteration.ll
@@ -1,16 +1,17 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
; RUN: opt -passes=loop-versioning -S %s | FileCheck %s
-define void @lver.check.unnecessary(ptr %arg, ptr %arg1, i1 %arg2) {
-; CHECK-LABEL: define void @lver.check.unnecessary(
-; CHECK-SAME: ptr [[ARG:%.*]], ptr [[ARG1:%.*]], i1 [[ARG2:%.*]]) {
+; Callers should not call LoopVersioning on single-iteration loops, but LoopVersioning faithfully versions the loop when the stride is unknown and there is just a single iteration.
+
+define double @single_iteration_unknown_stride(i32 %arg, ptr %arg1, i1 %arg2) {
+; CHECK-LABEL: define double @single_iteration_unknown_stride(
+; CHECK-SAME: i32 [[ARG:%.*]], ptr [[ARG1:%.*]], i1 [[ARG2:%.*]]) {
; CHECK-NEXT: [[ENTRY:.*:]]
-; CHECK-NEXT: [[LOAD:%.*]] = load i32, ptr [[ARG]], align 4
; CHECK-NEXT: br i1 [[ARG2]], label %[[NOLOOP_EXIT:.*]], label %[[LOOP_BODY_LVER_CHECK:.*]]
; CHECK: [[LOOP_BODY_LVER_CHECK]]:
-; CHECK-NEXT: [[SEXT7:%.*]] = sext i32 [[LOAD]] to i64
+; CHECK-NEXT: [[SEXT7:%.*]] = sext i32 [[ARG]] to i64
; CHECK-NEXT: [[GEP8:%.*]] = getelementptr i8, ptr [[ARG1]], i64 8
-; CHECK-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i32 [[LOAD]], 1
+; CHECK-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i32 [[ARG]], 1
; CHECK-NEXT: br i1 [[IDENT_CHECK]], label %[[LOOP_BODY_PH_LVER_ORIG:.*]], label %[[LOOP_BODY_PH:.*]]
; CHECK: [[LOOP_BODY_PH_LVER_ORIG]]:
; CHECK-NEXT: br label %[[LOOP_BODY_LVER_ORIG:.*]]
@@ -35,24 +36,25 @@ define void @lver.check.unnecessary(ptr %arg, ptr %arg1, i1 %arg2) {
; CHECK-NEXT: [[ICMP:%.*]] = icmp eq i64 [[PHI]], 0
; CHECK-NEXT: br i1 [[ICMP]], label %[[LOOP_EXIT_LOOPEXIT1:.*]], label %[[LOOP_BODY]]
; CHECK: [[NOLOOP_EXIT]]:
-; CHECK-NEXT: [[SEXT:%.*]] = sext i32 [[LOAD]] to i64
+; CHECK-NEXT: [[SEXT:%.*]] = sext i32 [[ARG]] to i64
; CHECK-NEXT: [[GEP:%.*]] = getelementptr double, ptr [[ARG1]], i64 [[SEXT]]
; CHECK-NEXT: [[LOAD5:%.*]] = load double, ptr [[GEP]], align 8
-; CHECK-NEXT: store double [[LOAD5]], ptr [[ARG]], align 8
-; CHECK-NEXT: ret void
+; CHECK-NEXT: ret double [[LOAD5]]
; CHECK: [[LOOP_EXIT_LOOPEXIT]]:
; CHECK-NEXT: br label %[[LOOP_EXIT:.*]]
; CHECK: [[LOOP_EXIT_LOOPEXIT1]]:
; CHECK-NEXT: br label %[[LOOP_EXIT]]
; CHECK: [[LOOP_EXIT]]:
-; CHECK-NEXT: ret void
+; CHECK-NEXT: [[SEXT2:%.*]] = sext i32 [[ARG]] to i64
+; CHECK-NEXT: [[GEP2:%.*]] = getelementptr double, ptr [[ARG1]], i64 [[SEXT2]]
+; CHECK-NEXT: [[LOAD6:%.*]] = load double, ptr [[GEP2]], align 8
+; CHECK-NEXT: ret double [[LOAD6]]
;
entry:
- %load = load i32, ptr %arg, align 4
br i1 %arg2, label %noloop.exit, label %loop.ph
loop.ph: ; preds = %entry
- %sext7 = sext i32 %load to i64
+ %sext7 = sext i32 %arg to i64
%gep8 = getelementptr i8, ptr %arg1, i64 8
br label %loop.body
@@ -67,12 +69,14 @@ loop.body: ; preds = %loop.body, %loop.ph
br i1 %icmp, label %loop.exit, label %loop.body
noloop.exit: ; preds = %entry
- %sext = sext i32 %load to i64
+ %sext = sext i32 %arg to i64
%gep = getelementptr double, ptr %arg1, i64 %sext
%load5 = load double, ptr %gep, align 8
- store double %load5, ptr %arg, align 8
- ret void
+ ret double %load5
loop.exit: ; preds = %loop.body
- ret void
+ %sext2 = sext i32 %arg to i64
+ %gep2 = getelementptr double, ptr %arg1, i64 %sext2
+ %load6 = load double, ptr %gep2, align 8
+ ret double %load6
}
More information about the llvm-commits
mailing list