[clang] [OpenMP] Restore loop variable values after loop-tranformation constructs. (PR #208533)
Zahira Ammarguellat via cfe-commits
cfe-commits at lists.llvm.org
Thu Jul 23 10:54:51 PDT 2026
https://github.com/zahiraam updated https://github.com/llvm/llvm-project/pull/208533
>From 444217556666673b9efe64e2edb40ff08d777f16 Mon Sep 17 00:00:00 2001
From: Zahira Ammarguellat <zahira.ammarguellat at intel.com>
Date: Thu, 9 Jul 2026 11:10:26 -0700
Subject: [PATCH 1/5] [OpenMP] Restore loop variable values after
loop-transformation constructs
---
clang/lib/Sema/SemaOpenMP.cpp | 44 ++++++-
clang/test/OpenMP/loop_transform_final_iv.c | 129 ++++++++++++++++++++
2 files changed, 172 insertions(+), 1 deletion(-)
create mode 100644 clang/test/OpenMP/loop_transform_final_iv.c
diff --git a/clang/lib/Sema/SemaOpenMP.cpp b/clang/lib/Sema/SemaOpenMP.cpp
index 0f45a21c5e461..b2605e41bb513 100644
--- a/clang/lib/Sema/SemaOpenMP.cpp
+++ b/clang/lib/Sema/SemaOpenMP.cpp
@@ -14960,6 +14960,31 @@ static Expr *makeFloorIVRef(Sema &SemaRef, ArrayRef<VarDecl *> FloorIndVars,
OrigCntVar->getExprLoc());
}
+/// Add loop variable finalization statements to restore original loop variable
+/// values after a loop-transformation construct, as required by OpenMP 6.0 spec
+/// (pg 371, lines 19-21).
+static Stmt *addLoopVarFinalization(
+ ASTContext &Context, Stmt *TransformedStmt,
+ ArrayRef<OMPLoopBasedDirective::HelperExprs> LoopHelpers) {
+ SmallVector<Stmt *, 8> FinalizationStmts;
+ FinalizationStmts.push_back(TransformedStmt);
+
+ for (const auto &LoopHelper : LoopHelpers) {
+ if (!LoopHelper.Finals.empty()) {
+ if (Expr *Final = LoopHelper.Finals[0]) {
+ FinalizationStmts.push_back(Final);
+ }
+ }
+ }
+
+ if (FinalizationStmts.size() > 1) {
+ return CompoundStmt::Create(Context, FinalizationStmts, FPOptionsOverride(),
+ TransformedStmt->getBeginLoc(),
+ TransformedStmt->getEndLoc());
+ }
+ return TransformedStmt;
+}
+
StmtResult SemaOpenMP::ActOnOpenMPTileDirective(ArrayRef<OMPClause *> Clauses,
Stmt *AStmt,
SourceLocation StartLoc,
@@ -15230,6 +15255,8 @@ StmtResult SemaOpenMP::ActOnOpenMPTileDirective(ArrayRef<OMPClause *> Clauses,
LoopHelper.Init->getBeginLoc(), LoopHelper.Inc->getEndLoc());
}
+ Inner = addLoopVarFinalization(Context, Inner, LoopHelpers);
+
return OMPTileDirective::Create(Context, StartLoc, EndLoc, Clauses, NumLoops,
AStmt, Inner,
buildPreInits(Context, PreInits));
@@ -15489,6 +15516,8 @@ StmtResult SemaOpenMP::ActOnOpenMPStripeDirective(ArrayRef<OMPClause *> Clauses,
LoopHelper.Init->getBeginLoc(), LoopHelper.Inc->getEndLoc());
}
+ Inner = addLoopVarFinalization(Context, Inner, LoopHelpers);
+
return OMPStripeDirective::Create(Context, StartLoc, EndLoc, Clauses,
NumLoops, AStmt, Inner,
buildPreInits(Context, PreInits));
@@ -15957,8 +15986,11 @@ StmtResult SemaOpenMP::ActOnOpenMPReverseDirective(Stmt *AStmt,
ForStmt(Context, Init.get(), Cond.get(), nullptr, Incr.get(),
ReversedBody, LoopHelper.Init->getBeginLoc(),
LoopHelper.Init->getBeginLoc(), LoopHelper.Inc->getEndLoc());
+
+ Stmt *Inner = addLoopVarFinalization(Context, ReversedFor, LoopHelpers);
+
return OMPReverseDirective::Create(Context, StartLoc, EndLoc, AStmt, NumLoops,
- ReversedFor,
+ Inner,
buildPreInits(Context, PreInits));
}
@@ -16381,6 +16413,8 @@ StmtResult SemaOpenMP::ActOnOpenMPInterchangeDirective(
SourceHelper.Inc->getEndLoc());
}
+ Inner = addLoopVarFinalization(Context, Inner, LoopHelpers);
+
return OMPInterchangeDirective::Create(Context, StartLoc, EndLoc, Clauses,
NumLoops, AStmt, Inner,
buildPreInits(Context, PreInits));
@@ -16859,6 +16893,14 @@ StmtResult SemaOpenMP::ActOnOpenMPFuseDirective(ArrayRef<OMPClause *> Clauses,
FusionStmt = CompoundStmt::Create(Context, FinalLoops, FPOptionsOverride(),
SourceLocation(), SourceLocation());
}
+
+ // Collect HelperExprs from all loops for finalization
+ SmallVector<OMPLoopBasedDirective::HelperExprs, 2> LoopHelpers;
+ for (const auto &Loop : SeqAnalysis.Loops) {
+ LoopHelpers.push_back(Loop.HelperExprs);
+ }
+ FusionStmt = addLoopVarFinalization(Context, FusionStmt, LoopHelpers);
+
return OMPFuseDirective::Create(Context, StartLoc, EndLoc, Clauses,
NumGeneratedTopLevelLoops, AStmt, FusionStmt,
buildPreInits(Context, PreInits));
diff --git a/clang/test/OpenMP/loop_transform_final_iv.c b/clang/test/OpenMP/loop_transform_final_iv.c
new file mode 100644
index 0000000000000..969badbe9636b
--- /dev/null
+++ b/clang/test/OpenMP/loop_transform_final_iv.c
@@ -0,0 +1,129 @@
+// RUN: %clang_cc1 -verify -fopenmp -fopenmp-version=60 -emit-llvm %s -o - | FileCheck %s
+// expected-no-diagnostics
+
+// Test that loop variables are correctly finalized after loop-transformation
+// constructs as required by OpenMP 6.0 spec (pg 371, lines 19-21).
+
+void test_tile(void) {
+ // CHECK-LABEL: define {{.*}} @test_tile
+ int i;
+ #pragma omp tile sizes(2)
+ for (i = 1; i <= 10; i++) {
+ }
+ // CHECK: store i32 11, ptr %i
+ // After loop: i should be 11 (lb + tripcount * stride = 1 + 10 * 1)
+}
+
+void test_stripe(void) {
+ // CHECK-LABEL: define {{.*}} @test_stripe
+ int i;
+ #pragma omp stripe sizes(3)
+ for (i = 0; i < 10; i++) {
+ }
+ // CHECK: store i32 10, ptr %i
+ // After loop: i should be 10 (lb + tripcount * stride = 0 + 10 * 1)
+}
+
+void test_tile_nested(void) {
+ // CHECK-LABEL: define {{.*}} @test_tile_nested
+ int i, j;
+ #pragma omp tile sizes(2, 3)
+ for (i = 0; i < 10; i++)
+ for (j = 0; j < 5; j++) {
+ }
+ // CHECK: store i32 10, ptr %i
+ // CHECK: store i32 5, ptr %j
+ // After loop: i should be 10, j should be 5
+}
+
+void test_tile_stride(void) {
+ // CHECK-LABEL: define {{.*}} @test_tile_stride
+ int i;
+ #pragma omp tile sizes(4)
+ for (i = 5; i <= 15; i += 2) {
+ }
+ // CHECK: store i32 17, ptr %i
+ // After loop: i should be 17 (lb + tripcount * stride = 5 + 6 * 2)
+}
+
+void test_tile_variable_bound(int n) {
+ // CHECK-LABEL: define {{.*}} @test_tile_variable_bound
+ int i;
+ #pragma omp tile sizes(3)
+ for (i = 0; i < n; i++) {
+ }
+ // CHECK: for.end14:
+ // CHECK-NEXT: %{{.*}} = load i32, ptr %.capture_expr.
+ // CHECK-NEXT: %{{.*}} = sub nsw i32 %{{.*}}, 0
+ // CHECK-NEXT: %{{.*}} = sdiv i32 %{{.*}}, 1
+ // CHECK-NEXT: %{{.*}} = mul nsw i32 %{{.*}}, 1
+ // CHECK-NEXT: %{{.*}} = add nsw i32 0, %{{.*}}
+ // CHECK-NEXT: store i32 %{{.*}}, ptr %i
+ // After loop: i should equal n (the runtime upper bound)
+}
+
+void test_reverse(void) {
+ // CHECK-LABEL: define {{.*}} @test_reverse
+ int i;
+ #pragma omp reverse
+ for (i = 0; i < 10; i++) {
+ }
+ // CHECK: store i32 10, ptr %i
+ // After loop: i should be 10 (lb + tripcount * stride = 0 + 10 * 1)
+}
+
+void test_interchange(void) {
+ // CHECK-LABEL: define {{.*}} @test_interchange
+ int i, j;
+ #pragma omp interchange permutation(2, 1)
+ for (i = 0; i < 10; i++)
+ for (j = 0; j < 5; j++) {
+ }
+ // CHECK: store i32 10, ptr %i
+ // CHECK: store i32 5, ptr %j
+ // After loop: i should be 10, j should be 5
+}
+
+void test_fuse(void) {
+ // CHECK-LABEL: define {{.*}} @test_fuse
+ int i, j;
+ #pragma omp fuse
+ {
+ for (i = 0; i < 10; i++) {
+ }
+ for (j = 0; j < 5; j++) {
+ }
+ }
+ // CHECK: store i32 10, ptr %i
+ // CHECK: store i32 5, ptr %j
+ // After loop: i should be 10, j should be 5
+}
+
+void test_for_workshare(void) {
+ // CHECK-LABEL: define {{.*}} @test_for_workshare
+ int i;
+ #pragma omp for
+ for (i = 5; i <= 10; i += 2) {
+ }
+ // CHECK: omp.loop.exit:
+ // CHECK-NEXT: call void @__kmpc_for_static_fini
+ // CHECK-NEXT: call void @__kmpc_barrier
+ // CHECK-NEXT: ret void
+ // omp for worksharing doesn't finalize the loop variable after the loop
+}
+
+void test_for_lastprivate(void) {
+ // CHECK-LABEL: define {{.*}} @test_for_lastprivate
+ int i;
+ #pragma omp parallel
+ #pragma omp for lastprivate(i)
+ for (i = 5; i <= 10; i += 2) {
+ }
+ // CHECK: .omp.lastprivate.then:
+ // CHECK-NEXT: store i32 11, ptr %i{{.*}}, align 4
+ // CHECK-NEXT: %{{.*}} = load i32, ptr %i{{.*}}, align 4
+ // CHECK-NEXT: store i32 %{{.*}}, ptr %{{.*}}, align 4
+ // omp for with lastprivate DOES finalize the loop variable
+ // i = last_iteration_value + stride = 9 + 2 = 11
+}
+
>From 9d704356a18c01cfbf4c236e3dd96f6a2bb9e575 Mon Sep 17 00:00:00 2001
From: Zahira Ammarguellat <zahira.ammarguellat at intel.com>
Date: Thu, 9 Jul 2026 11:53:07 -0700
Subject: [PATCH 2/5] Fix format
---
clang/lib/Sema/SemaOpenMP.cpp | 3 +--
1 file changed, 1 insertion(+), 2 deletions(-)
diff --git a/clang/lib/Sema/SemaOpenMP.cpp b/clang/lib/Sema/SemaOpenMP.cpp
index b2605e41bb513..e6cac635f9864 100644
--- a/clang/lib/Sema/SemaOpenMP.cpp
+++ b/clang/lib/Sema/SemaOpenMP.cpp
@@ -15990,8 +15990,7 @@ StmtResult SemaOpenMP::ActOnOpenMPReverseDirective(Stmt *AStmt,
Stmt *Inner = addLoopVarFinalization(Context, ReversedFor, LoopHelpers);
return OMPReverseDirective::Create(Context, StartLoc, EndLoc, AStmt, NumLoops,
- Inner,
- buildPreInits(Context, PreInits));
+ Inner, buildPreInits(Context, PreInits));
}
/// Build the AST for \#pragma omp split counts(c1, c2, ...).
>From d61e20b8c2ba2c05134924597b7b367a44910dc1 Mon Sep 17 00:00:00 2001
From: Zahira Ammarguellat <zahira.ammarguellat at intel.com>
Date: Wed, 22 Jul 2026 08:56:09 -0700
Subject: [PATCH 3/5] Fixed LIT tests
---
clang/include/clang/AST/StmtOpenMP.h | 59 +++-
clang/lib/AST/StmtOpenMP.cpp | 35 +-
clang/lib/CodeGen/CGStmtOpenMP.cpp | 25 ++
clang/lib/Sema/SemaOpenMP.cpp | 78 ++---
clang/test/OpenMP/interchange_codegen.cpp | 179 ++++++++---
clang/test/OpenMP/loop_transform_final_iv.c | 15 -
clang/test/OpenMP/reverse_codegen.cpp | 265 +++++++++++-----
clang/test/OpenMP/stripe_codegen.cpp | 334 +++++++++++++-------
clang/test/OpenMP/tile_codegen.cpp | 334 +++++++++++++-------
9 files changed, 894 insertions(+), 430 deletions(-)
diff --git a/clang/include/clang/AST/StmtOpenMP.h b/clang/include/clang/AST/StmtOpenMP.h
index dbc76e7df8ecd..5c374320bf018 100644
--- a/clang/include/clang/AST/StmtOpenMP.h
+++ b/clang/include/clang/AST/StmtOpenMP.h
@@ -5601,6 +5601,7 @@ class OMPTileDirective final
enum {
PreInitsOffset = 0,
TransformedStmtOffset,
+ FinalsOffset,
};
explicit OMPTileDirective(SourceLocation StartLoc, SourceLocation EndLoc,
@@ -5617,6 +5618,10 @@ class OMPTileDirective final
Data->getChildren()[TransformedStmtOffset] = S;
}
+ void setFinals(Stmt *Finals) {
+ Data->getChildren()[FinalsOffset] = Finals;
+ }
+
public:
/// Create a new AST node representation for '#pragma omp tile'.
///
@@ -5630,11 +5635,13 @@ class OMPTileDirective final
/// \param TransformedStmt The loop nest after tiling, or nullptr in
/// dependent contexts.
/// \param PreInits Helper preinits statements for the loop nest.
+ /// \param Finals Loop variable finalization statements.
static OMPTileDirective *Create(const ASTContext &C, SourceLocation StartLoc,
SourceLocation EndLoc,
ArrayRef<OMPClause *> Clauses,
unsigned NumLoops, Stmt *AssociatedStmt,
- Stmt *TransformedStmt, Stmt *PreInits);
+ Stmt *TransformedStmt, Stmt *PreInits,
+ Stmt *Finals);
/// Build an empty '#pragma omp tile' AST node for deserialization.
///
@@ -5662,6 +5669,9 @@ class OMPTileDirective final
/// Return preinits statement.
Stmt *getPreInits() const { return Data->getChildren()[PreInitsOffset]; }
+ /// Return finals statement (loop variable finalization).
+ Stmt *getFinals() const { return Data->getChildren()[FinalsOffset]; }
+
static bool classof(const Stmt *T) {
return T->getStmtClass() == OMPTileDirectiveClass;
}
@@ -5677,6 +5687,7 @@ class OMPStripeDirective final
enum {
PreInitsOffset = 0,
TransformedStmtOffset,
+ FinalsOffset,
};
explicit OMPStripeDirective(SourceLocation StartLoc, SourceLocation EndLoc,
@@ -5693,6 +5704,10 @@ class OMPStripeDirective final
Data->getChildren()[TransformedStmtOffset] = S;
}
+ void setFinals(Stmt *Finals) {
+ Data->getChildren()[FinalsOffset] = Finals;
+ }
+
public:
/// Create a new AST node representation for '#pragma omp stripe'.
///
@@ -5706,10 +5721,11 @@ class OMPStripeDirective final
/// \param TransformedStmt The loop nest after striping, or nullptr in
/// dependent contexts.
/// \param PreInits Helper preinits statements for the loop nest.
+ /// \param Finals Loop variable finalization statements.
static OMPStripeDirective *
Create(const ASTContext &C, SourceLocation StartLoc, SourceLocation EndLoc,
ArrayRef<OMPClause *> Clauses, unsigned NumLoops, Stmt *AssociatedStmt,
- Stmt *TransformedStmt, Stmt *PreInits);
+ Stmt *TransformedStmt, Stmt *PreInits, Stmt *Finals);
/// Build an empty '#pragma omp stripe' AST node for deserialization.
///
@@ -5736,6 +5752,9 @@ class OMPStripeDirective final
/// Return preinits statement.
Stmt *getPreInits() const { return Data->getChildren()[PreInitsOffset]; }
+ /// Return finals statement (loop variable finalization).
+ Stmt *getFinals() const { return Data->getChildren()[FinalsOffset]; }
+
static bool classof(const Stmt *T) {
return T->getStmtClass() == OMPStripeDirectiveClass;
}
@@ -5756,6 +5775,7 @@ class OMPUnrollDirective final
enum {
PreInitsOffset = 0,
TransformedStmtOffset,
+ FinalsOffset,
};
explicit OMPUnrollDirective(SourceLocation StartLoc, SourceLocation EndLoc)
@@ -5832,6 +5852,7 @@ class OMPReverseDirective final
enum {
PreInitsOffset = 0,
TransformedStmtOffset,
+ FinalsOffset,
};
explicit OMPReverseDirective(SourceLocation StartLoc, SourceLocation EndLoc,
@@ -5848,6 +5869,10 @@ class OMPReverseDirective final
Data->getChildren()[TransformedStmtOffset] = S;
}
+ void setFinals(Stmt *Finals) {
+ Data->getChildren()[FinalsOffset] = Finals;
+ }
+
public:
/// Create a new AST node representation for '#pragma omp reverse'.
///
@@ -5859,11 +5884,13 @@ class OMPReverseDirective final
/// \param TransformedStmt The loop nest after tiling, or nullptr in
/// dependent contexts.
/// \param PreInits Helper preinits statements for the loop nest.
+ /// \param Finals Loop variable finalization statements.
static OMPReverseDirective *Create(const ASTContext &C,
SourceLocation StartLoc,
SourceLocation EndLoc,
Stmt *AssociatedStmt, unsigned NumLoops,
- Stmt *TransformedStmt, Stmt *PreInits);
+ Stmt *TransformedStmt, Stmt *PreInits,
+ Stmt *Finals);
/// Build an empty '#pragma omp reverse' AST node for deserialization.
///
@@ -5881,6 +5908,9 @@ class OMPReverseDirective final
/// Return preinits statement.
Stmt *getPreInits() const { return Data->getChildren()[PreInitsOffset]; }
+ /// Return finals statement.
+ Stmt *getFinals() const { return Data->getChildren()[FinalsOffset]; }
+
static bool classof(const Stmt *T) {
return T->getStmtClass() == OMPReverseDirectiveClass;
}
@@ -5903,6 +5933,7 @@ class OMPInterchangeDirective final
enum {
PreInitsOffset = 0,
TransformedStmtOffset,
+ FinalsOffset,
};
explicit OMPInterchangeDirective(SourceLocation StartLoc,
@@ -5919,6 +5950,10 @@ class OMPInterchangeDirective final
Data->getChildren()[TransformedStmtOffset] = S;
}
+ void setFinals(Stmt *Finals) {
+ Data->getChildren()[FinalsOffset] = Finals;
+ }
+
public:
/// Create a new AST node representation for '#pragma omp interchange'.
///
@@ -5932,10 +5967,11 @@ class OMPInterchangeDirective final
/// \param TransformedStmt The loop nest after tiling, or nullptr in
/// dependent contexts.
/// \param PreInits Helper preinits statements for the loop nest.
+ /// \param Finals Loop variable finalization statements.
static OMPInterchangeDirective *
Create(const ASTContext &C, SourceLocation StartLoc, SourceLocation EndLoc,
ArrayRef<OMPClause *> Clauses, unsigned NumLoops, Stmt *AssociatedStmt,
- Stmt *TransformedStmt, Stmt *PreInits);
+ Stmt *TransformedStmt, Stmt *PreInits, Stmt *Finals);
/// Build an empty '#pragma omp interchange' AST node for deserialization.
///
@@ -5954,6 +5990,9 @@ class OMPInterchangeDirective final
/// Return preinits statement.
Stmt *getPreInits() const { return Data->getChildren()[PreInitsOffset]; }
+ /// Return finals statement.
+ Stmt *getFinals() const { return Data->getChildren()[FinalsOffset]; }
+
static bool classof(const Stmt *T) {
return T->getStmtClass() == OMPInterchangeDirectiveClass;
}
@@ -6009,6 +6048,7 @@ class OMPFuseDirective final
enum {
PreInitsOffset = 0,
TransformedStmtOffset,
+ FinalsOffset,
};
explicit OMPFuseDirective(SourceLocation StartLoc, SourceLocation EndLoc)
@@ -6023,6 +6063,10 @@ class OMPFuseDirective final
Data->getChildren()[TransformedStmtOffset] = S;
}
+ void setFinals(Stmt *Finals) {
+ Data->getChildren()[FinalsOffset] = Finals;
+ }
+
public:
/// Create a new AST node representation for #pragma omp fuse'
///
@@ -6038,10 +6082,12 @@ class OMPFuseDirective final
/// \param TransformedStmt The loop nest after fusion, or nullptr in
/// dependent
/// \param PreInits Helper preinits statements for the loop nest
+ /// \param Finals Loop variable finalization statements
static OMPFuseDirective *
Create(const ASTContext &C, SourceLocation StartLoc, SourceLocation EndLoc,
ArrayRef<OMPClause *> Clauses, unsigned NumGeneratedTopLevelLoops,
- Stmt *AssociatedStmt, Stmt *TransformedStmt, Stmt *PreInits);
+ Stmt *AssociatedStmt, Stmt *TransformedStmt, Stmt *PreInits,
+ Stmt *Finals);
/// Build an empty '#pragma omp fuse' AST node for deserialization
///
@@ -6060,6 +6106,9 @@ class OMPFuseDirective final
/// Return preinits statement.
Stmt *getPreInits() const { return Data->getChildren()[PreInitsOffset]; }
+ /// Return finals statement.
+ Stmt *getFinals() const { return Data->getChildren()[FinalsOffset]; }
+
static bool classof(const Stmt *T) {
return T->getStmtClass() == OMPFuseDirectiveClass;
}
diff --git a/clang/lib/AST/StmtOpenMP.cpp b/clang/lib/AST/StmtOpenMP.cpp
index 9d6b315effb41..cace1c992eedf 100644
--- a/clang/lib/AST/StmtOpenMP.cpp
+++ b/clang/lib/AST/StmtOpenMP.cpp
@@ -452,12 +452,13 @@ OMPTileDirective *
OMPTileDirective::Create(const ASTContext &C, SourceLocation StartLoc,
SourceLocation EndLoc, ArrayRef<OMPClause *> Clauses,
unsigned NumLoops, Stmt *AssociatedStmt,
- Stmt *TransformedStmt, Stmt *PreInits) {
+ Stmt *TransformedStmt, Stmt *PreInits, Stmt *Finals) {
OMPTileDirective *Dir = createDirective<OMPTileDirective>(
- C, Clauses, AssociatedStmt, TransformedStmtOffset + 1, StartLoc, EndLoc,
+ C, Clauses, AssociatedStmt, FinalsOffset + 1, StartLoc, EndLoc,
NumLoops);
Dir->setTransformedStmt(TransformedStmt);
Dir->setPreInits(PreInits);
+ Dir->setFinals(Finals);
return Dir;
}
@@ -465,7 +466,7 @@ OMPTileDirective *OMPTileDirective::CreateEmpty(const ASTContext &C,
unsigned NumClauses,
unsigned NumLoops) {
return createEmptyDirective<OMPTileDirective>(
- C, NumClauses, /*HasAssociatedStmt=*/true, TransformedStmtOffset + 1,
+ C, NumClauses, /*HasAssociatedStmt=*/true, FinalsOffset + 1,
SourceLocation(), SourceLocation(), NumLoops);
}
@@ -473,12 +474,13 @@ OMPStripeDirective *
OMPStripeDirective::Create(const ASTContext &C, SourceLocation StartLoc,
SourceLocation EndLoc, ArrayRef<OMPClause *> Clauses,
unsigned NumLoops, Stmt *AssociatedStmt,
- Stmt *TransformedStmt, Stmt *PreInits) {
+ Stmt *TransformedStmt, Stmt *PreInits, Stmt *Finals) {
OMPStripeDirective *Dir = createDirective<OMPStripeDirective>(
- C, Clauses, AssociatedStmt, TransformedStmtOffset + 1, StartLoc, EndLoc,
+ C, Clauses, AssociatedStmt, FinalsOffset + 1, StartLoc, EndLoc,
NumLoops);
Dir->setTransformedStmt(TransformedStmt);
Dir->setPreInits(PreInits);
+ Dir->setFinals(Finals);
return Dir;
}
@@ -486,7 +488,7 @@ OMPStripeDirective *OMPStripeDirective::CreateEmpty(const ASTContext &C,
unsigned NumClauses,
unsigned NumLoops) {
return createEmptyDirective<OMPStripeDirective>(
- C, NumClauses, /*HasAssociatedStmt=*/true, TransformedStmtOffset + 1,
+ C, NumClauses, /*HasAssociatedStmt=*/true, FinalsOffset + 1,
SourceLocation(), SourceLocation(), NumLoops);
}
@@ -516,12 +518,13 @@ OMPReverseDirective *
OMPReverseDirective::Create(const ASTContext &C, SourceLocation StartLoc,
SourceLocation EndLoc, Stmt *AssociatedStmt,
unsigned NumLoops, Stmt *TransformedStmt,
- Stmt *PreInits) {
+ Stmt *PreInits, Stmt *Finals) {
OMPReverseDirective *Dir = createDirective<OMPReverseDirective>(
- C, {}, AssociatedStmt, TransformedStmtOffset + 1, StartLoc, EndLoc,
+ C, {}, AssociatedStmt, FinalsOffset + 1, StartLoc, EndLoc,
NumLoops);
Dir->setTransformedStmt(TransformedStmt);
Dir->setPreInits(PreInits);
+ Dir->setFinals(Finals);
return Dir;
}
@@ -529,18 +532,19 @@ OMPReverseDirective *OMPReverseDirective::CreateEmpty(const ASTContext &C,
unsigned NumLoops) {
return createEmptyDirective<OMPReverseDirective>(
C, /*NumClauses=*/0, /*HasAssociatedStmt=*/true,
- TransformedStmtOffset + 1, SourceLocation(), SourceLocation(), NumLoops);
+ FinalsOffset + 1, SourceLocation(), SourceLocation(), NumLoops);
}
OMPInterchangeDirective *OMPInterchangeDirective::Create(
const ASTContext &C, SourceLocation StartLoc, SourceLocation EndLoc,
ArrayRef<OMPClause *> Clauses, unsigned NumLoops, Stmt *AssociatedStmt,
- Stmt *TransformedStmt, Stmt *PreInits) {
+ Stmt *TransformedStmt, Stmt *PreInits, Stmt *Finals) {
OMPInterchangeDirective *Dir = createDirective<OMPInterchangeDirective>(
- C, Clauses, AssociatedStmt, TransformedStmtOffset + 1, StartLoc, EndLoc,
+ C, Clauses, AssociatedStmt, FinalsOffset + 1, StartLoc, EndLoc,
NumLoops);
Dir->setTransformedStmt(TransformedStmt);
Dir->setPreInits(PreInits);
+ Dir->setFinals(Finals);
return Dir;
}
@@ -548,7 +552,7 @@ OMPInterchangeDirective *
OMPInterchangeDirective::CreateEmpty(const ASTContext &C, unsigned NumClauses,
unsigned NumLoops) {
return createEmptyDirective<OMPInterchangeDirective>(
- C, NumClauses, /*HasAssociatedStmt=*/true, TransformedStmtOffset + 1,
+ C, NumClauses, /*HasAssociatedStmt=*/true, FinalsOffset + 1,
SourceLocation(), SourceLocation(), NumLoops);
}
@@ -576,12 +580,13 @@ OMPSplitDirective *OMPSplitDirective::CreateEmpty(const ASTContext &C,
OMPFuseDirective *OMPFuseDirective::Create(
const ASTContext &C, SourceLocation StartLoc, SourceLocation EndLoc,
ArrayRef<OMPClause *> Clauses, unsigned NumGeneratedTopLevelLoops,
- Stmt *AssociatedStmt, Stmt *TransformedStmt, Stmt *PreInits) {
+ Stmt *AssociatedStmt, Stmt *TransformedStmt, Stmt *PreInits, Stmt *Finals) {
OMPFuseDirective *Dir = createDirective<OMPFuseDirective>(
- C, Clauses, AssociatedStmt, TransformedStmtOffset + 1, StartLoc, EndLoc);
+ C, Clauses, AssociatedStmt, FinalsOffset + 1, StartLoc, EndLoc);
Dir->setTransformedStmt(TransformedStmt);
Dir->setPreInits(PreInits);
+ Dir->setFinals(Finals);
Dir->setNumGeneratedTopLevelLoops(NumGeneratedTopLevelLoops);
return Dir;
}
@@ -589,7 +594,7 @@ OMPFuseDirective *OMPFuseDirective::Create(
OMPFuseDirective *OMPFuseDirective::CreateEmpty(const ASTContext &C,
unsigned NumClauses) {
OMPFuseDirective *Dir = createEmptyDirective<OMPFuseDirective>(
- C, NumClauses, /*HasAssociatedStmt=*/true, TransformedStmtOffset + 1,
+ C, NumClauses, /*HasAssociatedStmt=*/true, FinalsOffset + 1,
SourceLocation(), SourceLocation());
return Dir;
}
diff --git a/clang/lib/CodeGen/CGStmtOpenMP.cpp b/clang/lib/CodeGen/CGStmtOpenMP.cpp
index 88f698f38cce0..247cb65512b68 100644
--- a/clang/lib/CodeGen/CGStmtOpenMP.cpp
+++ b/clang/lib/CodeGen/CGStmtOpenMP.cpp
@@ -3221,18 +3221,33 @@ void CodeGenFunction::EmitOMPTileDirective(const OMPTileDirective &S) {
// Emit the de-sugared statement.
OMPTransformDirectiveScopeRAII TileScope(*this, &S);
EmitStmt(S.getTransformedStmt());
+
+ // Emit loop variable finalization as required by OpenMP 6.0 spec to restore
+ // original loop variable values after the loop-transformation construct.
+ if (auto *Finals = S.getFinals())
+ EmitStmt(Finals);
}
void CodeGenFunction::EmitOMPStripeDirective(const OMPStripeDirective &S) {
// Emit the de-sugared statement.
OMPTransformDirectiveScopeRAII StripeScope(*this, &S);
EmitStmt(S.getTransformedStmt());
+
+ // Emit loop variable finalization as required by OpenMP 6.0 spec to restore
+ // original loop variable values after the loop-transformation construct.
+ if (auto *Finals = S.getFinals())
+ EmitStmt(Finals);
}
void CodeGenFunction::EmitOMPReverseDirective(const OMPReverseDirective &S) {
// Emit the de-sugared statement.
OMPTransformDirectiveScopeRAII ReverseScope(*this, &S);
EmitStmt(S.getTransformedStmt());
+
+ // Emit loop variable finalization as required by OpenMP 6.0 spec to restore
+ // original loop variable values after the loop-transformation construct.
+ if (auto *Finals = S.getFinals())
+ EmitStmt(Finals);
}
void CodeGenFunction::EmitOMPSplitDirective(const OMPSplitDirective &S) {
@@ -3246,12 +3261,22 @@ void CodeGenFunction::EmitOMPInterchangeDirective(
// Emit the de-sugared statement.
OMPTransformDirectiveScopeRAII InterchangeScope(*this, &S);
EmitStmt(S.getTransformedStmt());
+
+ // Emit loop variable finalization as required by OpenMP 6.0 spec to restore
+ // original loop variable values after the loop-transformation construct.
+ if (auto *Finals = S.getFinals())
+ EmitStmt(Finals);
}
void CodeGenFunction::EmitOMPFuseDirective(const OMPFuseDirective &S) {
// Emit the de-sugared statement
OMPTransformDirectiveScopeRAII FuseScope(*this, &S);
EmitStmt(S.getTransformedStmt());
+
+ // Emit loop variable finalization as required by OpenMP 6.0 spec to restore
+ // original loop variable values after the loop-transformation construct.
+ if (auto *Finals = S.getFinals())
+ EmitStmt(Finals);
}
void CodeGenFunction::EmitOMPUnrollDirective(const OMPUnrollDirective &S) {
diff --git a/clang/lib/Sema/SemaOpenMP.cpp b/clang/lib/Sema/SemaOpenMP.cpp
index e6cac635f9864..0e5cac6e08b2d 100644
--- a/clang/lib/Sema/SemaOpenMP.cpp
+++ b/clang/lib/Sema/SemaOpenMP.cpp
@@ -14960,29 +14960,22 @@ static Expr *makeFloorIVRef(Sema &SemaRef, ArrayRef<VarDecl *> FloorIndVars,
OrigCntVar->getExprLoc());
}
-/// Add loop variable finalization statements to restore original loop variable
-/// values after a loop-transformation construct, as required by OpenMP 6.0 spec
-/// (pg 371, lines 19-21).
-static Stmt *addLoopVarFinalization(
- ASTContext &Context, Stmt *TransformedStmt,
+/// Build loop variable finalization statement from HelperExprs.Finals.
+/// Returns a CompoundStmt containing all finalization statements, or nullptr
+/// if there are no finalization statements.
+static Stmt *buildLoopFinalization(
+ ASTContext &Context,
ArrayRef<OMPLoopBasedDirective::HelperExprs> LoopHelpers) {
SmallVector<Stmt *, 8> FinalizationStmts;
- FinalizationStmts.push_back(TransformedStmt);
-
- for (const auto &LoopHelper : LoopHelpers) {
- if (!LoopHelper.Finals.empty()) {
- if (Expr *Final = LoopHelper.Finals[0]) {
- FinalizationStmts.push_back(Final);
- }
+ for (const auto &Helper : LoopHelpers) {
+ if (!Helper.Finals.empty() && Helper.Finals[0]) {
+ FinalizationStmts.push_back(Helper.Finals[0]);
}
}
-
- if (FinalizationStmts.size() > 1) {
- return CompoundStmt::Create(Context, FinalizationStmts, FPOptionsOverride(),
- TransformedStmt->getBeginLoc(),
- TransformedStmt->getEndLoc());
- }
- return TransformedStmt;
+ if (FinalizationStmts.empty())
+ return nullptr;
+ return CompoundStmt::Create(Context, FinalizationStmts, FPOptionsOverride(),
+ SourceLocation(), SourceLocation());
}
StmtResult SemaOpenMP::ActOnOpenMPTileDirective(ArrayRef<OMPClause *> Clauses,
@@ -15014,7 +15007,7 @@ StmtResult SemaOpenMP::ActOnOpenMPTileDirective(ArrayRef<OMPClause *> Clauses,
// Delay tiling to when template is completely instantiated.
if (SemaRef.CurContext->isDependentContext())
return OMPTileDirective::Create(Context, StartLoc, EndLoc, Clauses,
- NumLoops, AStmt, nullptr, nullptr);
+ NumLoops, AStmt, nullptr, nullptr, nullptr);
assert(LoopHelpers.size() == NumLoops &&
"Expecting loop iteration space dimensionality to match number of "
@@ -15255,11 +15248,9 @@ StmtResult SemaOpenMP::ActOnOpenMPTileDirective(ArrayRef<OMPClause *> Clauses,
LoopHelper.Init->getBeginLoc(), LoopHelper.Inc->getEndLoc());
}
- Inner = addLoopVarFinalization(Context, Inner, LoopHelpers);
-
return OMPTileDirective::Create(Context, StartLoc, EndLoc, Clauses, NumLoops,
- AStmt, Inner,
- buildPreInits(Context, PreInits));
+ AStmt, Inner, buildPreInits(Context, PreInits),
+ buildLoopFinalization(Context, LoopHelpers));
}
StmtResult SemaOpenMP::ActOnOpenMPStripeDirective(ArrayRef<OMPClause *> Clauses,
@@ -15293,7 +15284,7 @@ StmtResult SemaOpenMP::ActOnOpenMPStripeDirective(ArrayRef<OMPClause *> Clauses,
// Delay striping to when template is completely instantiated.
if (SemaRef.CurContext->isDependentContext())
return OMPStripeDirective::Create(Context, StartLoc, EndLoc, Clauses,
- NumLoops, AStmt, nullptr, nullptr);
+ NumLoops, AStmt, nullptr, nullptr, nullptr);
assert(LoopHelpers.size() == NumLoops &&
"Expecting loop iteration space dimensionality to match number of "
@@ -15516,11 +15507,10 @@ StmtResult SemaOpenMP::ActOnOpenMPStripeDirective(ArrayRef<OMPClause *> Clauses,
LoopHelper.Init->getBeginLoc(), LoopHelper.Inc->getEndLoc());
}
- Inner = addLoopVarFinalization(Context, Inner, LoopHelpers);
-
return OMPStripeDirective::Create(Context, StartLoc, EndLoc, Clauses,
NumLoops, AStmt, Inner,
- buildPreInits(Context, PreInits));
+ buildPreInits(Context, PreInits),
+ buildLoopFinalization(Context, LoopHelpers));
}
StmtResult SemaOpenMP::ActOnOpenMPUnrollDirective(ArrayRef<OMPClause *> Clauses,
@@ -15827,7 +15817,7 @@ StmtResult SemaOpenMP::ActOnOpenMPReverseDirective(Stmt *AStmt,
// instantiated.
if (SemaRef.CurContext->isDependentContext())
return OMPReverseDirective::Create(Context, StartLoc, EndLoc, AStmt,
- NumLoops, nullptr, nullptr);
+ NumLoops, nullptr, nullptr, nullptr);
assert(LoopHelpers.size() == NumLoops &&
"Expecting a single-dimensional loop iteration space");
@@ -15987,10 +15977,9 @@ StmtResult SemaOpenMP::ActOnOpenMPReverseDirective(Stmt *AStmt,
ReversedBody, LoopHelper.Init->getBeginLoc(),
LoopHelper.Init->getBeginLoc(), LoopHelper.Inc->getEndLoc());
- Stmt *Inner = addLoopVarFinalization(Context, ReversedFor, LoopHelpers);
-
return OMPReverseDirective::Create(Context, StartLoc, EndLoc, AStmt, NumLoops,
- Inner, buildPreInits(Context, PreInits));
+ ReversedFor, buildPreInits(Context, PreInits),
+ buildLoopFinalization(Context, LoopHelpers));
}
/// Build the AST for \#pragma omp split counts(c1, c2, ...).
@@ -16249,7 +16238,7 @@ StmtResult SemaOpenMP::ActOnOpenMPInterchangeDirective(
// Delay interchange to when template is completely instantiated.
if (CurContext->isDependentContext())
return OMPInterchangeDirective::Create(Context, StartLoc, EndLoc, Clauses,
- NumLoops, AStmt, nullptr, nullptr);
+ NumLoops, AStmt, nullptr, nullptr, nullptr);
// An invalid expression in the permutation clause is set to nullptr in
// ActOnOpenMPPermutationClause.
@@ -16302,7 +16291,7 @@ StmtResult SemaOpenMP::ActOnOpenMPInterchangeDirective(
return Idx == Arg;
}))
return OMPInterchangeDirective::Create(Context, StartLoc, EndLoc, Clauses,
- NumLoops, AStmt, AStmt, nullptr);
+ NumLoops, AStmt, AStmt, nullptr, nullptr);
// Find the affected loops.
SmallVector<Stmt *> LoopStmts(NumLoops, nullptr);
@@ -16412,11 +16401,10 @@ StmtResult SemaOpenMP::ActOnOpenMPInterchangeDirective(
SourceHelper.Inc->getEndLoc());
}
- Inner = addLoopVarFinalization(Context, Inner, LoopHelpers);
-
return OMPInterchangeDirective::Create(Context, StartLoc, EndLoc, Clauses,
NumLoops, AStmt, Inner,
- buildPreInits(Context, PreInits));
+ buildPreInits(Context, PreInits),
+ buildLoopFinalization(Context, LoopHelpers));
}
StmtResult SemaOpenMP::ActOnOpenMPFuseDirective(ArrayRef<OMPClause *> Clauses,
@@ -16439,7 +16427,7 @@ StmtResult SemaOpenMP::ActOnOpenMPFuseDirective(ArrayRef<OMPClause *> Clauses,
// because a dependent context could prevent determining its true value
if (CurrContext->isDependentContext())
return OMPFuseDirective::Create(Context, StartLoc, EndLoc, Clauses,
- /* NumLoops */ 1, AStmt, nullptr, nullptr);
+ /* NumLoops */ 1, AStmt, nullptr, nullptr, nullptr);
// Validate that the potential loop sequence is transformable for fusion
// Also collect the HelperExprs, Loop Stmts, Inits, and Number of loops
@@ -16893,16 +16881,16 @@ StmtResult SemaOpenMP::ActOnOpenMPFuseDirective(ArrayRef<OMPClause *> Clauses,
SourceLocation(), SourceLocation());
}
- // Collect HelperExprs from all loops for finalization
- SmallVector<OMPLoopBasedDirective::HelperExprs, 2> LoopHelpers;
- for (const auto &Loop : SeqAnalysis.Loops) {
- LoopHelpers.push_back(Loop.HelperExprs);
- }
- FusionStmt = addLoopVarFinalization(Context, FusionStmt, LoopHelpers);
+ // TODO: For fuse directive, loop variable finalization is complex because:
+ // 1. The fused loops' original variables are replaced by new fusion variables
+ // 2. The scope and structure change significantly after fusion
+ // 3. Pre-computed Finals reference variables that may not exist after fusion
+ // For now, we pass nullptr for Finals to avoid scope issues.
+ // Proper implementation requires computing Finals after fusion transformation.
return OMPFuseDirective::Create(Context, StartLoc, EndLoc, Clauses,
NumGeneratedTopLevelLoops, AStmt, FusionStmt,
- buildPreInits(Context, PreInits));
+ buildPreInits(Context, PreInits), nullptr);
}
OMPClause *SemaOpenMP::ActOnOpenMPSingleExprClause(OpenMPClauseKind Kind,
diff --git a/clang/test/OpenMP/interchange_codegen.cpp b/clang/test/OpenMP/interchange_codegen.cpp
index 8e833c9df324c..656d0fbff7cad 100644
--- a/clang/test/OpenMP/interchange_codegen.cpp
+++ b/clang/test/OpenMP/interchange_codegen.cpp
@@ -123,6 +123,7 @@ extern "C" void foo10() {
#endif /* HEADER */
+
// CHECK1-LABEL: define {{[^@]+}}@body
// CHECK1-SAME: (...) #[[ATTR0:[0-9]+]] {
// CHECK1-NEXT: entry:
@@ -156,7 +157,7 @@ extern "C" void foo10() {
// CHECK1-NEXT: [[TMP5:%.*]] = load i32, ptr [[I]], align 4
// CHECK1-NEXT: [[ADD:%.*]] = add nsw i32 [[TMP5]], [[TMP4]]
// CHECK1-NEXT: store i32 [[ADD]], ptr [[I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP3:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP2:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: ret void
//
@@ -262,15 +263,41 @@ extern "C" void foo10() {
// CHECK1-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTPERMUTED_1_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add i32 [[TMP28]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTPERMUTED_1_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND16]], !llvm.loop [[LOOP5:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND16]], !llvm.loop [[LOOP4:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC22:%.*]]
// CHECK1: for.inc22:
// CHECK1-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTPERMUTED_0_IV_J]], align 4
// CHECK1-NEXT: [[INC23:%.*]] = add i32 [[TMP29]], 1
// CHECK1-NEXT: store i32 [[INC23]], ptr [[DOTPERMUTED_0_IV_J]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP5:![0-9]+]]
// CHECK1: for.end24:
+// CHECK1-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK1-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[SUB25:%.*]] = sub i32 [[TMP31]], [[TMP32]]
+// CHECK1-NEXT: [[SUB26:%.*]] = sub i32 [[SUB25]], 1
+// CHECK1-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK1-NEXT: [[ADD27:%.*]] = add i32 [[SUB26]], [[TMP33]]
+// CHECK1-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK1-NEXT: [[DIV28:%.*]] = udiv i32 [[ADD27]], [[TMP34]]
+// CHECK1-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK1-NEXT: [[MUL29:%.*]] = mul i32 [[DIV28]], [[TMP35]]
+// CHECK1-NEXT: [[ADD30:%.*]] = add i32 [[TMP30]], [[MUL29]]
+// CHECK1-NEXT: store i32 [[ADD30]], ptr [[I]], align 4
+// CHECK1-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_5]], align 4
+// CHECK1-NEXT: [[TMP37:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
+// CHECK1-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_5]], align 4
+// CHECK1-NEXT: [[SUB31:%.*]] = sub i32 [[TMP37]], [[TMP38]]
+// CHECK1-NEXT: [[SUB32:%.*]] = sub i32 [[SUB31]], 1
+// CHECK1-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTNEW_STEP7]], align 4
+// CHECK1-NEXT: [[ADD33:%.*]] = add i32 [[SUB32]], [[TMP39]]
+// CHECK1-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTNEW_STEP7]], align 4
+// CHECK1-NEXT: [[DIV34:%.*]] = udiv i32 [[ADD33]], [[TMP40]]
+// CHECK1-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTNEW_STEP7]], align 4
+// CHECK1-NEXT: [[MUL35:%.*]] = mul i32 [[DIV34]], [[TMP41]]
+// CHECK1-NEXT: [[ADD36:%.*]] = add i32 [[TMP36]], [[MUL35]]
+// CHECK1-NEXT: store i32 [[ADD36]], ptr [[J]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -342,7 +369,7 @@ extern "C" void foo10() {
// CHECK1-NEXT: [[TMP12:%.*]] = load i32, ptr [[DOTPERMUTED_1_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP12]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTPERMUTED_1_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP7:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK1: omp.body.continue:
@@ -439,7 +466,7 @@ extern "C" void foo10() {
// CHECK1-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTPERMUTED_1_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP14]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTPERMUTED_1_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP7:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK1: omp.body.continue:
@@ -754,29 +781,33 @@ extern "C" void foo10() {
// CHECK1-NEXT: [[TMP12:%.*]] = load i32, ptr [[DOTPERMUTED_3_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP12]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTPERMUTED_3_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND11]], !llvm.loop [[LOOP9:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND11]], !llvm.loop [[LOOP8:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC16:%.*]]
// CHECK1: for.inc16:
// CHECK1-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTPERMUTED_2_IV_L]], align 4
// CHECK1-NEXT: [[INC17:%.*]] = add nsw i32 [[TMP13]], 1
// CHECK1-NEXT: store i32 [[INC17]], ptr [[DOTPERMUTED_2_IV_L]], align 4
-// CHECK1-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP10:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP9:![0-9]+]]
// CHECK1: for.end18:
// CHECK1-NEXT: br label [[FOR_INC19:%.*]]
// CHECK1: for.inc19:
// CHECK1-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTPERMUTED_1_IV_K]], align 4
// CHECK1-NEXT: [[INC20:%.*]] = add nsw i32 [[TMP14]], 1
// CHECK1-NEXT: store i32 [[INC20]], ptr [[DOTPERMUTED_1_IV_K]], align 4
-// CHECK1-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP11:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP10:![0-9]+]]
// CHECK1: for.end21:
// CHECK1-NEXT: br label [[FOR_INC22:%.*]]
// CHECK1: for.inc22:
// CHECK1-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTPERMUTED_0_IV_J]], align 4
// CHECK1-NEXT: [[INC23:%.*]] = add nsw i32 [[TMP15]], 1
// CHECK1-NEXT: store i32 [[INC23]], ptr [[DOTPERMUTED_0_IV_J]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP12:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP11:![0-9]+]]
// CHECK1: for.end24:
+// CHECK1-NEXT: store i32 19, ptr [[I]], align 4
+// CHECK1-NEXT: store i32 19, ptr [[J]], align 4
+// CHECK1-NEXT: store i32 19, ptr [[K]], align 4
+// CHECK1-NEXT: store i32 19, ptr [[L]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -2225,14 +2256,14 @@ extern "C" void foo10() {
// CHECK1-NEXT: [[V:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: store double 4.200000e+01, ptr [[C]], align 8
// CHECK1-NEXT: store ptr [[ARR]], ptr [[__RANGE2]], align 8
-// CHECK1-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META12:![0-9]+]], !align [[META13:![0-9]+]]
// CHECK1-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP0]], i64 0, i64 0
// CHECK1-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 128
// CHECK1-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META12]], !align [[META13]]
// CHECK1-NEXT: [[ARRAYDECAY1:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY1]], ptr [[__BEGIN2]], align 8
-// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META12]], !align [[META13]]
// CHECK1-NEXT: [[ARRAYDECAY2:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY2]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK1-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__END2]], align 8
@@ -2277,7 +2308,7 @@ extern "C" void foo10() {
// CHECK1-NEXT: [[TMP12:%.*]] = load ptr, ptr [[__BEGIN2]], align 8
// CHECK1-NEXT: store ptr [[TMP12]], ptr [[V]], align 8
// CHECK1-NEXT: [[TMP13:%.*]] = load double, ptr [[C]], align 8
-// CHECK1-NEXT: [[TMP14:%.*]] = load ptr, ptr [[V]], align 8
+// CHECK1-NEXT: [[TMP14:%.*]] = load ptr, ptr [[V]], align 8, !nonnull [[META12]], !align [[META13]]
// CHECK1-NEXT: [[TMP15:%.*]] = load double, ptr [[TMP14]], align 8
// CHECK1-NEXT: [[TMP16:%.*]] = load i32, ptr [[I]], align 4
// CHECK1-NEXT: call void (...) @body(double noundef [[TMP13]], double noundef [[TMP15]], i32 noundef [[TMP16]])
@@ -2286,15 +2317,29 @@ extern "C" void foo10() {
// CHECK1-NEXT: [[TMP17:%.*]] = load i64, ptr [[DOTPERMUTED_1_IV___BEGIN2]], align 8
// CHECK1-NEXT: [[INC:%.*]] = add nsw i64 [[TMP17]], 1
// CHECK1-NEXT: store i64 [[INC]], ptr [[DOTPERMUTED_1_IV___BEGIN2]], align 8
-// CHECK1-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP13:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP14:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC13:%.*]]
// CHECK1: for.inc13:
// CHECK1-NEXT: [[TMP18:%.*]] = load i32, ptr [[DOTPERMUTED_0_IV_I]], align 4
// CHECK1-NEXT: [[INC14:%.*]] = add nsw i32 [[TMP18]], 1
// CHECK1-NEXT: store i32 [[INC14]], ptr [[DOTPERMUTED_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP14:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP15:![0-9]+]]
// CHECK1: for.end15:
+// CHECK1-NEXT: [[TMP19:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK1-NEXT: [[TMP20:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_3]], align 8
+// CHECK1-NEXT: [[TMP21:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK1-NEXT: [[SUB_PTR_LHS_CAST16:%.*]] = ptrtoint ptr [[TMP20]] to i64
+// CHECK1-NEXT: [[SUB_PTR_RHS_CAST17:%.*]] = ptrtoint ptr [[TMP21]] to i64
+// CHECK1-NEXT: [[SUB_PTR_SUB18:%.*]] = sub i64 [[SUB_PTR_LHS_CAST16]], [[SUB_PTR_RHS_CAST17]]
+// CHECK1-NEXT: [[SUB_PTR_DIV19:%.*]] = sdiv exact i64 [[SUB_PTR_SUB18]], 8
+// CHECK1-NEXT: [[SUB20:%.*]] = sub nsw i64 [[SUB_PTR_DIV19]], 1
+// CHECK1-NEXT: [[ADD21:%.*]] = add nsw i64 [[SUB20]], 1
+// CHECK1-NEXT: [[DIV22:%.*]] = sdiv i64 [[ADD21]], 1
+// CHECK1-NEXT: [[MUL23:%.*]] = mul nsw i64 [[DIV22]], 1
+// CHECK1-NEXT: [[ADD_PTR24:%.*]] = getelementptr inbounds double, ptr [[TMP19]], i64 [[MUL23]]
+// CHECK1-NEXT: store ptr [[ADD_PTR24]], ptr [[__BEGIN2]], align 8
+// CHECK1-NEXT: store i32 42, ptr [[I]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -2342,14 +2387,14 @@ extern "C" void foo10() {
// CHECK1-NEXT: [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB2]])
// CHECK1-NEXT: store double 4.200000e+01, ptr [[C]], align 8
// CHECK1-NEXT: store ptr [[A]], ptr [[__RANGE3]], align 8
-// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE3]], align 8
+// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE3]], align 8, !nonnull [[META12]], !align [[META13]]
// CHECK1-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK1-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 128
// CHECK1-NEXT: store ptr [[ADD_PTR]], ptr [[__END3]], align 8
-// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE3]], align 8
+// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE3]], align 8, !nonnull [[META12]], !align [[META13]]
// CHECK1-NEXT: [[ARRAYDECAY4:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY4]], ptr [[__BEGIN3]], align 8
-// CHECK1-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__RANGE3]], align 8
+// CHECK1-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__RANGE3]], align 8, !nonnull [[META12]], !align [[META13]]
// CHECK1-NEXT: [[ARRAYDECAY5:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP3]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY5]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = load ptr, ptr [[__END3]], align 8
@@ -2367,14 +2412,14 @@ extern "C" void foo10() {
// CHECK1-NEXT: store i64 [[SUB8]], ptr [[DOTCAPTURE_EXPR_7]], align 8
// CHECK1-NEXT: store double 4.200000e+01, ptr [[D]], align 8
// CHECK1-NEXT: store ptr [[B]], ptr [[__RANGE4]], align 8
-// CHECK1-NEXT: [[TMP7:%.*]] = load ptr, ptr [[__RANGE4]], align 8
+// CHECK1-NEXT: [[TMP7:%.*]] = load ptr, ptr [[__RANGE4]], align 8, !nonnull [[META12]], !align [[META13]]
// CHECK1-NEXT: [[ARRAYDECAY9:%.*]] = getelementptr inbounds [16 x double], ptr [[TMP7]], i64 0, i64 0
// CHECK1-NEXT: [[ADD_PTR10:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY9]], i64 16
// CHECK1-NEXT: store ptr [[ADD_PTR10]], ptr [[__END4]], align 8
-// CHECK1-NEXT: [[TMP8:%.*]] = load ptr, ptr [[__RANGE4]], align 8
+// CHECK1-NEXT: [[TMP8:%.*]] = load ptr, ptr [[__RANGE4]], align 8, !nonnull [[META12]], !align [[META13]]
// CHECK1-NEXT: [[ARRAYDECAY11:%.*]] = getelementptr inbounds [16 x double], ptr [[TMP8]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY11]], ptr [[__BEGIN4]], align 8
-// CHECK1-NEXT: [[TMP9:%.*]] = load ptr, ptr [[__RANGE4]], align 8
+// CHECK1-NEXT: [[TMP9:%.*]] = load ptr, ptr [[__RANGE4]], align 8, !nonnull [[META12]], !align [[META13]]
// CHECK1-NEXT: [[ARRAYDECAY13:%.*]] = getelementptr inbounds [16 x double], ptr [[TMP9]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY13]], ptr [[DOTCAPTURE_EXPR_12]], align 8
// CHECK1-NEXT: [[TMP10:%.*]] = load ptr, ptr [[__END4]], align 8
@@ -2701,7 +2746,7 @@ extern "C" void foo10() {
// CHECK1-NEXT: [[TMP87:%.*]] = load double, ptr [[C]], align 8
// CHECK1-NEXT: [[TMP88:%.*]] = load double, ptr [[AA]], align 8
// CHECK1-NEXT: [[TMP89:%.*]] = load double, ptr [[D]], align 8
-// CHECK1-NEXT: [[TMP90:%.*]] = load ptr, ptr [[BB]], align 8
+// CHECK1-NEXT: [[TMP90:%.*]] = load ptr, ptr [[BB]], align 8, !nonnull [[META12]], !align [[META13]]
// CHECK1-NEXT: [[TMP91:%.*]] = load double, ptr [[TMP90]], align 8
// CHECK1-NEXT: [[TMP92:%.*]] = load i32, ptr [[J40]], align 4
// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP86]], double noundef [[TMP87]], double noundef [[TMP88]], double noundef [[TMP89]], double noundef [[TMP91]], i32 noundef [[TMP92]])
@@ -2756,7 +2801,7 @@ extern "C" void foo10() {
// CHECK2-NEXT: [[TMP5:%.*]] = load i32, ptr [[I]], align 4
// CHECK2-NEXT: [[ADD:%.*]] = add nsw i32 [[TMP5]], [[TMP4]]
// CHECK2-NEXT: store i32 [[ADD]], ptr [[I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP3:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP2:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: ret void
//
@@ -2805,14 +2850,14 @@ extern "C" void foo10() {
// CHECK2-NEXT: [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB2:[0-9]+]])
// CHECK2-NEXT: store double 4.200000e+01, ptr [[C]], align 8
// CHECK2-NEXT: store ptr [[A]], ptr [[__RANGE3]], align 8
-// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE3]], align 8
+// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE3]], align 8, !nonnull [[META4:![0-9]+]], !align [[META5:![0-9]+]]
// CHECK2-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK2-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 128
// CHECK2-NEXT: store ptr [[ADD_PTR]], ptr [[__END3]], align 8
-// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE3]], align 8
+// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE3]], align 8, !nonnull [[META4]], !align [[META5]]
// CHECK2-NEXT: [[ARRAYDECAY4:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY4]], ptr [[__BEGIN3]], align 8
-// CHECK2-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__RANGE3]], align 8
+// CHECK2-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__RANGE3]], align 8, !nonnull [[META4]], !align [[META5]]
// CHECK2-NEXT: [[ARRAYDECAY5:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP3]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY5]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = load ptr, ptr [[__END3]], align 8
@@ -2830,14 +2875,14 @@ extern "C" void foo10() {
// CHECK2-NEXT: store i64 [[SUB8]], ptr [[DOTCAPTURE_EXPR_7]], align 8
// CHECK2-NEXT: store double 4.200000e+01, ptr [[D]], align 8
// CHECK2-NEXT: store ptr [[B]], ptr [[__RANGE4]], align 8
-// CHECK2-NEXT: [[TMP7:%.*]] = load ptr, ptr [[__RANGE4]], align 8
+// CHECK2-NEXT: [[TMP7:%.*]] = load ptr, ptr [[__RANGE4]], align 8, !nonnull [[META4]], !align [[META5]]
// CHECK2-NEXT: [[ARRAYDECAY9:%.*]] = getelementptr inbounds [16 x double], ptr [[TMP7]], i64 0, i64 0
// CHECK2-NEXT: [[ADD_PTR10:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY9]], i64 16
// CHECK2-NEXT: store ptr [[ADD_PTR10]], ptr [[__END4]], align 8
-// CHECK2-NEXT: [[TMP8:%.*]] = load ptr, ptr [[__RANGE4]], align 8
+// CHECK2-NEXT: [[TMP8:%.*]] = load ptr, ptr [[__RANGE4]], align 8, !nonnull [[META4]], !align [[META5]]
// CHECK2-NEXT: [[ARRAYDECAY11:%.*]] = getelementptr inbounds [16 x double], ptr [[TMP8]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY11]], ptr [[__BEGIN4]], align 8
-// CHECK2-NEXT: [[TMP9:%.*]] = load ptr, ptr [[__RANGE4]], align 8
+// CHECK2-NEXT: [[TMP9:%.*]] = load ptr, ptr [[__RANGE4]], align 8, !nonnull [[META4]], !align [[META5]]
// CHECK2-NEXT: [[ARRAYDECAY13:%.*]] = getelementptr inbounds [16 x double], ptr [[TMP9]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY13]], ptr [[DOTCAPTURE_EXPR_12]], align 8
// CHECK2-NEXT: [[TMP10:%.*]] = load ptr, ptr [[__END4]], align 8
@@ -3164,7 +3209,7 @@ extern "C" void foo10() {
// CHECK2-NEXT: [[TMP87:%.*]] = load double, ptr [[C]], align 8
// CHECK2-NEXT: [[TMP88:%.*]] = load double, ptr [[AA]], align 8
// CHECK2-NEXT: [[TMP89:%.*]] = load double, ptr [[D]], align 8
-// CHECK2-NEXT: [[TMP90:%.*]] = load ptr, ptr [[BB]], align 8
+// CHECK2-NEXT: [[TMP90:%.*]] = load ptr, ptr [[BB]], align 8, !nonnull [[META4]], !align [[META5]]
// CHECK2-NEXT: [[TMP91:%.*]] = load double, ptr [[TMP90]], align 8
// CHECK2-NEXT: [[TMP92:%.*]] = load i32, ptr [[J40]], align 4
// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP86]], double noundef [[TMP87]], double noundef [[TMP88]], double noundef [[TMP89]], double noundef [[TMP91]], i32 noundef [[TMP92]])
@@ -3287,15 +3332,41 @@ extern "C" void foo10() {
// CHECK2-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTPERMUTED_1_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add i32 [[TMP28]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTPERMUTED_1_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND16]], !llvm.loop [[LOOP5:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND16]], !llvm.loop [[LOOP6:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC22:%.*]]
// CHECK2: for.inc22:
// CHECK2-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTPERMUTED_0_IV_J]], align 4
// CHECK2-NEXT: [[INC23:%.*]] = add i32 [[TMP29]], 1
// CHECK2-NEXT: store i32 [[INC23]], ptr [[DOTPERMUTED_0_IV_J]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP7:![0-9]+]]
// CHECK2: for.end24:
+// CHECK2-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK2-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[SUB25:%.*]] = sub i32 [[TMP31]], [[TMP32]]
+// CHECK2-NEXT: [[SUB26:%.*]] = sub i32 [[SUB25]], 1
+// CHECK2-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK2-NEXT: [[ADD27:%.*]] = add i32 [[SUB26]], [[TMP33]]
+// CHECK2-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK2-NEXT: [[DIV28:%.*]] = udiv i32 [[ADD27]], [[TMP34]]
+// CHECK2-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK2-NEXT: [[MUL29:%.*]] = mul i32 [[DIV28]], [[TMP35]]
+// CHECK2-NEXT: [[ADD30:%.*]] = add i32 [[TMP30]], [[MUL29]]
+// CHECK2-NEXT: store i32 [[ADD30]], ptr [[I]], align 4
+// CHECK2-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_5]], align 4
+// CHECK2-NEXT: [[TMP37:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
+// CHECK2-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_5]], align 4
+// CHECK2-NEXT: [[SUB31:%.*]] = sub i32 [[TMP37]], [[TMP38]]
+// CHECK2-NEXT: [[SUB32:%.*]] = sub i32 [[SUB31]], 1
+// CHECK2-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTNEW_STEP7]], align 4
+// CHECK2-NEXT: [[ADD33:%.*]] = add i32 [[SUB32]], [[TMP39]]
+// CHECK2-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTNEW_STEP7]], align 4
+// CHECK2-NEXT: [[DIV34:%.*]] = udiv i32 [[ADD33]], [[TMP40]]
+// CHECK2-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTNEW_STEP7]], align 4
+// CHECK2-NEXT: [[MUL35:%.*]] = mul i32 [[DIV34]], [[TMP41]]
+// CHECK2-NEXT: [[ADD36:%.*]] = add i32 [[TMP36]], [[MUL35]]
+// CHECK2-NEXT: store i32 [[ADD36]], ptr [[J]], align 4
// CHECK2-NEXT: ret void
//
//
@@ -3367,7 +3438,7 @@ extern "C" void foo10() {
// CHECK2-NEXT: [[TMP12:%.*]] = load i32, ptr [[DOTPERMUTED_1_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP12]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTPERMUTED_1_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP7:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK2: omp.body.continue:
@@ -3464,7 +3535,7 @@ extern "C" void foo10() {
// CHECK2-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTPERMUTED_1_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP14]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTPERMUTED_1_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP9:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK2: omp.body.continue:
@@ -3779,29 +3850,33 @@ extern "C" void foo10() {
// CHECK2-NEXT: [[TMP12:%.*]] = load i32, ptr [[DOTPERMUTED_3_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP12]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTPERMUTED_3_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND11]], !llvm.loop [[LOOP9:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND11]], !llvm.loop [[LOOP10:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC16:%.*]]
// CHECK2: for.inc16:
// CHECK2-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTPERMUTED_2_IV_L]], align 4
// CHECK2-NEXT: [[INC17:%.*]] = add nsw i32 [[TMP13]], 1
// CHECK2-NEXT: store i32 [[INC17]], ptr [[DOTPERMUTED_2_IV_L]], align 4
-// CHECK2-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP10:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP11:![0-9]+]]
// CHECK2: for.end18:
// CHECK2-NEXT: br label [[FOR_INC19:%.*]]
// CHECK2: for.inc19:
// CHECK2-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTPERMUTED_1_IV_K]], align 4
// CHECK2-NEXT: [[INC20:%.*]] = add nsw i32 [[TMP14]], 1
// CHECK2-NEXT: store i32 [[INC20]], ptr [[DOTPERMUTED_1_IV_K]], align 4
-// CHECK2-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP11:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP12:![0-9]+]]
// CHECK2: for.end21:
// CHECK2-NEXT: br label [[FOR_INC22:%.*]]
// CHECK2: for.inc22:
// CHECK2-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTPERMUTED_0_IV_J]], align 4
// CHECK2-NEXT: [[INC23:%.*]] = add nsw i32 [[TMP15]], 1
// CHECK2-NEXT: store i32 [[INC23]], ptr [[DOTPERMUTED_0_IV_J]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP12:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP13:![0-9]+]]
// CHECK2: for.end24:
+// CHECK2-NEXT: store i32 19, ptr [[I]], align 4
+// CHECK2-NEXT: store i32 19, ptr [[J]], align 4
+// CHECK2-NEXT: store i32 19, ptr [[K]], align 4
+// CHECK2-NEXT: store i32 19, ptr [[L]], align 4
// CHECK2-NEXT: ret void
//
//
@@ -3822,14 +3897,14 @@ extern "C" void foo10() {
// CHECK2-NEXT: [[V:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: store double 4.200000e+01, ptr [[C]], align 8
// CHECK2-NEXT: store ptr [[ARR]], ptr [[__RANGE2]], align 8
-// CHECK2-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META4]], !align [[META5]]
// CHECK2-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP0]], i64 0, i64 0
// CHECK2-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 128
// CHECK2-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META4]], !align [[META5]]
// CHECK2-NEXT: [[ARRAYDECAY1:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY1]], ptr [[__BEGIN2]], align 8
-// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META4]], !align [[META5]]
// CHECK2-NEXT: [[ARRAYDECAY2:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY2]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK2-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__END2]], align 8
@@ -3874,7 +3949,7 @@ extern "C" void foo10() {
// CHECK2-NEXT: [[TMP12:%.*]] = load ptr, ptr [[__BEGIN2]], align 8
// CHECK2-NEXT: store ptr [[TMP12]], ptr [[V]], align 8
// CHECK2-NEXT: [[TMP13:%.*]] = load double, ptr [[C]], align 8
-// CHECK2-NEXT: [[TMP14:%.*]] = load ptr, ptr [[V]], align 8
+// CHECK2-NEXT: [[TMP14:%.*]] = load ptr, ptr [[V]], align 8, !nonnull [[META4]], !align [[META5]]
// CHECK2-NEXT: [[TMP15:%.*]] = load double, ptr [[TMP14]], align 8
// CHECK2-NEXT: [[TMP16:%.*]] = load i32, ptr [[I]], align 4
// CHECK2-NEXT: call void (...) @body(double noundef [[TMP13]], double noundef [[TMP15]], i32 noundef [[TMP16]])
@@ -3883,15 +3958,29 @@ extern "C" void foo10() {
// CHECK2-NEXT: [[TMP17:%.*]] = load i64, ptr [[DOTPERMUTED_1_IV___BEGIN2]], align 8
// CHECK2-NEXT: [[INC:%.*]] = add nsw i64 [[TMP17]], 1
// CHECK2-NEXT: store i64 [[INC]], ptr [[DOTPERMUTED_1_IV___BEGIN2]], align 8
-// CHECK2-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP13:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP14:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC13:%.*]]
// CHECK2: for.inc13:
// CHECK2-NEXT: [[TMP18:%.*]] = load i32, ptr [[DOTPERMUTED_0_IV_I]], align 4
// CHECK2-NEXT: [[INC14:%.*]] = add nsw i32 [[TMP18]], 1
// CHECK2-NEXT: store i32 [[INC14]], ptr [[DOTPERMUTED_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP14:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP15:![0-9]+]]
// CHECK2: for.end15:
+// CHECK2-NEXT: [[TMP19:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK2-NEXT: [[TMP20:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_3]], align 8
+// CHECK2-NEXT: [[TMP21:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK2-NEXT: [[SUB_PTR_LHS_CAST16:%.*]] = ptrtoint ptr [[TMP20]] to i64
+// CHECK2-NEXT: [[SUB_PTR_RHS_CAST17:%.*]] = ptrtoint ptr [[TMP21]] to i64
+// CHECK2-NEXT: [[SUB_PTR_SUB18:%.*]] = sub i64 [[SUB_PTR_LHS_CAST16]], [[SUB_PTR_RHS_CAST17]]
+// CHECK2-NEXT: [[SUB_PTR_DIV19:%.*]] = sdiv exact i64 [[SUB_PTR_SUB18]], 8
+// CHECK2-NEXT: [[SUB20:%.*]] = sub nsw i64 [[SUB_PTR_DIV19]], 1
+// CHECK2-NEXT: [[ADD21:%.*]] = add nsw i64 [[SUB20]], 1
+// CHECK2-NEXT: [[DIV22:%.*]] = sdiv i64 [[ADD21]], 1
+// CHECK2-NEXT: [[MUL23:%.*]] = mul nsw i64 [[DIV22]], 1
+// CHECK2-NEXT: [[ADD_PTR24:%.*]] = getelementptr inbounds double, ptr [[TMP19]], i64 [[MUL23]]
+// CHECK2-NEXT: store ptr [[ADD_PTR24]], ptr [[__BEGIN2]], align 8
+// CHECK2-NEXT: store i32 42, ptr [[I]], align 4
// CHECK2-NEXT: ret void
//
//
@@ -5321,4 +5410,4 @@ extern "C" void foo10() {
// CHECK2: omp.precond.end:
// CHECK2-NEXT: call void @__kmpc_barrier(ptr @[[GLOB3]], i32 [[TMP0]])
// CHECK2-NEXT: ret void
-
+//
diff --git a/clang/test/OpenMP/loop_transform_final_iv.c b/clang/test/OpenMP/loop_transform_final_iv.c
index 969badbe9636b..99175a1415b66 100644
--- a/clang/test/OpenMP/loop_transform_final_iv.c
+++ b/clang/test/OpenMP/loop_transform_final_iv.c
@@ -84,21 +84,6 @@ void test_interchange(void) {
// After loop: i should be 10, j should be 5
}
-void test_fuse(void) {
- // CHECK-LABEL: define {{.*}} @test_fuse
- int i, j;
- #pragma omp fuse
- {
- for (i = 0; i < 10; i++) {
- }
- for (j = 0; j < 5; j++) {
- }
- }
- // CHECK: store i32 10, ptr %i
- // CHECK: store i32 5, ptr %j
- // After loop: i should be 10, j should be 5
-}
-
void test_for_workshare(void) {
// CHECK-LABEL: define {{.*}} @test_for_workshare
int i;
diff --git a/clang/test/OpenMP/reverse_codegen.cpp b/clang/test/OpenMP/reverse_codegen.cpp
index 3915b8ab9c74f..d28686836c6c6 100644
--- a/clang/test/OpenMP/reverse_codegen.cpp
+++ b/clang/test/OpenMP/reverse_codegen.cpp
@@ -1,4 +1,4 @@
-// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --function-signature --include-generated-funcs --replace-value-regex "__omp_offloading_[0-9a-z]+_[0-9a-z]+" "reduction_size[.].+[.]" "pl_cond[.].+[.|,]" --prefix-filecheck-ir-name _
+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --include-generated-funcs --replace-value-regex "__omp_offloading_[0-9a-z]+_[0-9a-z]+" "reduction_size[.].+[.]" "pl_cond[.].+[.|,]" --prefix-filecheck-ir-name _ --version 4
// expected-no-diagnostics
@@ -94,21 +94,21 @@ extern "C" void foo7() {
#endif /* HEADER */
-// CHECK1-LABEL: define {{[^@]+}}@body
-// CHECK1-SAME: (...) #[[ATTR0:[0-9]+]] {
+// CHECK1-LABEL: define dso_local void @body(
+// CHECK1-SAME: ...) #[[ATTR0:[0-9]+]] {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: ret void
//
//
-// CHECK1-LABEL: define {{[^@]+}}@__cxx_global_var_init
-// CHECK1-SAME: () #[[ATTR1:[0-9]+]] section ".text.startup" {
+// CHECK1-LABEL: define internal void @__cxx_global_var_init(
+// CHECK1-SAME: ) #[[ATTR1:[0-9]+]] section ".text.startup" {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: call void @_ZN1SC1Ev(ptr noundef nonnull align 4 dereferenceable(4) @s)
// CHECK1-NEXT: ret void
//
//
-// CHECK1-LABEL: define {{[^@]+}}@_ZN1SC1Ev
-// CHECK1-SAME: (ptr noundef nonnull align 4 dereferenceable(4) [[THIS:%.*]]) unnamed_addr #[[ATTR0]] comdat align 2 {
+// CHECK1-LABEL: define linkonce_odr void @_ZN1SC1Ev(
+// CHECK1-SAME: ptr noundef nonnull align 4 dereferenceable(4) [[THIS:%.*]]) unnamed_addr #[[ATTR0]] comdat align 2 {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: store ptr [[THIS]], ptr [[THIS_ADDR]], align 8
@@ -117,8 +117,8 @@ extern "C" void foo7() {
// CHECK1-NEXT: ret void
//
//
-// CHECK1-LABEL: define {{[^@]+}}@_ZN1SC2Ev
-// CHECK1-SAME: (ptr noundef nonnull align 4 dereferenceable(4) [[THIS:%.*]]) unnamed_addr #[[ATTR0]] comdat align 2 {
+// CHECK1-LABEL: define linkonce_odr void @_ZN1SC2Ev(
+// CHECK1-SAME: ptr noundef nonnull align 4 dereferenceable(4) [[THIS:%.*]]) unnamed_addr #[[ATTR0]] comdat align 2 {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: [[I2:%.*]] = alloca ptr, align 8
@@ -143,9 +143,9 @@ extern "C" void foo7() {
// CHECK1-NEXT: [[TMP2:%.*]] = load i32, ptr [[DOTREVERSED_IV_I]], align 4
// CHECK1-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP2]], 3
// CHECK1-NEXT: [[ADD:%.*]] = add nsw i32 7, [[MUL]]
-// CHECK1-NEXT: [[TMP3:%.*]] = load ptr, ptr [[I2]], align 8
+// CHECK1-NEXT: [[TMP3:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2:![0-9]+]], !align [[META3:![0-9]+]]
// CHECK1-NEXT: store i32 [[ADD]], ptr [[TMP3]], align 4
-// CHECK1-NEXT: [[TMP4:%.*]] = load ptr, ptr [[I2]], align 8
+// CHECK1-NEXT: [[TMP4:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2]], !align [[META3]]
// CHECK1-NEXT: [[TMP5:%.*]] = load i32, ptr [[TMP4]], align 4
// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP5]])
// CHECK1-NEXT: br label [[FOR_INC:%.*]]
@@ -153,13 +153,15 @@ extern "C" void foo7() {
// CHECK1-NEXT: [[TMP6:%.*]] = load i32, ptr [[DOTFORWARD_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP6]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTFORWARD_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP3:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP4:![0-9]+]]
// CHECK1: for.end:
+// CHECK1-NEXT: [[TMP7:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2]], !align [[META3]]
+// CHECK1-NEXT: store i32 19, ptr [[TMP7]], align 4
// CHECK1-NEXT: ret void
//
//
-// CHECK1-LABEL: define {{[^@]+}}@foo1
-// CHECK1-SAME: (i32 noundef [[START:%.*]], i32 noundef [[END:%.*]], i32 noundef [[STEP:%.*]]) #[[ATTR0]] {
+// CHECK1-LABEL: define dso_local void @foo1(
+// CHECK1-SAME: i32 noundef [[START:%.*]], i32 noundef [[END:%.*]], i32 noundef [[STEP:%.*]]) #[[ATTR0]] {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[START_ADDR:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[END_ADDR:%.*]] = alloca i32, align 4
@@ -220,13 +222,26 @@ extern "C" void foo7() {
// CHECK1-NEXT: [[TMP16:%.*]] = load i32, ptr [[DOTFORWARD_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add i32 [[TMP16]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTFORWARD_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP5:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
// CHECK1: for.end:
+// CHECK1-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[TMP18:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK1-NEXT: [[TMP19:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[SUB10:%.*]] = sub i32 [[TMP18]], [[TMP19]]
+// CHECK1-NEXT: [[SUB11:%.*]] = sub i32 [[SUB10]], 1
+// CHECK1-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK1-NEXT: [[ADD12:%.*]] = add i32 [[SUB11]], [[TMP20]]
+// CHECK1-NEXT: [[TMP21:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK1-NEXT: [[DIV13:%.*]] = udiv i32 [[ADD12]], [[TMP21]]
+// CHECK1-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK1-NEXT: [[MUL14:%.*]] = mul i32 [[DIV13]], [[TMP22]]
+// CHECK1-NEXT: [[ADD15:%.*]] = add i32 [[TMP17]], [[MUL14]]
+// CHECK1-NEXT: store i32 [[ADD15]], ptr [[I]], align 4
// CHECK1-NEXT: ret void
//
//
-// CHECK1-LABEL: define {{[^@]+}}@foo2
-// CHECK1-SAME: () #[[ATTR0]] {
+// CHECK1-LABEL: define dso_local void @foo2(
+// CHECK1-SAME: ) #[[ATTR0]] {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[TMP:%.*]] = alloca i32, align 4
@@ -293,8 +308,8 @@ extern "C" void foo7() {
// CHECK1-NEXT: ret void
//
//
-// CHECK1-LABEL: define {{[^@]+}}@foo3
-// CHECK1-SAME: () #[[ATTR0]] {
+// CHECK1-LABEL: define dso_local void @foo3(
+// CHECK1-SAME: ) #[[ATTR0]] {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[TMP:%.*]] = alloca i32, align 4
@@ -393,15 +408,15 @@ extern "C" void foo7() {
// CHECK1-NEXT: ret void
//
//
-// CHECK1-LABEL: define {{[^@]+}}@foo4
-// CHECK1-SAME: () #[[ATTR0]] {
+// CHECK1-LABEL: define dso_local void @foo4(
+// CHECK1-SAME: ) #[[ATTR0]] {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: call void (ptr, i32, ptr, ...) @__kmpc_fork_call(ptr @[[GLOB2]], i32 0, ptr @foo4.omp_outlined)
// CHECK1-NEXT: ret void
//
//
-// CHECK1-LABEL: define {{[^@]+}}@foo4.omp_outlined
-// CHECK1-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]]) #[[ATTR4:[0-9]+]] {
+// CHECK1-LABEL: define internal void @foo4.omp_outlined(
+// CHECK1-SAME: ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]]) #[[ATTR4:[0-9]+]] {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 8
@@ -472,15 +487,15 @@ extern "C" void foo7() {
// CHECK1-NEXT: ret void
//
//
-// CHECK1-LABEL: define {{[^@]+}}@tfoo5
-// CHECK1-SAME: () #[[ATTR0]] {
+// CHECK1-LABEL: define dso_local void @tfoo5(
+// CHECK1-SAME: ) #[[ATTR0]] {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: call void @_Z4foo5IiTnT_Li3EEvS0_S0_(i32 noundef 0, i32 noundef 42)
// CHECK1-NEXT: ret void
//
//
-// CHECK1-LABEL: define {{[^@]+}}@_Z4foo5IiTnT_Li3EEvS0_S0_
-// CHECK1-SAME: (i32 noundef [[START:%.*]], i32 noundef [[END:%.*]]) #[[ATTR0]] comdat {
+// CHECK1-LABEL: define linkonce_odr void @_Z4foo5IiTnT_Li3EEvS0_S0_(
+// CHECK1-SAME: i32 noundef [[START:%.*]], i32 noundef [[END:%.*]]) #[[ATTR0]] comdat {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[START_ADDR:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[END_ADDR:%.*]] = alloca i32, align 4
@@ -533,13 +548,23 @@ extern "C" void foo7() {
// CHECK1-NEXT: [[TMP12:%.*]] = load i32, ptr [[DOTFORWARD_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add i32 [[TMP12]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTFORWARD_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP9:![0-9]+]]
// CHECK1: for.end:
+// CHECK1-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK1-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[SUB10:%.*]] = sub i32 [[TMP14]], [[TMP15]]
+// CHECK1-NEXT: [[SUB11:%.*]] = sub i32 [[SUB10]], 1
+// CHECK1-NEXT: [[ADD12:%.*]] = add i32 [[SUB11]], 3
+// CHECK1-NEXT: [[DIV13:%.*]] = udiv i32 [[ADD12]], 3
+// CHECK1-NEXT: [[MUL14:%.*]] = mul i32 [[DIV13]], 3
+// CHECK1-NEXT: [[ADD15:%.*]] = add i32 [[TMP13]], [[MUL14]]
+// CHECK1-NEXT: store i32 [[ADD15]], ptr [[I]], align 4
// CHECK1-NEXT: ret void
//
//
-// CHECK1-LABEL: define {{[^@]+}}@foo6
-// CHECK1-SAME: () #[[ATTR0]] {
+// CHECK1-LABEL: define dso_local void @foo6(
+// CHECK1-SAME: ) #[[ATTR0]] {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[ARR:%.*]] = alloca [128 x double], align 16
// CHECK1-NEXT: [[C:%.*]] = alloca i32, align 4
@@ -554,14 +579,14 @@ extern "C" void foo7() {
// CHECK1-NEXT: [[V:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: store i32 42, ptr [[C]], align 4
// CHECK1-NEXT: store ptr [[ARR]], ptr [[__RANGE2]], align 8
-// CHECK1-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META10:![0-9]+]]
// CHECK1-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP0]], i64 0, i64 0
// CHECK1-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 128
// CHECK1-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META10]]
// CHECK1-NEXT: [[ARRAYDECAY1:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY1]], ptr [[__BEGIN2]], align 8
-// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META10]]
// CHECK1-NEXT: [[ARRAYDECAY2:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY2]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK1-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__END2]], align 8
@@ -599,7 +624,7 @@ extern "C" void foo7() {
// CHECK1-NEXT: store ptr [[ADD_PTR10]], ptr [[__BEGIN2]], align 8
// CHECK1-NEXT: [[TMP12:%.*]] = load ptr, ptr [[__BEGIN2]], align 8
// CHECK1-NEXT: store ptr [[TMP12]], ptr [[V]], align 8
-// CHECK1-NEXT: [[TMP13:%.*]] = load ptr, ptr [[V]], align 8
+// CHECK1-NEXT: [[TMP13:%.*]] = load ptr, ptr [[V]], align 8, !nonnull [[META2]], !align [[META10]]
// CHECK1-NEXT: [[TMP14:%.*]] = load double, ptr [[TMP13]], align 8
// CHECK1-NEXT: [[TMP15:%.*]] = load i32, ptr [[C]], align 4
// CHECK1-NEXT: call void (...) @body(double noundef [[TMP14]], i32 noundef [[TMP15]])
@@ -608,13 +633,26 @@ extern "C" void foo7() {
// CHECK1-NEXT: [[TMP16:%.*]] = load i64, ptr [[DOTFORWARD_IV___BEGIN2]], align 8
// CHECK1-NEXT: [[INC:%.*]] = add nsw i64 [[TMP16]], 1
// CHECK1-NEXT: store i64 [[INC]], ptr [[DOTFORWARD_IV___BEGIN2]], align 8
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP9:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP11:![0-9]+]]
// CHECK1: for.end:
+// CHECK1-NEXT: [[TMP17:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK1-NEXT: [[TMP18:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_3]], align 8
+// CHECK1-NEXT: [[TMP19:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK1-NEXT: [[SUB_PTR_LHS_CAST11:%.*]] = ptrtoint ptr [[TMP18]] to i64
+// CHECK1-NEXT: [[SUB_PTR_RHS_CAST12:%.*]] = ptrtoint ptr [[TMP19]] to i64
+// CHECK1-NEXT: [[SUB_PTR_SUB13:%.*]] = sub i64 [[SUB_PTR_LHS_CAST11]], [[SUB_PTR_RHS_CAST12]]
+// CHECK1-NEXT: [[SUB_PTR_DIV14:%.*]] = sdiv exact i64 [[SUB_PTR_SUB13]], 8
+// CHECK1-NEXT: [[SUB15:%.*]] = sub nsw i64 [[SUB_PTR_DIV14]], 1
+// CHECK1-NEXT: [[ADD16:%.*]] = add nsw i64 [[SUB15]], 1
+// CHECK1-NEXT: [[DIV17:%.*]] = sdiv i64 [[ADD16]], 1
+// CHECK1-NEXT: [[MUL18:%.*]] = mul nsw i64 [[DIV17]], 1
+// CHECK1-NEXT: [[ADD_PTR19:%.*]] = getelementptr inbounds double, ptr [[TMP17]], i64 [[MUL18]]
+// CHECK1-NEXT: store ptr [[ADD_PTR19]], ptr [[__BEGIN2]], align 8
// CHECK1-NEXT: ret void
//
//
-// CHECK1-LABEL: define {{[^@]+}}@foo7
-// CHECK1-SAME: () #[[ATTR0]] {
+// CHECK1-LABEL: define dso_local void @foo7(
+// CHECK1-SAME: ) #[[ATTR0]] {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[A:%.*]] = alloca [128 x double], align 16
// CHECK1-NEXT: [[DOTOMP_IV:%.*]] = alloca i64, align 8
@@ -645,14 +683,14 @@ extern "C" void foo7() {
// CHECK1-NEXT: [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB2]])
// CHECK1-NEXT: store i32 42, ptr [[C]], align 4
// CHECK1-NEXT: store ptr [[A]], ptr [[__RANGE3]], align 8
-// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE3]], align 8
+// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE3]], align 8, !nonnull [[META2]], !align [[META10]]
// CHECK1-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK1-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 128
// CHECK1-NEXT: store ptr [[ADD_PTR]], ptr [[__END3]], align 8
-// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE3]], align 8
+// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE3]], align 8, !nonnull [[META2]], !align [[META10]]
// CHECK1-NEXT: [[ARRAYDECAY3:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY3]], ptr [[__BEGIN3]], align 8
-// CHECK1-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__RANGE3]], align 8
+// CHECK1-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__RANGE3]], align 8, !nonnull [[META2]], !align [[META10]]
// CHECK1-NEXT: [[ARRAYDECAY4:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP3]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY4]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = load ptr, ptr [[__END3]], align 8
@@ -795,7 +833,7 @@ extern "C" void foo7() {
// CHECK1-NEXT: store ptr [[TMP36]], ptr [[V]], align 8
// CHECK1-NEXT: [[TMP37:%.*]] = load i32, ptr [[K15]], align 4
// CHECK1-NEXT: [[TMP38:%.*]] = load i32, ptr [[C]], align 4
-// CHECK1-NEXT: [[TMP39:%.*]] = load ptr, ptr [[V]], align 8
+// CHECK1-NEXT: [[TMP39:%.*]] = load ptr, ptr [[V]], align 8, !nonnull [[META2]], !align [[META10]]
// CHECK1-NEXT: [[TMP40:%.*]] = load double, ptr [[TMP39]], align 8
// CHECK1-NEXT: [[TMP41:%.*]] = load i32, ptr [[J17]], align 4
// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP37]], i32 noundef [[TMP38]], double noundef [[TMP40]], i32 noundef [[TMP41]])
@@ -817,22 +855,22 @@ extern "C" void foo7() {
// CHECK1-NEXT: ret void
//
//
-// CHECK1-LABEL: define {{[^@]+}}@_GLOBAL__sub_I_reverse_codegen.cpp
-// CHECK1-SAME: () #[[ATTR1]] section ".text.startup" {
+// CHECK1-LABEL: define internal void @_GLOBAL__sub_I_reverse_codegen.cpp(
+// CHECK1-SAME: ) #[[ATTR1]] section ".text.startup" {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: call void @__cxx_global_var_init()
// CHECK1-NEXT: ret void
//
//
-// CHECK2-LABEL: define {{[^@]+}}@__cxx_global_var_init
-// CHECK2-SAME: () #[[ATTR0:[0-9]+]] section ".text.startup" {
+// CHECK2-LABEL: define internal void @__cxx_global_var_init(
+// CHECK2-SAME: ) #[[ATTR0:[0-9]+]] section ".text.startup" {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: call void @_ZN1SC1Ev(ptr noundef nonnull align 4 dereferenceable(4) @s)
// CHECK2-NEXT: ret void
//
//
-// CHECK2-LABEL: define {{[^@]+}}@_ZN1SC1Ev
-// CHECK2-SAME: (ptr noundef nonnull align 4 dereferenceable(4) [[THIS:%.*]]) unnamed_addr #[[ATTR1:[0-9]+]] comdat align 2 {
+// CHECK2-LABEL: define linkonce_odr void @_ZN1SC1Ev(
+// CHECK2-SAME: ptr noundef nonnull align 4 dereferenceable(4) [[THIS:%.*]]) unnamed_addr #[[ATTR1:[0-9]+]] comdat align 2 {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: store ptr [[THIS]], ptr [[THIS_ADDR]], align 8
@@ -841,8 +879,8 @@ extern "C" void foo7() {
// CHECK2-NEXT: ret void
//
//
-// CHECK2-LABEL: define {{[^@]+}}@_ZN1SC2Ev
-// CHECK2-SAME: (ptr noundef nonnull align 4 dereferenceable(4) [[THIS:%.*]]) unnamed_addr #[[ATTR1]] comdat align 2 {
+// CHECK2-LABEL: define linkonce_odr void @_ZN1SC2Ev(
+// CHECK2-SAME: ptr noundef nonnull align 4 dereferenceable(4) [[THIS:%.*]]) unnamed_addr #[[ATTR1]] comdat align 2 {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: [[I2:%.*]] = alloca ptr, align 8
@@ -867,9 +905,9 @@ extern "C" void foo7() {
// CHECK2-NEXT: [[TMP2:%.*]] = load i32, ptr [[DOTREVERSED_IV_I]], align 4
// CHECK2-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP2]], 3
// CHECK2-NEXT: [[ADD:%.*]] = add nsw i32 7, [[MUL]]
-// CHECK2-NEXT: [[TMP3:%.*]] = load ptr, ptr [[I2]], align 8
+// CHECK2-NEXT: [[TMP3:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2:![0-9]+]], !align [[META3:![0-9]+]]
// CHECK2-NEXT: store i32 [[ADD]], ptr [[TMP3]], align 4
-// CHECK2-NEXT: [[TMP4:%.*]] = load ptr, ptr [[I2]], align 8
+// CHECK2-NEXT: [[TMP4:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2]], !align [[META3]]
// CHECK2-NEXT: [[TMP5:%.*]] = load i32, ptr [[TMP4]], align 4
// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP5]])
// CHECK2-NEXT: br label [[FOR_INC:%.*]]
@@ -877,19 +915,21 @@ extern "C" void foo7() {
// CHECK2-NEXT: [[TMP6:%.*]] = load i32, ptr [[DOTFORWARD_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP6]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTFORWARD_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP3:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP4:![0-9]+]]
// CHECK2: for.end:
+// CHECK2-NEXT: [[TMP7:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2]], !align [[META3]]
+// CHECK2-NEXT: store i32 19, ptr [[TMP7]], align 4
// CHECK2-NEXT: ret void
//
//
-// CHECK2-LABEL: define {{[^@]+}}@body
-// CHECK2-SAME: (...) #[[ATTR1]] {
+// CHECK2-LABEL: define dso_local void @body(
+// CHECK2-SAME: ...) #[[ATTR1]] {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: ret void
//
//
-// CHECK2-LABEL: define {{[^@]+}}@foo1
-// CHECK2-SAME: (i32 noundef [[START:%.*]], i32 noundef [[END:%.*]], i32 noundef [[STEP:%.*]]) #[[ATTR1]] {
+// CHECK2-LABEL: define dso_local void @foo1(
+// CHECK2-SAME: i32 noundef [[START:%.*]], i32 noundef [[END:%.*]], i32 noundef [[STEP:%.*]]) #[[ATTR1]] {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[START_ADDR:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[END_ADDR:%.*]] = alloca i32, align 4
@@ -950,13 +990,26 @@ extern "C" void foo7() {
// CHECK2-NEXT: [[TMP16:%.*]] = load i32, ptr [[DOTFORWARD_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add i32 [[TMP16]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTFORWARD_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP5:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
// CHECK2: for.end:
+// CHECK2-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[TMP18:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK2-NEXT: [[TMP19:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[SUB10:%.*]] = sub i32 [[TMP18]], [[TMP19]]
+// CHECK2-NEXT: [[SUB11:%.*]] = sub i32 [[SUB10]], 1
+// CHECK2-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK2-NEXT: [[ADD12:%.*]] = add i32 [[SUB11]], [[TMP20]]
+// CHECK2-NEXT: [[TMP21:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK2-NEXT: [[DIV13:%.*]] = udiv i32 [[ADD12]], [[TMP21]]
+// CHECK2-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK2-NEXT: [[MUL14:%.*]] = mul i32 [[DIV13]], [[TMP22]]
+// CHECK2-NEXT: [[ADD15:%.*]] = add i32 [[TMP17]], [[MUL14]]
+// CHECK2-NEXT: store i32 [[ADD15]], ptr [[I]], align 4
// CHECK2-NEXT: ret void
//
//
-// CHECK2-LABEL: define {{[^@]+}}@foo2
-// CHECK2-SAME: () #[[ATTR1]] {
+// CHECK2-LABEL: define dso_local void @foo2(
+// CHECK2-SAME: ) #[[ATTR1]] {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[TMP:%.*]] = alloca i32, align 4
@@ -1023,8 +1076,8 @@ extern "C" void foo7() {
// CHECK2-NEXT: ret void
//
//
-// CHECK2-LABEL: define {{[^@]+}}@foo3
-// CHECK2-SAME: () #[[ATTR1]] {
+// CHECK2-LABEL: define dso_local void @foo3(
+// CHECK2-SAME: ) #[[ATTR1]] {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[TMP:%.*]] = alloca i32, align 4
@@ -1123,15 +1176,15 @@ extern "C" void foo7() {
// CHECK2-NEXT: ret void
//
//
-// CHECK2-LABEL: define {{[^@]+}}@foo4
-// CHECK2-SAME: () #[[ATTR1]] {
+// CHECK2-LABEL: define dso_local void @foo4(
+// CHECK2-SAME: ) #[[ATTR1]] {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: call void (ptr, i32, ptr, ...) @__kmpc_fork_call(ptr @[[GLOB2]], i32 0, ptr @foo4.omp_outlined)
// CHECK2-NEXT: ret void
//
//
-// CHECK2-LABEL: define {{[^@]+}}@foo4.omp_outlined
-// CHECK2-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]]) #[[ATTR4:[0-9]+]] {
+// CHECK2-LABEL: define internal void @foo4.omp_outlined(
+// CHECK2-SAME: ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]]) #[[ATTR4:[0-9]+]] {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 8
@@ -1202,8 +1255,8 @@ extern "C" void foo7() {
// CHECK2-NEXT: ret void
//
//
-// CHECK2-LABEL: define {{[^@]+}}@foo6
-// CHECK2-SAME: () #[[ATTR1]] {
+// CHECK2-LABEL: define dso_local void @foo6(
+// CHECK2-SAME: ) #[[ATTR1]] {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[ARR:%.*]] = alloca [128 x double], align 16
// CHECK2-NEXT: [[C:%.*]] = alloca i32, align 4
@@ -1218,14 +1271,14 @@ extern "C" void foo7() {
// CHECK2-NEXT: [[V:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: store i32 42, ptr [[C]], align 4
// CHECK2-NEXT: store ptr [[ARR]], ptr [[__RANGE2]], align 8
-// CHECK2-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META9:![0-9]+]]
// CHECK2-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP0]], i64 0, i64 0
// CHECK2-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 128
// CHECK2-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[ARRAYDECAY1:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY1]], ptr [[__BEGIN2]], align 8
-// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[ARRAYDECAY2:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY2]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK2-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__END2]], align 8
@@ -1263,7 +1316,7 @@ extern "C" void foo7() {
// CHECK2-NEXT: store ptr [[ADD_PTR10]], ptr [[__BEGIN2]], align 8
// CHECK2-NEXT: [[TMP12:%.*]] = load ptr, ptr [[__BEGIN2]], align 8
// CHECK2-NEXT: store ptr [[TMP12]], ptr [[V]], align 8
-// CHECK2-NEXT: [[TMP13:%.*]] = load ptr, ptr [[V]], align 8
+// CHECK2-NEXT: [[TMP13:%.*]] = load ptr, ptr [[V]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[TMP14:%.*]] = load double, ptr [[TMP13]], align 8
// CHECK2-NEXT: [[TMP15:%.*]] = load i32, ptr [[C]], align 4
// CHECK2-NEXT: call void (...) @body(double noundef [[TMP14]], i32 noundef [[TMP15]])
@@ -1272,13 +1325,26 @@ extern "C" void foo7() {
// CHECK2-NEXT: [[TMP16:%.*]] = load i64, ptr [[DOTFORWARD_IV___BEGIN2]], align 8
// CHECK2-NEXT: [[INC:%.*]] = add nsw i64 [[TMP16]], 1
// CHECK2-NEXT: store i64 [[INC]], ptr [[DOTFORWARD_IV___BEGIN2]], align 8
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP10:![0-9]+]]
// CHECK2: for.end:
+// CHECK2-NEXT: [[TMP17:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK2-NEXT: [[TMP18:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_3]], align 8
+// CHECK2-NEXT: [[TMP19:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK2-NEXT: [[SUB_PTR_LHS_CAST11:%.*]] = ptrtoint ptr [[TMP18]] to i64
+// CHECK2-NEXT: [[SUB_PTR_RHS_CAST12:%.*]] = ptrtoint ptr [[TMP19]] to i64
+// CHECK2-NEXT: [[SUB_PTR_SUB13:%.*]] = sub i64 [[SUB_PTR_LHS_CAST11]], [[SUB_PTR_RHS_CAST12]]
+// CHECK2-NEXT: [[SUB_PTR_DIV14:%.*]] = sdiv exact i64 [[SUB_PTR_SUB13]], 8
+// CHECK2-NEXT: [[SUB15:%.*]] = sub nsw i64 [[SUB_PTR_DIV14]], 1
+// CHECK2-NEXT: [[ADD16:%.*]] = add nsw i64 [[SUB15]], 1
+// CHECK2-NEXT: [[DIV17:%.*]] = sdiv i64 [[ADD16]], 1
+// CHECK2-NEXT: [[MUL18:%.*]] = mul nsw i64 [[DIV17]], 1
+// CHECK2-NEXT: [[ADD_PTR19:%.*]] = getelementptr inbounds double, ptr [[TMP17]], i64 [[MUL18]]
+// CHECK2-NEXT: store ptr [[ADD_PTR19]], ptr [[__BEGIN2]], align 8
// CHECK2-NEXT: ret void
//
//
-// CHECK2-LABEL: define {{[^@]+}}@foo7
-// CHECK2-SAME: () #[[ATTR1]] {
+// CHECK2-LABEL: define dso_local void @foo7(
+// CHECK2-SAME: ) #[[ATTR1]] {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[A:%.*]] = alloca [128 x double], align 16
// CHECK2-NEXT: [[DOTOMP_IV:%.*]] = alloca i64, align 8
@@ -1309,14 +1375,14 @@ extern "C" void foo7() {
// CHECK2-NEXT: [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB2]])
// CHECK2-NEXT: store i32 42, ptr [[C]], align 4
// CHECK2-NEXT: store ptr [[A]], ptr [[__RANGE3]], align 8
-// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE3]], align 8
+// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE3]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK2-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 128
// CHECK2-NEXT: store ptr [[ADD_PTR]], ptr [[__END3]], align 8
-// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE3]], align 8
+// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE3]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[ARRAYDECAY3:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY3]], ptr [[__BEGIN3]], align 8
-// CHECK2-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__RANGE3]], align 8
+// CHECK2-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__RANGE3]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[ARRAYDECAY4:%.*]] = getelementptr inbounds [128 x double], ptr [[TMP3]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY4]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = load ptr, ptr [[__END3]], align 8
@@ -1459,7 +1525,7 @@ extern "C" void foo7() {
// CHECK2-NEXT: store ptr [[TMP36]], ptr [[V]], align 8
// CHECK2-NEXT: [[TMP37:%.*]] = load i32, ptr [[K15]], align 4
// CHECK2-NEXT: [[TMP38:%.*]] = load i32, ptr [[C]], align 4
-// CHECK2-NEXT: [[TMP39:%.*]] = load ptr, ptr [[V]], align 8
+// CHECK2-NEXT: [[TMP39:%.*]] = load ptr, ptr [[V]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[TMP40:%.*]] = load double, ptr [[TMP39]], align 8
// CHECK2-NEXT: [[TMP41:%.*]] = load i32, ptr [[J17]], align 4
// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP37]], i32 noundef [[TMP38]], double noundef [[TMP40]], i32 noundef [[TMP41]])
@@ -1481,15 +1547,15 @@ extern "C" void foo7() {
// CHECK2-NEXT: ret void
//
//
-// CHECK2-LABEL: define {{[^@]+}}@tfoo5
-// CHECK2-SAME: () #[[ATTR1]] {
+// CHECK2-LABEL: define dso_local void @tfoo5(
+// CHECK2-SAME: ) #[[ATTR1]] {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: call void @_Z4foo5IiTnT_Li3EEvS0_S0_(i32 noundef 0, i32 noundef 42)
// CHECK2-NEXT: ret void
//
//
-// CHECK2-LABEL: define {{[^@]+}}@_Z4foo5IiTnT_Li3EEvS0_S0_
-// CHECK2-SAME: (i32 noundef [[START:%.*]], i32 noundef [[END:%.*]]) #[[ATTR1]] comdat {
+// CHECK2-LABEL: define linkonce_odr void @_Z4foo5IiTnT_Li3EEvS0_S0_(
+// CHECK2-SAME: i32 noundef [[START:%.*]], i32 noundef [[END:%.*]]) #[[ATTR1]] comdat {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[START_ADDR:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[END_ADDR:%.*]] = alloca i32, align 4
@@ -1542,14 +1608,43 @@ extern "C" void foo7() {
// CHECK2-NEXT: [[TMP12:%.*]] = load i32, ptr [[DOTFORWARD_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add i32 [[TMP12]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTFORWARD_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP9:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP11:![0-9]+]]
// CHECK2: for.end:
+// CHECK2-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK2-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[SUB10:%.*]] = sub i32 [[TMP14]], [[TMP15]]
+// CHECK2-NEXT: [[SUB11:%.*]] = sub i32 [[SUB10]], 1
+// CHECK2-NEXT: [[ADD12:%.*]] = add i32 [[SUB11]], 3
+// CHECK2-NEXT: [[DIV13:%.*]] = udiv i32 [[ADD12]], 3
+// CHECK2-NEXT: [[MUL14:%.*]] = mul i32 [[DIV13]], 3
+// CHECK2-NEXT: [[ADD15:%.*]] = add i32 [[TMP13]], [[MUL14]]
+// CHECK2-NEXT: store i32 [[ADD15]], ptr [[I]], align 4
// CHECK2-NEXT: ret void
//
//
-// CHECK2-LABEL: define {{[^@]+}}@_GLOBAL__sub_I_reverse_codegen.cpp
-// CHECK2-SAME: () #[[ATTR0]] section ".text.startup" {
+// CHECK2-LABEL: define internal void @_GLOBAL__sub_I_reverse_codegen.cpp(
+// CHECK2-SAME: ) #[[ATTR0]] section ".text.startup" {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: call void @__cxx_global_var_init()
// CHECK2-NEXT: ret void
//
+//.
+// CHECK1: [[META2]] = !{}
+// CHECK1: [[META3]] = !{i64 4}
+// CHECK1: [[LOOP4]] = distinct !{[[LOOP4]], [[META5:![0-9]+]]}
+// CHECK1: [[META5]] = !{!"llvm.loop.mustprogress"}
+// CHECK1: [[LOOP6]] = distinct !{[[LOOP6]], [[META5]]}
+// CHECK1: [[LOOP9]] = distinct !{[[LOOP9]], [[META5]]}
+// CHECK1: [[META10]] = !{i64 8}
+// CHECK1: [[LOOP11]] = distinct !{[[LOOP11]], [[META5]]}
+//.
+// CHECK2: [[META2]] = !{}
+// CHECK2: [[META3]] = !{i64 4}
+// CHECK2: [[LOOP4]] = distinct !{[[LOOP4]], [[META5:![0-9]+]]}
+// CHECK2: [[META5]] = !{!"llvm.loop.mustprogress"}
+// CHECK2: [[LOOP6]] = distinct !{[[LOOP6]], [[META5]]}
+// CHECK2: [[META9]] = !{i64 8}
+// CHECK2: [[LOOP10]] = distinct !{[[LOOP10]], [[META5]]}
+// CHECK2: [[LOOP11]] = distinct !{[[LOOP11]], [[META5]]}
+//.
diff --git a/clang/test/OpenMP/stripe_codegen.cpp b/clang/test/OpenMP/stripe_codegen.cpp
index d2a58a2c3addc..5c88f1818efa8 100644
--- a/clang/test/OpenMP/stripe_codegen.cpp
+++ b/clang/test/OpenMP/stripe_codegen.cpp
@@ -179,9 +179,9 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP5:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK1-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP5]], 3
// CHECK1-NEXT: [[ADD9:%.*]] = add nsw i32 7, [[MUL]]
-// CHECK1-NEXT: [[TMP6:%.*]] = load ptr, ptr [[I2]], align 8
+// CHECK1-NEXT: [[TMP6:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2:![0-9]+]], !align [[META3:![0-9]+]]
// CHECK1-NEXT: store i32 [[ADD9]], ptr [[TMP6]], align 4
-// CHECK1-NEXT: [[TMP7:%.*]] = load ptr, ptr [[I2]], align 8
+// CHECK1-NEXT: [[TMP7:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2]], !align [[META3]]
// CHECK1-NEXT: [[TMP8:%.*]] = load i32, ptr [[TMP7]], align 4
// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP8]])
// CHECK1-NEXT: br label [[FOR_INC:%.*]]
@@ -189,15 +189,17 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP9]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP3:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP4:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC10:%.*]]
// CHECK1: for.inc10:
// CHECK1-NEXT: [[TMP10:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK1-NEXT: [[ADD11:%.*]] = add nsw i32 [[TMP10]], 5
// CHECK1-NEXT: store i32 [[ADD11]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP5:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
// CHECK1: for.end12:
+// CHECK1-NEXT: [[TMP11:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2]], !align [[META3]]
+// CHECK1-NEXT: store i32 19, ptr [[TMP11]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -281,15 +283,28 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add i32 [[TMP20]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP6:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP7:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC15:%.*]]
// CHECK1: for.inc15:
// CHECK1-NEXT: [[TMP21:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK1-NEXT: [[ADD16:%.*]] = add i32 [[TMP21]], 5
// CHECK1-NEXT: store i32 [[ADD16]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP7:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
// CHECK1: for.end17:
+// CHECK1-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[TMP23:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK1-NEXT: [[TMP24:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[SUB18:%.*]] = sub i32 [[TMP23]], [[TMP24]]
+// CHECK1-NEXT: [[SUB19:%.*]] = sub i32 [[SUB18]], 1
+// CHECK1-NEXT: [[TMP25:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK1-NEXT: [[ADD20:%.*]] = add i32 [[SUB19]], [[TMP25]]
+// CHECK1-NEXT: [[TMP26:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK1-NEXT: [[DIV21:%.*]] = udiv i32 [[ADD20]], [[TMP26]]
+// CHECK1-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK1-NEXT: [[MUL22:%.*]] = mul i32 [[DIV21]], [[TMP27]]
+// CHECK1-NEXT: [[ADD23:%.*]] = add i32 [[TMP22]], [[MUL22]]
+// CHECK1-NEXT: store i32 [[ADD23]], ptr [[I]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -380,29 +395,31 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTSTRIPE_1_IV_J]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP14]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_1_IV_J]], align 4
-// CHECK1-NEXT: br label [[FOR_COND10]], !llvm.loop [[LOOP8:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND10]], !llvm.loop [[LOOP9:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC22:%.*]]
// CHECK1: for.inc22:
// CHECK1-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC23:%.*]] = add nsw i32 [[TMP15]], 1
// CHECK1-NEXT: store i32 [[INC23]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP9:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP10:![0-9]+]]
// CHECK1: for.end24:
// CHECK1-NEXT: br label [[FOR_INC25:%.*]]
// CHECK1: for.inc25:
// CHECK1-NEXT: [[TMP16:%.*]] = load i32, ptr [[DOTFLOOR_1_IV_J]], align 4
// CHECK1-NEXT: [[ADD26:%.*]] = add nsw i32 [[TMP16]], 5
// CHECK1-NEXT: store i32 [[ADD26]], ptr [[DOTFLOOR_1_IV_J]], align 4
-// CHECK1-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP10:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP11:![0-9]+]]
// CHECK1: for.end27:
// CHECK1-NEXT: br label [[FOR_INC28:%.*]]
// CHECK1: for.inc28:
// CHECK1-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK1-NEXT: [[ADD29:%.*]] = add nsw i32 [[TMP17]], 5
// CHECK1-NEXT: store i32 [[ADD29]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP11:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP12:![0-9]+]]
// CHECK1: for.end30:
+// CHECK1-NEXT: store i32 19, ptr [[I]], align 4
+// CHECK1-NEXT: store i32 19, ptr [[J]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -516,21 +533,21 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTSTRIPE_1_IV_J]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP20]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_1_IV_J]], align 4
-// CHECK1-NEXT: br label [[FOR_COND15]], !llvm.loop [[LOOP12:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND15]], !llvm.loop [[LOOP13:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC27:%.*]]
// CHECK1: for.inc27:
// CHECK1-NEXT: [[TMP21:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC28:%.*]] = add nsw i32 [[TMP21]], 1
// CHECK1-NEXT: store i32 [[INC28]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND3]], !llvm.loop [[LOOP13:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND3]], !llvm.loop [[LOOP14:![0-9]+]]
// CHECK1: for.end29:
// CHECK1-NEXT: br label [[FOR_INC30:%.*]]
// CHECK1: for.inc30:
// CHECK1-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTFLOOR_1_IV_J]], align 4
// CHECK1-NEXT: [[ADD31:%.*]] = add nsw i32 [[TMP22]], 5
// CHECK1-NEXT: store i32 [[ADD31]], ptr [[DOTFLOOR_1_IV_J]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP14:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP15:![0-9]+]]
// CHECK1: for.end32:
// CHECK1-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK1: omp.body.continue:
@@ -669,21 +686,21 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTSTRIPE_1_IV_J]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP22]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_1_IV_J]], align 4
-// CHECK1-NEXT: br label [[FOR_COND20]], !llvm.loop [[LOOP15:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND20]], !llvm.loop [[LOOP16:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC32:%.*]]
// CHECK1: for.inc32:
// CHECK1-NEXT: [[TMP23:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC33:%.*]] = add nsw i32 [[TMP23]], 1
// CHECK1-NEXT: store i32 [[INC33]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND8]], !llvm.loop [[LOOP16:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND8]], !llvm.loop [[LOOP17:![0-9]+]]
// CHECK1: for.end34:
// CHECK1-NEXT: br label [[FOR_INC35:%.*]]
// CHECK1: for.inc35:
// CHECK1-NEXT: [[TMP24:%.*]] = load i32, ptr [[DOTFLOOR_1_IV_J]], align 4
// CHECK1-NEXT: [[ADD36:%.*]] = add nsw i32 [[TMP24]], 5
// CHECK1-NEXT: store i32 [[ADD36]], ptr [[DOTFLOOR_1_IV_J]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP17:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP18:![0-9]+]]
// CHECK1: for.end37:
// CHECK1-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK1: omp.body.continue:
@@ -996,7 +1013,7 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP14]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP18:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP19:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK1: omp.body.continue:
@@ -1092,15 +1109,25 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP16:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add i32 [[TMP16]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP21:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP22:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC15:%.*]]
// CHECK1: for.inc15:
// CHECK1-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK1-NEXT: [[ADD16:%.*]] = add i32 [[TMP17]], 5
// CHECK1-NEXT: store i32 [[ADD16]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP22:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP23:![0-9]+]]
// CHECK1: for.end17:
+// CHECK1-NEXT: [[TMP18:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[TMP19:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK1-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[SUB18:%.*]] = sub i32 [[TMP19]], [[TMP20]]
+// CHECK1-NEXT: [[SUB19:%.*]] = sub i32 [[SUB18]], 1
+// CHECK1-NEXT: [[ADD20:%.*]] = add i32 [[SUB19]], 3
+// CHECK1-NEXT: [[DIV21:%.*]] = udiv i32 [[ADD20]], 3
+// CHECK1-NEXT: [[MUL22:%.*]] = mul i32 [[DIV21]], 3
+// CHECK1-NEXT: [[ADD23:%.*]] = add i32 [[TMP18]], [[MUL22]]
+// CHECK1-NEXT: store i32 [[ADD23]], ptr [[I]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -1171,7 +1198,7 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP11:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP11]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP23:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP24:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC17:%.*]]
// CHECK1: for.inc17:
@@ -1188,8 +1215,9 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK1-NEXT: [[ADD23:%.*]] = add nsw i32 [[TMP14]], [[COND22]]
// CHECK1-NEXT: store i32 [[ADD23]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP24:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP25:![0-9]+]]
// CHECK1: for.end24:
+// CHECK1-NEXT: store i32 19, ptr [[I]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -1207,14 +1235,14 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[V:%.*]] = alloca double, align 8
// CHECK1-NEXT: [[ARRAY:%.*]] = getelementptr inbounds nuw [[STRUCT_DATA_T]], ptr [[DATA]], i32 0, i32 0
// CHECK1-NEXT: store ptr [[ARRAY]], ptr [[__RANGE2]], align 8
-// CHECK1-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META26:![0-9]+]]
// CHECK1-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP0]], i64 0, i64 0
// CHECK1-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 12
// CHECK1-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META26]]
// CHECK1-NEXT: [[ARRAYDECAY1:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY1]], ptr [[__BEGIN2]], align 8
-// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META26]]
// CHECK1-NEXT: [[ARRAYDECAY2:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY2]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK1-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__END2]], align 8
@@ -1278,15 +1306,28 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP19:%.*]] = load i64, ptr [[DOTSTRIPE_0_IV___BEGIN2]], align 8
// CHECK1-NEXT: [[INC:%.*]] = add nsw i64 [[TMP19]], 1
// CHECK1-NEXT: store i64 [[INC]], ptr [[DOTSTRIPE_0_IV___BEGIN2]], align 8
-// CHECK1-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP25:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP27:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC16:%.*]]
// CHECK1: for.inc16:
// CHECK1-NEXT: [[TMP20:%.*]] = load i64, ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
// CHECK1-NEXT: [[ADD17:%.*]] = add nsw i64 [[TMP20]], 5
// CHECK1-NEXT: store i64 [[ADD17]], ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP26:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP28:![0-9]+]]
// CHECK1: for.end18:
+// CHECK1-NEXT: [[TMP21:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK1-NEXT: [[TMP22:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_3]], align 8
+// CHECK1-NEXT: [[TMP23:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK1-NEXT: [[SUB_PTR_LHS_CAST19:%.*]] = ptrtoint ptr [[TMP22]] to i64
+// CHECK1-NEXT: [[SUB_PTR_RHS_CAST20:%.*]] = ptrtoint ptr [[TMP23]] to i64
+// CHECK1-NEXT: [[SUB_PTR_SUB21:%.*]] = sub i64 [[SUB_PTR_LHS_CAST19]], [[SUB_PTR_RHS_CAST20]]
+// CHECK1-NEXT: [[SUB_PTR_DIV22:%.*]] = sdiv exact i64 [[SUB_PTR_SUB21]], 8
+// CHECK1-NEXT: [[SUB23:%.*]] = sub nsw i64 [[SUB_PTR_DIV22]], 1
+// CHECK1-NEXT: [[ADD24:%.*]] = add nsw i64 [[SUB23]], 1
+// CHECK1-NEXT: [[DIV25:%.*]] = sdiv i64 [[ADD24]], 1
+// CHECK1-NEXT: [[MUL26:%.*]] = mul nsw i64 [[DIV25]], 1
+// CHECK1-NEXT: [[ADD_PTR27:%.*]] = getelementptr inbounds double, ptr [[TMP21]], i64 [[MUL26]]
+// CHECK1-NEXT: store ptr [[ADD_PTR27]], ptr [[__BEGIN2]], align 8
// CHECK1-NEXT: ret void
//
//
@@ -1306,14 +1347,14 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: store double 4.200000e+01, ptr [[C]], align 8
// CHECK1-NEXT: [[ARRAY:%.*]] = getelementptr inbounds nuw [[STRUCT_DATA_T]], ptr [[DATA]], i32 0, i32 0
// CHECK1-NEXT: store ptr [[ARRAY]], ptr [[__RANGE2]], align 8
-// CHECK1-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META26]]
// CHECK1-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP0]], i64 0, i64 0
// CHECK1-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 12
// CHECK1-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META26]]
// CHECK1-NEXT: [[ARRAYDECAY1:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY1]], ptr [[__BEGIN2]], align 8
-// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META26]]
// CHECK1-NEXT: [[ARRAYDECAY2:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY2]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK1-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__END2]], align 8
@@ -1378,15 +1419,28 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP20:%.*]] = load i64, ptr [[DOTSTRIPE_0_IV___BEGIN2]], align 8
// CHECK1-NEXT: [[INC:%.*]] = add nsw i64 [[TMP20]], 1
// CHECK1-NEXT: store i64 [[INC]], ptr [[DOTSTRIPE_0_IV___BEGIN2]], align 8
-// CHECK1-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP27:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP29:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC16:%.*]]
// CHECK1: for.inc16:
// CHECK1-NEXT: [[TMP21:%.*]] = load i64, ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
// CHECK1-NEXT: [[ADD17:%.*]] = add nsw i64 [[TMP21]], 5
// CHECK1-NEXT: store i64 [[ADD17]], ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP28:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP30:![0-9]+]]
// CHECK1: for.end18:
+// CHECK1-NEXT: [[TMP22:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK1-NEXT: [[TMP23:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_3]], align 8
+// CHECK1-NEXT: [[TMP24:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK1-NEXT: [[SUB_PTR_LHS_CAST19:%.*]] = ptrtoint ptr [[TMP23]] to i64
+// CHECK1-NEXT: [[SUB_PTR_RHS_CAST20:%.*]] = ptrtoint ptr [[TMP24]] to i64
+// CHECK1-NEXT: [[SUB_PTR_SUB21:%.*]] = sub i64 [[SUB_PTR_LHS_CAST19]], [[SUB_PTR_RHS_CAST20]]
+// CHECK1-NEXT: [[SUB_PTR_DIV22:%.*]] = sdiv exact i64 [[SUB_PTR_SUB21]], 8
+// CHECK1-NEXT: [[SUB23:%.*]] = sub nsw i64 [[SUB_PTR_DIV22]], 1
+// CHECK1-NEXT: [[ADD24:%.*]] = add nsw i64 [[SUB23]], 1
+// CHECK1-NEXT: [[DIV25:%.*]] = sdiv i64 [[ADD24]], 1
+// CHECK1-NEXT: [[MUL26:%.*]] = mul nsw i64 [[DIV25]], 1
+// CHECK1-NEXT: [[ADD_PTR27:%.*]] = getelementptr inbounds double, ptr [[TMP22]], i64 [[MUL26]]
+// CHECK1-NEXT: store ptr [[ADD_PTR27]], ptr [[__BEGIN2]], align 8
// CHECK1-NEXT: ret void
//
//
@@ -1457,9 +1511,9 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP5:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK2-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP5]], 3
// CHECK2-NEXT: [[ADD9:%.*]] = add nsw i32 7, [[MUL]]
-// CHECK2-NEXT: [[TMP6:%.*]] = load ptr, ptr [[I2]], align 8
+// CHECK2-NEXT: [[TMP6:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2:![0-9]+]], !align [[META3:![0-9]+]]
// CHECK2-NEXT: store i32 [[ADD9]], ptr [[TMP6]], align 4
-// CHECK2-NEXT: [[TMP7:%.*]] = load ptr, ptr [[I2]], align 8
+// CHECK2-NEXT: [[TMP7:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2]], !align [[META3]]
// CHECK2-NEXT: [[TMP8:%.*]] = load i32, ptr [[TMP7]], align 4
// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP8]])
// CHECK2-NEXT: br label [[FOR_INC:%.*]]
@@ -1467,15 +1521,17 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP9]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP3:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP4:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC10:%.*]]
// CHECK2: for.inc10:
// CHECK2-NEXT: [[TMP10:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK2-NEXT: [[ADD11:%.*]] = add nsw i32 [[TMP10]], 5
// CHECK2-NEXT: store i32 [[ADD11]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP5:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
// CHECK2: for.end12:
+// CHECK2-NEXT: [[TMP11:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2]], !align [[META3]]
+// CHECK2-NEXT: store i32 19, ptr [[TMP11]], align 4
// CHECK2-NEXT: ret void
//
//
@@ -1565,15 +1621,28 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add i32 [[TMP20]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP6:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP7:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC15:%.*]]
// CHECK2: for.inc15:
// CHECK2-NEXT: [[TMP21:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK2-NEXT: [[ADD16:%.*]] = add i32 [[TMP21]], 5
// CHECK2-NEXT: store i32 [[ADD16]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP7:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
// CHECK2: for.end17:
+// CHECK2-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[TMP23:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK2-NEXT: [[TMP24:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[SUB18:%.*]] = sub i32 [[TMP23]], [[TMP24]]
+// CHECK2-NEXT: [[SUB19:%.*]] = sub i32 [[SUB18]], 1
+// CHECK2-NEXT: [[TMP25:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK2-NEXT: [[ADD20:%.*]] = add i32 [[SUB19]], [[TMP25]]
+// CHECK2-NEXT: [[TMP26:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK2-NEXT: [[DIV21:%.*]] = udiv i32 [[ADD20]], [[TMP26]]
+// CHECK2-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK2-NEXT: [[MUL22:%.*]] = mul i32 [[DIV21]], [[TMP27]]
+// CHECK2-NEXT: [[ADD23:%.*]] = add i32 [[TMP22]], [[MUL22]]
+// CHECK2-NEXT: store i32 [[ADD23]], ptr [[I]], align 4
// CHECK2-NEXT: ret void
//
//
@@ -1593,14 +1662,14 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: store double 4.200000e+01, ptr [[C]], align 8
// CHECK2-NEXT: [[ARRAY:%.*]] = getelementptr inbounds nuw [[STRUCT_DATA_T]], ptr [[DATA]], i32 0, i32 0
// CHECK2-NEXT: store ptr [[ARRAY]], ptr [[__RANGE2]], align 8
-// CHECK2-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META9:![0-9]+]]
// CHECK2-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP0]], i64 0, i64 0
// CHECK2-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 12
// CHECK2-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[ARRAYDECAY1:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY1]], ptr [[__BEGIN2]], align 8
-// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[ARRAYDECAY2:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY2]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK2-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__END2]], align 8
@@ -1665,15 +1734,28 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP20:%.*]] = load i64, ptr [[DOTSTRIPE_0_IV___BEGIN2]], align 8
// CHECK2-NEXT: [[INC:%.*]] = add nsw i64 [[TMP20]], 1
// CHECK2-NEXT: store i64 [[INC]], ptr [[DOTSTRIPE_0_IV___BEGIN2]], align 8
-// CHECK2-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP8:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP10:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC16:%.*]]
// CHECK2: for.inc16:
// CHECK2-NEXT: [[TMP21:%.*]] = load i64, ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
// CHECK2-NEXT: [[ADD17:%.*]] = add nsw i64 [[TMP21]], 5
// CHECK2-NEXT: store i64 [[ADD17]], ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP9:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP11:![0-9]+]]
// CHECK2: for.end18:
+// CHECK2-NEXT: [[TMP22:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK2-NEXT: [[TMP23:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_3]], align 8
+// CHECK2-NEXT: [[TMP24:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK2-NEXT: [[SUB_PTR_LHS_CAST19:%.*]] = ptrtoint ptr [[TMP23]] to i64
+// CHECK2-NEXT: [[SUB_PTR_RHS_CAST20:%.*]] = ptrtoint ptr [[TMP24]] to i64
+// CHECK2-NEXT: [[SUB_PTR_SUB21:%.*]] = sub i64 [[SUB_PTR_LHS_CAST19]], [[SUB_PTR_RHS_CAST20]]
+// CHECK2-NEXT: [[SUB_PTR_DIV22:%.*]] = sdiv exact i64 [[SUB_PTR_SUB21]], 8
+// CHECK2-NEXT: [[SUB23:%.*]] = sub nsw i64 [[SUB_PTR_DIV22]], 1
+// CHECK2-NEXT: [[ADD24:%.*]] = add nsw i64 [[SUB23]], 1
+// CHECK2-NEXT: [[DIV25:%.*]] = sdiv i64 [[ADD24]], 1
+// CHECK2-NEXT: [[MUL26:%.*]] = mul nsw i64 [[DIV25]], 1
+// CHECK2-NEXT: [[ADD_PTR27:%.*]] = getelementptr inbounds double, ptr [[TMP22]], i64 [[MUL26]]
+// CHECK2-NEXT: store ptr [[ADD_PTR27]], ptr [[__BEGIN2]], align 8
// CHECK2-NEXT: ret void
//
//
@@ -1764,29 +1846,31 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTSTRIPE_1_IV_J]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP14]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_1_IV_J]], align 4
-// CHECK2-NEXT: br label [[FOR_COND10]], !llvm.loop [[LOOP10:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND10]], !llvm.loop [[LOOP12:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC22:%.*]]
// CHECK2: for.inc22:
// CHECK2-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC23:%.*]] = add nsw i32 [[TMP15]], 1
// CHECK2-NEXT: store i32 [[INC23]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP11:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP13:![0-9]+]]
// CHECK2: for.end24:
// CHECK2-NEXT: br label [[FOR_INC25:%.*]]
// CHECK2: for.inc25:
// CHECK2-NEXT: [[TMP16:%.*]] = load i32, ptr [[DOTFLOOR_1_IV_J]], align 4
// CHECK2-NEXT: [[ADD26:%.*]] = add nsw i32 [[TMP16]], 5
// CHECK2-NEXT: store i32 [[ADD26]], ptr [[DOTFLOOR_1_IV_J]], align 4
-// CHECK2-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP12:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP14:![0-9]+]]
// CHECK2: for.end27:
// CHECK2-NEXT: br label [[FOR_INC28:%.*]]
// CHECK2: for.inc28:
// CHECK2-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK2-NEXT: [[ADD29:%.*]] = add nsw i32 [[TMP17]], 5
// CHECK2-NEXT: store i32 [[ADD29]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP13:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP15:![0-9]+]]
// CHECK2: for.end30:
+// CHECK2-NEXT: store i32 19, ptr [[I]], align 4
+// CHECK2-NEXT: store i32 19, ptr [[J]], align 4
// CHECK2-NEXT: ret void
//
//
@@ -1900,21 +1984,21 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTSTRIPE_1_IV_J]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP20]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_1_IV_J]], align 4
-// CHECK2-NEXT: br label [[FOR_COND15]], !llvm.loop [[LOOP14:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND15]], !llvm.loop [[LOOP16:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC27:%.*]]
// CHECK2: for.inc27:
// CHECK2-NEXT: [[TMP21:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC28:%.*]] = add nsw i32 [[TMP21]], 1
// CHECK2-NEXT: store i32 [[INC28]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND3]], !llvm.loop [[LOOP15:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND3]], !llvm.loop [[LOOP17:![0-9]+]]
// CHECK2: for.end29:
// CHECK2-NEXT: br label [[FOR_INC30:%.*]]
// CHECK2: for.inc30:
// CHECK2-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTFLOOR_1_IV_J]], align 4
// CHECK2-NEXT: [[ADD31:%.*]] = add nsw i32 [[TMP22]], 5
// CHECK2-NEXT: store i32 [[ADD31]], ptr [[DOTFLOOR_1_IV_J]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP16:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP18:![0-9]+]]
// CHECK2: for.end32:
// CHECK2-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK2: omp.body.continue:
@@ -2053,21 +2137,21 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTSTRIPE_1_IV_J]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP22]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_1_IV_J]], align 4
-// CHECK2-NEXT: br label [[FOR_COND20]], !llvm.loop [[LOOP17:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND20]], !llvm.loop [[LOOP19:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC32:%.*]]
// CHECK2: for.inc32:
// CHECK2-NEXT: [[TMP23:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC33:%.*]] = add nsw i32 [[TMP23]], 1
// CHECK2-NEXT: store i32 [[INC33]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND8]], !llvm.loop [[LOOP18:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND8]], !llvm.loop [[LOOP20:![0-9]+]]
// CHECK2: for.end34:
// CHECK2-NEXT: br label [[FOR_INC35:%.*]]
// CHECK2: for.inc35:
// CHECK2-NEXT: [[TMP24:%.*]] = load i32, ptr [[DOTFLOOR_1_IV_J]], align 4
// CHECK2-NEXT: [[ADD36:%.*]] = add nsw i32 [[TMP24]], 5
// CHECK2-NEXT: store i32 [[ADD36]], ptr [[DOTFLOOR_1_IV_J]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP19:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP21:![0-9]+]]
// CHECK2: for.end37:
// CHECK2-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK2: omp.body.continue:
@@ -2380,7 +2464,7 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP14]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP20:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP22:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK2: omp.body.continue:
@@ -2464,7 +2548,7 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP11:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP11]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP23:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP25:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC17:%.*]]
// CHECK2: for.inc17:
@@ -2481,8 +2565,9 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK2-NEXT: [[ADD23:%.*]] = add nsw i32 [[TMP14]], [[COND22]]
// CHECK2-NEXT: store i32 [[ADD23]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP24:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP26:![0-9]+]]
// CHECK2: for.end24:
+// CHECK2-NEXT: store i32 19, ptr [[I]], align 4
// CHECK2-NEXT: ret void
//
//
@@ -2500,14 +2585,14 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[V:%.*]] = alloca double, align 8
// CHECK2-NEXT: [[ARRAY:%.*]] = getelementptr inbounds nuw [[STRUCT_DATA_T]], ptr [[DATA]], i32 0, i32 0
// CHECK2-NEXT: store ptr [[ARRAY]], ptr [[__RANGE2]], align 8
-// CHECK2-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP0]], i64 0, i64 0
// CHECK2-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 12
// CHECK2-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[ARRAYDECAY1:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY1]], ptr [[__BEGIN2]], align 8
-// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[ARRAYDECAY2:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY2]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK2-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__END2]], align 8
@@ -2571,15 +2656,28 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP19:%.*]] = load i64, ptr [[DOTSTRIPE_0_IV___BEGIN2]], align 8
// CHECK2-NEXT: [[INC:%.*]] = add nsw i64 [[TMP19]], 1
// CHECK2-NEXT: store i64 [[INC]], ptr [[DOTSTRIPE_0_IV___BEGIN2]], align 8
-// CHECK2-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP25:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP27:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC16:%.*]]
// CHECK2: for.inc16:
// CHECK2-NEXT: [[TMP20:%.*]] = load i64, ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
// CHECK2-NEXT: [[ADD17:%.*]] = add nsw i64 [[TMP20]], 5
// CHECK2-NEXT: store i64 [[ADD17]], ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP26:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP28:![0-9]+]]
// CHECK2: for.end18:
+// CHECK2-NEXT: [[TMP21:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK2-NEXT: [[TMP22:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_3]], align 8
+// CHECK2-NEXT: [[TMP23:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK2-NEXT: [[SUB_PTR_LHS_CAST19:%.*]] = ptrtoint ptr [[TMP22]] to i64
+// CHECK2-NEXT: [[SUB_PTR_RHS_CAST20:%.*]] = ptrtoint ptr [[TMP23]] to i64
+// CHECK2-NEXT: [[SUB_PTR_SUB21:%.*]] = sub i64 [[SUB_PTR_LHS_CAST19]], [[SUB_PTR_RHS_CAST20]]
+// CHECK2-NEXT: [[SUB_PTR_DIV22:%.*]] = sdiv exact i64 [[SUB_PTR_SUB21]], 8
+// CHECK2-NEXT: [[SUB23:%.*]] = sub nsw i64 [[SUB_PTR_DIV22]], 1
+// CHECK2-NEXT: [[ADD24:%.*]] = add nsw i64 [[SUB23]], 1
+// CHECK2-NEXT: [[DIV25:%.*]] = sdiv i64 [[ADD24]], 1
+// CHECK2-NEXT: [[MUL26:%.*]] = mul nsw i64 [[DIV25]], 1
+// CHECK2-NEXT: [[ADD_PTR27:%.*]] = getelementptr inbounds double, ptr [[TMP21]], i64 [[MUL26]]
+// CHECK2-NEXT: store ptr [[ADD_PTR27]], ptr [[__BEGIN2]], align 8
// CHECK2-NEXT: ret void
//
//
@@ -2662,15 +2760,25 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP16:%.*]] = load i32, ptr [[DOTSTRIPE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add i32 [[TMP16]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTSTRIPE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP27:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP29:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC15:%.*]]
// CHECK2: for.inc15:
// CHECK2-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK2-NEXT: [[ADD16:%.*]] = add i32 [[TMP17]], 5
// CHECK2-NEXT: store i32 [[ADD16]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP28:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP30:![0-9]+]]
// CHECK2: for.end17:
+// CHECK2-NEXT: [[TMP18:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[TMP19:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK2-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[SUB18:%.*]] = sub i32 [[TMP19]], [[TMP20]]
+// CHECK2-NEXT: [[SUB19:%.*]] = sub i32 [[SUB18]], 1
+// CHECK2-NEXT: [[ADD20:%.*]] = add i32 [[SUB19]], 3
+// CHECK2-NEXT: [[DIV21:%.*]] = udiv i32 [[ADD20]], 3
+// CHECK2-NEXT: [[MUL22:%.*]] = mul i32 [[DIV21]], 3
+// CHECK2-NEXT: [[ADD23:%.*]] = add i32 [[TMP18]], [[MUL22]]
+// CHECK2-NEXT: store i32 [[ADD23]], ptr [[I]], align 4
// CHECK2-NEXT: ret void
//
//
@@ -2681,53 +2789,59 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: ret void
//
//.
-// CHECK1: [[LOOP3]] = distinct !{[[LOOP3]], [[META4:![0-9]+]]}
-// CHECK1: [[META4]] = !{!"llvm.loop.mustprogress"}
-// CHECK1: [[LOOP5]] = distinct !{[[LOOP5]], [[META4]]}
-// CHECK1: [[LOOP6]] = distinct !{[[LOOP6]], [[META4]]}
-// CHECK1: [[LOOP7]] = distinct !{[[LOOP7]], [[META4]]}
-// CHECK1: [[LOOP8]] = distinct !{[[LOOP8]], [[META4]]}
-// CHECK1: [[LOOP9]] = distinct !{[[LOOP9]], [[META4]]}
-// CHECK1: [[LOOP10]] = distinct !{[[LOOP10]], [[META4]]}
-// CHECK1: [[LOOP11]] = distinct !{[[LOOP11]], [[META4]]}
-// CHECK1: [[LOOP12]] = distinct !{[[LOOP12]], [[META4]]}
-// CHECK1: [[LOOP13]] = distinct !{[[LOOP13]], [[META4]]}
-// CHECK1: [[LOOP14]] = distinct !{[[LOOP14]], [[META4]]}
-// CHECK1: [[LOOP15]] = distinct !{[[LOOP15]], [[META4]]}
-// CHECK1: [[LOOP16]] = distinct !{[[LOOP16]], [[META4]]}
-// CHECK1: [[LOOP17]] = distinct !{[[LOOP17]], [[META4]]}
-// CHECK1: [[LOOP18]] = distinct !{[[LOOP18]], [[META4]]}
-// CHECK1: [[LOOP21]] = distinct !{[[LOOP21]], [[META4]]}
-// CHECK1: [[LOOP22]] = distinct !{[[LOOP22]], [[META4]]}
-// CHECK1: [[LOOP23]] = distinct !{[[LOOP23]], [[META4]]}
-// CHECK1: [[LOOP24]] = distinct !{[[LOOP24]], [[META4]]}
-// CHECK1: [[LOOP25]] = distinct !{[[LOOP25]], [[META4]]}
-// CHECK1: [[LOOP26]] = distinct !{[[LOOP26]], [[META4]]}
-// CHECK1: [[LOOP27]] = distinct !{[[LOOP27]], [[META4]]}
-// CHECK1: [[LOOP28]] = distinct !{[[LOOP28]], [[META4]]}
+// CHECK1: [[META2]] = !{}
+// CHECK1: [[META3]] = !{i64 4}
+// CHECK1: [[LOOP4]] = distinct !{[[LOOP4]], [[META5:![0-9]+]]}
+// CHECK1: [[META5]] = !{!"llvm.loop.mustprogress"}
+// CHECK1: [[LOOP6]] = distinct !{[[LOOP6]], [[META5]]}
+// CHECK1: [[LOOP7]] = distinct !{[[LOOP7]], [[META5]]}
+// CHECK1: [[LOOP8]] = distinct !{[[LOOP8]], [[META5]]}
+// CHECK1: [[LOOP9]] = distinct !{[[LOOP9]], [[META5]]}
+// CHECK1: [[LOOP10]] = distinct !{[[LOOP10]], [[META5]]}
+// CHECK1: [[LOOP11]] = distinct !{[[LOOP11]], [[META5]]}
+// CHECK1: [[LOOP12]] = distinct !{[[LOOP12]], [[META5]]}
+// CHECK1: [[LOOP13]] = distinct !{[[LOOP13]], [[META5]]}
+// CHECK1: [[LOOP14]] = distinct !{[[LOOP14]], [[META5]]}
+// CHECK1: [[LOOP15]] = distinct !{[[LOOP15]], [[META5]]}
+// CHECK1: [[LOOP16]] = distinct !{[[LOOP16]], [[META5]]}
+// CHECK1: [[LOOP17]] = distinct !{[[LOOP17]], [[META5]]}
+// CHECK1: [[LOOP18]] = distinct !{[[LOOP18]], [[META5]]}
+// CHECK1: [[LOOP19]] = distinct !{[[LOOP19]], [[META5]]}
+// CHECK1: [[LOOP22]] = distinct !{[[LOOP22]], [[META5]]}
+// CHECK1: [[LOOP23]] = distinct !{[[LOOP23]], [[META5]]}
+// CHECK1: [[LOOP24]] = distinct !{[[LOOP24]], [[META5]]}
+// CHECK1: [[LOOP25]] = distinct !{[[LOOP25]], [[META5]]}
+// CHECK1: [[META26]] = !{i64 8}
+// CHECK1: [[LOOP27]] = distinct !{[[LOOP27]], [[META5]]}
+// CHECK1: [[LOOP28]] = distinct !{[[LOOP28]], [[META5]]}
+// CHECK1: [[LOOP29]] = distinct !{[[LOOP29]], [[META5]]}
+// CHECK1: [[LOOP30]] = distinct !{[[LOOP30]], [[META5]]}
//.
-// CHECK2: [[LOOP3]] = distinct !{[[LOOP3]], [[META4:![0-9]+]]}
-// CHECK2: [[META4]] = !{!"llvm.loop.mustprogress"}
-// CHECK2: [[LOOP5]] = distinct !{[[LOOP5]], [[META4]]}
-// CHECK2: [[LOOP6]] = distinct !{[[LOOP6]], [[META4]]}
-// CHECK2: [[LOOP7]] = distinct !{[[LOOP7]], [[META4]]}
-// CHECK2: [[LOOP8]] = distinct !{[[LOOP8]], [[META4]]}
-// CHECK2: [[LOOP9]] = distinct !{[[LOOP9]], [[META4]]}
-// CHECK2: [[LOOP10]] = distinct !{[[LOOP10]], [[META4]]}
-// CHECK2: [[LOOP11]] = distinct !{[[LOOP11]], [[META4]]}
-// CHECK2: [[LOOP12]] = distinct !{[[LOOP12]], [[META4]]}
-// CHECK2: [[LOOP13]] = distinct !{[[LOOP13]], [[META4]]}
-// CHECK2: [[LOOP14]] = distinct !{[[LOOP14]], [[META4]]}
-// CHECK2: [[LOOP15]] = distinct !{[[LOOP15]], [[META4]]}
-// CHECK2: [[LOOP16]] = distinct !{[[LOOP16]], [[META4]]}
-// CHECK2: [[LOOP17]] = distinct !{[[LOOP17]], [[META4]]}
-// CHECK2: [[LOOP18]] = distinct !{[[LOOP18]], [[META4]]}
-// CHECK2: [[LOOP19]] = distinct !{[[LOOP19]], [[META4]]}
-// CHECK2: [[LOOP20]] = distinct !{[[LOOP20]], [[META4]]}
-// CHECK2: [[LOOP23]] = distinct !{[[LOOP23]], [[META4]]}
-// CHECK2: [[LOOP24]] = distinct !{[[LOOP24]], [[META4]]}
-// CHECK2: [[LOOP25]] = distinct !{[[LOOP25]], [[META4]]}
-// CHECK2: [[LOOP26]] = distinct !{[[LOOP26]], [[META4]]}
-// CHECK2: [[LOOP27]] = distinct !{[[LOOP27]], [[META4]]}
-// CHECK2: [[LOOP28]] = distinct !{[[LOOP28]], [[META4]]}
+// CHECK2: [[META2]] = !{}
+// CHECK2: [[META3]] = !{i64 4}
+// CHECK2: [[LOOP4]] = distinct !{[[LOOP4]], [[META5:![0-9]+]]}
+// CHECK2: [[META5]] = !{!"llvm.loop.mustprogress"}
+// CHECK2: [[LOOP6]] = distinct !{[[LOOP6]], [[META5]]}
+// CHECK2: [[LOOP7]] = distinct !{[[LOOP7]], [[META5]]}
+// CHECK2: [[LOOP8]] = distinct !{[[LOOP8]], [[META5]]}
+// CHECK2: [[META9]] = !{i64 8}
+// CHECK2: [[LOOP10]] = distinct !{[[LOOP10]], [[META5]]}
+// CHECK2: [[LOOP11]] = distinct !{[[LOOP11]], [[META5]]}
+// CHECK2: [[LOOP12]] = distinct !{[[LOOP12]], [[META5]]}
+// CHECK2: [[LOOP13]] = distinct !{[[LOOP13]], [[META5]]}
+// CHECK2: [[LOOP14]] = distinct !{[[LOOP14]], [[META5]]}
+// CHECK2: [[LOOP15]] = distinct !{[[LOOP15]], [[META5]]}
+// CHECK2: [[LOOP16]] = distinct !{[[LOOP16]], [[META5]]}
+// CHECK2: [[LOOP17]] = distinct !{[[LOOP17]], [[META5]]}
+// CHECK2: [[LOOP18]] = distinct !{[[LOOP18]], [[META5]]}
+// CHECK2: [[LOOP19]] = distinct !{[[LOOP19]], [[META5]]}
+// CHECK2: [[LOOP20]] = distinct !{[[LOOP20]], [[META5]]}
+// CHECK2: [[LOOP21]] = distinct !{[[LOOP21]], [[META5]]}
+// CHECK2: [[LOOP22]] = distinct !{[[LOOP22]], [[META5]]}
+// CHECK2: [[LOOP25]] = distinct !{[[LOOP25]], [[META5]]}
+// CHECK2: [[LOOP26]] = distinct !{[[LOOP26]], [[META5]]}
+// CHECK2: [[LOOP27]] = distinct !{[[LOOP27]], [[META5]]}
+// CHECK2: [[LOOP28]] = distinct !{[[LOOP28]], [[META5]]}
+// CHECK2: [[LOOP29]] = distinct !{[[LOOP29]], [[META5]]}
+// CHECK2: [[LOOP30]] = distinct !{[[LOOP30]], [[META5]]}
//.
diff --git a/clang/test/OpenMP/tile_codegen.cpp b/clang/test/OpenMP/tile_codegen.cpp
index 85b004dc73f4f..93159c5e96661 100644
--- a/clang/test/OpenMP/tile_codegen.cpp
+++ b/clang/test/OpenMP/tile_codegen.cpp
@@ -174,9 +174,9 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP5:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK1-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP5]], 3
// CHECK1-NEXT: [[ADD9:%.*]] = add nsw i32 7, [[MUL]]
-// CHECK1-NEXT: [[TMP6:%.*]] = load ptr, ptr [[I2]], align 8
+// CHECK1-NEXT: [[TMP6:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2:![0-9]+]], !align [[META3:![0-9]+]]
// CHECK1-NEXT: store i32 [[ADD9]], ptr [[TMP6]], align 4
-// CHECK1-NEXT: [[TMP7:%.*]] = load ptr, ptr [[I2]], align 8
+// CHECK1-NEXT: [[TMP7:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2]], !align [[META3]]
// CHECK1-NEXT: [[TMP8:%.*]] = load i32, ptr [[TMP7]], align 4
// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP8]])
// CHECK1-NEXT: br label [[FOR_INC:%.*]]
@@ -184,15 +184,17 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP9]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP3:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP4:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC10:%.*]]
// CHECK1: for.inc10:
// CHECK1-NEXT: [[TMP10:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK1-NEXT: [[ADD11:%.*]] = add nsw i32 [[TMP10]], 5
// CHECK1-NEXT: store i32 [[ADD11]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP5:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
// CHECK1: for.end12:
+// CHECK1-NEXT: [[TMP11:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2]], !align [[META3]]
+// CHECK1-NEXT: store i32 19, ptr [[TMP11]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -276,15 +278,28 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add i32 [[TMP20]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP6:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP7:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC15:%.*]]
// CHECK1: for.inc15:
// CHECK1-NEXT: [[TMP21:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK1-NEXT: [[ADD16:%.*]] = add i32 [[TMP21]], 5
// CHECK1-NEXT: store i32 [[ADD16]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP7:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
// CHECK1: for.end17:
+// CHECK1-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[TMP23:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK1-NEXT: [[TMP24:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[SUB18:%.*]] = sub i32 [[TMP23]], [[TMP24]]
+// CHECK1-NEXT: [[SUB19:%.*]] = sub i32 [[SUB18]], 1
+// CHECK1-NEXT: [[TMP25:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK1-NEXT: [[ADD20:%.*]] = add i32 [[SUB19]], [[TMP25]]
+// CHECK1-NEXT: [[TMP26:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK1-NEXT: [[DIV21:%.*]] = udiv i32 [[ADD20]], [[TMP26]]
+// CHECK1-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK1-NEXT: [[MUL22:%.*]] = mul i32 [[DIV21]], [[TMP27]]
+// CHECK1-NEXT: [[ADD23:%.*]] = add i32 [[TMP22]], [[MUL22]]
+// CHECK1-NEXT: store i32 [[ADD23]], ptr [[I]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -375,29 +390,31 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTTILE_1_IV_J]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP14]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTTILE_1_IV_J]], align 4
-// CHECK1-NEXT: br label [[FOR_COND10]], !llvm.loop [[LOOP8:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND10]], !llvm.loop [[LOOP9:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC22:%.*]]
// CHECK1: for.inc22:
// CHECK1-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC23:%.*]] = add nsw i32 [[TMP15]], 1
// CHECK1-NEXT: store i32 [[INC23]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP9:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP10:![0-9]+]]
// CHECK1: for.end24:
// CHECK1-NEXT: br label [[FOR_INC25:%.*]]
// CHECK1: for.inc25:
// CHECK1-NEXT: [[TMP16:%.*]] = load i32, ptr [[DOTFLOOR_1_IV_J]], align 4
// CHECK1-NEXT: [[ADD26:%.*]] = add nsw i32 [[TMP16]], 5
// CHECK1-NEXT: store i32 [[ADD26]], ptr [[DOTFLOOR_1_IV_J]], align 4
-// CHECK1-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP10:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP11:![0-9]+]]
// CHECK1: for.end27:
// CHECK1-NEXT: br label [[FOR_INC28:%.*]]
// CHECK1: for.inc28:
// CHECK1-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK1-NEXT: [[ADD29:%.*]] = add nsw i32 [[TMP17]], 5
// CHECK1-NEXT: store i32 [[ADD29]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP11:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP12:![0-9]+]]
// CHECK1: for.end30:
+// CHECK1-NEXT: store i32 19, ptr [[I]], align 4
+// CHECK1-NEXT: store i32 19, ptr [[J]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -511,21 +528,21 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTTILE_1_IV_J]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP20]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTTILE_1_IV_J]], align 4
-// CHECK1-NEXT: br label [[FOR_COND15]], !llvm.loop [[LOOP12:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND15]], !llvm.loop [[LOOP13:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC27:%.*]]
// CHECK1: for.inc27:
// CHECK1-NEXT: [[TMP21:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC28:%.*]] = add nsw i32 [[TMP21]], 1
// CHECK1-NEXT: store i32 [[INC28]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND3]], !llvm.loop [[LOOP13:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND3]], !llvm.loop [[LOOP14:![0-9]+]]
// CHECK1: for.end29:
// CHECK1-NEXT: br label [[FOR_INC30:%.*]]
// CHECK1: for.inc30:
// CHECK1-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTFLOOR_1_IV_J]], align 4
// CHECK1-NEXT: [[ADD31:%.*]] = add nsw i32 [[TMP22]], 5
// CHECK1-NEXT: store i32 [[ADD31]], ptr [[DOTFLOOR_1_IV_J]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP14:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP15:![0-9]+]]
// CHECK1: for.end32:
// CHECK1-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK1: omp.body.continue:
@@ -664,21 +681,21 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTTILE_1_IV_J]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP22]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTTILE_1_IV_J]], align 4
-// CHECK1-NEXT: br label [[FOR_COND20]], !llvm.loop [[LOOP15:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND20]], !llvm.loop [[LOOP16:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC32:%.*]]
// CHECK1: for.inc32:
// CHECK1-NEXT: [[TMP23:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC33:%.*]] = add nsw i32 [[TMP23]], 1
// CHECK1-NEXT: store i32 [[INC33]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND8]], !llvm.loop [[LOOP16:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND8]], !llvm.loop [[LOOP17:![0-9]+]]
// CHECK1: for.end34:
// CHECK1-NEXT: br label [[FOR_INC35:%.*]]
// CHECK1: for.inc35:
// CHECK1-NEXT: [[TMP24:%.*]] = load i32, ptr [[DOTFLOOR_1_IV_J]], align 4
// CHECK1-NEXT: [[ADD36:%.*]] = add nsw i32 [[TMP24]], 5
// CHECK1-NEXT: store i32 [[ADD36]], ptr [[DOTFLOOR_1_IV_J]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP17:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP18:![0-9]+]]
// CHECK1: for.end37:
// CHECK1-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK1: omp.body.continue:
@@ -991,7 +1008,7 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP14]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP18:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP19:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK1: omp.body.continue:
@@ -1087,15 +1104,25 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP16:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add i32 [[TMP16]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP21:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP22:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC15:%.*]]
// CHECK1: for.inc15:
// CHECK1-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK1-NEXT: [[ADD16:%.*]] = add i32 [[TMP17]], 5
// CHECK1-NEXT: store i32 [[ADD16]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP22:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP23:![0-9]+]]
// CHECK1: for.end17:
+// CHECK1-NEXT: [[TMP18:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[TMP19:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK1-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[SUB18:%.*]] = sub i32 [[TMP19]], [[TMP20]]
+// CHECK1-NEXT: [[SUB19:%.*]] = sub i32 [[SUB18]], 1
+// CHECK1-NEXT: [[ADD20:%.*]] = add i32 [[SUB19]], 3
+// CHECK1-NEXT: [[DIV21:%.*]] = udiv i32 [[ADD20]], 3
+// CHECK1-NEXT: [[MUL22:%.*]] = mul i32 [[DIV21]], 3
+// CHECK1-NEXT: [[ADD23:%.*]] = add i32 [[TMP18]], [[MUL22]]
+// CHECK1-NEXT: store i32 [[ADD23]], ptr [[I]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -1166,7 +1193,7 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP11:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP11]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP23:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP24:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC17:%.*]]
// CHECK1: for.inc17:
@@ -1183,8 +1210,9 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK1-NEXT: [[ADD23:%.*]] = add nsw i32 [[TMP14]], [[COND22]]
// CHECK1-NEXT: store i32 [[ADD23]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP24:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP25:![0-9]+]]
// CHECK1: for.end24:
+// CHECK1-NEXT: store i32 19, ptr [[I]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -1202,14 +1230,14 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[V:%.*]] = alloca double, align 8
// CHECK1-NEXT: [[ARRAY:%.*]] = getelementptr inbounds nuw [[STRUCT_DATA_T]], ptr [[DATA]], i32 0, i32 0
// CHECK1-NEXT: store ptr [[ARRAY]], ptr [[__RANGE2]], align 8
-// CHECK1-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META26:![0-9]+]]
// CHECK1-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP0]], i64 0, i64 0
// CHECK1-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 12
// CHECK1-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META26]]
// CHECK1-NEXT: [[ARRAYDECAY1:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY1]], ptr [[__BEGIN2]], align 8
-// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META26]]
// CHECK1-NEXT: [[ARRAYDECAY2:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY2]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK1-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__END2]], align 8
@@ -1273,15 +1301,28 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP19:%.*]] = load i64, ptr [[DOTTILE_0_IV___BEGIN2]], align 8
// CHECK1-NEXT: [[INC:%.*]] = add nsw i64 [[TMP19]], 1
// CHECK1-NEXT: store i64 [[INC]], ptr [[DOTTILE_0_IV___BEGIN2]], align 8
-// CHECK1-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP25:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP27:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC16:%.*]]
// CHECK1: for.inc16:
// CHECK1-NEXT: [[TMP20:%.*]] = load i64, ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
// CHECK1-NEXT: [[ADD17:%.*]] = add nsw i64 [[TMP20]], 5
// CHECK1-NEXT: store i64 [[ADD17]], ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP26:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP28:![0-9]+]]
// CHECK1: for.end18:
+// CHECK1-NEXT: [[TMP21:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK1-NEXT: [[TMP22:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_3]], align 8
+// CHECK1-NEXT: [[TMP23:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK1-NEXT: [[SUB_PTR_LHS_CAST19:%.*]] = ptrtoint ptr [[TMP22]] to i64
+// CHECK1-NEXT: [[SUB_PTR_RHS_CAST20:%.*]] = ptrtoint ptr [[TMP23]] to i64
+// CHECK1-NEXT: [[SUB_PTR_SUB21:%.*]] = sub i64 [[SUB_PTR_LHS_CAST19]], [[SUB_PTR_RHS_CAST20]]
+// CHECK1-NEXT: [[SUB_PTR_DIV22:%.*]] = sdiv exact i64 [[SUB_PTR_SUB21]], 8
+// CHECK1-NEXT: [[SUB23:%.*]] = sub nsw i64 [[SUB_PTR_DIV22]], 1
+// CHECK1-NEXT: [[ADD24:%.*]] = add nsw i64 [[SUB23]], 1
+// CHECK1-NEXT: [[DIV25:%.*]] = sdiv i64 [[ADD24]], 1
+// CHECK1-NEXT: [[MUL26:%.*]] = mul nsw i64 [[DIV25]], 1
+// CHECK1-NEXT: [[ADD_PTR27:%.*]] = getelementptr inbounds double, ptr [[TMP21]], i64 [[MUL26]]
+// CHECK1-NEXT: store ptr [[ADD_PTR27]], ptr [[__BEGIN2]], align 8
// CHECK1-NEXT: ret void
//
//
@@ -1301,14 +1342,14 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: store double 4.200000e+01, ptr [[C]], align 8
// CHECK1-NEXT: [[ARRAY:%.*]] = getelementptr inbounds nuw [[STRUCT_DATA_T]], ptr [[DATA]], i32 0, i32 0
// CHECK1-NEXT: store ptr [[ARRAY]], ptr [[__RANGE2]], align 8
-// CHECK1-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META26]]
// CHECK1-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP0]], i64 0, i64 0
// CHECK1-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 12
// CHECK1-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META26]]
// CHECK1-NEXT: [[ARRAYDECAY1:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY1]], ptr [[__BEGIN2]], align 8
-// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META26]]
// CHECK1-NEXT: [[ARRAYDECAY2:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY2]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK1-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__END2]], align 8
@@ -1373,15 +1414,28 @@ extern "C" void foo10(data_t data) {
// CHECK1-NEXT: [[TMP20:%.*]] = load i64, ptr [[DOTTILE_0_IV___BEGIN2]], align 8
// CHECK1-NEXT: [[INC:%.*]] = add nsw i64 [[TMP20]], 1
// CHECK1-NEXT: store i64 [[INC]], ptr [[DOTTILE_0_IV___BEGIN2]], align 8
-// CHECK1-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP27:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP29:![0-9]+]]
// CHECK1: for.end:
// CHECK1-NEXT: br label [[FOR_INC16:%.*]]
// CHECK1: for.inc16:
// CHECK1-NEXT: [[TMP21:%.*]] = load i64, ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
// CHECK1-NEXT: [[ADD17:%.*]] = add nsw i64 [[TMP21]], 5
// CHECK1-NEXT: store i64 [[ADD17]], ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
-// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP28:![0-9]+]]
+// CHECK1-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP30:![0-9]+]]
// CHECK1: for.end18:
+// CHECK1-NEXT: [[TMP22:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK1-NEXT: [[TMP23:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_3]], align 8
+// CHECK1-NEXT: [[TMP24:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK1-NEXT: [[SUB_PTR_LHS_CAST19:%.*]] = ptrtoint ptr [[TMP23]] to i64
+// CHECK1-NEXT: [[SUB_PTR_RHS_CAST20:%.*]] = ptrtoint ptr [[TMP24]] to i64
+// CHECK1-NEXT: [[SUB_PTR_SUB21:%.*]] = sub i64 [[SUB_PTR_LHS_CAST19]], [[SUB_PTR_RHS_CAST20]]
+// CHECK1-NEXT: [[SUB_PTR_DIV22:%.*]] = sdiv exact i64 [[SUB_PTR_SUB21]], 8
+// CHECK1-NEXT: [[SUB23:%.*]] = sub nsw i64 [[SUB_PTR_DIV22]], 1
+// CHECK1-NEXT: [[ADD24:%.*]] = add nsw i64 [[SUB23]], 1
+// CHECK1-NEXT: [[DIV25:%.*]] = sdiv i64 [[ADD24]], 1
+// CHECK1-NEXT: [[MUL26:%.*]] = mul nsw i64 [[DIV25]], 1
+// CHECK1-NEXT: [[ADD_PTR27:%.*]] = getelementptr inbounds double, ptr [[TMP22]], i64 [[MUL26]]
+// CHECK1-NEXT: store ptr [[ADD_PTR27]], ptr [[__BEGIN2]], align 8
// CHECK1-NEXT: ret void
//
//
@@ -1452,9 +1506,9 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP5:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK2-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP5]], 3
// CHECK2-NEXT: [[ADD9:%.*]] = add nsw i32 7, [[MUL]]
-// CHECK2-NEXT: [[TMP6:%.*]] = load ptr, ptr [[I2]], align 8
+// CHECK2-NEXT: [[TMP6:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2:![0-9]+]], !align [[META3:![0-9]+]]
// CHECK2-NEXT: store i32 [[ADD9]], ptr [[TMP6]], align 4
-// CHECK2-NEXT: [[TMP7:%.*]] = load ptr, ptr [[I2]], align 8
+// CHECK2-NEXT: [[TMP7:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2]], !align [[META3]]
// CHECK2-NEXT: [[TMP8:%.*]] = load i32, ptr [[TMP7]], align 4
// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP8]])
// CHECK2-NEXT: br label [[FOR_INC:%.*]]
@@ -1462,15 +1516,17 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP9]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP3:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP4:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC10:%.*]]
// CHECK2: for.inc10:
// CHECK2-NEXT: [[TMP10:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK2-NEXT: [[ADD11:%.*]] = add nsw i32 [[TMP10]], 5
// CHECK2-NEXT: store i32 [[ADD11]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP5:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
// CHECK2: for.end12:
+// CHECK2-NEXT: [[TMP11:%.*]] = load ptr, ptr [[I2]], align 8, !nonnull [[META2]], !align [[META3]]
+// CHECK2-NEXT: store i32 19, ptr [[TMP11]], align 4
// CHECK2-NEXT: ret void
//
//
@@ -1560,15 +1616,28 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add i32 [[TMP20]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP6:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP7:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC15:%.*]]
// CHECK2: for.inc15:
// CHECK2-NEXT: [[TMP21:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK2-NEXT: [[ADD16:%.*]] = add i32 [[TMP21]], 5
// CHECK2-NEXT: store i32 [[ADD16]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP7:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
// CHECK2: for.end17:
+// CHECK2-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[TMP23:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK2-NEXT: [[TMP24:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[SUB18:%.*]] = sub i32 [[TMP23]], [[TMP24]]
+// CHECK2-NEXT: [[SUB19:%.*]] = sub i32 [[SUB18]], 1
+// CHECK2-NEXT: [[TMP25:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK2-NEXT: [[ADD20:%.*]] = add i32 [[SUB19]], [[TMP25]]
+// CHECK2-NEXT: [[TMP26:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK2-NEXT: [[DIV21:%.*]] = udiv i32 [[ADD20]], [[TMP26]]
+// CHECK2-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK2-NEXT: [[MUL22:%.*]] = mul i32 [[DIV21]], [[TMP27]]
+// CHECK2-NEXT: [[ADD23:%.*]] = add i32 [[TMP22]], [[MUL22]]
+// CHECK2-NEXT: store i32 [[ADD23]], ptr [[I]], align 4
// CHECK2-NEXT: ret void
//
//
@@ -1588,14 +1657,14 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: store double 4.200000e+01, ptr [[C]], align 8
// CHECK2-NEXT: [[ARRAY:%.*]] = getelementptr inbounds nuw [[STRUCT_DATA_T]], ptr [[DATA]], i32 0, i32 0
// CHECK2-NEXT: store ptr [[ARRAY]], ptr [[__RANGE2]], align 8
-// CHECK2-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META9:![0-9]+]]
// CHECK2-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP0]], i64 0, i64 0
// CHECK2-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 12
// CHECK2-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[ARRAYDECAY1:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY1]], ptr [[__BEGIN2]], align 8
-// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[ARRAYDECAY2:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY2]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK2-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__END2]], align 8
@@ -1660,15 +1729,28 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP20:%.*]] = load i64, ptr [[DOTTILE_0_IV___BEGIN2]], align 8
// CHECK2-NEXT: [[INC:%.*]] = add nsw i64 [[TMP20]], 1
// CHECK2-NEXT: store i64 [[INC]], ptr [[DOTTILE_0_IV___BEGIN2]], align 8
-// CHECK2-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP8:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP10:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC16:%.*]]
// CHECK2: for.inc16:
// CHECK2-NEXT: [[TMP21:%.*]] = load i64, ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
// CHECK2-NEXT: [[ADD17:%.*]] = add nsw i64 [[TMP21]], 5
// CHECK2-NEXT: store i64 [[ADD17]], ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP9:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP11:![0-9]+]]
// CHECK2: for.end18:
+// CHECK2-NEXT: [[TMP22:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK2-NEXT: [[TMP23:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_3]], align 8
+// CHECK2-NEXT: [[TMP24:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK2-NEXT: [[SUB_PTR_LHS_CAST19:%.*]] = ptrtoint ptr [[TMP23]] to i64
+// CHECK2-NEXT: [[SUB_PTR_RHS_CAST20:%.*]] = ptrtoint ptr [[TMP24]] to i64
+// CHECK2-NEXT: [[SUB_PTR_SUB21:%.*]] = sub i64 [[SUB_PTR_LHS_CAST19]], [[SUB_PTR_RHS_CAST20]]
+// CHECK2-NEXT: [[SUB_PTR_DIV22:%.*]] = sdiv exact i64 [[SUB_PTR_SUB21]], 8
+// CHECK2-NEXT: [[SUB23:%.*]] = sub nsw i64 [[SUB_PTR_DIV22]], 1
+// CHECK2-NEXT: [[ADD24:%.*]] = add nsw i64 [[SUB23]], 1
+// CHECK2-NEXT: [[DIV25:%.*]] = sdiv i64 [[ADD24]], 1
+// CHECK2-NEXT: [[MUL26:%.*]] = mul nsw i64 [[DIV25]], 1
+// CHECK2-NEXT: [[ADD_PTR27:%.*]] = getelementptr inbounds double, ptr [[TMP22]], i64 [[MUL26]]
+// CHECK2-NEXT: store ptr [[ADD_PTR27]], ptr [[__BEGIN2]], align 8
// CHECK2-NEXT: ret void
//
//
@@ -1759,29 +1841,31 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTTILE_1_IV_J]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP14]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTTILE_1_IV_J]], align 4
-// CHECK2-NEXT: br label [[FOR_COND10]], !llvm.loop [[LOOP10:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND10]], !llvm.loop [[LOOP12:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC22:%.*]]
// CHECK2: for.inc22:
// CHECK2-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC23:%.*]] = add nsw i32 [[TMP15]], 1
// CHECK2-NEXT: store i32 [[INC23]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP11:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND4]], !llvm.loop [[LOOP13:![0-9]+]]
// CHECK2: for.end24:
// CHECK2-NEXT: br label [[FOR_INC25:%.*]]
// CHECK2: for.inc25:
// CHECK2-NEXT: [[TMP16:%.*]] = load i32, ptr [[DOTFLOOR_1_IV_J]], align 4
// CHECK2-NEXT: [[ADD26:%.*]] = add nsw i32 [[TMP16]], 5
// CHECK2-NEXT: store i32 [[ADD26]], ptr [[DOTFLOOR_1_IV_J]], align 4
-// CHECK2-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP12:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP14:![0-9]+]]
// CHECK2: for.end27:
// CHECK2-NEXT: br label [[FOR_INC28:%.*]]
// CHECK2: for.inc28:
// CHECK2-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK2-NEXT: [[ADD29:%.*]] = add nsw i32 [[TMP17]], 5
// CHECK2-NEXT: store i32 [[ADD29]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP13:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP15:![0-9]+]]
// CHECK2: for.end30:
+// CHECK2-NEXT: store i32 19, ptr [[I]], align 4
+// CHECK2-NEXT: store i32 19, ptr [[J]], align 4
// CHECK2-NEXT: ret void
//
//
@@ -1895,21 +1979,21 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTTILE_1_IV_J]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP20]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTTILE_1_IV_J]], align 4
-// CHECK2-NEXT: br label [[FOR_COND15]], !llvm.loop [[LOOP14:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND15]], !llvm.loop [[LOOP16:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC27:%.*]]
// CHECK2: for.inc27:
// CHECK2-NEXT: [[TMP21:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC28:%.*]] = add nsw i32 [[TMP21]], 1
// CHECK2-NEXT: store i32 [[INC28]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND3]], !llvm.loop [[LOOP15:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND3]], !llvm.loop [[LOOP17:![0-9]+]]
// CHECK2: for.end29:
// CHECK2-NEXT: br label [[FOR_INC30:%.*]]
// CHECK2: for.inc30:
// CHECK2-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTFLOOR_1_IV_J]], align 4
// CHECK2-NEXT: [[ADD31:%.*]] = add nsw i32 [[TMP22]], 5
// CHECK2-NEXT: store i32 [[ADD31]], ptr [[DOTFLOOR_1_IV_J]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP16:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP18:![0-9]+]]
// CHECK2: for.end32:
// CHECK2-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK2: omp.body.continue:
@@ -2048,21 +2132,21 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTTILE_1_IV_J]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP22]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTTILE_1_IV_J]], align 4
-// CHECK2-NEXT: br label [[FOR_COND20]], !llvm.loop [[LOOP17:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND20]], !llvm.loop [[LOOP19:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC32:%.*]]
// CHECK2: for.inc32:
// CHECK2-NEXT: [[TMP23:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC33:%.*]] = add nsw i32 [[TMP23]], 1
// CHECK2-NEXT: store i32 [[INC33]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND8]], !llvm.loop [[LOOP18:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND8]], !llvm.loop [[LOOP20:![0-9]+]]
// CHECK2: for.end34:
// CHECK2-NEXT: br label [[FOR_INC35:%.*]]
// CHECK2: for.inc35:
// CHECK2-NEXT: [[TMP24:%.*]] = load i32, ptr [[DOTFLOOR_1_IV_J]], align 4
// CHECK2-NEXT: [[ADD36:%.*]] = add nsw i32 [[TMP24]], 5
// CHECK2-NEXT: store i32 [[ADD36]], ptr [[DOTFLOOR_1_IV_J]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP19:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP21:![0-9]+]]
// CHECK2: for.end37:
// CHECK2-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK2: omp.body.continue:
@@ -2375,7 +2459,7 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP14]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP20:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP22:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK2: omp.body.continue:
@@ -2459,7 +2543,7 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP11:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP11]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP23:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND1]], !llvm.loop [[LOOP25:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC17:%.*]]
// CHECK2: for.inc17:
@@ -2476,8 +2560,9 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK2-NEXT: [[ADD23:%.*]] = add nsw i32 [[TMP14]], [[COND22]]
// CHECK2-NEXT: store i32 [[ADD23]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP24:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP26:![0-9]+]]
// CHECK2: for.end24:
+// CHECK2-NEXT: store i32 19, ptr [[I]], align 4
// CHECK2-NEXT: ret void
//
//
@@ -2495,14 +2580,14 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[V:%.*]] = alloca double, align 8
// CHECK2-NEXT: [[ARRAY:%.*]] = getelementptr inbounds nuw [[STRUCT_DATA_T]], ptr [[DATA]], i32 0, i32 0
// CHECK2-NEXT: store ptr [[ARRAY]], ptr [[__RANGE2]], align 8
-// CHECK2-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP0:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP0]], i64 0, i64 0
// CHECK2-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 12
// CHECK2-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP1:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[ARRAYDECAY1:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP1]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY1]], ptr [[__BEGIN2]], align 8
-// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP2:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META2]], !align [[META9]]
// CHECK2-NEXT: [[ARRAYDECAY2:%.*]] = getelementptr inbounds [12 x double], ptr [[TMP2]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY2]], ptr [[DOTCAPTURE_EXPR_]], align 8
// CHECK2-NEXT: [[TMP3:%.*]] = load ptr, ptr [[__END2]], align 8
@@ -2566,15 +2651,28 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP19:%.*]] = load i64, ptr [[DOTTILE_0_IV___BEGIN2]], align 8
// CHECK2-NEXT: [[INC:%.*]] = add nsw i64 [[TMP19]], 1
// CHECK2-NEXT: store i64 [[INC]], ptr [[DOTTILE_0_IV___BEGIN2]], align 8
-// CHECK2-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP25:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND7]], !llvm.loop [[LOOP27:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC16:%.*]]
// CHECK2: for.inc16:
// CHECK2-NEXT: [[TMP20:%.*]] = load i64, ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
// CHECK2-NEXT: [[ADD17:%.*]] = add nsw i64 [[TMP20]], 5
// CHECK2-NEXT: store i64 [[ADD17]], ptr [[DOTFLOOR_0_IV___BEGIN2]], align 8
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP26:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP28:![0-9]+]]
// CHECK2: for.end18:
+// CHECK2-NEXT: [[TMP21:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK2-NEXT: [[TMP22:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_3]], align 8
+// CHECK2-NEXT: [[TMP23:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_]], align 8
+// CHECK2-NEXT: [[SUB_PTR_LHS_CAST19:%.*]] = ptrtoint ptr [[TMP22]] to i64
+// CHECK2-NEXT: [[SUB_PTR_RHS_CAST20:%.*]] = ptrtoint ptr [[TMP23]] to i64
+// CHECK2-NEXT: [[SUB_PTR_SUB21:%.*]] = sub i64 [[SUB_PTR_LHS_CAST19]], [[SUB_PTR_RHS_CAST20]]
+// CHECK2-NEXT: [[SUB_PTR_DIV22:%.*]] = sdiv exact i64 [[SUB_PTR_SUB21]], 8
+// CHECK2-NEXT: [[SUB23:%.*]] = sub nsw i64 [[SUB_PTR_DIV22]], 1
+// CHECK2-NEXT: [[ADD24:%.*]] = add nsw i64 [[SUB23]], 1
+// CHECK2-NEXT: [[DIV25:%.*]] = sdiv i64 [[ADD24]], 1
+// CHECK2-NEXT: [[MUL26:%.*]] = mul nsw i64 [[DIV25]], 1
+// CHECK2-NEXT: [[ADD_PTR27:%.*]] = getelementptr inbounds double, ptr [[TMP21]], i64 [[MUL26]]
+// CHECK2-NEXT: store ptr [[ADD_PTR27]], ptr [[__BEGIN2]], align 8
// CHECK2-NEXT: ret void
//
//
@@ -2657,15 +2755,25 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: [[TMP16:%.*]] = load i32, ptr [[DOTTILE_0_IV_I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add i32 [[TMP16]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTTILE_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP27:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND6]], !llvm.loop [[LOOP29:![0-9]+]]
// CHECK2: for.end:
// CHECK2-NEXT: br label [[FOR_INC15:%.*]]
// CHECK2: for.inc15:
// CHECK2-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTFLOOR_0_IV_I]], align 4
// CHECK2-NEXT: [[ADD16:%.*]] = add i32 [[TMP17]], 5
// CHECK2-NEXT: store i32 [[ADD16]], ptr [[DOTFLOOR_0_IV_I]], align 4
-// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP28:![0-9]+]]
+// CHECK2-NEXT: br label [[FOR_COND]], !llvm.loop [[LOOP30:![0-9]+]]
// CHECK2: for.end17:
+// CHECK2-NEXT: [[TMP18:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[TMP19:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK2-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[SUB18:%.*]] = sub i32 [[TMP19]], [[TMP20]]
+// CHECK2-NEXT: [[SUB19:%.*]] = sub i32 [[SUB18]], 1
+// CHECK2-NEXT: [[ADD20:%.*]] = add i32 [[SUB19]], 3
+// CHECK2-NEXT: [[DIV21:%.*]] = udiv i32 [[ADD20]], 3
+// CHECK2-NEXT: [[MUL22:%.*]] = mul i32 [[DIV21]], 3
+// CHECK2-NEXT: [[ADD23:%.*]] = add i32 [[TMP18]], [[MUL22]]
+// CHECK2-NEXT: store i32 [[ADD23]], ptr [[I]], align 4
// CHECK2-NEXT: ret void
//
//
@@ -2676,53 +2784,59 @@ extern "C" void foo10(data_t data) {
// CHECK2-NEXT: ret void
//
//.
-// CHECK1: [[LOOP3]] = distinct !{[[LOOP3]], [[META4:![0-9]+]]}
-// CHECK1: [[META4]] = !{!"llvm.loop.mustprogress"}
-// CHECK1: [[LOOP5]] = distinct !{[[LOOP5]], [[META4]]}
-// CHECK1: [[LOOP6]] = distinct !{[[LOOP6]], [[META4]]}
-// CHECK1: [[LOOP7]] = distinct !{[[LOOP7]], [[META4]]}
-// CHECK1: [[LOOP8]] = distinct !{[[LOOP8]], [[META4]]}
-// CHECK1: [[LOOP9]] = distinct !{[[LOOP9]], [[META4]]}
-// CHECK1: [[LOOP10]] = distinct !{[[LOOP10]], [[META4]]}
-// CHECK1: [[LOOP11]] = distinct !{[[LOOP11]], [[META4]]}
-// CHECK1: [[LOOP12]] = distinct !{[[LOOP12]], [[META4]]}
-// CHECK1: [[LOOP13]] = distinct !{[[LOOP13]], [[META4]]}
-// CHECK1: [[LOOP14]] = distinct !{[[LOOP14]], [[META4]]}
-// CHECK1: [[LOOP15]] = distinct !{[[LOOP15]], [[META4]]}
-// CHECK1: [[LOOP16]] = distinct !{[[LOOP16]], [[META4]]}
-// CHECK1: [[LOOP17]] = distinct !{[[LOOP17]], [[META4]]}
-// CHECK1: [[LOOP18]] = distinct !{[[LOOP18]], [[META4]]}
-// CHECK1: [[LOOP21]] = distinct !{[[LOOP21]], [[META4]]}
-// CHECK1: [[LOOP22]] = distinct !{[[LOOP22]], [[META4]]}
-// CHECK1: [[LOOP23]] = distinct !{[[LOOP23]], [[META4]]}
-// CHECK1: [[LOOP24]] = distinct !{[[LOOP24]], [[META4]]}
-// CHECK1: [[LOOP25]] = distinct !{[[LOOP25]], [[META4]]}
-// CHECK1: [[LOOP26]] = distinct !{[[LOOP26]], [[META4]]}
-// CHECK1: [[LOOP27]] = distinct !{[[LOOP27]], [[META4]]}
-// CHECK1: [[LOOP28]] = distinct !{[[LOOP28]], [[META4]]}
+// CHECK1: [[META2]] = !{}
+// CHECK1: [[META3]] = !{i64 4}
+// CHECK1: [[LOOP4]] = distinct !{[[LOOP4]], [[META5:![0-9]+]]}
+// CHECK1: [[META5]] = !{!"llvm.loop.mustprogress"}
+// CHECK1: [[LOOP6]] = distinct !{[[LOOP6]], [[META5]]}
+// CHECK1: [[LOOP7]] = distinct !{[[LOOP7]], [[META5]]}
+// CHECK1: [[LOOP8]] = distinct !{[[LOOP8]], [[META5]]}
+// CHECK1: [[LOOP9]] = distinct !{[[LOOP9]], [[META5]]}
+// CHECK1: [[LOOP10]] = distinct !{[[LOOP10]], [[META5]]}
+// CHECK1: [[LOOP11]] = distinct !{[[LOOP11]], [[META5]]}
+// CHECK1: [[LOOP12]] = distinct !{[[LOOP12]], [[META5]]}
+// CHECK1: [[LOOP13]] = distinct !{[[LOOP13]], [[META5]]}
+// CHECK1: [[LOOP14]] = distinct !{[[LOOP14]], [[META5]]}
+// CHECK1: [[LOOP15]] = distinct !{[[LOOP15]], [[META5]]}
+// CHECK1: [[LOOP16]] = distinct !{[[LOOP16]], [[META5]]}
+// CHECK1: [[LOOP17]] = distinct !{[[LOOP17]], [[META5]]}
+// CHECK1: [[LOOP18]] = distinct !{[[LOOP18]], [[META5]]}
+// CHECK1: [[LOOP19]] = distinct !{[[LOOP19]], [[META5]]}
+// CHECK1: [[LOOP22]] = distinct !{[[LOOP22]], [[META5]]}
+// CHECK1: [[LOOP23]] = distinct !{[[LOOP23]], [[META5]]}
+// CHECK1: [[LOOP24]] = distinct !{[[LOOP24]], [[META5]]}
+// CHECK1: [[LOOP25]] = distinct !{[[LOOP25]], [[META5]]}
+// CHECK1: [[META26]] = !{i64 8}
+// CHECK1: [[LOOP27]] = distinct !{[[LOOP27]], [[META5]]}
+// CHECK1: [[LOOP28]] = distinct !{[[LOOP28]], [[META5]]}
+// CHECK1: [[LOOP29]] = distinct !{[[LOOP29]], [[META5]]}
+// CHECK1: [[LOOP30]] = distinct !{[[LOOP30]], [[META5]]}
//.
-// CHECK2: [[LOOP3]] = distinct !{[[LOOP3]], [[META4:![0-9]+]]}
-// CHECK2: [[META4]] = !{!"llvm.loop.mustprogress"}
-// CHECK2: [[LOOP5]] = distinct !{[[LOOP5]], [[META4]]}
-// CHECK2: [[LOOP6]] = distinct !{[[LOOP6]], [[META4]]}
-// CHECK2: [[LOOP7]] = distinct !{[[LOOP7]], [[META4]]}
-// CHECK2: [[LOOP8]] = distinct !{[[LOOP8]], [[META4]]}
-// CHECK2: [[LOOP9]] = distinct !{[[LOOP9]], [[META4]]}
-// CHECK2: [[LOOP10]] = distinct !{[[LOOP10]], [[META4]]}
-// CHECK2: [[LOOP11]] = distinct !{[[LOOP11]], [[META4]]}
-// CHECK2: [[LOOP12]] = distinct !{[[LOOP12]], [[META4]]}
-// CHECK2: [[LOOP13]] = distinct !{[[LOOP13]], [[META4]]}
-// CHECK2: [[LOOP14]] = distinct !{[[LOOP14]], [[META4]]}
-// CHECK2: [[LOOP15]] = distinct !{[[LOOP15]], [[META4]]}
-// CHECK2: [[LOOP16]] = distinct !{[[LOOP16]], [[META4]]}
-// CHECK2: [[LOOP17]] = distinct !{[[LOOP17]], [[META4]]}
-// CHECK2: [[LOOP18]] = distinct !{[[LOOP18]], [[META4]]}
-// CHECK2: [[LOOP19]] = distinct !{[[LOOP19]], [[META4]]}
-// CHECK2: [[LOOP20]] = distinct !{[[LOOP20]], [[META4]]}
-// CHECK2: [[LOOP23]] = distinct !{[[LOOP23]], [[META4]]}
-// CHECK2: [[LOOP24]] = distinct !{[[LOOP24]], [[META4]]}
-// CHECK2: [[LOOP25]] = distinct !{[[LOOP25]], [[META4]]}
-// CHECK2: [[LOOP26]] = distinct !{[[LOOP26]], [[META4]]}
-// CHECK2: [[LOOP27]] = distinct !{[[LOOP27]], [[META4]]}
-// CHECK2: [[LOOP28]] = distinct !{[[LOOP28]], [[META4]]}
+// CHECK2: [[META2]] = !{}
+// CHECK2: [[META3]] = !{i64 4}
+// CHECK2: [[LOOP4]] = distinct !{[[LOOP4]], [[META5:![0-9]+]]}
+// CHECK2: [[META5]] = !{!"llvm.loop.mustprogress"}
+// CHECK2: [[LOOP6]] = distinct !{[[LOOP6]], [[META5]]}
+// CHECK2: [[LOOP7]] = distinct !{[[LOOP7]], [[META5]]}
+// CHECK2: [[LOOP8]] = distinct !{[[LOOP8]], [[META5]]}
+// CHECK2: [[META9]] = !{i64 8}
+// CHECK2: [[LOOP10]] = distinct !{[[LOOP10]], [[META5]]}
+// CHECK2: [[LOOP11]] = distinct !{[[LOOP11]], [[META5]]}
+// CHECK2: [[LOOP12]] = distinct !{[[LOOP12]], [[META5]]}
+// CHECK2: [[LOOP13]] = distinct !{[[LOOP13]], [[META5]]}
+// CHECK2: [[LOOP14]] = distinct !{[[LOOP14]], [[META5]]}
+// CHECK2: [[LOOP15]] = distinct !{[[LOOP15]], [[META5]]}
+// CHECK2: [[LOOP16]] = distinct !{[[LOOP16]], [[META5]]}
+// CHECK2: [[LOOP17]] = distinct !{[[LOOP17]], [[META5]]}
+// CHECK2: [[LOOP18]] = distinct !{[[LOOP18]], [[META5]]}
+// CHECK2: [[LOOP19]] = distinct !{[[LOOP19]], [[META5]]}
+// CHECK2: [[LOOP20]] = distinct !{[[LOOP20]], [[META5]]}
+// CHECK2: [[LOOP21]] = distinct !{[[LOOP21]], [[META5]]}
+// CHECK2: [[LOOP22]] = distinct !{[[LOOP22]], [[META5]]}
+// CHECK2: [[LOOP25]] = distinct !{[[LOOP25]], [[META5]]}
+// CHECK2: [[LOOP26]] = distinct !{[[LOOP26]], [[META5]]}
+// CHECK2: [[LOOP27]] = distinct !{[[LOOP27]], [[META5]]}
+// CHECK2: [[LOOP28]] = distinct !{[[LOOP28]], [[META5]]}
+// CHECK2: [[LOOP29]] = distinct !{[[LOOP29]], [[META5]]}
+// CHECK2: [[LOOP30]] = distinct !{[[LOOP30]], [[META5]]}
//.
>From c141f6d99231410c61839bed96ad78879547b63c Mon Sep 17 00:00:00 2001
From: Zahira Ammarguellat <zahira.ammarguellat at intel.com>
Date: Thu, 23 Jul 2026 10:19:02 -0700
Subject: [PATCH 4/5] Added the finalization for fuse
---
clang/include/clang/AST/StmtOpenMP.h | 51 +-
clang/lib/AST/StmtOpenMP.cpp | 25 +-
clang/lib/Sema/SemaOpenMP.cpp | 93 +-
clang/test/OpenMP/fuse_codegen.cpp | 2484 ++++++++++++++------------
4 files changed, 1432 insertions(+), 1221 deletions(-)
diff --git a/clang/include/clang/AST/StmtOpenMP.h b/clang/include/clang/AST/StmtOpenMP.h
index 5c374320bf018..2b1a4b4fa72af 100644
--- a/clang/include/clang/AST/StmtOpenMP.h
+++ b/clang/include/clang/AST/StmtOpenMP.h
@@ -5618,9 +5618,7 @@ class OMPTileDirective final
Data->getChildren()[TransformedStmtOffset] = S;
}
- void setFinals(Stmt *Finals) {
- Data->getChildren()[FinalsOffset] = Finals;
- }
+ void setFinals(Stmt *Finals) { Data->getChildren()[FinalsOffset] = Finals; }
public:
/// Create a new AST node representation for '#pragma omp tile'.
@@ -5636,12 +5634,10 @@ class OMPTileDirective final
/// dependent contexts.
/// \param PreInits Helper preinits statements for the loop nest.
/// \param Finals Loop variable finalization statements.
- static OMPTileDirective *Create(const ASTContext &C, SourceLocation StartLoc,
- SourceLocation EndLoc,
- ArrayRef<OMPClause *> Clauses,
- unsigned NumLoops, Stmt *AssociatedStmt,
- Stmt *TransformedStmt, Stmt *PreInits,
- Stmt *Finals);
+ static OMPTileDirective *
+ Create(const ASTContext &C, SourceLocation StartLoc, SourceLocation EndLoc,
+ ArrayRef<OMPClause *> Clauses, unsigned NumLoops, Stmt *AssociatedStmt,
+ Stmt *TransformedStmt, Stmt *PreInits, Stmt *Finals);
/// Build an empty '#pragma omp tile' AST node for deserialization.
///
@@ -5704,9 +5700,7 @@ class OMPStripeDirective final
Data->getChildren()[TransformedStmtOffset] = S;
}
- void setFinals(Stmt *Finals) {
- Data->getChildren()[FinalsOffset] = Finals;
- }
+ void setFinals(Stmt *Finals) { Data->getChildren()[FinalsOffset] = Finals; }
public:
/// Create a new AST node representation for '#pragma omp stripe'.
@@ -5869,9 +5863,7 @@ class OMPReverseDirective final
Data->getChildren()[TransformedStmtOffset] = S;
}
- void setFinals(Stmt *Finals) {
- Data->getChildren()[FinalsOffset] = Finals;
- }
+ void setFinals(Stmt *Finals) { Data->getChildren()[FinalsOffset] = Finals; }
public:
/// Create a new AST node representation for '#pragma omp reverse'.
@@ -5885,12 +5877,10 @@ class OMPReverseDirective final
/// dependent contexts.
/// \param PreInits Helper preinits statements for the loop nest.
/// \param Finals Loop variable finalization statements.
- static OMPReverseDirective *Create(const ASTContext &C,
- SourceLocation StartLoc,
- SourceLocation EndLoc,
- Stmt *AssociatedStmt, unsigned NumLoops,
- Stmt *TransformedStmt, Stmt *PreInits,
- Stmt *Finals);
+ static OMPReverseDirective
+ Create(const ASTContext &C, SourceLocation StartLoc, SourceLocation EndLoc,
+ Stmt *AssociatedStmt, unsigned NumLoops, Stmt *TransformedStmt,
+ Stmt *PreInits, Stmt *Finals);
/// Build an empty '#pragma omp reverse' AST node for deserialization.
///
@@ -5950,9 +5940,7 @@ class OMPInterchangeDirective final
Data->getChildren()[TransformedStmtOffset] = S;
}
- void setFinals(Stmt *Finals) {
- Data->getChildren()[FinalsOffset] = Finals;
- }
+ void setFinals(Stmt *Finals) { Data->getChildren()[FinalsOffset] = Finals; }
public:
/// Create a new AST node representation for '#pragma omp interchange'.
@@ -6063,9 +6051,7 @@ class OMPFuseDirective final
Data->getChildren()[TransformedStmtOffset] = S;
}
- void setFinals(Stmt *Finals) {
- Data->getChildren()[FinalsOffset] = Finals;
- }
+ void setFinals(Stmt *Finals) { Data->getChildren()[FinalsOffset] = Finals; }
public:
/// Create a new AST node representation for #pragma omp fuse'
@@ -6083,11 +6069,12 @@ class OMPFuseDirective final
/// dependent
/// \param PreInits Helper preinits statements for the loop nest
/// \param Finals Loop variable finalization statements
- static OMPFuseDirective *
- Create(const ASTContext &C, SourceLocation StartLoc, SourceLocation EndLoc,
- ArrayRef<OMPClause *> Clauses, unsigned NumGeneratedTopLevelLoops,
- Stmt *AssociatedStmt, Stmt *TransformedStmt, Stmt *PreInits,
- Stmt *Finals);
+ static OMPFuseDirective * Create(const ASTContext &C, SourceLocation StartLoc,
+ SourceLocation EndLoc,
+ ArrayRef<OMPClause *> Clauses,
+ unsigned NumGeneratedTopLevelLoops,
+ Stmt *AssociatedStmt, Stmt *TransformedStmt,
+ Stmt *PreInits, Stmt *Finals);
/// Build an empty '#pragma omp fuse' AST node for deserialization
///
diff --git a/clang/lib/AST/StmtOpenMP.cpp b/clang/lib/AST/StmtOpenMP.cpp
index cace1c992eedf..13f1d2d94aa85 100644
--- a/clang/lib/AST/StmtOpenMP.cpp
+++ b/clang/lib/AST/StmtOpenMP.cpp
@@ -454,8 +454,7 @@ OMPTileDirective::Create(const ASTContext &C, SourceLocation StartLoc,
unsigned NumLoops, Stmt *AssociatedStmt,
Stmt *TransformedStmt, Stmt *PreInits, Stmt *Finals) {
OMPTileDirective *Dir = createDirective<OMPTileDirective>(
- C, Clauses, AssociatedStmt, FinalsOffset + 1, StartLoc, EndLoc,
- NumLoops);
+ C, Clauses, AssociatedStmt, FinalsOffset + 1, StartLoc, EndLoc, NumLoops);
Dir->setTransformedStmt(TransformedStmt);
Dir->setPreInits(PreInits);
Dir->setFinals(Finals);
@@ -470,14 +469,12 @@ OMPTileDirective *OMPTileDirective::CreateEmpty(const ASTContext &C,
SourceLocation(), SourceLocation(), NumLoops);
}
-OMPStripeDirective *
-OMPStripeDirective::Create(const ASTContext &C, SourceLocation StartLoc,
- SourceLocation EndLoc, ArrayRef<OMPClause *> Clauses,
- unsigned NumLoops, Stmt *AssociatedStmt,
- Stmt *TransformedStmt, Stmt *PreInits, Stmt *Finals) {
+OMPStripeDirective *OMPStripeDirective::Create(
+ const ASTContext &C, SourceLocation StartLoc, SourceLocation EndLoc,
+ ArrayRef<OMPClause *> Clauses, unsigned NumLoops, Stmt *AssociatedStmt,
+ Stmt *TransformedStmt, Stmt *PreInits, Stmt *Finals) {
OMPStripeDirective *Dir = createDirective<OMPStripeDirective>(
- C, Clauses, AssociatedStmt, FinalsOffset + 1, StartLoc, EndLoc,
- NumLoops);
+ C, Clauses, AssociatedStmt, FinalsOffset + 1, StartLoc, EndLoc, NumLoops);
Dir->setTransformedStmt(TransformedStmt);
Dir->setPreInits(PreInits);
Dir->setFinals(Finals);
@@ -520,8 +517,7 @@ OMPReverseDirective::Create(const ASTContext &C, SourceLocation StartLoc,
unsigned NumLoops, Stmt *TransformedStmt,
Stmt *PreInits, Stmt *Finals) {
OMPReverseDirective *Dir = createDirective<OMPReverseDirective>(
- C, {}, AssociatedStmt, FinalsOffset + 1, StartLoc, EndLoc,
- NumLoops);
+ C, {}, AssociatedStmt, FinalsOffset + 1, StartLoc, EndLoc, NumLoops);
Dir->setTransformedStmt(TransformedStmt);
Dir->setPreInits(PreInits);
Dir->setFinals(Finals);
@@ -531,8 +527,8 @@ OMPReverseDirective::Create(const ASTContext &C, SourceLocation StartLoc,
OMPReverseDirective *OMPReverseDirective::CreateEmpty(const ASTContext &C,
unsigned NumLoops) {
return createEmptyDirective<OMPReverseDirective>(
- C, /*NumClauses=*/0, /*HasAssociatedStmt=*/true,
- FinalsOffset + 1, SourceLocation(), SourceLocation(), NumLoops);
+ C, /*NumClauses=*/0, /*HasAssociatedStmt=*/true, FinalsOffset + 1,
+ SourceLocation(), SourceLocation(), NumLoops);
}
OMPInterchangeDirective *OMPInterchangeDirective::Create(
@@ -540,8 +536,7 @@ OMPInterchangeDirective *OMPInterchangeDirective::Create(
ArrayRef<OMPClause *> Clauses, unsigned NumLoops, Stmt *AssociatedStmt,
Stmt *TransformedStmt, Stmt *PreInits, Stmt *Finals) {
OMPInterchangeDirective *Dir = createDirective<OMPInterchangeDirective>(
- C, Clauses, AssociatedStmt, FinalsOffset + 1, StartLoc, EndLoc,
- NumLoops);
+ C, Clauses, AssociatedStmt, FinalsOffset + 1, StartLoc, EndLoc, NumLoops);
Dir->setTransformedStmt(TransformedStmt);
Dir->setPreInits(PreInits);
Dir->setFinals(Finals);
diff --git a/clang/lib/Sema/SemaOpenMP.cpp b/clang/lib/Sema/SemaOpenMP.cpp
index 0e5cac6e08b2d..1ddf8db56ada6 100644
--- a/clang/lib/Sema/SemaOpenMP.cpp
+++ b/clang/lib/Sema/SemaOpenMP.cpp
@@ -15006,8 +15006,8 @@ StmtResult SemaOpenMP::ActOnOpenMPTileDirective(ArrayRef<OMPClause *> Clauses,
// Delay tiling to when template is completely instantiated.
if (SemaRef.CurContext->isDependentContext())
- return OMPTileDirective::Create(Context, StartLoc, EndLoc, Clauses,
- NumLoops, AStmt, nullptr, nullptr, nullptr);
+ return OMPTileDirective::Create(Context, StartLoc, EndLoc, Clauses, NumLoops,
+ AStmt, nullptr, nullptr, nullptr);
assert(LoopHelpers.size() == NumLoops &&
"Expecting loop iteration space dimensionality to match number of "
@@ -15284,7 +15284,8 @@ StmtResult SemaOpenMP::ActOnOpenMPStripeDirective(ArrayRef<OMPClause *> Clauses,
// Delay striping to when template is completely instantiated.
if (SemaRef.CurContext->isDependentContext())
return OMPStripeDirective::Create(Context, StartLoc, EndLoc, Clauses,
- NumLoops, AStmt, nullptr, nullptr, nullptr);
+ NumLoops, AStmt, nullptr, nullptr,
+ nullptr);
assert(LoopHelpers.size() == NumLoops &&
"Expecting loop iteration space dimensionality to match number of "
@@ -15507,10 +15508,10 @@ StmtResult SemaOpenMP::ActOnOpenMPStripeDirective(ArrayRef<OMPClause *> Clauses,
LoopHelper.Init->getBeginLoc(), LoopHelper.Inc->getEndLoc());
}
- return OMPStripeDirective::Create(Context, StartLoc, EndLoc, Clauses,
- NumLoops, AStmt, Inner,
- buildPreInits(Context, PreInits),
- buildLoopFinalization(Context, LoopHelpers));
+ return OMPStripeDirective::Create(
+ Context, StartLoc, EndLoc, Clauses, NumLoops, AStmt, Inner,
+ buildPreInits(Context, PreInits),
+ buildLoopFinalization(Context, LoopHelpers));
}
StmtResult SemaOpenMP::ActOnOpenMPUnrollDirective(ArrayRef<OMPClause *> Clauses,
@@ -15977,9 +15978,10 @@ StmtResult SemaOpenMP::ActOnOpenMPReverseDirective(Stmt *AStmt,
ReversedBody, LoopHelper.Init->getBeginLoc(),
LoopHelper.Init->getBeginLoc(), LoopHelper.Inc->getEndLoc());
- return OMPReverseDirective::Create(Context, StartLoc, EndLoc, AStmt, NumLoops,
- ReversedFor, buildPreInits(Context, PreInits),
- buildLoopFinalization(Context, LoopHelpers));
+ return OMPReverseDirective::Create(
+ Context, StartLoc, EndLoc, AStmt, NumLoops, ReversedFor,
+ buildPreInits(Context, PreInits),
+ buildLoopFinalization(Context, LoopHelpers));
}
/// Build the AST for \#pragma omp split counts(c1, c2, ...).
@@ -16238,7 +16240,8 @@ StmtResult SemaOpenMP::ActOnOpenMPInterchangeDirective(
// Delay interchange to when template is completely instantiated.
if (CurContext->isDependentContext())
return OMPInterchangeDirective::Create(Context, StartLoc, EndLoc, Clauses,
- NumLoops, AStmt, nullptr, nullptr, nullptr);
+ NumLoops, AStmt, nullptr, nullptr,
+ nullptr);
// An invalid expression in the permutation clause is set to nullptr in
// ActOnOpenMPPermutationClause.
@@ -16621,6 +16624,8 @@ StmtResult SemaOpenMP::ActOnOpenMPFuseDirective(ArrayRef<OMPClause *> Clauses,
NIVarDecls.push_back(NIVD);
IVVarDecls.push_back(IVVD);
+ if (UBDStmt.isUsable())
+ PreInits.push_back(UBDStmt.get());
PreInits.push_back(LBDStmt.get());
PreInits.push_back(STDStmt.get());
PreInits.push_back(NIDStmt.get());
@@ -16833,6 +16838,56 @@ StmtResult SemaOpenMP::ActOnOpenMPFuseDirective(ArrayRef<OMPClause *> Clauses,
FusedBody, InitStmt.get()->getBeginLoc(), SourceLocation(),
IncrExpr.get()->getEndLoc());
+ // 8. Build finalization statements for fused loops.
+ // For each fused loop i: final_value_i = lb_i + ni_i * st_i.
+ SmallVector<Stmt *, 8> FusedLoopFinalization;
+ for (unsigned I = FirstVal - 1, J = 0; I < LastVal; ++I, ++J) {
+ // Skip non-regular loops (transformations, range-based loops, etc.).
+ if (!SeqAnalysis.Loops[I].isRegularLoop())
+ continue;
+
+ // Check if this is a ForStmt.
+ if (!isa<ForStmt>(SeqAnalysis.Loops[I].TheForStmt))
+ continue;
+
+ // Get the user's actual loop variable.
+ assert(!SeqAnalysis.Loops[I].HelperExprs.Counters.empty() &&
+ "Expected at least one counter for the loop");
+ DeclRefExpr *UserLoopVarRef = cast<DeclRefExpr>(
+ SeqAnalysis.Loops[I].HelperExprs.Counters[0]);
+ VarDecl *UserLoopVarDecl = cast<VarDecl>(UserLoopVarRef->getDecl());
+
+ // Skip internal OpenMP variables (they start with .omp.).
+ if (UserLoopVarDecl->getName().starts_with(".omp."))
+ continue;
+
+ // Build: user_var = lb + ni * st.
+ // ni * st.
+ ExprResult FinalValue = SemaRef.BuildBinOp(
+ CurScope, SourceLocation(), BO_Mul,
+ MakeVarDeclRef(NIVarDecls[J]), MakeVarDeclRef(STVarDecls[J]));
+ if (!FinalValue.isUsable())
+ return StmtError();
+
+ // lb + (ni * st).
+ FinalValue = SemaRef.BuildBinOp(
+ CurScope, SourceLocation(), BO_Add,
+ MakeVarDeclRef(LBVarDecls[J]), FinalValue.get());
+ if (!FinalValue.isUsable())
+ return StmtError();
+
+ // user_var = final_value.
+ FinalValue = SemaRef.BuildBinOp(
+ CurScope, SourceLocation(), BO_Assign,
+ buildDeclRefExpr(SemaRef, UserLoopVarDecl, UserLoopVarDecl->getType(),
+ UserLoopVarDecl->getLocation(), false),
+ FinalValue.get());
+ if (!FinalValue.isUsable())
+ return StmtError();
+
+ FusedLoopFinalization.push_back(FinalValue.get());
+ }
+
// In the case of looprange, the result of fuse won't simply
// be a single loop (ForStmt), but rather a loop sequence
// (CompoundStmt) of 3 parts: the pre-fusion loops, the fused loop
@@ -16876,21 +16931,23 @@ StmtResult SemaOpenMP::ActOnOpenMPFuseDirective(ArrayRef<OMPClause *> Clauses,
FinalLoops.push_back(SeqAnalysis.Loops[I].TheForStmt);
}
+ // Insert the fused loop.
FinalLoops.insert(FinalLoops.begin() + (FirstVal - 1), FusedForStmt);
+
FusionStmt = CompoundStmt::Create(Context, FinalLoops, FPOptionsOverride(),
SourceLocation(), SourceLocation());
}
- // TODO: For fuse directive, loop variable finalization is complex because:
- // 1. The fused loops' original variables are replaced by new fusion variables
- // 2. The scope and structure change significantly after fusion
- // 3. Pre-computed Finals reference variables that may not exist after fusion
- // For now, we pass nullptr for Finals to avoid scope issues.
- // Proper implementation requires computing Finals after fusion transformation.
+ Stmt *Finals = nullptr;
+ if (!FusedLoopFinalization.empty()) {
+ Finals = CompoundStmt::Create(Context, FusedLoopFinalization,
+ FPOptionsOverride(),
+ SourceLocation(), SourceLocation());
+ }
return OMPFuseDirective::Create(Context, StartLoc, EndLoc, Clauses,
NumGeneratedTopLevelLoops, AStmt, FusionStmt,
- buildPreInits(Context, PreInits), nullptr);
+ buildPreInits(Context, PreInits), Finals);
}
OMPClause *SemaOpenMP::ActOnOpenMPSingleExprClause(OpenMPClauseKind Kind,
diff --git a/clang/test/OpenMP/fuse_codegen.cpp b/clang/test/OpenMP/fuse_codegen.cpp
index 742c280ed0172..2a7d410f9543d 100644
--- a/clang/test/OpenMP/fuse_codegen.cpp
+++ b/clang/test/OpenMP/fuse_codegen.cpp
@@ -105,6 +105,7 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[DOTCAPTURE_EXPR_1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTNEW_STEP:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_UB0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_LB0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_ST0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_NI0:%.*]] = alloca i32, align 4
@@ -113,6 +114,7 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[DOTCAPTURE_EXPR_7:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTNEW_STEP8:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_UB1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_LB1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_ST1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_NI1:%.*]] = alloca i32, align 4
@@ -144,106 +146,122 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[DIV:%.*]] = udiv i32 [[ADD]], [[TMP7]]
// CHECK1-NEXT: [[SUB4:%.*]] = sub i32 [[DIV]], 1
// CHECK1-NEXT: store i32 [[SUB4]], ptr [[DOTCAPTURE_EXPR_2]], align 4
+// CHECK1-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
+// CHECK1-NEXT: store i32 [[TMP8]], ptr [[DOTOMP_UB0]], align 4
// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_LB0]], align 4
// CHECK1-NEXT: store i32 1, ptr [[DOTOMP_ST0]], align 4
-// CHECK1-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
-// CHECK1-NEXT: [[ADD5:%.*]] = add i32 [[TMP8]], 1
+// CHECK1-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
+// CHECK1-NEXT: [[ADD5:%.*]] = add i32 [[TMP9]], 1
// CHECK1-NEXT: store i32 [[ADD5]], ptr [[DOTOMP_NI0]], align 4
-// CHECK1-NEXT: [[TMP9:%.*]] = load i32, ptr [[START2_ADDR]], align 4
-// CHECK1-NEXT: store i32 [[TMP9]], ptr [[J]], align 4
// CHECK1-NEXT: [[TMP10:%.*]] = load i32, ptr [[START2_ADDR]], align 4
-// CHECK1-NEXT: store i32 [[TMP10]], ptr [[DOTCAPTURE_EXPR_6]], align 4
-// CHECK1-NEXT: [[TMP11:%.*]] = load i32, ptr [[END2_ADDR]], align 4
-// CHECK1-NEXT: store i32 [[TMP11]], ptr [[DOTCAPTURE_EXPR_7]], align 4
-// CHECK1-NEXT: [[TMP12:%.*]] = load i32, ptr [[STEP2_ADDR]], align 4
-// CHECK1-NEXT: store i32 [[TMP12]], ptr [[DOTNEW_STEP8]], align 4
-// CHECK1-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_7]], align 4
-// CHECK1-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
-// CHECK1-NEXT: [[SUB10:%.*]] = sub i32 [[TMP13]], [[TMP14]]
+// CHECK1-NEXT: store i32 [[TMP10]], ptr [[J]], align 4
+// CHECK1-NEXT: [[TMP11:%.*]] = load i32, ptr [[START2_ADDR]], align 4
+// CHECK1-NEXT: store i32 [[TMP11]], ptr [[DOTCAPTURE_EXPR_6]], align 4
+// CHECK1-NEXT: [[TMP12:%.*]] = load i32, ptr [[END2_ADDR]], align 4
+// CHECK1-NEXT: store i32 [[TMP12]], ptr [[DOTCAPTURE_EXPR_7]], align 4
+// CHECK1-NEXT: [[TMP13:%.*]] = load i32, ptr [[STEP2_ADDR]], align 4
+// CHECK1-NEXT: store i32 [[TMP13]], ptr [[DOTNEW_STEP8]], align 4
+// CHECK1-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_7]], align 4
+// CHECK1-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
+// CHECK1-NEXT: [[SUB10:%.*]] = sub i32 [[TMP14]], [[TMP15]]
// CHECK1-NEXT: [[SUB11:%.*]] = sub i32 [[SUB10]], 1
-// CHECK1-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
-// CHECK1-NEXT: [[ADD12:%.*]] = add i32 [[SUB11]], [[TMP15]]
// CHECK1-NEXT: [[TMP16:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
-// CHECK1-NEXT: [[DIV13:%.*]] = udiv i32 [[ADD12]], [[TMP16]]
+// CHECK1-NEXT: [[ADD12:%.*]] = add i32 [[SUB11]], [[TMP16]]
+// CHECK1-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
+// CHECK1-NEXT: [[DIV13:%.*]] = udiv i32 [[ADD12]], [[TMP17]]
// CHECK1-NEXT: [[SUB14:%.*]] = sub i32 [[DIV13]], 1
// CHECK1-NEXT: store i32 [[SUB14]], ptr [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK1-NEXT: [[TMP18:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK1-NEXT: store i32 [[TMP18]], ptr [[DOTOMP_UB1]], align 4
// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_LB1]], align 4
// CHECK1-NEXT: store i32 1, ptr [[DOTOMP_ST1]], align 4
-// CHECK1-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK1-NEXT: [[ADD15:%.*]] = add i32 [[TMP17]], 1
+// CHECK1-NEXT: [[TMP19:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK1-NEXT: [[ADD15:%.*]] = add i32 [[TMP19]], 1
// CHECK1-NEXT: store i32 [[ADD15]], ptr [[DOTOMP_NI1]], align 4
-// CHECK1-NEXT: [[TMP18:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
-// CHECK1-NEXT: store i32 [[TMP18]], ptr [[DOTOMP_TEMP_1]], align 4
-// CHECK1-NEXT: [[TMP19:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
-// CHECK1-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
-// CHECK1-NEXT: [[CMP:%.*]] = icmp ugt i32 [[TMP19]], [[TMP20]]
+// CHECK1-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK1-NEXT: store i32 [[TMP20]], ptr [[DOTOMP_TEMP_1]], align 4
+// CHECK1-NEXT: [[TMP21:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
+// CHECK1-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK1-NEXT: [[CMP:%.*]] = icmp ugt i32 [[TMP21]], [[TMP22]]
// CHECK1-NEXT: br i1 [[CMP]], label %[[COND_TRUE:.*]], label %[[COND_FALSE:.*]]
// CHECK1: [[COND_TRUE]]:
-// CHECK1-NEXT: [[TMP21:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
+// CHECK1-NEXT: [[TMP23:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
// CHECK1-NEXT: br label %[[COND_END:.*]]
// CHECK1: [[COND_FALSE]]:
-// CHECK1-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK1-NEXT: [[TMP24:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
// CHECK1-NEXT: br label %[[COND_END]]
// CHECK1: [[COND_END]]:
-// CHECK1-NEXT: [[COND:%.*]] = phi i32 [ [[TMP21]], %[[COND_TRUE]] ], [ [[TMP22]], %[[COND_FALSE]] ]
+// CHECK1-NEXT: [[COND:%.*]] = phi i32 [ [[TMP23]], %[[COND_TRUE]] ], [ [[TMP24]], %[[COND_FALSE]] ]
// CHECK1-NEXT: store i32 [[COND]], ptr [[DOTOMP_FUSE_MAX]], align 4
// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_FUSE_INDEX]], align 4
// CHECK1-NEXT: br label %[[FOR_COND:.*]]
// CHECK1: [[FOR_COND]]:
-// CHECK1-NEXT: [[TMP23:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[TMP24:%.*]] = load i32, ptr [[DOTOMP_FUSE_MAX]], align 4
-// CHECK1-NEXT: [[CMP16:%.*]] = icmp ult i32 [[TMP23]], [[TMP24]]
+// CHECK1-NEXT: [[TMP25:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[TMP26:%.*]] = load i32, ptr [[DOTOMP_FUSE_MAX]], align 4
+// CHECK1-NEXT: [[CMP16:%.*]] = icmp ult i32 [[TMP25]], [[TMP26]]
// CHECK1-NEXT: br i1 [[CMP16]], label %[[FOR_BODY:.*]], label %[[FOR_END:.*]]
// CHECK1: [[FOR_BODY]]:
-// CHECK1-NEXT: [[TMP25:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[TMP26:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
-// CHECK1-NEXT: [[CMP17:%.*]] = icmp ult i32 [[TMP25]], [[TMP26]]
+// CHECK1-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK1-NEXT: [[CMP17:%.*]] = icmp ult i32 [[TMP27]], [[TMP28]]
// CHECK1-NEXT: br i1 [[CMP17]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
// CHECK1: [[IF_THEN]]:
-// CHECK1-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
-// CHECK1-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
-// CHECK1-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[MUL:%.*]] = mul i32 [[TMP28]], [[TMP29]]
-// CHECK1-NEXT: [[ADD18:%.*]] = add i32 [[TMP27]], [[MUL]]
+// CHECK1-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
+// CHECK1-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
+// CHECK1-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[MUL:%.*]] = mul i32 [[TMP30]], [[TMP31]]
+// CHECK1-NEXT: [[ADD18:%.*]] = add i32 [[TMP29]], [[MUL]]
// CHECK1-NEXT: store i32 [[ADD18]], ptr [[DOTOMP_IV0]], align 4
-// CHECK1-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
-// CHECK1-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
-// CHECK1-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
-// CHECK1-NEXT: [[MUL19:%.*]] = mul i32 [[TMP31]], [[TMP32]]
-// CHECK1-NEXT: [[ADD20:%.*]] = add i32 [[TMP30]], [[MUL19]]
+// CHECK1-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
+// CHECK1-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK1-NEXT: [[MUL19:%.*]] = mul i32 [[TMP33]], [[TMP34]]
+// CHECK1-NEXT: [[ADD20:%.*]] = add i32 [[TMP32]], [[MUL19]]
// CHECK1-NEXT: store i32 [[ADD20]], ptr [[I]], align 4
-// CHECK1-NEXT: [[TMP33:%.*]] = load i32, ptr [[I]], align 4
-// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP33]])
+// CHECK1-NEXT: [[TMP35:%.*]] = load i32, ptr [[I]], align 4
+// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP35]])
// CHECK1-NEXT: br label %[[IF_END]]
// CHECK1: [[IF_END]]:
-// CHECK1-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
-// CHECK1-NEXT: [[CMP21:%.*]] = icmp ult i32 [[TMP34]], [[TMP35]]
+// CHECK1-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[TMP37:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK1-NEXT: [[CMP21:%.*]] = icmp ult i32 [[TMP36]], [[TMP37]]
// CHECK1-NEXT: br i1 [[CMP21]], label %[[IF_THEN22:.*]], label %[[IF_END27:.*]]
// CHECK1: [[IF_THEN22]]:
-// CHECK1-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
-// CHECK1-NEXT: [[TMP37:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
-// CHECK1-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[MUL23:%.*]] = mul i32 [[TMP37]], [[TMP38]]
-// CHECK1-NEXT: [[ADD24:%.*]] = add i32 [[TMP36]], [[MUL23]]
+// CHECK1-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
+// CHECK1-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
+// CHECK1-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[MUL23:%.*]] = mul i32 [[TMP39]], [[TMP40]]
+// CHECK1-NEXT: [[ADD24:%.*]] = add i32 [[TMP38]], [[MUL23]]
// CHECK1-NEXT: store i32 [[ADD24]], ptr [[DOTOMP_IV1]], align 4
-// CHECK1-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
-// CHECK1-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
-// CHECK1-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
-// CHECK1-NEXT: [[MUL25:%.*]] = mul i32 [[TMP40]], [[TMP41]]
-// CHECK1-NEXT: [[ADD26:%.*]] = add i32 [[TMP39]], [[MUL25]]
+// CHECK1-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
+// CHECK1-NEXT: [[TMP42:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
+// CHECK1-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
+// CHECK1-NEXT: [[MUL25:%.*]] = mul i32 [[TMP42]], [[TMP43]]
+// CHECK1-NEXT: [[ADD26:%.*]] = add i32 [[TMP41]], [[MUL25]]
// CHECK1-NEXT: store i32 [[ADD26]], ptr [[J]], align 4
-// CHECK1-NEXT: [[TMP42:%.*]] = load i32, ptr [[J]], align 4
-// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP42]])
+// CHECK1-NEXT: [[TMP44:%.*]] = load i32, ptr [[J]], align 4
+// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP44]])
// CHECK1-NEXT: br label %[[IF_END27]]
// CHECK1: [[IF_END27]]:
// CHECK1-NEXT: br label %[[FOR_INC:.*]]
// CHECK1: [[FOR_INC]]:
-// CHECK1-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[INC:%.*]] = add i32 [[TMP43]], 1
+// CHECK1-NEXT: [[TMP45:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[INC:%.*]] = add i32 [[TMP45]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP3:![0-9]+]]
+// CHECK1-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP2:![0-9]+]]
// CHECK1: [[FOR_END]]:
+// CHECK1-NEXT: [[TMP46:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
+// CHECK1-NEXT: [[TMP47:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK1-NEXT: [[TMP48:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
+// CHECK1-NEXT: [[MUL28:%.*]] = mul i32 [[TMP47]], [[TMP48]]
+// CHECK1-NEXT: [[ADD29:%.*]] = add i32 [[TMP46]], [[MUL28]]
+// CHECK1-NEXT: store i32 [[ADD29]], ptr [[I]], align 4
+// CHECK1-NEXT: [[TMP49:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
+// CHECK1-NEXT: [[TMP50:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK1-NEXT: [[TMP51:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
+// CHECK1-NEXT: [[MUL30:%.*]] = mul i32 [[TMP50]], [[TMP51]]
+// CHECK1-NEXT: [[ADD31:%.*]] = add i32 [[TMP49]], [[MUL30]]
+// CHECK1-NEXT: store i32 [[ADD31]], ptr [[J]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -267,6 +285,7 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[DOTCAPTURE_EXPR_1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTNEW_STEP:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_UB0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_LB0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_ST0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_NI0:%.*]] = alloca i32, align 4
@@ -275,6 +294,7 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[DOTCAPTURE_EXPR_7:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTNEW_STEP8:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_UB1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_LB1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_ST1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_NI1:%.*]] = alloca i32, align 4
@@ -283,6 +303,7 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTNEW_STEP21:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTCAPTURE_EXPR_22:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_UB2:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_LB2:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_ST2:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_NI2:%.*]] = alloca i32, align 4
@@ -312,169 +333,193 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[DIV:%.*]] = udiv i32 [[ADD]], [[TMP7]]
// CHECK1-NEXT: [[SUB4:%.*]] = sub i32 [[DIV]], 1
// CHECK1-NEXT: store i32 [[SUB4]], ptr [[DOTCAPTURE_EXPR_2]], align 4
+// CHECK1-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
+// CHECK1-NEXT: store i32 [[TMP8]], ptr [[DOTOMP_UB0]], align 4
// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_LB0]], align 4
// CHECK1-NEXT: store i32 1, ptr [[DOTOMP_ST0]], align 4
-// CHECK1-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
-// CHECK1-NEXT: [[ADD5:%.*]] = add i32 [[TMP8]], 1
+// CHECK1-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
+// CHECK1-NEXT: [[ADD5:%.*]] = add i32 [[TMP9]], 1
// CHECK1-NEXT: store i32 [[ADD5]], ptr [[DOTOMP_NI0]], align 4
-// CHECK1-NEXT: [[TMP9:%.*]] = load i32, ptr [[END_ADDR]], align 4
-// CHECK1-NEXT: store i32 [[TMP9]], ptr [[J]], align 4
// CHECK1-NEXT: [[TMP10:%.*]] = load i32, ptr [[END_ADDR]], align 4
-// CHECK1-NEXT: store i32 [[TMP10]], ptr [[DOTCAPTURE_EXPR_6]], align 4
-// CHECK1-NEXT: [[TMP11:%.*]] = load i32, ptr [[START_ADDR]], align 4
-// CHECK1-NEXT: store i32 [[TMP11]], ptr [[DOTCAPTURE_EXPR_7]], align 4
-// CHECK1-NEXT: [[TMP12:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
-// CHECK1-NEXT: store i32 [[TMP12]], ptr [[DOTNEW_STEP8]], align 4
-// CHECK1-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
-// CHECK1-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_7]], align 4
-// CHECK1-NEXT: [[SUB10:%.*]] = sub i32 [[TMP13]], [[TMP14]]
+// CHECK1-NEXT: store i32 [[TMP10]], ptr [[J]], align 4
+// CHECK1-NEXT: [[TMP11:%.*]] = load i32, ptr [[END_ADDR]], align 4
+// CHECK1-NEXT: store i32 [[TMP11]], ptr [[DOTCAPTURE_EXPR_6]], align 4
+// CHECK1-NEXT: [[TMP12:%.*]] = load i32, ptr [[START_ADDR]], align 4
+// CHECK1-NEXT: store i32 [[TMP12]], ptr [[DOTCAPTURE_EXPR_7]], align 4
+// CHECK1-NEXT: [[TMP13:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
+// CHECK1-NEXT: store i32 [[TMP13]], ptr [[DOTNEW_STEP8]], align 4
+// CHECK1-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
+// CHECK1-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_7]], align 4
+// CHECK1-NEXT: [[SUB10:%.*]] = sub i32 [[TMP14]], [[TMP15]]
// CHECK1-NEXT: [[SUB11:%.*]] = sub i32 [[SUB10]], 1
-// CHECK1-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
-// CHECK1-NEXT: [[ADD12:%.*]] = add i32 [[SUB11]], [[TMP15]]
// CHECK1-NEXT: [[TMP16:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
-// CHECK1-NEXT: [[DIV13:%.*]] = udiv i32 [[ADD12]], [[TMP16]]
+// CHECK1-NEXT: [[ADD12:%.*]] = add i32 [[SUB11]], [[TMP16]]
+// CHECK1-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
+// CHECK1-NEXT: [[DIV13:%.*]] = udiv i32 [[ADD12]], [[TMP17]]
// CHECK1-NEXT: [[SUB14:%.*]] = sub i32 [[DIV13]], 1
// CHECK1-NEXT: store i32 [[SUB14]], ptr [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK1-NEXT: [[TMP18:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK1-NEXT: store i32 [[TMP18]], ptr [[DOTOMP_UB1]], align 4
// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_LB1]], align 4
// CHECK1-NEXT: store i32 1, ptr [[DOTOMP_ST1]], align 4
-// CHECK1-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK1-NEXT: [[ADD15:%.*]] = add i32 [[TMP17]], 1
+// CHECK1-NEXT: [[TMP19:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK1-NEXT: [[ADD15:%.*]] = add i32 [[TMP19]], 1
// CHECK1-NEXT: store i32 [[ADD15]], ptr [[DOTOMP_NI1]], align 4
-// CHECK1-NEXT: [[TMP18:%.*]] = load i32, ptr [[START_ADDR]], align 4
-// CHECK1-NEXT: [[TMP19:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
-// CHECK1-NEXT: [[ADD16:%.*]] = add nsw i32 [[TMP18]], [[TMP19]]
-// CHECK1-NEXT: store i32 [[ADD16]], ptr [[K]], align 4
// CHECK1-NEXT: [[TMP20:%.*]] = load i32, ptr [[START_ADDR]], align 4
// CHECK1-NEXT: [[TMP21:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
-// CHECK1-NEXT: [[ADD18:%.*]] = add nsw i32 [[TMP20]], [[TMP21]]
-// CHECK1-NEXT: store i32 [[ADD18]], ptr [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK1-NEXT: [[TMP22:%.*]] = load i32, ptr [[END_ADDR]], align 4
+// CHECK1-NEXT: [[ADD16:%.*]] = add nsw i32 [[TMP20]], [[TMP21]]
+// CHECK1-NEXT: store i32 [[ADD16]], ptr [[K]], align 4
+// CHECK1-NEXT: [[TMP22:%.*]] = load i32, ptr [[START_ADDR]], align 4
// CHECK1-NEXT: [[TMP23:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
-// CHECK1-NEXT: [[ADD20:%.*]] = add nsw i32 [[TMP22]], [[TMP23]]
+// CHECK1-NEXT: [[ADD18:%.*]] = add nsw i32 [[TMP22]], [[TMP23]]
+// CHECK1-NEXT: store i32 [[ADD18]], ptr [[DOTCAPTURE_EXPR_17]], align 4
+// CHECK1-NEXT: [[TMP24:%.*]] = load i32, ptr [[END_ADDR]], align 4
+// CHECK1-NEXT: [[TMP25:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
+// CHECK1-NEXT: [[ADD20:%.*]] = add nsw i32 [[TMP24]], [[TMP25]]
// CHECK1-NEXT: store i32 [[ADD20]], ptr [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK1-NEXT: [[TMP24:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
-// CHECK1-NEXT: store i32 [[TMP24]], ptr [[DOTNEW_STEP21]], align 4
-// CHECK1-NEXT: [[TMP25:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK1-NEXT: [[TMP26:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK1-NEXT: [[SUB23:%.*]] = sub i32 [[TMP25]], [[TMP26]]
+// CHECK1-NEXT: [[TMP26:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
+// CHECK1-NEXT: store i32 [[TMP26]], ptr [[DOTNEW_STEP21]], align 4
+// CHECK1-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK1-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_17]], align 4
+// CHECK1-NEXT: [[SUB23:%.*]] = sub i32 [[TMP27]], [[TMP28]]
// CHECK1-NEXT: [[SUB24:%.*]] = sub i32 [[SUB23]], 1
-// CHECK1-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTNEW_STEP21]], align 4
-// CHECK1-NEXT: [[ADD25:%.*]] = add i32 [[SUB24]], [[TMP27]]
-// CHECK1-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTNEW_STEP21]], align 4
-// CHECK1-NEXT: [[DIV26:%.*]] = udiv i32 [[ADD25]], [[TMP28]]
+// CHECK1-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTNEW_STEP21]], align 4
+// CHECK1-NEXT: [[ADD25:%.*]] = add i32 [[SUB24]], [[TMP29]]
+// CHECK1-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTNEW_STEP21]], align 4
+// CHECK1-NEXT: [[DIV26:%.*]] = udiv i32 [[ADD25]], [[TMP30]]
// CHECK1-NEXT: [[SUB27:%.*]] = sub i32 [[DIV26]], 1
// CHECK1-NEXT: store i32 [[SUB27]], ptr [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK1-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK1-NEXT: store i32 [[TMP31]], ptr [[DOTOMP_UB2]], align 4
// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_LB2]], align 4
// CHECK1-NEXT: store i32 1, ptr [[DOTOMP_ST2]], align 4
-// CHECK1-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_22]], align 4
-// CHECK1-NEXT: [[ADD28:%.*]] = add i32 [[TMP29]], 1
+// CHECK1-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK1-NEXT: [[ADD28:%.*]] = add i32 [[TMP32]], 1
// CHECK1-NEXT: store i32 [[ADD28]], ptr [[DOTOMP_NI2]], align 4
-// CHECK1-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
-// CHECK1-NEXT: store i32 [[TMP30]], ptr [[DOTOMP_TEMP_1]], align 4
-// CHECK1-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
-// CHECK1-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
-// CHECK1-NEXT: [[CMP:%.*]] = icmp ugt i32 [[TMP31]], [[TMP32]]
+// CHECK1-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK1-NEXT: store i32 [[TMP33]], ptr [[DOTOMP_TEMP_1]], align 4
+// CHECK1-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
+// CHECK1-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK1-NEXT: [[CMP:%.*]] = icmp ugt i32 [[TMP34]], [[TMP35]]
// CHECK1-NEXT: br i1 [[CMP]], label %[[COND_TRUE:.*]], label %[[COND_FALSE:.*]]
// CHECK1: [[COND_TRUE]]:
-// CHECK1-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
+// CHECK1-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
// CHECK1-NEXT: br label %[[COND_END:.*]]
// CHECK1: [[COND_FALSE]]:
-// CHECK1-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK1-NEXT: [[TMP37:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
// CHECK1-NEXT: br label %[[COND_END]]
// CHECK1: [[COND_END]]:
-// CHECK1-NEXT: [[COND:%.*]] = phi i32 [ [[TMP33]], %[[COND_TRUE]] ], [ [[TMP34]], %[[COND_FALSE]] ]
+// CHECK1-NEXT: [[COND:%.*]] = phi i32 [ [[TMP36]], %[[COND_TRUE]] ], [ [[TMP37]], %[[COND_FALSE]] ]
// CHECK1-NEXT: store i32 [[COND]], ptr [[DOTOMP_TEMP_2]], align 4
-// CHECK1-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTOMP_TEMP_2]], align 4
-// CHECK1-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_NI2]], align 4
-// CHECK1-NEXT: [[CMP29:%.*]] = icmp ugt i32 [[TMP35]], [[TMP36]]
+// CHECK1-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_TEMP_2]], align 4
+// CHECK1-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTOMP_NI2]], align 4
+// CHECK1-NEXT: [[CMP29:%.*]] = icmp ugt i32 [[TMP38]], [[TMP39]]
// CHECK1-NEXT: br i1 [[CMP29]], label %[[COND_TRUE30:.*]], label %[[COND_FALSE31:.*]]
// CHECK1: [[COND_TRUE30]]:
-// CHECK1-NEXT: [[TMP37:%.*]] = load i32, ptr [[DOTOMP_TEMP_2]], align 4
+// CHECK1-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTOMP_TEMP_2]], align 4
// CHECK1-NEXT: br label %[[COND_END32:.*]]
// CHECK1: [[COND_FALSE31]]:
-// CHECK1-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_NI2]], align 4
+// CHECK1-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTOMP_NI2]], align 4
// CHECK1-NEXT: br label %[[COND_END32]]
// CHECK1: [[COND_END32]]:
-// CHECK1-NEXT: [[COND33:%.*]] = phi i32 [ [[TMP37]], %[[COND_TRUE30]] ], [ [[TMP38]], %[[COND_FALSE31]] ]
+// CHECK1-NEXT: [[COND33:%.*]] = phi i32 [ [[TMP40]], %[[COND_TRUE30]] ], [ [[TMP41]], %[[COND_FALSE31]] ]
// CHECK1-NEXT: store i32 [[COND33]], ptr [[DOTOMP_FUSE_MAX]], align 4
// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_FUSE_INDEX]], align 4
// CHECK1-NEXT: br label %[[FOR_COND:.*]]
// CHECK1: [[FOR_COND]]:
-// CHECK1-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTOMP_FUSE_MAX]], align 4
-// CHECK1-NEXT: [[CMP34:%.*]] = icmp ult i32 [[TMP39]], [[TMP40]]
+// CHECK1-NEXT: [[TMP42:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTOMP_FUSE_MAX]], align 4
+// CHECK1-NEXT: [[CMP34:%.*]] = icmp ult i32 [[TMP42]], [[TMP43]]
// CHECK1-NEXT: br i1 [[CMP34]], label %[[FOR_BODY:.*]], label %[[FOR_END:.*]]
// CHECK1: [[FOR_BODY]]:
-// CHECK1-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[TMP42:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
-// CHECK1-NEXT: [[CMP35:%.*]] = icmp ult i32 [[TMP41]], [[TMP42]]
+// CHECK1-NEXT: [[TMP44:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[TMP45:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK1-NEXT: [[CMP35:%.*]] = icmp ult i32 [[TMP44]], [[TMP45]]
// CHECK1-NEXT: br i1 [[CMP35]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
// CHECK1: [[IF_THEN]]:
-// CHECK1-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
-// CHECK1-NEXT: [[TMP44:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
-// CHECK1-NEXT: [[TMP45:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[MUL:%.*]] = mul i32 [[TMP44]], [[TMP45]]
-// CHECK1-NEXT: [[ADD36:%.*]] = add i32 [[TMP43]], [[MUL]]
+// CHECK1-NEXT: [[TMP46:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
+// CHECK1-NEXT: [[TMP47:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
+// CHECK1-NEXT: [[TMP48:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[MUL:%.*]] = mul i32 [[TMP47]], [[TMP48]]
+// CHECK1-NEXT: [[ADD36:%.*]] = add i32 [[TMP46]], [[MUL]]
// CHECK1-NEXT: store i32 [[ADD36]], ptr [[DOTOMP_IV0]], align 4
-// CHECK1-NEXT: [[TMP46:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
-// CHECK1-NEXT: [[TMP47:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
-// CHECK1-NEXT: [[TMP48:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
-// CHECK1-NEXT: [[MUL37:%.*]] = mul i32 [[TMP47]], [[TMP48]]
-// CHECK1-NEXT: [[ADD38:%.*]] = add i32 [[TMP46]], [[MUL37]]
+// CHECK1-NEXT: [[TMP49:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[TMP50:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
+// CHECK1-NEXT: [[TMP51:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK1-NEXT: [[MUL37:%.*]] = mul i32 [[TMP50]], [[TMP51]]
+// CHECK1-NEXT: [[ADD38:%.*]] = add i32 [[TMP49]], [[MUL37]]
// CHECK1-NEXT: store i32 [[ADD38]], ptr [[I]], align 4
-// CHECK1-NEXT: [[TMP49:%.*]] = load i32, ptr [[I]], align 4
-// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP49]])
+// CHECK1-NEXT: [[TMP52:%.*]] = load i32, ptr [[I]], align 4
+// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP52]])
// CHECK1-NEXT: br label %[[IF_END]]
// CHECK1: [[IF_END]]:
-// CHECK1-NEXT: [[TMP50:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[TMP51:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
-// CHECK1-NEXT: [[CMP39:%.*]] = icmp ult i32 [[TMP50]], [[TMP51]]
+// CHECK1-NEXT: [[TMP53:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[TMP54:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK1-NEXT: [[CMP39:%.*]] = icmp ult i32 [[TMP53]], [[TMP54]]
// CHECK1-NEXT: br i1 [[CMP39]], label %[[IF_THEN40:.*]], label %[[IF_END45:.*]]
// CHECK1: [[IF_THEN40]]:
-// CHECK1-NEXT: [[TMP52:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
-// CHECK1-NEXT: [[TMP53:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
-// CHECK1-NEXT: [[TMP54:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[MUL41:%.*]] = mul i32 [[TMP53]], [[TMP54]]
-// CHECK1-NEXT: [[ADD42:%.*]] = add i32 [[TMP52]], [[MUL41]]
+// CHECK1-NEXT: [[TMP55:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
+// CHECK1-NEXT: [[TMP56:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
+// CHECK1-NEXT: [[TMP57:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[MUL41:%.*]] = mul i32 [[TMP56]], [[TMP57]]
+// CHECK1-NEXT: [[ADD42:%.*]] = add i32 [[TMP55]], [[MUL41]]
// CHECK1-NEXT: store i32 [[ADD42]], ptr [[DOTOMP_IV1]], align 4
-// CHECK1-NEXT: [[TMP55:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
-// CHECK1-NEXT: [[TMP56:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
-// CHECK1-NEXT: [[TMP57:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
-// CHECK1-NEXT: [[MUL43:%.*]] = mul i32 [[TMP56]], [[TMP57]]
-// CHECK1-NEXT: [[SUB44:%.*]] = sub i32 [[TMP55]], [[MUL43]]
+// CHECK1-NEXT: [[TMP58:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
+// CHECK1-NEXT: [[TMP59:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
+// CHECK1-NEXT: [[TMP60:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
+// CHECK1-NEXT: [[MUL43:%.*]] = mul i32 [[TMP59]], [[TMP60]]
+// CHECK1-NEXT: [[SUB44:%.*]] = sub i32 [[TMP58]], [[MUL43]]
// CHECK1-NEXT: store i32 [[SUB44]], ptr [[J]], align 4
-// CHECK1-NEXT: [[TMP58:%.*]] = load i32, ptr [[J]], align 4
-// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP58]])
+// CHECK1-NEXT: [[TMP61:%.*]] = load i32, ptr [[J]], align 4
+// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP61]])
// CHECK1-NEXT: br label %[[IF_END45]]
// CHECK1: [[IF_END45]]:
-// CHECK1-NEXT: [[TMP59:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[TMP60:%.*]] = load i32, ptr [[DOTOMP_NI2]], align 4
-// CHECK1-NEXT: [[CMP46:%.*]] = icmp ult i32 [[TMP59]], [[TMP60]]
+// CHECK1-NEXT: [[TMP62:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[TMP63:%.*]] = load i32, ptr [[DOTOMP_NI2]], align 4
+// CHECK1-NEXT: [[CMP46:%.*]] = icmp ult i32 [[TMP62]], [[TMP63]]
// CHECK1-NEXT: br i1 [[CMP46]], label %[[IF_THEN47:.*]], label %[[IF_END52:.*]]
// CHECK1: [[IF_THEN47]]:
-// CHECK1-NEXT: [[TMP61:%.*]] = load i32, ptr [[DOTOMP_LB2]], align 4
-// CHECK1-NEXT: [[TMP62:%.*]] = load i32, ptr [[DOTOMP_ST2]], align 4
-// CHECK1-NEXT: [[TMP63:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[MUL48:%.*]] = mul i32 [[TMP62]], [[TMP63]]
-// CHECK1-NEXT: [[ADD49:%.*]] = add i32 [[TMP61]], [[MUL48]]
+// CHECK1-NEXT: [[TMP64:%.*]] = load i32, ptr [[DOTOMP_LB2]], align 4
+// CHECK1-NEXT: [[TMP65:%.*]] = load i32, ptr [[DOTOMP_ST2]], align 4
+// CHECK1-NEXT: [[TMP66:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[MUL48:%.*]] = mul i32 [[TMP65]], [[TMP66]]
+// CHECK1-NEXT: [[ADD49:%.*]] = add i32 [[TMP64]], [[MUL48]]
// CHECK1-NEXT: store i32 [[ADD49]], ptr [[DOTOMP_IV2]], align 4
-// CHECK1-NEXT: [[TMP64:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK1-NEXT: [[TMP65:%.*]] = load i32, ptr [[DOTOMP_IV2]], align 4
-// CHECK1-NEXT: [[TMP66:%.*]] = load i32, ptr [[DOTNEW_STEP21]], align 4
-// CHECK1-NEXT: [[MUL50:%.*]] = mul i32 [[TMP65]], [[TMP66]]
-// CHECK1-NEXT: [[ADD51:%.*]] = add i32 [[TMP64]], [[MUL50]]
+// CHECK1-NEXT: [[TMP67:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_17]], align 4
+// CHECK1-NEXT: [[TMP68:%.*]] = load i32, ptr [[DOTOMP_IV2]], align 4
+// CHECK1-NEXT: [[TMP69:%.*]] = load i32, ptr [[DOTNEW_STEP21]], align 4
+// CHECK1-NEXT: [[MUL50:%.*]] = mul i32 [[TMP68]], [[TMP69]]
+// CHECK1-NEXT: [[ADD51:%.*]] = add i32 [[TMP67]], [[MUL50]]
// CHECK1-NEXT: store i32 [[ADD51]], ptr [[K]], align 4
-// CHECK1-NEXT: [[TMP67:%.*]] = load i32, ptr [[K]], align 4
-// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP67]])
+// CHECK1-NEXT: [[TMP70:%.*]] = load i32, ptr [[K]], align 4
+// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP70]])
// CHECK1-NEXT: br label %[[IF_END52]]
// CHECK1: [[IF_END52]]:
// CHECK1-NEXT: br label %[[FOR_INC:.*]]
// CHECK1: [[FOR_INC]]:
-// CHECK1-NEXT: [[TMP68:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[INC:%.*]] = add i32 [[TMP68]], 1
+// CHECK1-NEXT: [[TMP71:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[INC:%.*]] = add i32 [[TMP71]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP5:![0-9]+]]
+// CHECK1-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP4:![0-9]+]]
// CHECK1: [[FOR_END]]:
+// CHECK1-NEXT: [[TMP72:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
+// CHECK1-NEXT: [[TMP73:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK1-NEXT: [[TMP74:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
+// CHECK1-NEXT: [[MUL53:%.*]] = mul i32 [[TMP73]], [[TMP74]]
+// CHECK1-NEXT: [[ADD54:%.*]] = add i32 [[TMP72]], [[MUL53]]
+// CHECK1-NEXT: store i32 [[ADD54]], ptr [[I]], align 4
+// CHECK1-NEXT: [[TMP75:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
+// CHECK1-NEXT: [[TMP76:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK1-NEXT: [[TMP77:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
+// CHECK1-NEXT: [[MUL55:%.*]] = mul i32 [[TMP76]], [[TMP77]]
+// CHECK1-NEXT: [[ADD56:%.*]] = add i32 [[TMP75]], [[MUL55]]
+// CHECK1-NEXT: store i32 [[ADD56]], ptr [[J]], align 4
+// CHECK1-NEXT: [[TMP78:%.*]] = load i32, ptr [[DOTOMP_LB2]], align 4
+// CHECK1-NEXT: [[TMP79:%.*]] = load i32, ptr [[DOTOMP_NI2]], align 4
+// CHECK1-NEXT: [[TMP80:%.*]] = load i32, ptr [[DOTOMP_ST2]], align 4
+// CHECK1-NEXT: [[MUL57:%.*]] = mul i32 [[TMP79]], [[TMP80]]
+// CHECK1-NEXT: [[ADD58:%.*]] = add i32 [[TMP78]], [[MUL57]]
+// CHECK1-NEXT: store i32 [[ADD58]], ptr [[K]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -483,11 +528,13 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[ENTRY:.*:]]
// CHECK1-NEXT: [[ARR:%.*]] = alloca [256 x double], align 16
// CHECK1-NEXT: [[I:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_UB0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_LB0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_ST0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_NI0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_IV0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[J:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_UB1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_LB1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_ST1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_NI1:%.*]] = alloca i32, align 4
@@ -497,43 +544,48 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[DOTOMP_FUSE_INDEX:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTCAPTURE_EXPR_1:%.*]] = alloca i32, align 4
-// CHECK1-NEXT: [[DOTOMP_LB03:%.*]] = alloca i32, align 4
-// CHECK1-NEXT: [[DOTOMP_ST04:%.*]] = alloca i32, align 4
-// CHECK1-NEXT: [[DOTOMP_NI05:%.*]] = alloca i64, align 8
-// CHECK1-NEXT: [[DOTOMP_IV06:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_UB03:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_LB04:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_ST05:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_NI06:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_IV07:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[C:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[__RANGE2:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: [[__END2:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: [[__BEGIN2:%.*]] = alloca ptr, align 8
-// CHECK1-NEXT: [[DOTCAPTURE_EXPR_8:%.*]] = alloca ptr, align 8
-// CHECK1-NEXT: [[DOTCAPTURE_EXPR_10:%.*]] = alloca ptr, align 8
-// CHECK1-NEXT: [[DOTCAPTURE_EXPR_11:%.*]] = alloca i64, align 8
-// CHECK1-NEXT: [[DOTOMP_LB116:%.*]] = alloca i64, align 8
-// CHECK1-NEXT: [[DOTOMP_ST117:%.*]] = alloca i64, align 8
-// CHECK1-NEXT: [[DOTOMP_NI118:%.*]] = alloca i64, align 8
-// CHECK1-NEXT: [[DOTOMP_IV120:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTCAPTURE_EXPR_9:%.*]] = alloca ptr, align 8
+// CHECK1-NEXT: [[DOTCAPTURE_EXPR_11:%.*]] = alloca ptr, align 8
+// CHECK1-NEXT: [[DOTCAPTURE_EXPR_12:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_UB117:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_LB118:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_ST119:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_NI120:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_IV122:%.*]] = alloca i64, align 8
// CHECK1-NEXT: [[CC:%.*]] = alloca i32, align 4
-// CHECK1-NEXT: [[__RANGE221:%.*]] = alloca ptr, align 8
-// CHECK1-NEXT: [[__END222:%.*]] = alloca ptr, align 8
-// CHECK1-NEXT: [[__BEGIN225:%.*]] = alloca ptr, align 8
-// CHECK1-NEXT: [[DOTCAPTURE_EXPR_27:%.*]] = alloca ptr, align 8
+// CHECK1-NEXT: [[__RANGE223:%.*]] = alloca ptr, align 8
+// CHECK1-NEXT: [[__END224:%.*]] = alloca ptr, align 8
+// CHECK1-NEXT: [[__BEGIN227:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: [[DOTCAPTURE_EXPR_29:%.*]] = alloca ptr, align 8
-// CHECK1-NEXT: [[DOTCAPTURE_EXPR_30:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTCAPTURE_EXPR_31:%.*]] = alloca ptr, align 8
+// CHECK1-NEXT: [[DOTCAPTURE_EXPR_32:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_UB2:%.*]] = alloca i64, align 8
// CHECK1-NEXT: [[DOTOMP_LB2:%.*]] = alloca i64, align 8
// CHECK1-NEXT: [[DOTOMP_ST2:%.*]] = alloca i64, align 8
// CHECK1-NEXT: [[DOTOMP_NI2:%.*]] = alloca i64, align 8
// CHECK1-NEXT: [[DOTOMP_IV2:%.*]] = alloca i64, align 8
-// CHECK1-NEXT: [[DOTOMP_TEMP_140:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_TEMP_142:%.*]] = alloca i64, align 8
// CHECK1-NEXT: [[DOTOMP_TEMP_2:%.*]] = alloca i64, align 8
-// CHECK1-NEXT: [[DOTOMP_FUSE_MAX46:%.*]] = alloca i64, align 8
-// CHECK1-NEXT: [[DOTOMP_FUSE_INDEX52:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_FUSE_MAX48:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_FUSE_INDEX54:%.*]] = alloca i64, align 8
// CHECK1-NEXT: [[V:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: [[VV:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: store i32 0, ptr [[I]], align 4
+// CHECK1-NEXT: store i32 127, ptr [[DOTOMP_UB0]], align 4
// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_LB0]], align 4
// CHECK1-NEXT: store i32 1, ptr [[DOTOMP_ST0]], align 4
// CHECK1-NEXT: store i32 128, ptr [[DOTOMP_NI0]], align 4
// CHECK1-NEXT: store i32 0, ptr [[J]], align 4
+// CHECK1-NEXT: store i32 127, ptr [[DOTOMP_UB1]], align 4
// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_LB1]], align 4
// CHECK1-NEXT: store i32 1, ptr [[DOTOMP_ST1]], align 4
// CHECK1-NEXT: store i32 128, ptr [[DOTOMP_NI1]], align 4
@@ -560,220 +612,226 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[DIV:%.*]] = sdiv i32 [[SUB]], 1
// CHECK1-NEXT: [[SUB2:%.*]] = sub nsw i32 [[DIV]], 1
// CHECK1-NEXT: store i32 [[SUB2]], ptr [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_LB03]], align 4
-// CHECK1-NEXT: store i32 1, ptr [[DOTOMP_ST04]], align 4
// CHECK1-NEXT: [[TMP7:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK1-NEXT: [[ADD:%.*]] = add nsw i32 [[TMP7]], 1
+// CHECK1-NEXT: store i32 [[TMP7]], ptr [[DOTOMP_UB03]], align 4
+// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_LB04]], align 4
+// CHECK1-NEXT: store i32 1, ptr [[DOTOMP_ST05]], align 4
+// CHECK1-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK1-NEXT: [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
// CHECK1-NEXT: [[CONV:%.*]] = sext i32 [[ADD]] to i64
-// CHECK1-NEXT: store i64 [[CONV]], ptr [[DOTOMP_NI05]], align 8
+// CHECK1-NEXT: store i64 [[CONV]], ptr [[DOTOMP_NI06]], align 8
// CHECK1-NEXT: store i32 42, ptr [[C]], align 4
// CHECK1-NEXT: store ptr [[ARR]], ptr [[__RANGE2]], align 8
-// CHECK1-NEXT: [[TMP8:%.*]] = load ptr, ptr [[__RANGE2]], align 8
-// CHECK1-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP8]], i64 0, i64 0
+// CHECK1-NEXT: [[TMP9:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META5:![0-9]+]], !align [[META6:![0-9]+]]
+// CHECK1-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP9]], i64 0, i64 0
// CHECK1-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 256
// CHECK1-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK1-NEXT: [[TMP9:%.*]] = load ptr, ptr [[__RANGE2]], align 8
-// CHECK1-NEXT: [[ARRAYDECAY7:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP9]], i64 0, i64 0
-// CHECK1-NEXT: store ptr [[ARRAYDECAY7]], ptr [[__BEGIN2]], align 8
-// CHECK1-NEXT: [[TMP10:%.*]] = load ptr, ptr [[__RANGE2]], align 8
-// CHECK1-NEXT: [[ARRAYDECAY9:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP10]], i64 0, i64 0
-// CHECK1-NEXT: store ptr [[ARRAYDECAY9]], ptr [[DOTCAPTURE_EXPR_8]], align 8
-// CHECK1-NEXT: [[TMP11:%.*]] = load ptr, ptr [[__END2]], align 8
-// CHECK1-NEXT: store ptr [[TMP11]], ptr [[DOTCAPTURE_EXPR_10]], align 8
-// CHECK1-NEXT: [[TMP12:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_10]], align 8
-// CHECK1-NEXT: [[TMP13:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_8]], align 8
-// CHECK1-NEXT: [[SUB_PTR_LHS_CAST:%.*]] = ptrtoint ptr [[TMP12]] to i64
-// CHECK1-NEXT: [[SUB_PTR_RHS_CAST:%.*]] = ptrtoint ptr [[TMP13]] to i64
+// CHECK1-NEXT: [[TMP10:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META5]], !align [[META6]]
+// CHECK1-NEXT: [[ARRAYDECAY8:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP10]], i64 0, i64 0
+// CHECK1-NEXT: store ptr [[ARRAYDECAY8]], ptr [[__BEGIN2]], align 8
+// CHECK1-NEXT: [[TMP11:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META5]], !align [[META6]]
+// CHECK1-NEXT: [[ARRAYDECAY10:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP11]], i64 0, i64 0
+// CHECK1-NEXT: store ptr [[ARRAYDECAY10]], ptr [[DOTCAPTURE_EXPR_9]], align 8
+// CHECK1-NEXT: [[TMP12:%.*]] = load ptr, ptr [[__END2]], align 8
+// CHECK1-NEXT: store ptr [[TMP12]], ptr [[DOTCAPTURE_EXPR_11]], align 8
+// CHECK1-NEXT: [[TMP13:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_11]], align 8
+// CHECK1-NEXT: [[TMP14:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_9]], align 8
+// CHECK1-NEXT: [[SUB_PTR_LHS_CAST:%.*]] = ptrtoint ptr [[TMP13]] to i64
+// CHECK1-NEXT: [[SUB_PTR_RHS_CAST:%.*]] = ptrtoint ptr [[TMP14]] to i64
// CHECK1-NEXT: [[SUB_PTR_SUB:%.*]] = sub i64 [[SUB_PTR_LHS_CAST]], [[SUB_PTR_RHS_CAST]]
// CHECK1-NEXT: [[SUB_PTR_DIV:%.*]] = sdiv exact i64 [[SUB_PTR_SUB]], 8
-// CHECK1-NEXT: [[SUB12:%.*]] = sub nsw i64 [[SUB_PTR_DIV]], 1
-// CHECK1-NEXT: [[ADD13:%.*]] = add nsw i64 [[SUB12]], 1
-// CHECK1-NEXT: [[DIV14:%.*]] = sdiv i64 [[ADD13]], 1
-// CHECK1-NEXT: [[SUB15:%.*]] = sub nsw i64 [[DIV14]], 1
-// CHECK1-NEXT: store i64 [[SUB15]], ptr [[DOTCAPTURE_EXPR_11]], align 8
-// CHECK1-NEXT: store i64 0, ptr [[DOTOMP_LB116]], align 8
-// CHECK1-NEXT: store i64 1, ptr [[DOTOMP_ST117]], align 8
-// CHECK1-NEXT: [[TMP14:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_11]], align 8
-// CHECK1-NEXT: [[ADD19:%.*]] = add nsw i64 [[TMP14]], 1
-// CHECK1-NEXT: store i64 [[ADD19]], ptr [[DOTOMP_NI118]], align 8
+// CHECK1-NEXT: [[SUB13:%.*]] = sub nsw i64 [[SUB_PTR_DIV]], 1
+// CHECK1-NEXT: [[ADD14:%.*]] = add nsw i64 [[SUB13]], 1
+// CHECK1-NEXT: [[DIV15:%.*]] = sdiv i64 [[ADD14]], 1
+// CHECK1-NEXT: [[SUB16:%.*]] = sub nsw i64 [[DIV15]], 1
+// CHECK1-NEXT: store i64 [[SUB16]], ptr [[DOTCAPTURE_EXPR_12]], align 8
+// CHECK1-NEXT: [[TMP15:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_12]], align 8
+// CHECK1-NEXT: store i64 [[TMP15]], ptr [[DOTOMP_UB117]], align 8
+// CHECK1-NEXT: store i64 0, ptr [[DOTOMP_LB118]], align 8
+// CHECK1-NEXT: store i64 1, ptr [[DOTOMP_ST119]], align 8
+// CHECK1-NEXT: [[TMP16:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_12]], align 8
+// CHECK1-NEXT: [[ADD21:%.*]] = add nsw i64 [[TMP16]], 1
+// CHECK1-NEXT: store i64 [[ADD21]], ptr [[DOTOMP_NI120]], align 8
// CHECK1-NEXT: store i32 37, ptr [[CC]], align 4
-// CHECK1-NEXT: store ptr [[ARR]], ptr [[__RANGE221]], align 8
-// CHECK1-NEXT: [[TMP15:%.*]] = load ptr, ptr [[__RANGE221]], align 8
-// CHECK1-NEXT: [[ARRAYDECAY23:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP15]], i64 0, i64 0
-// CHECK1-NEXT: [[ADD_PTR24:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY23]], i64 256
-// CHECK1-NEXT: store ptr [[ADD_PTR24]], ptr [[__END222]], align 8
-// CHECK1-NEXT: [[TMP16:%.*]] = load ptr, ptr [[__RANGE221]], align 8
-// CHECK1-NEXT: [[ARRAYDECAY26:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP16]], i64 0, i64 0
-// CHECK1-NEXT: store ptr [[ARRAYDECAY26]], ptr [[__BEGIN225]], align 8
-// CHECK1-NEXT: [[TMP17:%.*]] = load ptr, ptr [[__RANGE221]], align 8
-// CHECK1-NEXT: [[ARRAYDECAY28:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP17]], i64 0, i64 0
-// CHECK1-NEXT: store ptr [[ARRAYDECAY28]], ptr [[DOTCAPTURE_EXPR_27]], align 8
-// CHECK1-NEXT: [[TMP18:%.*]] = load ptr, ptr [[__END222]], align 8
-// CHECK1-NEXT: store ptr [[TMP18]], ptr [[DOTCAPTURE_EXPR_29]], align 8
-// CHECK1-NEXT: [[TMP19:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_29]], align 8
-// CHECK1-NEXT: [[TMP20:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_27]], align 8
-// CHECK1-NEXT: [[SUB_PTR_LHS_CAST31:%.*]] = ptrtoint ptr [[TMP19]] to i64
-// CHECK1-NEXT: [[SUB_PTR_RHS_CAST32:%.*]] = ptrtoint ptr [[TMP20]] to i64
-// CHECK1-NEXT: [[SUB_PTR_SUB33:%.*]] = sub i64 [[SUB_PTR_LHS_CAST31]], [[SUB_PTR_RHS_CAST32]]
-// CHECK1-NEXT: [[SUB_PTR_DIV34:%.*]] = sdiv exact i64 [[SUB_PTR_SUB33]], 8
-// CHECK1-NEXT: [[SUB35:%.*]] = sub nsw i64 [[SUB_PTR_DIV34]], 1
-// CHECK1-NEXT: [[ADD36:%.*]] = add nsw i64 [[SUB35]], 1
-// CHECK1-NEXT: [[DIV37:%.*]] = sdiv i64 [[ADD36]], 1
-// CHECK1-NEXT: [[SUB38:%.*]] = sub nsw i64 [[DIV37]], 1
-// CHECK1-NEXT: store i64 [[SUB38]], ptr [[DOTCAPTURE_EXPR_30]], align 8
+// CHECK1-NEXT: store ptr [[ARR]], ptr [[__RANGE223]], align 8
+// CHECK1-NEXT: [[TMP17:%.*]] = load ptr, ptr [[__RANGE223]], align 8, !nonnull [[META5]], !align [[META6]]
+// CHECK1-NEXT: [[ARRAYDECAY25:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP17]], i64 0, i64 0
+// CHECK1-NEXT: [[ADD_PTR26:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY25]], i64 256
+// CHECK1-NEXT: store ptr [[ADD_PTR26]], ptr [[__END224]], align 8
+// CHECK1-NEXT: [[TMP18:%.*]] = load ptr, ptr [[__RANGE223]], align 8, !nonnull [[META5]], !align [[META6]]
+// CHECK1-NEXT: [[ARRAYDECAY28:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP18]], i64 0, i64 0
+// CHECK1-NEXT: store ptr [[ARRAYDECAY28]], ptr [[__BEGIN227]], align 8
+// CHECK1-NEXT: [[TMP19:%.*]] = load ptr, ptr [[__RANGE223]], align 8, !nonnull [[META5]], !align [[META6]]
+// CHECK1-NEXT: [[ARRAYDECAY30:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP19]], i64 0, i64 0
+// CHECK1-NEXT: store ptr [[ARRAYDECAY30]], ptr [[DOTCAPTURE_EXPR_29]], align 8
+// CHECK1-NEXT: [[TMP20:%.*]] = load ptr, ptr [[__END224]], align 8
+// CHECK1-NEXT: store ptr [[TMP20]], ptr [[DOTCAPTURE_EXPR_31]], align 8
+// CHECK1-NEXT: [[TMP21:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_31]], align 8
+// CHECK1-NEXT: [[TMP22:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_29]], align 8
+// CHECK1-NEXT: [[SUB_PTR_LHS_CAST33:%.*]] = ptrtoint ptr [[TMP21]] to i64
+// CHECK1-NEXT: [[SUB_PTR_RHS_CAST34:%.*]] = ptrtoint ptr [[TMP22]] to i64
+// CHECK1-NEXT: [[SUB_PTR_SUB35:%.*]] = sub i64 [[SUB_PTR_LHS_CAST33]], [[SUB_PTR_RHS_CAST34]]
+// CHECK1-NEXT: [[SUB_PTR_DIV36:%.*]] = sdiv exact i64 [[SUB_PTR_SUB35]], 8
+// CHECK1-NEXT: [[SUB37:%.*]] = sub nsw i64 [[SUB_PTR_DIV36]], 1
+// CHECK1-NEXT: [[ADD38:%.*]] = add nsw i64 [[SUB37]], 1
+// CHECK1-NEXT: [[DIV39:%.*]] = sdiv i64 [[ADD38]], 1
+// CHECK1-NEXT: [[SUB40:%.*]] = sub nsw i64 [[DIV39]], 1
+// CHECK1-NEXT: store i64 [[SUB40]], ptr [[DOTCAPTURE_EXPR_32]], align 8
+// CHECK1-NEXT: [[TMP23:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_32]], align 8
+// CHECK1-NEXT: store i64 [[TMP23]], ptr [[DOTOMP_UB2]], align 8
// CHECK1-NEXT: store i64 0, ptr [[DOTOMP_LB2]], align 8
// CHECK1-NEXT: store i64 1, ptr [[DOTOMP_ST2]], align 8
-// CHECK1-NEXT: [[TMP21:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_30]], align 8
-// CHECK1-NEXT: [[ADD39:%.*]] = add nsw i64 [[TMP21]], 1
-// CHECK1-NEXT: store i64 [[ADD39]], ptr [[DOTOMP_NI2]], align 8
-// CHECK1-NEXT: [[TMP22:%.*]] = load i64, ptr [[DOTOMP_NI05]], align 8
-// CHECK1-NEXT: store i64 [[TMP22]], ptr [[DOTOMP_TEMP_140]], align 8
-// CHECK1-NEXT: [[TMP23:%.*]] = load i64, ptr [[DOTOMP_TEMP_140]], align 8
-// CHECK1-NEXT: [[TMP24:%.*]] = load i64, ptr [[DOTOMP_NI118]], align 8
-// CHECK1-NEXT: [[CMP41:%.*]] = icmp sgt i64 [[TMP23]], [[TMP24]]
-// CHECK1-NEXT: br i1 [[CMP41]], label %[[COND_TRUE42:.*]], label %[[COND_FALSE43:.*]]
-// CHECK1: [[COND_TRUE42]]:
-// CHECK1-NEXT: [[TMP25:%.*]] = load i64, ptr [[DOTOMP_TEMP_140]], align 8
-// CHECK1-NEXT: br label %[[COND_END44:.*]]
-// CHECK1: [[COND_FALSE43]]:
-// CHECK1-NEXT: [[TMP26:%.*]] = load i64, ptr [[DOTOMP_NI118]], align 8
-// CHECK1-NEXT: br label %[[COND_END44]]
-// CHECK1: [[COND_END44]]:
-// CHECK1-NEXT: [[COND45:%.*]] = phi i64 [ [[TMP25]], %[[COND_TRUE42]] ], [ [[TMP26]], %[[COND_FALSE43]] ]
-// CHECK1-NEXT: store i64 [[COND45]], ptr [[DOTOMP_TEMP_2]], align 8
-// CHECK1-NEXT: [[TMP27:%.*]] = load i64, ptr [[DOTOMP_TEMP_2]], align 8
-// CHECK1-NEXT: [[TMP28:%.*]] = load i64, ptr [[DOTOMP_NI2]], align 8
-// CHECK1-NEXT: [[CMP47:%.*]] = icmp sgt i64 [[TMP27]], [[TMP28]]
-// CHECK1-NEXT: br i1 [[CMP47]], label %[[COND_TRUE48:.*]], label %[[COND_FALSE49:.*]]
-// CHECK1: [[COND_TRUE48]]:
-// CHECK1-NEXT: [[TMP29:%.*]] = load i64, ptr [[DOTOMP_TEMP_2]], align 8
-// CHECK1-NEXT: br label %[[COND_END50:.*]]
-// CHECK1: [[COND_FALSE49]]:
-// CHECK1-NEXT: [[TMP30:%.*]] = load i64, ptr [[DOTOMP_NI2]], align 8
-// CHECK1-NEXT: br label %[[COND_END50]]
-// CHECK1: [[COND_END50]]:
-// CHECK1-NEXT: [[COND51:%.*]] = phi i64 [ [[TMP29]], %[[COND_TRUE48]] ], [ [[TMP30]], %[[COND_FALSE49]] ]
-// CHECK1-NEXT: store i64 [[COND51]], ptr [[DOTOMP_FUSE_MAX46]], align 8
-// CHECK1-NEXT: store i64 0, ptr [[DOTOMP_FUSE_INDEX52]], align 8
+// CHECK1-NEXT: [[TMP24:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_32]], align 8
+// CHECK1-NEXT: [[ADD41:%.*]] = add nsw i64 [[TMP24]], 1
+// CHECK1-NEXT: store i64 [[ADD41]], ptr [[DOTOMP_NI2]], align 8
+// CHECK1-NEXT: [[TMP25:%.*]] = load i64, ptr [[DOTOMP_NI06]], align 8
+// CHECK1-NEXT: store i64 [[TMP25]], ptr [[DOTOMP_TEMP_142]], align 8
+// CHECK1-NEXT: [[TMP26:%.*]] = load i64, ptr [[DOTOMP_TEMP_142]], align 8
+// CHECK1-NEXT: [[TMP27:%.*]] = load i64, ptr [[DOTOMP_NI120]], align 8
+// CHECK1-NEXT: [[CMP43:%.*]] = icmp sgt i64 [[TMP26]], [[TMP27]]
+// CHECK1-NEXT: br i1 [[CMP43]], label %[[COND_TRUE44:.*]], label %[[COND_FALSE45:.*]]
+// CHECK1: [[COND_TRUE44]]:
+// CHECK1-NEXT: [[TMP28:%.*]] = load i64, ptr [[DOTOMP_TEMP_142]], align 8
+// CHECK1-NEXT: br label %[[COND_END46:.*]]
+// CHECK1: [[COND_FALSE45]]:
+// CHECK1-NEXT: [[TMP29:%.*]] = load i64, ptr [[DOTOMP_NI120]], align 8
+// CHECK1-NEXT: br label %[[COND_END46]]
+// CHECK1: [[COND_END46]]:
+// CHECK1-NEXT: [[COND47:%.*]] = phi i64 [ [[TMP28]], %[[COND_TRUE44]] ], [ [[TMP29]], %[[COND_FALSE45]] ]
+// CHECK1-NEXT: store i64 [[COND47]], ptr [[DOTOMP_TEMP_2]], align 8
+// CHECK1-NEXT: [[TMP30:%.*]] = load i64, ptr [[DOTOMP_TEMP_2]], align 8
+// CHECK1-NEXT: [[TMP31:%.*]] = load i64, ptr [[DOTOMP_NI2]], align 8
+// CHECK1-NEXT: [[CMP49:%.*]] = icmp sgt i64 [[TMP30]], [[TMP31]]
+// CHECK1-NEXT: br i1 [[CMP49]], label %[[COND_TRUE50:.*]], label %[[COND_FALSE51:.*]]
+// CHECK1: [[COND_TRUE50]]:
+// CHECK1-NEXT: [[TMP32:%.*]] = load i64, ptr [[DOTOMP_TEMP_2]], align 8
+// CHECK1-NEXT: br label %[[COND_END52:.*]]
+// CHECK1: [[COND_FALSE51]]:
+// CHECK1-NEXT: [[TMP33:%.*]] = load i64, ptr [[DOTOMP_NI2]], align 8
+// CHECK1-NEXT: br label %[[COND_END52]]
+// CHECK1: [[COND_END52]]:
+// CHECK1-NEXT: [[COND53:%.*]] = phi i64 [ [[TMP32]], %[[COND_TRUE50]] ], [ [[TMP33]], %[[COND_FALSE51]] ]
+// CHECK1-NEXT: store i64 [[COND53]], ptr [[DOTOMP_FUSE_MAX48]], align 8
+// CHECK1-NEXT: store i64 0, ptr [[DOTOMP_FUSE_INDEX54]], align 8
// CHECK1-NEXT: br label %[[FOR_COND:.*]]
// CHECK1: [[FOR_COND]]:
-// CHECK1-NEXT: [[TMP31:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK1-NEXT: [[TMP32:%.*]] = load i64, ptr [[DOTOMP_FUSE_MAX46]], align 8
-// CHECK1-NEXT: [[CMP53:%.*]] = icmp slt i64 [[TMP31]], [[TMP32]]
-// CHECK1-NEXT: br i1 [[CMP53]], label %[[FOR_BODY:.*]], label %[[FOR_END:.*]]
+// CHECK1-NEXT: [[TMP34:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK1-NEXT: [[TMP35:%.*]] = load i64, ptr [[DOTOMP_FUSE_MAX48]], align 8
+// CHECK1-NEXT: [[CMP55:%.*]] = icmp slt i64 [[TMP34]], [[TMP35]]
+// CHECK1-NEXT: br i1 [[CMP55]], label %[[FOR_BODY:.*]], label %[[FOR_END:.*]]
// CHECK1: [[FOR_BODY]]:
-// CHECK1-NEXT: [[TMP33:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK1-NEXT: [[TMP34:%.*]] = load i64, ptr [[DOTOMP_NI05]], align 8
-// CHECK1-NEXT: [[CMP54:%.*]] = icmp slt i64 [[TMP33]], [[TMP34]]
-// CHECK1-NEXT: br i1 [[CMP54]], label %[[IF_THEN:.*]], label %[[IF_END74:.*]]
+// CHECK1-NEXT: [[TMP36:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK1-NEXT: [[TMP37:%.*]] = load i64, ptr [[DOTOMP_NI06]], align 8
+// CHECK1-NEXT: [[CMP56:%.*]] = icmp slt i64 [[TMP36]], [[TMP37]]
+// CHECK1-NEXT: br i1 [[CMP56]], label %[[IF_THEN:.*]], label %[[IF_END76:.*]]
// CHECK1: [[IF_THEN]]:
-// CHECK1-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTOMP_LB03]], align 4
-// CHECK1-NEXT: [[CONV55:%.*]] = sext i32 [[TMP35]] to i64
-// CHECK1-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_ST04]], align 4
-// CHECK1-NEXT: [[CONV56:%.*]] = sext i32 [[TMP36]] to i64
-// CHECK1-NEXT: [[TMP37:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK1-NEXT: [[MUL:%.*]] = mul nsw i64 [[CONV56]], [[TMP37]]
-// CHECK1-NEXT: [[ADD57:%.*]] = add nsw i64 [[CONV55]], [[MUL]]
-// CHECK1-NEXT: [[CONV58:%.*]] = trunc i64 [[ADD57]] to i32
-// CHECK1-NEXT: store i32 [[CONV58]], ptr [[DOTOMP_IV06]], align 4
-// CHECK1-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_IV06]], align 4
-// CHECK1-NEXT: [[MUL59:%.*]] = mul nsw i32 [[TMP38]], 1
-// CHECK1-NEXT: [[ADD60:%.*]] = add nsw i32 0, [[MUL59]]
-// CHECK1-NEXT: store i32 [[ADD60]], ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
-// CHECK1-NEXT: [[CMP61:%.*]] = icmp slt i32 [[TMP39]], [[TMP40]]
-// CHECK1-NEXT: br i1 [[CMP61]], label %[[IF_THEN62:.*]], label %[[IF_END:.*]]
-// CHECK1: [[IF_THEN62]]:
-// CHECK1-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
-// CHECK1-NEXT: [[TMP42:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
-// CHECK1-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[MUL63:%.*]] = mul nsw i32 [[TMP42]], [[TMP43]]
-// CHECK1-NEXT: [[ADD64:%.*]] = add nsw i32 [[TMP41]], [[MUL63]]
-// CHECK1-NEXT: store i32 [[ADD64]], ptr [[DOTOMP_IV0]], align 4
-// CHECK1-NEXT: [[TMP44:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
-// CHECK1-NEXT: [[MUL65:%.*]] = mul nsw i32 [[TMP44]], 1
-// CHECK1-NEXT: [[ADD66:%.*]] = add nsw i32 0, [[MUL65]]
-// CHECK1-NEXT: store i32 [[ADD66]], ptr [[I]], align 4
-// CHECK1-NEXT: [[TMP45:%.*]] = load i32, ptr [[I]], align 4
-// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP45]])
+// CHECK1-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_LB04]], align 4
+// CHECK1-NEXT: [[CONV57:%.*]] = sext i32 [[TMP38]] to i64
+// CHECK1-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTOMP_ST05]], align 4
+// CHECK1-NEXT: [[CONV58:%.*]] = sext i32 [[TMP39]] to i64
+// CHECK1-NEXT: [[TMP40:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK1-NEXT: [[MUL:%.*]] = mul nsw i64 [[CONV58]], [[TMP40]]
+// CHECK1-NEXT: [[ADD59:%.*]] = add nsw i64 [[CONV57]], [[MUL]]
+// CHECK1-NEXT: [[CONV60:%.*]] = trunc i64 [[ADD59]] to i32
+// CHECK1-NEXT: store i32 [[CONV60]], ptr [[DOTOMP_IV07]], align 4
+// CHECK1-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTOMP_IV07]], align 4
+// CHECK1-NEXT: [[MUL61:%.*]] = mul nsw i32 [[TMP41]], 1
+// CHECK1-NEXT: [[ADD62:%.*]] = add nsw i32 0, [[MUL61]]
+// CHECK1-NEXT: store i32 [[ADD62]], ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[TMP42:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK1-NEXT: [[CMP63:%.*]] = icmp slt i32 [[TMP42]], [[TMP43]]
+// CHECK1-NEXT: br i1 [[CMP63]], label %[[IF_THEN64:.*]], label %[[IF_END:.*]]
+// CHECK1: [[IF_THEN64]]:
+// CHECK1-NEXT: [[TMP44:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
+// CHECK1-NEXT: [[TMP45:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
+// CHECK1-NEXT: [[TMP46:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[MUL65:%.*]] = mul nsw i32 [[TMP45]], [[TMP46]]
+// CHECK1-NEXT: [[ADD66:%.*]] = add nsw i32 [[TMP44]], [[MUL65]]
+// CHECK1-NEXT: store i32 [[ADD66]], ptr [[DOTOMP_IV0]], align 4
+// CHECK1-NEXT: [[TMP47:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
+// CHECK1-NEXT: [[MUL67:%.*]] = mul nsw i32 [[TMP47]], 1
+// CHECK1-NEXT: [[ADD68:%.*]] = add nsw i32 0, [[MUL67]]
+// CHECK1-NEXT: store i32 [[ADD68]], ptr [[I]], align 4
+// CHECK1-NEXT: [[TMP48:%.*]] = load i32, ptr [[I]], align 4
+// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP48]])
// CHECK1-NEXT: br label %[[IF_END]]
// CHECK1: [[IF_END]]:
-// CHECK1-NEXT: [[TMP46:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[TMP47:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
-// CHECK1-NEXT: [[CMP67:%.*]] = icmp slt i32 [[TMP46]], [[TMP47]]
-// CHECK1-NEXT: br i1 [[CMP67]], label %[[IF_THEN68:.*]], label %[[IF_END73:.*]]
-// CHECK1: [[IF_THEN68]]:
-// CHECK1-NEXT: [[TMP48:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
-// CHECK1-NEXT: [[TMP49:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
-// CHECK1-NEXT: [[TMP50:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[MUL69:%.*]] = mul nsw i32 [[TMP49]], [[TMP50]]
-// CHECK1-NEXT: [[ADD70:%.*]] = add nsw i32 [[TMP48]], [[MUL69]]
-// CHECK1-NEXT: store i32 [[ADD70]], ptr [[DOTOMP_IV1]], align 4
-// CHECK1-NEXT: [[TMP51:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
-// CHECK1-NEXT: [[MUL71:%.*]] = mul nsw i32 [[TMP51]], 2
-// CHECK1-NEXT: [[ADD72:%.*]] = add nsw i32 0, [[MUL71]]
-// CHECK1-NEXT: store i32 [[ADD72]], ptr [[J]], align 4
-// CHECK1-NEXT: [[TMP52:%.*]] = load i32, ptr [[J]], align 4
-// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP52]])
-// CHECK1-NEXT: br label %[[IF_END73]]
-// CHECK1: [[IF_END73]]:
-// CHECK1-NEXT: br label %[[IF_END74]]
-// CHECK1: [[IF_END74]]:
-// CHECK1-NEXT: [[TMP53:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK1-NEXT: [[TMP54:%.*]] = load i64, ptr [[DOTOMP_NI118]], align 8
-// CHECK1-NEXT: [[CMP75:%.*]] = icmp slt i64 [[TMP53]], [[TMP54]]
-// CHECK1-NEXT: br i1 [[CMP75]], label %[[IF_THEN76:.*]], label %[[IF_END81:.*]]
-// CHECK1: [[IF_THEN76]]:
-// CHECK1-NEXT: [[TMP55:%.*]] = load i64, ptr [[DOTOMP_LB116]], align 8
-// CHECK1-NEXT: [[TMP56:%.*]] = load i64, ptr [[DOTOMP_ST117]], align 8
-// CHECK1-NEXT: [[TMP57:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK1-NEXT: [[MUL77:%.*]] = mul nsw i64 [[TMP56]], [[TMP57]]
-// CHECK1-NEXT: [[ADD78:%.*]] = add nsw i64 [[TMP55]], [[MUL77]]
-// CHECK1-NEXT: store i64 [[ADD78]], ptr [[DOTOMP_IV120]], align 8
-// CHECK1-NEXT: [[TMP58:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_8]], align 8
-// CHECK1-NEXT: [[TMP59:%.*]] = load i64, ptr [[DOTOMP_IV120]], align 8
-// CHECK1-NEXT: [[MUL79:%.*]] = mul nsw i64 [[TMP59]], 1
-// CHECK1-NEXT: [[ADD_PTR80:%.*]] = getelementptr inbounds double, ptr [[TMP58]], i64 [[MUL79]]
-// CHECK1-NEXT: store ptr [[ADD_PTR80]], ptr [[__BEGIN2]], align 8
-// CHECK1-NEXT: [[TMP60:%.*]] = load ptr, ptr [[__BEGIN2]], align 8
-// CHECK1-NEXT: store ptr [[TMP60]], ptr [[V]], align 8
-// CHECK1-NEXT: [[TMP61:%.*]] = load i32, ptr [[C]], align 4
-// CHECK1-NEXT: [[TMP62:%.*]] = load ptr, ptr [[V]], align 8
-// CHECK1-NEXT: [[TMP63:%.*]] = load double, ptr [[TMP62]], align 8
-// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP61]], double noundef [[TMP63]])
-// CHECK1-NEXT: br label %[[IF_END81]]
-// CHECK1: [[IF_END81]]:
-// CHECK1-NEXT: [[TMP64:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK1-NEXT: [[TMP65:%.*]] = load i64, ptr [[DOTOMP_NI2]], align 8
-// CHECK1-NEXT: [[CMP82:%.*]] = icmp slt i64 [[TMP64]], [[TMP65]]
-// CHECK1-NEXT: br i1 [[CMP82]], label %[[IF_THEN83:.*]], label %[[IF_END88:.*]]
-// CHECK1: [[IF_THEN83]]:
-// CHECK1-NEXT: [[TMP66:%.*]] = load i64, ptr [[DOTOMP_LB2]], align 8
-// CHECK1-NEXT: [[TMP67:%.*]] = load i64, ptr [[DOTOMP_ST2]], align 8
-// CHECK1-NEXT: [[TMP68:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK1-NEXT: [[MUL84:%.*]] = mul nsw i64 [[TMP67]], [[TMP68]]
-// CHECK1-NEXT: [[ADD85:%.*]] = add nsw i64 [[TMP66]], [[MUL84]]
-// CHECK1-NEXT: store i64 [[ADD85]], ptr [[DOTOMP_IV2]], align 8
-// CHECK1-NEXT: [[TMP69:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_27]], align 8
-// CHECK1-NEXT: [[TMP70:%.*]] = load i64, ptr [[DOTOMP_IV2]], align 8
-// CHECK1-NEXT: [[MUL86:%.*]] = mul nsw i64 [[TMP70]], 1
-// CHECK1-NEXT: [[ADD_PTR87:%.*]] = getelementptr inbounds double, ptr [[TMP69]], i64 [[MUL86]]
-// CHECK1-NEXT: store ptr [[ADD_PTR87]], ptr [[__BEGIN225]], align 8
-// CHECK1-NEXT: [[TMP71:%.*]] = load ptr, ptr [[__BEGIN225]], align 8
-// CHECK1-NEXT: store ptr [[TMP71]], ptr [[VV]], align 8
-// CHECK1-NEXT: [[TMP72:%.*]] = load i32, ptr [[CC]], align 4
-// CHECK1-NEXT: [[TMP73:%.*]] = load ptr, ptr [[VV]], align 8
-// CHECK1-NEXT: [[TMP74:%.*]] = load double, ptr [[TMP73]], align 8
-// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP72]], double noundef [[TMP74]])
-// CHECK1-NEXT: br label %[[IF_END88]]
-// CHECK1: [[IF_END88]]:
+// CHECK1-NEXT: [[TMP49:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[TMP50:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK1-NEXT: [[CMP69:%.*]] = icmp slt i32 [[TMP49]], [[TMP50]]
+// CHECK1-NEXT: br i1 [[CMP69]], label %[[IF_THEN70:.*]], label %[[IF_END75:.*]]
+// CHECK1: [[IF_THEN70]]:
+// CHECK1-NEXT: [[TMP51:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
+// CHECK1-NEXT: [[TMP52:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
+// CHECK1-NEXT: [[TMP53:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[MUL71:%.*]] = mul nsw i32 [[TMP52]], [[TMP53]]
+// CHECK1-NEXT: [[ADD72:%.*]] = add nsw i32 [[TMP51]], [[MUL71]]
+// CHECK1-NEXT: store i32 [[ADD72]], ptr [[DOTOMP_IV1]], align 4
+// CHECK1-NEXT: [[TMP54:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
+// CHECK1-NEXT: [[MUL73:%.*]] = mul nsw i32 [[TMP54]], 2
+// CHECK1-NEXT: [[ADD74:%.*]] = add nsw i32 0, [[MUL73]]
+// CHECK1-NEXT: store i32 [[ADD74]], ptr [[J]], align 4
+// CHECK1-NEXT: [[TMP55:%.*]] = load i32, ptr [[J]], align 4
+// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP55]])
+// CHECK1-NEXT: br label %[[IF_END75]]
+// CHECK1: [[IF_END75]]:
+// CHECK1-NEXT: br label %[[IF_END76]]
+// CHECK1: [[IF_END76]]:
+// CHECK1-NEXT: [[TMP56:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK1-NEXT: [[TMP57:%.*]] = load i64, ptr [[DOTOMP_NI120]], align 8
+// CHECK1-NEXT: [[CMP77:%.*]] = icmp slt i64 [[TMP56]], [[TMP57]]
+// CHECK1-NEXT: br i1 [[CMP77]], label %[[IF_THEN78:.*]], label %[[IF_END83:.*]]
+// CHECK1: [[IF_THEN78]]:
+// CHECK1-NEXT: [[TMP58:%.*]] = load i64, ptr [[DOTOMP_LB118]], align 8
+// CHECK1-NEXT: [[TMP59:%.*]] = load i64, ptr [[DOTOMP_ST119]], align 8
+// CHECK1-NEXT: [[TMP60:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK1-NEXT: [[MUL79:%.*]] = mul nsw i64 [[TMP59]], [[TMP60]]
+// CHECK1-NEXT: [[ADD80:%.*]] = add nsw i64 [[TMP58]], [[MUL79]]
+// CHECK1-NEXT: store i64 [[ADD80]], ptr [[DOTOMP_IV122]], align 8
+// CHECK1-NEXT: [[TMP61:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_9]], align 8
+// CHECK1-NEXT: [[TMP62:%.*]] = load i64, ptr [[DOTOMP_IV122]], align 8
+// CHECK1-NEXT: [[MUL81:%.*]] = mul nsw i64 [[TMP62]], 1
+// CHECK1-NEXT: [[ADD_PTR82:%.*]] = getelementptr inbounds double, ptr [[TMP61]], i64 [[MUL81]]
+// CHECK1-NEXT: store ptr [[ADD_PTR82]], ptr [[__BEGIN2]], align 8
+// CHECK1-NEXT: [[TMP63:%.*]] = load ptr, ptr [[__BEGIN2]], align 8
+// CHECK1-NEXT: store ptr [[TMP63]], ptr [[V]], align 8
+// CHECK1-NEXT: [[TMP64:%.*]] = load i32, ptr [[C]], align 4
+// CHECK1-NEXT: [[TMP65:%.*]] = load ptr, ptr [[V]], align 8, !nonnull [[META5]], !align [[META6]]
+// CHECK1-NEXT: [[TMP66:%.*]] = load double, ptr [[TMP65]], align 8
+// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP64]], double noundef [[TMP66]])
+// CHECK1-NEXT: br label %[[IF_END83]]
+// CHECK1: [[IF_END83]]:
+// CHECK1-NEXT: [[TMP67:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK1-NEXT: [[TMP68:%.*]] = load i64, ptr [[DOTOMP_NI2]], align 8
+// CHECK1-NEXT: [[CMP84:%.*]] = icmp slt i64 [[TMP67]], [[TMP68]]
+// CHECK1-NEXT: br i1 [[CMP84]], label %[[IF_THEN85:.*]], label %[[IF_END90:.*]]
+// CHECK1: [[IF_THEN85]]:
+// CHECK1-NEXT: [[TMP69:%.*]] = load i64, ptr [[DOTOMP_LB2]], align 8
+// CHECK1-NEXT: [[TMP70:%.*]] = load i64, ptr [[DOTOMP_ST2]], align 8
+// CHECK1-NEXT: [[TMP71:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK1-NEXT: [[MUL86:%.*]] = mul nsw i64 [[TMP70]], [[TMP71]]
+// CHECK1-NEXT: [[ADD87:%.*]] = add nsw i64 [[TMP69]], [[MUL86]]
+// CHECK1-NEXT: store i64 [[ADD87]], ptr [[DOTOMP_IV2]], align 8
+// CHECK1-NEXT: [[TMP72:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_29]], align 8
+// CHECK1-NEXT: [[TMP73:%.*]] = load i64, ptr [[DOTOMP_IV2]], align 8
+// CHECK1-NEXT: [[MUL88:%.*]] = mul nsw i64 [[TMP73]], 1
+// CHECK1-NEXT: [[ADD_PTR89:%.*]] = getelementptr inbounds double, ptr [[TMP72]], i64 [[MUL88]]
+// CHECK1-NEXT: store ptr [[ADD_PTR89]], ptr [[__BEGIN227]], align 8
+// CHECK1-NEXT: [[TMP74:%.*]] = load ptr, ptr [[__BEGIN227]], align 8
+// CHECK1-NEXT: store ptr [[TMP74]], ptr [[VV]], align 8
+// CHECK1-NEXT: [[TMP75:%.*]] = load i32, ptr [[CC]], align 4
+// CHECK1-NEXT: [[TMP76:%.*]] = load ptr, ptr [[VV]], align 8, !nonnull [[META5]], !align [[META6]]
+// CHECK1-NEXT: [[TMP77:%.*]] = load double, ptr [[TMP76]], align 8
+// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP75]], double noundef [[TMP77]])
+// CHECK1-NEXT: br label %[[IF_END90]]
+// CHECK1: [[IF_END90]]:
// CHECK1-NEXT: br label %[[FOR_INC:.*]]
// CHECK1: [[FOR_INC]]:
-// CHECK1-NEXT: [[TMP75:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK1-NEXT: [[INC:%.*]] = add nsw i64 [[TMP75]], 1
-// CHECK1-NEXT: store i64 [[INC]], ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK1-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
+// CHECK1-NEXT: [[TMP78:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK1-NEXT: [[INC:%.*]] = add nsw i64 [[TMP78]], 1
+// CHECK1-NEXT: store i64 [[INC]], ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK1-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP7:![0-9]+]]
// CHECK1: [[FOR_END]]:
// CHECK1-NEXT: ret void
//
@@ -783,11 +841,13 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[ENTRY:.*:]]
// CHECK1-NEXT: [[ARR:%.*]] = alloca [256 x double], align 16
// CHECK1-NEXT: [[J:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_UB0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_LB0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_ST0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_NI0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_IV0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[K:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_UB1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_LB1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_ST1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_NI1:%.*]] = alloca i32, align 4
@@ -802,10 +862,12 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[__END2:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: [[V:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: store i32 0, ptr [[J]], align 4
+// CHECK1-NEXT: store i32 127, ptr [[DOTOMP_UB0]], align 4
// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_LB0]], align 4
// CHECK1-NEXT: store i32 1, ptr [[DOTOMP_ST0]], align 4
// CHECK1-NEXT: store i32 128, ptr [[DOTOMP_NI0]], align 4
// CHECK1-NEXT: store i32 0, ptr [[K]], align 4
+// CHECK1-NEXT: store i32 63, ptr [[DOTOMP_UB1]], align 4
// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_LB1]], align 4
// CHECK1-NEXT: store i32 1, ptr [[DOTOMP_ST1]], align 4
// CHECK1-NEXT: store i32 64, ptr [[DOTOMP_NI1]], align 4
@@ -838,7 +900,7 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[TMP7:%.*]] = load i32, ptr [[I]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP7]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[I]], align 4
-// CHECK1-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP7:![0-9]+]]
+// CHECK1-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
// CHECK1: [[FOR_END]]:
// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_FUSE_INDEX]], align 4
// CHECK1-NEXT: br label %[[FOR_COND2:.*]]
@@ -891,14 +953,14 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[TMP24:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
// CHECK1-NEXT: [[INC16:%.*]] = add nsw i32 [[TMP24]], 1
// CHECK1-NEXT: store i32 [[INC16]], ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: br label %[[FOR_COND2]], !llvm.loop [[LOOP8:![0-9]+]]
+// CHECK1-NEXT: br label %[[FOR_COND2]], !llvm.loop [[LOOP9:![0-9]+]]
// CHECK1: [[FOR_END17]]:
// CHECK1-NEXT: store i32 42, ptr [[C]], align 4
// CHECK1-NEXT: store ptr [[ARR]], ptr [[__RANGE2]], align 8
-// CHECK1-NEXT: [[TMP25:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP25:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META5]], !align [[META6]]
// CHECK1-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP25]], i64 0, i64 0
// CHECK1-NEXT: store ptr [[ARRAYDECAY]], ptr [[__BEGIN2]], align 8
-// CHECK1-NEXT: [[TMP26:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK1-NEXT: [[TMP26:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META5]], !align [[META6]]
// CHECK1-NEXT: [[ARRAYDECAY18:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP26]], i64 0, i64 0
// CHECK1-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY18]], i64 256
// CHECK1-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
@@ -912,7 +974,7 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[TMP29:%.*]] = load ptr, ptr [[__BEGIN2]], align 8
// CHECK1-NEXT: store ptr [[TMP29]], ptr [[V]], align 8
// CHECK1-NEXT: [[TMP30:%.*]] = load i32, ptr [[C]], align 4
-// CHECK1-NEXT: [[TMP31:%.*]] = load ptr, ptr [[V]], align 8
+// CHECK1-NEXT: [[TMP31:%.*]] = load ptr, ptr [[V]], align 8, !nonnull [[META5]], !align [[META6]]
// CHECK1-NEXT: [[TMP32:%.*]] = load double, ptr [[TMP31]], align 8
// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP30]], double noundef [[TMP32]])
// CHECK1-NEXT: br label %[[FOR_INC22:.*]]
@@ -922,6 +984,18 @@ extern "C" void foo5() {
// CHECK1-NEXT: store ptr [[INCDEC_PTR]], ptr [[__BEGIN2]], align 8
// CHECK1-NEXT: br label %[[FOR_COND19]]
// CHECK1: [[FOR_END23]]:
+// CHECK1-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
+// CHECK1-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK1-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
+// CHECK1-NEXT: [[MUL24:%.*]] = mul nsw i32 [[TMP35]], [[TMP36]]
+// CHECK1-NEXT: [[ADD25:%.*]] = add nsw i32 [[TMP34]], [[MUL24]]
+// CHECK1-NEXT: store i32 [[ADD25]], ptr [[J]], align 4
+// CHECK1-NEXT: [[TMP37:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
+// CHECK1-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK1-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
+// CHECK1-NEXT: [[MUL26:%.*]] = mul nsw i32 [[TMP38]], [[TMP39]]
+// CHECK1-NEXT: [[ADD27:%.*]] = add nsw i32 [[TMP37]], [[MUL26]]
+// CHECK1-NEXT: store i32 [[ADD27]], ptr [[K]], align 4
// CHECK1-NEXT: ret void
//
//
@@ -930,11 +1004,13 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[ENTRY:.*:]]
// CHECK1-NEXT: [[ARR:%.*]] = alloca [256 x double], align 16
// CHECK1-NEXT: [[J:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_UB0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_LB0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_ST0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_NI0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_IV0:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[K:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_UB1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_LB1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_ST1:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_NI1:%.*]] = alloca i32, align 4
@@ -944,36 +1020,40 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[DOTOMP_FUSE_INDEX:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTCAPTURE_EXPR_1:%.*]] = alloca i32, align 4
-// CHECK1-NEXT: [[DOTOMP_LB03:%.*]] = alloca i32, align 4
-// CHECK1-NEXT: [[DOTOMP_ST04:%.*]] = alloca i32, align 4
-// CHECK1-NEXT: [[DOTOMP_NI05:%.*]] = alloca i64, align 8
-// CHECK1-NEXT: [[DOTOMP_IV06:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_UB03:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_LB04:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_ST05:%.*]] = alloca i32, align 4
+// CHECK1-NEXT: [[DOTOMP_NI06:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_IV07:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[C:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[__RANGE2:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: [[__END2:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: [[__BEGIN2:%.*]] = alloca ptr, align 8
-// CHECK1-NEXT: [[DOTCAPTURE_EXPR_8:%.*]] = alloca ptr, align 8
-// CHECK1-NEXT: [[DOTCAPTURE_EXPR_10:%.*]] = alloca ptr, align 8
-// CHECK1-NEXT: [[DOTCAPTURE_EXPR_11:%.*]] = alloca i64, align 8
-// CHECK1-NEXT: [[DOTOMP_LB116:%.*]] = alloca i64, align 8
-// CHECK1-NEXT: [[DOTOMP_ST117:%.*]] = alloca i64, align 8
-// CHECK1-NEXT: [[DOTOMP_NI118:%.*]] = alloca i64, align 8
-// CHECK1-NEXT: [[DOTOMP_IV120:%.*]] = alloca i64, align 8
-// CHECK1-NEXT: [[DOTOMP_TEMP_121:%.*]] = alloca i64, align 8
-// CHECK1-NEXT: [[DOTOMP_FUSE_MAX22:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTCAPTURE_EXPR_9:%.*]] = alloca ptr, align 8
+// CHECK1-NEXT: [[DOTCAPTURE_EXPR_11:%.*]] = alloca ptr, align 8
+// CHECK1-NEXT: [[DOTCAPTURE_EXPR_12:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_UB117:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_LB118:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_ST119:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_NI120:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_IV122:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_TEMP_123:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_FUSE_MAX24:%.*]] = alloca i64, align 8
// CHECK1-NEXT: [[I:%.*]] = alloca i32, align 4
-// CHECK1-NEXT: [[DOTOMP_FUSE_INDEX29:%.*]] = alloca i64, align 8
+// CHECK1-NEXT: [[DOTOMP_FUSE_INDEX31:%.*]] = alloca i64, align 8
// CHECK1-NEXT: [[V:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: [[CC:%.*]] = alloca i32, align 4
-// CHECK1-NEXT: [[__RANGE264:%.*]] = alloca ptr, align 8
-// CHECK1-NEXT: [[__BEGIN265:%.*]] = alloca ptr, align 8
-// CHECK1-NEXT: [[__END267:%.*]] = alloca ptr, align 8
+// CHECK1-NEXT: [[__RANGE266:%.*]] = alloca ptr, align 8
+// CHECK1-NEXT: [[__BEGIN267:%.*]] = alloca ptr, align 8
+// CHECK1-NEXT: [[__END269:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: [[VV:%.*]] = alloca ptr, align 8
// CHECK1-NEXT: store i32 0, ptr [[J]], align 4
+// CHECK1-NEXT: store i32 127, ptr [[DOTOMP_UB0]], align 4
// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_LB0]], align 4
// CHECK1-NEXT: store i32 1, ptr [[DOTOMP_ST0]], align 4
// CHECK1-NEXT: store i32 128, ptr [[DOTOMP_NI0]], align 4
// CHECK1-NEXT: store i32 0, ptr [[K]], align 4
+// CHECK1-NEXT: store i32 511, ptr [[DOTOMP_UB1]], align 4
// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_LB1]], align 4
// CHECK1-NEXT: store i32 1, ptr [[DOTOMP_ST1]], align 4
// CHECK1-NEXT: store i32 512, ptr [[DOTOMP_NI1]], align 4
@@ -1000,199 +1080,203 @@ extern "C" void foo5() {
// CHECK1-NEXT: [[DIV:%.*]] = sdiv i32 [[SUB]], 1
// CHECK1-NEXT: [[SUB2:%.*]] = sub nsw i32 [[DIV]], 1
// CHECK1-NEXT: store i32 [[SUB2]], ptr [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_LB03]], align 4
-// CHECK1-NEXT: store i32 1, ptr [[DOTOMP_ST04]], align 4
// CHECK1-NEXT: [[TMP7:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK1-NEXT: [[ADD:%.*]] = add nsw i32 [[TMP7]], 1
+// CHECK1-NEXT: store i32 [[TMP7]], ptr [[DOTOMP_UB03]], align 4
+// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_LB04]], align 4
+// CHECK1-NEXT: store i32 1, ptr [[DOTOMP_ST05]], align 4
+// CHECK1-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK1-NEXT: [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
// CHECK1-NEXT: [[CONV:%.*]] = sext i32 [[ADD]] to i64
-// CHECK1-NEXT: store i64 [[CONV]], ptr [[DOTOMP_NI05]], align 8
+// CHECK1-NEXT: store i64 [[CONV]], ptr [[DOTOMP_NI06]], align 8
// CHECK1-NEXT: store i32 42, ptr [[C]], align 4
// CHECK1-NEXT: store ptr [[ARR]], ptr [[__RANGE2]], align 8
-// CHECK1-NEXT: [[TMP8:%.*]] = load ptr, ptr [[__RANGE2]], align 8
-// CHECK1-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP8]], i64 0, i64 0
+// CHECK1-NEXT: [[TMP9:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META5]], !align [[META6]]
+// CHECK1-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP9]], i64 0, i64 0
// CHECK1-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 256
// CHECK1-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK1-NEXT: [[TMP9:%.*]] = load ptr, ptr [[__RANGE2]], align 8
-// CHECK1-NEXT: [[ARRAYDECAY7:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP9]], i64 0, i64 0
-// CHECK1-NEXT: store ptr [[ARRAYDECAY7]], ptr [[__BEGIN2]], align 8
-// CHECK1-NEXT: [[TMP10:%.*]] = load ptr, ptr [[__RANGE2]], align 8
-// CHECK1-NEXT: [[ARRAYDECAY9:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP10]], i64 0, i64 0
-// CHECK1-NEXT: store ptr [[ARRAYDECAY9]], ptr [[DOTCAPTURE_EXPR_8]], align 8
-// CHECK1-NEXT: [[TMP11:%.*]] = load ptr, ptr [[__END2]], align 8
-// CHECK1-NEXT: store ptr [[TMP11]], ptr [[DOTCAPTURE_EXPR_10]], align 8
-// CHECK1-NEXT: [[TMP12:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_10]], align 8
-// CHECK1-NEXT: [[TMP13:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_8]], align 8
-// CHECK1-NEXT: [[SUB_PTR_LHS_CAST:%.*]] = ptrtoint ptr [[TMP12]] to i64
-// CHECK1-NEXT: [[SUB_PTR_RHS_CAST:%.*]] = ptrtoint ptr [[TMP13]] to i64
+// CHECK1-NEXT: [[TMP10:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META5]], !align [[META6]]
+// CHECK1-NEXT: [[ARRAYDECAY8:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP10]], i64 0, i64 0
+// CHECK1-NEXT: store ptr [[ARRAYDECAY8]], ptr [[__BEGIN2]], align 8
+// CHECK1-NEXT: [[TMP11:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META5]], !align [[META6]]
+// CHECK1-NEXT: [[ARRAYDECAY10:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP11]], i64 0, i64 0
+// CHECK1-NEXT: store ptr [[ARRAYDECAY10]], ptr [[DOTCAPTURE_EXPR_9]], align 8
+// CHECK1-NEXT: [[TMP12:%.*]] = load ptr, ptr [[__END2]], align 8
+// CHECK1-NEXT: store ptr [[TMP12]], ptr [[DOTCAPTURE_EXPR_11]], align 8
+// CHECK1-NEXT: [[TMP13:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_11]], align 8
+// CHECK1-NEXT: [[TMP14:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_9]], align 8
+// CHECK1-NEXT: [[SUB_PTR_LHS_CAST:%.*]] = ptrtoint ptr [[TMP13]] to i64
+// CHECK1-NEXT: [[SUB_PTR_RHS_CAST:%.*]] = ptrtoint ptr [[TMP14]] to i64
// CHECK1-NEXT: [[SUB_PTR_SUB:%.*]] = sub i64 [[SUB_PTR_LHS_CAST]], [[SUB_PTR_RHS_CAST]]
// CHECK1-NEXT: [[SUB_PTR_DIV:%.*]] = sdiv exact i64 [[SUB_PTR_SUB]], 8
-// CHECK1-NEXT: [[SUB12:%.*]] = sub nsw i64 [[SUB_PTR_DIV]], 1
-// CHECK1-NEXT: [[ADD13:%.*]] = add nsw i64 [[SUB12]], 1
-// CHECK1-NEXT: [[DIV14:%.*]] = sdiv i64 [[ADD13]], 1
-// CHECK1-NEXT: [[SUB15:%.*]] = sub nsw i64 [[DIV14]], 1
-// CHECK1-NEXT: store i64 [[SUB15]], ptr [[DOTCAPTURE_EXPR_11]], align 8
-// CHECK1-NEXT: store i64 0, ptr [[DOTOMP_LB116]], align 8
-// CHECK1-NEXT: store i64 1, ptr [[DOTOMP_ST117]], align 8
-// CHECK1-NEXT: [[TMP14:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_11]], align 8
-// CHECK1-NEXT: [[ADD19:%.*]] = add nsw i64 [[TMP14]], 1
-// CHECK1-NEXT: store i64 [[ADD19]], ptr [[DOTOMP_NI118]], align 8
-// CHECK1-NEXT: [[TMP15:%.*]] = load i64, ptr [[DOTOMP_NI05]], align 8
-// CHECK1-NEXT: store i64 [[TMP15]], ptr [[DOTOMP_TEMP_121]], align 8
-// CHECK1-NEXT: [[TMP16:%.*]] = load i64, ptr [[DOTOMP_TEMP_121]], align 8
-// CHECK1-NEXT: [[TMP17:%.*]] = load i64, ptr [[DOTOMP_NI118]], align 8
-// CHECK1-NEXT: [[CMP23:%.*]] = icmp sgt i64 [[TMP16]], [[TMP17]]
-// CHECK1-NEXT: br i1 [[CMP23]], label %[[COND_TRUE24:.*]], label %[[COND_FALSE25:.*]]
-// CHECK1: [[COND_TRUE24]]:
-// CHECK1-NEXT: [[TMP18:%.*]] = load i64, ptr [[DOTOMP_TEMP_121]], align 8
-// CHECK1-NEXT: br label %[[COND_END26:.*]]
-// CHECK1: [[COND_FALSE25]]:
-// CHECK1-NEXT: [[TMP19:%.*]] = load i64, ptr [[DOTOMP_NI118]], align 8
-// CHECK1-NEXT: br label %[[COND_END26]]
-// CHECK1: [[COND_END26]]:
-// CHECK1-NEXT: [[COND27:%.*]] = phi i64 [ [[TMP18]], %[[COND_TRUE24]] ], [ [[TMP19]], %[[COND_FALSE25]] ]
-// CHECK1-NEXT: store i64 [[COND27]], ptr [[DOTOMP_FUSE_MAX22]], align 8
+// CHECK1-NEXT: [[SUB13:%.*]] = sub nsw i64 [[SUB_PTR_DIV]], 1
+// CHECK1-NEXT: [[ADD14:%.*]] = add nsw i64 [[SUB13]], 1
+// CHECK1-NEXT: [[DIV15:%.*]] = sdiv i64 [[ADD14]], 1
+// CHECK1-NEXT: [[SUB16:%.*]] = sub nsw i64 [[DIV15]], 1
+// CHECK1-NEXT: store i64 [[SUB16]], ptr [[DOTCAPTURE_EXPR_12]], align 8
+// CHECK1-NEXT: [[TMP15:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_12]], align 8
+// CHECK1-NEXT: store i64 [[TMP15]], ptr [[DOTOMP_UB117]], align 8
+// CHECK1-NEXT: store i64 0, ptr [[DOTOMP_LB118]], align 8
+// CHECK1-NEXT: store i64 1, ptr [[DOTOMP_ST119]], align 8
+// CHECK1-NEXT: [[TMP16:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_12]], align 8
+// CHECK1-NEXT: [[ADD21:%.*]] = add nsw i64 [[TMP16]], 1
+// CHECK1-NEXT: store i64 [[ADD21]], ptr [[DOTOMP_NI120]], align 8
+// CHECK1-NEXT: [[TMP17:%.*]] = load i64, ptr [[DOTOMP_NI06]], align 8
+// CHECK1-NEXT: store i64 [[TMP17]], ptr [[DOTOMP_TEMP_123]], align 8
+// CHECK1-NEXT: [[TMP18:%.*]] = load i64, ptr [[DOTOMP_TEMP_123]], align 8
+// CHECK1-NEXT: [[TMP19:%.*]] = load i64, ptr [[DOTOMP_NI120]], align 8
+// CHECK1-NEXT: [[CMP25:%.*]] = icmp sgt i64 [[TMP18]], [[TMP19]]
+// CHECK1-NEXT: br i1 [[CMP25]], label %[[COND_TRUE26:.*]], label %[[COND_FALSE27:.*]]
+// CHECK1: [[COND_TRUE26]]:
+// CHECK1-NEXT: [[TMP20:%.*]] = load i64, ptr [[DOTOMP_TEMP_123]], align 8
+// CHECK1-NEXT: br label %[[COND_END28:.*]]
+// CHECK1: [[COND_FALSE27]]:
+// CHECK1-NEXT: [[TMP21:%.*]] = load i64, ptr [[DOTOMP_NI120]], align 8
+// CHECK1-NEXT: br label %[[COND_END28]]
+// CHECK1: [[COND_END28]]:
+// CHECK1-NEXT: [[COND29:%.*]] = phi i64 [ [[TMP20]], %[[COND_TRUE26]] ], [ [[TMP21]], %[[COND_FALSE27]] ]
+// CHECK1-NEXT: store i64 [[COND29]], ptr [[DOTOMP_FUSE_MAX24]], align 8
// CHECK1-NEXT: store i32 0, ptr [[I]], align 4
// CHECK1-NEXT: br label %[[FOR_COND:.*]]
// CHECK1: [[FOR_COND]]:
-// CHECK1-NEXT: [[TMP20:%.*]] = load i32, ptr [[I]], align 4
-// CHECK1-NEXT: [[CMP28:%.*]] = icmp slt i32 [[TMP20]], 128
-// CHECK1-NEXT: br i1 [[CMP28]], label %[[FOR_BODY:.*]], label %[[FOR_END:.*]]
+// CHECK1-NEXT: [[TMP22:%.*]] = load i32, ptr [[I]], align 4
+// CHECK1-NEXT: [[CMP30:%.*]] = icmp slt i32 [[TMP22]], 128
+// CHECK1-NEXT: br i1 [[CMP30]], label %[[FOR_BODY:.*]], label %[[FOR_END:.*]]
// CHECK1: [[FOR_BODY]]:
-// CHECK1-NEXT: [[TMP21:%.*]] = load i32, ptr [[I]], align 4
-// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP21]])
+// CHECK1-NEXT: [[TMP23:%.*]] = load i32, ptr [[I]], align 4
+// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP23]])
// CHECK1-NEXT: br label %[[FOR_INC:.*]]
// CHECK1: [[FOR_INC]]:
-// CHECK1-NEXT: [[TMP22:%.*]] = load i32, ptr [[I]], align 4
-// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP22]], 1
+// CHECK1-NEXT: [[TMP24:%.*]] = load i32, ptr [[I]], align 4
+// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP24]], 1
// CHECK1-NEXT: store i32 [[INC]], ptr [[I]], align 4
-// CHECK1-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP9:![0-9]+]]
+// CHECK1-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP10:![0-9]+]]
// CHECK1: [[FOR_END]]:
-// CHECK1-NEXT: store i64 0, ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK1-NEXT: br label %[[FOR_COND30:.*]]
-// CHECK1: [[FOR_COND30]]:
-// CHECK1-NEXT: [[TMP23:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK1-NEXT: [[TMP24:%.*]] = load i64, ptr [[DOTOMP_FUSE_MAX22]], align 8
-// CHECK1-NEXT: [[CMP31:%.*]] = icmp slt i64 [[TMP23]], [[TMP24]]
-// CHECK1-NEXT: br i1 [[CMP31]], label %[[FOR_BODY32:.*]], label %[[FOR_END63:.*]]
-// CHECK1: [[FOR_BODY32]]:
-// CHECK1-NEXT: [[TMP25:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK1-NEXT: [[TMP26:%.*]] = load i64, ptr [[DOTOMP_NI05]], align 8
+// CHECK1-NEXT: store i64 0, ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK1-NEXT: br label %[[FOR_COND32:.*]]
+// CHECK1: [[FOR_COND32]]:
+// CHECK1-NEXT: [[TMP25:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK1-NEXT: [[TMP26:%.*]] = load i64, ptr [[DOTOMP_FUSE_MAX24]], align 8
// CHECK1-NEXT: [[CMP33:%.*]] = icmp slt i64 [[TMP25]], [[TMP26]]
-// CHECK1-NEXT: br i1 [[CMP33]], label %[[IF_THEN:.*]], label %[[IF_END53:.*]]
+// CHECK1-NEXT: br i1 [[CMP33]], label %[[FOR_BODY34:.*]], label %[[FOR_END65:.*]]
+// CHECK1: [[FOR_BODY34]]:
+// CHECK1-NEXT: [[TMP27:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK1-NEXT: [[TMP28:%.*]] = load i64, ptr [[DOTOMP_NI06]], align 8
+// CHECK1-NEXT: [[CMP35:%.*]] = icmp slt i64 [[TMP27]], [[TMP28]]
+// CHECK1-NEXT: br i1 [[CMP35]], label %[[IF_THEN:.*]], label %[[IF_END55:.*]]
// CHECK1: [[IF_THEN]]:
-// CHECK1-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTOMP_LB03]], align 4
-// CHECK1-NEXT: [[CONV34:%.*]] = sext i32 [[TMP27]] to i64
-// CHECK1-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTOMP_ST04]], align 4
-// CHECK1-NEXT: [[CONV35:%.*]] = sext i32 [[TMP28]] to i64
-// CHECK1-NEXT: [[TMP29:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK1-NEXT: [[MUL:%.*]] = mul nsw i64 [[CONV35]], [[TMP29]]
-// CHECK1-NEXT: [[ADD36:%.*]] = add nsw i64 [[CONV34]], [[MUL]]
-// CHECK1-NEXT: [[CONV37:%.*]] = trunc i64 [[ADD36]] to i32
-// CHECK1-NEXT: store i32 [[CONV37]], ptr [[DOTOMP_IV06]], align 4
-// CHECK1-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTOMP_IV06]], align 4
-// CHECK1-NEXT: [[MUL38:%.*]] = mul nsw i32 [[TMP30]], 1
-// CHECK1-NEXT: [[ADD39:%.*]] = add nsw i32 0, [[MUL38]]
-// CHECK1-NEXT: store i32 [[ADD39]], ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
-// CHECK1-NEXT: [[CMP40:%.*]] = icmp slt i32 [[TMP31]], [[TMP32]]
-// CHECK1-NEXT: br i1 [[CMP40]], label %[[IF_THEN41:.*]], label %[[IF_END:.*]]
-// CHECK1: [[IF_THEN41]]:
-// CHECK1-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
-// CHECK1-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
-// CHECK1-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[MUL42:%.*]] = mul nsw i32 [[TMP34]], [[TMP35]]
-// CHECK1-NEXT: [[ADD43:%.*]] = add nsw i32 [[TMP33]], [[MUL42]]
-// CHECK1-NEXT: store i32 [[ADD43]], ptr [[DOTOMP_IV0]], align 4
-// CHECK1-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
-// CHECK1-NEXT: [[MUL44:%.*]] = mul nsw i32 [[TMP36]], 2
-// CHECK1-NEXT: [[ADD45:%.*]] = add nsw i32 0, [[MUL44]]
-// CHECK1-NEXT: store i32 [[ADD45]], ptr [[J]], align 4
-// CHECK1-NEXT: [[TMP37:%.*]] = load i32, ptr [[J]], align 4
-// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP37]])
+// CHECK1-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTOMP_LB04]], align 4
+// CHECK1-NEXT: [[CONV36:%.*]] = sext i32 [[TMP29]] to i64
+// CHECK1-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTOMP_ST05]], align 4
+// CHECK1-NEXT: [[CONV37:%.*]] = sext i32 [[TMP30]] to i64
+// CHECK1-NEXT: [[TMP31:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK1-NEXT: [[MUL:%.*]] = mul nsw i64 [[CONV37]], [[TMP31]]
+// CHECK1-NEXT: [[ADD38:%.*]] = add nsw i64 [[CONV36]], [[MUL]]
+// CHECK1-NEXT: [[CONV39:%.*]] = trunc i64 [[ADD38]] to i32
+// CHECK1-NEXT: store i32 [[CONV39]], ptr [[DOTOMP_IV07]], align 4
+// CHECK1-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTOMP_IV07]], align 4
+// CHECK1-NEXT: [[MUL40:%.*]] = mul nsw i32 [[TMP32]], 1
+// CHECK1-NEXT: [[ADD41:%.*]] = add nsw i32 0, [[MUL40]]
+// CHECK1-NEXT: store i32 [[ADD41]], ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK1-NEXT: [[CMP42:%.*]] = icmp slt i32 [[TMP33]], [[TMP34]]
+// CHECK1-NEXT: br i1 [[CMP42]], label %[[IF_THEN43:.*]], label %[[IF_END:.*]]
+// CHECK1: [[IF_THEN43]]:
+// CHECK1-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
+// CHECK1-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
+// CHECK1-NEXT: [[TMP37:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[MUL44:%.*]] = mul nsw i32 [[TMP36]], [[TMP37]]
+// CHECK1-NEXT: [[ADD45:%.*]] = add nsw i32 [[TMP35]], [[MUL44]]
+// CHECK1-NEXT: store i32 [[ADD45]], ptr [[DOTOMP_IV0]], align 4
+// CHECK1-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
+// CHECK1-NEXT: [[MUL46:%.*]] = mul nsw i32 [[TMP38]], 2
+// CHECK1-NEXT: [[ADD47:%.*]] = add nsw i32 0, [[MUL46]]
+// CHECK1-NEXT: store i32 [[ADD47]], ptr [[J]], align 4
+// CHECK1-NEXT: [[TMP39:%.*]] = load i32, ptr [[J]], align 4
+// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP39]])
// CHECK1-NEXT: br label %[[IF_END]]
// CHECK1: [[IF_END]]:
-// CHECK1-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
-// CHECK1-NEXT: [[CMP46:%.*]] = icmp slt i32 [[TMP38]], [[TMP39]]
-// CHECK1-NEXT: br i1 [[CMP46]], label %[[IF_THEN47:.*]], label %[[IF_END52:.*]]
-// CHECK1: [[IF_THEN47]]:
-// CHECK1-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
-// CHECK1-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
-// CHECK1-NEXT: [[TMP42:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK1-NEXT: [[MUL48:%.*]] = mul nsw i32 [[TMP41]], [[TMP42]]
-// CHECK1-NEXT: [[ADD49:%.*]] = add nsw i32 [[TMP40]], [[MUL48]]
-// CHECK1-NEXT: store i32 [[ADD49]], ptr [[DOTOMP_IV1]], align 4
-// CHECK1-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
-// CHECK1-NEXT: [[MUL50:%.*]] = mul nsw i32 [[TMP43]], 1
-// CHECK1-NEXT: [[ADD51:%.*]] = add nsw i32 0, [[MUL50]]
-// CHECK1-NEXT: store i32 [[ADD51]], ptr [[K]], align 4
-// CHECK1-NEXT: [[TMP44:%.*]] = load i32, ptr [[K]], align 4
-// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP44]])
-// CHECK1-NEXT: br label %[[IF_END52]]
-// CHECK1: [[IF_END52]]:
-// CHECK1-NEXT: br label %[[IF_END53]]
-// CHECK1: [[IF_END53]]:
-// CHECK1-NEXT: [[TMP45:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK1-NEXT: [[TMP46:%.*]] = load i64, ptr [[DOTOMP_NI118]], align 8
-// CHECK1-NEXT: [[CMP54:%.*]] = icmp slt i64 [[TMP45]], [[TMP46]]
-// CHECK1-NEXT: br i1 [[CMP54]], label %[[IF_THEN55:.*]], label %[[IF_END60:.*]]
-// CHECK1: [[IF_THEN55]]:
-// CHECK1-NEXT: [[TMP47:%.*]] = load i64, ptr [[DOTOMP_LB116]], align 8
-// CHECK1-NEXT: [[TMP48:%.*]] = load i64, ptr [[DOTOMP_ST117]], align 8
-// CHECK1-NEXT: [[TMP49:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK1-NEXT: [[MUL56:%.*]] = mul nsw i64 [[TMP48]], [[TMP49]]
-// CHECK1-NEXT: [[ADD57:%.*]] = add nsw i64 [[TMP47]], [[MUL56]]
-// CHECK1-NEXT: store i64 [[ADD57]], ptr [[DOTOMP_IV120]], align 8
-// CHECK1-NEXT: [[TMP50:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_8]], align 8
-// CHECK1-NEXT: [[TMP51:%.*]] = load i64, ptr [[DOTOMP_IV120]], align 8
-// CHECK1-NEXT: [[MUL58:%.*]] = mul nsw i64 [[TMP51]], 1
-// CHECK1-NEXT: [[ADD_PTR59:%.*]] = getelementptr inbounds double, ptr [[TMP50]], i64 [[MUL58]]
-// CHECK1-NEXT: store ptr [[ADD_PTR59]], ptr [[__BEGIN2]], align 8
-// CHECK1-NEXT: [[TMP52:%.*]] = load ptr, ptr [[__BEGIN2]], align 8
-// CHECK1-NEXT: store ptr [[TMP52]], ptr [[V]], align 8
-// CHECK1-NEXT: [[TMP53:%.*]] = load i32, ptr [[C]], align 4
-// CHECK1-NEXT: [[TMP54:%.*]] = load ptr, ptr [[V]], align 8
-// CHECK1-NEXT: [[TMP55:%.*]] = load double, ptr [[TMP54]], align 8
-// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP53]], double noundef [[TMP55]])
-// CHECK1-NEXT: br label %[[IF_END60]]
-// CHECK1: [[IF_END60]]:
-// CHECK1-NEXT: br label %[[FOR_INC61:.*]]
-// CHECK1: [[FOR_INC61]]:
-// CHECK1-NEXT: [[TMP56:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK1-NEXT: [[INC62:%.*]] = add nsw i64 [[TMP56]], 1
-// CHECK1-NEXT: store i64 [[INC62]], ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK1-NEXT: br label %[[FOR_COND30]], !llvm.loop [[LOOP10:![0-9]+]]
-// CHECK1: [[FOR_END63]]:
+// CHECK1-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK1-NEXT: [[CMP48:%.*]] = icmp slt i32 [[TMP40]], [[TMP41]]
+// CHECK1-NEXT: br i1 [[CMP48]], label %[[IF_THEN49:.*]], label %[[IF_END54:.*]]
+// CHECK1: [[IF_THEN49]]:
+// CHECK1-NEXT: [[TMP42:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
+// CHECK1-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
+// CHECK1-NEXT: [[TMP44:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK1-NEXT: [[MUL50:%.*]] = mul nsw i32 [[TMP43]], [[TMP44]]
+// CHECK1-NEXT: [[ADD51:%.*]] = add nsw i32 [[TMP42]], [[MUL50]]
+// CHECK1-NEXT: store i32 [[ADD51]], ptr [[DOTOMP_IV1]], align 4
+// CHECK1-NEXT: [[TMP45:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
+// CHECK1-NEXT: [[MUL52:%.*]] = mul nsw i32 [[TMP45]], 1
+// CHECK1-NEXT: [[ADD53:%.*]] = add nsw i32 0, [[MUL52]]
+// CHECK1-NEXT: store i32 [[ADD53]], ptr [[K]], align 4
+// CHECK1-NEXT: [[TMP46:%.*]] = load i32, ptr [[K]], align 4
+// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP46]])
+// CHECK1-NEXT: br label %[[IF_END54]]
+// CHECK1: [[IF_END54]]:
+// CHECK1-NEXT: br label %[[IF_END55]]
+// CHECK1: [[IF_END55]]:
+// CHECK1-NEXT: [[TMP47:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK1-NEXT: [[TMP48:%.*]] = load i64, ptr [[DOTOMP_NI120]], align 8
+// CHECK1-NEXT: [[CMP56:%.*]] = icmp slt i64 [[TMP47]], [[TMP48]]
+// CHECK1-NEXT: br i1 [[CMP56]], label %[[IF_THEN57:.*]], label %[[IF_END62:.*]]
+// CHECK1: [[IF_THEN57]]:
+// CHECK1-NEXT: [[TMP49:%.*]] = load i64, ptr [[DOTOMP_LB118]], align 8
+// CHECK1-NEXT: [[TMP50:%.*]] = load i64, ptr [[DOTOMP_ST119]], align 8
+// CHECK1-NEXT: [[TMP51:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK1-NEXT: [[MUL58:%.*]] = mul nsw i64 [[TMP50]], [[TMP51]]
+// CHECK1-NEXT: [[ADD59:%.*]] = add nsw i64 [[TMP49]], [[MUL58]]
+// CHECK1-NEXT: store i64 [[ADD59]], ptr [[DOTOMP_IV122]], align 8
+// CHECK1-NEXT: [[TMP52:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_9]], align 8
+// CHECK1-NEXT: [[TMP53:%.*]] = load i64, ptr [[DOTOMP_IV122]], align 8
+// CHECK1-NEXT: [[MUL60:%.*]] = mul nsw i64 [[TMP53]], 1
+// CHECK1-NEXT: [[ADD_PTR61:%.*]] = getelementptr inbounds double, ptr [[TMP52]], i64 [[MUL60]]
+// CHECK1-NEXT: store ptr [[ADD_PTR61]], ptr [[__BEGIN2]], align 8
+// CHECK1-NEXT: [[TMP54:%.*]] = load ptr, ptr [[__BEGIN2]], align 8
+// CHECK1-NEXT: store ptr [[TMP54]], ptr [[V]], align 8
+// CHECK1-NEXT: [[TMP55:%.*]] = load i32, ptr [[C]], align 4
+// CHECK1-NEXT: [[TMP56:%.*]] = load ptr, ptr [[V]], align 8, !nonnull [[META5]], !align [[META6]]
+// CHECK1-NEXT: [[TMP57:%.*]] = load double, ptr [[TMP56]], align 8
+// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP55]], double noundef [[TMP57]])
+// CHECK1-NEXT: br label %[[IF_END62]]
+// CHECK1: [[IF_END62]]:
+// CHECK1-NEXT: br label %[[FOR_INC63:.*]]
+// CHECK1: [[FOR_INC63]]:
+// CHECK1-NEXT: [[TMP58:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK1-NEXT: [[INC64:%.*]] = add nsw i64 [[TMP58]], 1
+// CHECK1-NEXT: store i64 [[INC64]], ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK1-NEXT: br label %[[FOR_COND32]], !llvm.loop [[LOOP11:![0-9]+]]
+// CHECK1: [[FOR_END65]]:
// CHECK1-NEXT: store i32 37, ptr [[CC]], align 4
-// CHECK1-NEXT: store ptr [[ARR]], ptr [[__RANGE264]], align 8
-// CHECK1-NEXT: [[TMP57:%.*]] = load ptr, ptr [[__RANGE264]], align 8
-// CHECK1-NEXT: [[ARRAYDECAY66:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP57]], i64 0, i64 0
-// CHECK1-NEXT: store ptr [[ARRAYDECAY66]], ptr [[__BEGIN265]], align 8
-// CHECK1-NEXT: [[TMP58:%.*]] = load ptr, ptr [[__RANGE264]], align 8
-// CHECK1-NEXT: [[ARRAYDECAY68:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP58]], i64 0, i64 0
-// CHECK1-NEXT: [[ADD_PTR69:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY68]], i64 256
-// CHECK1-NEXT: store ptr [[ADD_PTR69]], ptr [[__END267]], align 8
-// CHECK1-NEXT: br label %[[FOR_COND70:.*]]
-// CHECK1: [[FOR_COND70]]:
-// CHECK1-NEXT: [[TMP59:%.*]] = load ptr, ptr [[__BEGIN265]], align 8
-// CHECK1-NEXT: [[TMP60:%.*]] = load ptr, ptr [[__END267]], align 8
-// CHECK1-NEXT: [[CMP71:%.*]] = icmp ne ptr [[TMP59]], [[TMP60]]
-// CHECK1-NEXT: br i1 [[CMP71]], label %[[FOR_BODY72:.*]], label %[[FOR_END74:.*]]
-// CHECK1: [[FOR_BODY72]]:
-// CHECK1-NEXT: [[TMP61:%.*]] = load ptr, ptr [[__BEGIN265]], align 8
-// CHECK1-NEXT: store ptr [[TMP61]], ptr [[VV]], align 8
-// CHECK1-NEXT: [[TMP62:%.*]] = load i32, ptr [[CC]], align 4
-// CHECK1-NEXT: [[TMP63:%.*]] = load ptr, ptr [[VV]], align 8
-// CHECK1-NEXT: [[TMP64:%.*]] = load double, ptr [[TMP63]], align 8
-// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP62]], double noundef [[TMP64]])
-// CHECK1-NEXT: br label %[[FOR_INC73:.*]]
-// CHECK1: [[FOR_INC73]]:
-// CHECK1-NEXT: [[TMP65:%.*]] = load ptr, ptr [[__BEGIN265]], align 8
-// CHECK1-NEXT: [[INCDEC_PTR:%.*]] = getelementptr inbounds nuw double, ptr [[TMP65]], i32 1
-// CHECK1-NEXT: store ptr [[INCDEC_PTR]], ptr [[__BEGIN265]], align 8
-// CHECK1-NEXT: br label %[[FOR_COND70]]
-// CHECK1: [[FOR_END74]]:
+// CHECK1-NEXT: store ptr [[ARR]], ptr [[__RANGE266]], align 8
+// CHECK1-NEXT: [[TMP59:%.*]] = load ptr, ptr [[__RANGE266]], align 8, !nonnull [[META5]], !align [[META6]]
+// CHECK1-NEXT: [[ARRAYDECAY68:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP59]], i64 0, i64 0
+// CHECK1-NEXT: store ptr [[ARRAYDECAY68]], ptr [[__BEGIN267]], align 8
+// CHECK1-NEXT: [[TMP60:%.*]] = load ptr, ptr [[__RANGE266]], align 8, !nonnull [[META5]], !align [[META6]]
+// CHECK1-NEXT: [[ARRAYDECAY70:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP60]], i64 0, i64 0
+// CHECK1-NEXT: [[ADD_PTR71:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY70]], i64 256
+// CHECK1-NEXT: store ptr [[ADD_PTR71]], ptr [[__END269]], align 8
+// CHECK1-NEXT: br label %[[FOR_COND72:.*]]
+// CHECK1: [[FOR_COND72]]:
+// CHECK1-NEXT: [[TMP61:%.*]] = load ptr, ptr [[__BEGIN267]], align 8
+// CHECK1-NEXT: [[TMP62:%.*]] = load ptr, ptr [[__END269]], align 8
+// CHECK1-NEXT: [[CMP73:%.*]] = icmp ne ptr [[TMP61]], [[TMP62]]
+// CHECK1-NEXT: br i1 [[CMP73]], label %[[FOR_BODY74:.*]], label %[[FOR_END76:.*]]
+// CHECK1: [[FOR_BODY74]]:
+// CHECK1-NEXT: [[TMP63:%.*]] = load ptr, ptr [[__BEGIN267]], align 8
+// CHECK1-NEXT: store ptr [[TMP63]], ptr [[VV]], align 8
+// CHECK1-NEXT: [[TMP64:%.*]] = load i32, ptr [[CC]], align 4
+// CHECK1-NEXT: [[TMP65:%.*]] = load ptr, ptr [[VV]], align 8, !nonnull [[META5]], !align [[META6]]
+// CHECK1-NEXT: [[TMP66:%.*]] = load double, ptr [[TMP65]], align 8
+// CHECK1-NEXT: call void (...) @body(i32 noundef [[TMP64]], double noundef [[TMP66]])
+// CHECK1-NEXT: br label %[[FOR_INC75:.*]]
+// CHECK1: [[FOR_INC75]]:
+// CHECK1-NEXT: [[TMP67:%.*]] = load ptr, ptr [[__BEGIN267]], align 8
+// CHECK1-NEXT: [[INCDEC_PTR:%.*]] = getelementptr inbounds nuw double, ptr [[TMP67]], i32 1
+// CHECK1-NEXT: store ptr [[INCDEC_PTR]], ptr [[__BEGIN267]], align 8
+// CHECK1-NEXT: br label %[[FOR_COND72]]
+// CHECK1: [[FOR_END76]]:
// CHECK1-NEXT: ret void
//
//
@@ -1217,6 +1301,7 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[DOTCAPTURE_EXPR_1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTNEW_STEP:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_UB0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_LB0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_ST0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_NI0:%.*]] = alloca i32, align 4
@@ -1225,6 +1310,7 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[DOTCAPTURE_EXPR_7:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTNEW_STEP8:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_UB1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_LB1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_ST1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_NI1:%.*]] = alloca i32, align 4
@@ -1256,106 +1342,122 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[DIV:%.*]] = udiv i32 [[ADD]], [[TMP7]]
// CHECK2-NEXT: [[SUB4:%.*]] = sub i32 [[DIV]], 1
// CHECK2-NEXT: store i32 [[SUB4]], ptr [[DOTCAPTURE_EXPR_2]], align 4
+// CHECK2-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
+// CHECK2-NEXT: store i32 [[TMP8]], ptr [[DOTOMP_UB0]], align 4
// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_LB0]], align 4
// CHECK2-NEXT: store i32 1, ptr [[DOTOMP_ST0]], align 4
-// CHECK2-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
-// CHECK2-NEXT: [[ADD5:%.*]] = add i32 [[TMP8]], 1
+// CHECK2-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
+// CHECK2-NEXT: [[ADD5:%.*]] = add i32 [[TMP9]], 1
// CHECK2-NEXT: store i32 [[ADD5]], ptr [[DOTOMP_NI0]], align 4
-// CHECK2-NEXT: [[TMP9:%.*]] = load i32, ptr [[START2_ADDR]], align 4
-// CHECK2-NEXT: store i32 [[TMP9]], ptr [[J]], align 4
// CHECK2-NEXT: [[TMP10:%.*]] = load i32, ptr [[START2_ADDR]], align 4
-// CHECK2-NEXT: store i32 [[TMP10]], ptr [[DOTCAPTURE_EXPR_6]], align 4
-// CHECK2-NEXT: [[TMP11:%.*]] = load i32, ptr [[END2_ADDR]], align 4
-// CHECK2-NEXT: store i32 [[TMP11]], ptr [[DOTCAPTURE_EXPR_7]], align 4
-// CHECK2-NEXT: [[TMP12:%.*]] = load i32, ptr [[STEP2_ADDR]], align 4
-// CHECK2-NEXT: store i32 [[TMP12]], ptr [[DOTNEW_STEP8]], align 4
-// CHECK2-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_7]], align 4
-// CHECK2-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
-// CHECK2-NEXT: [[SUB10:%.*]] = sub i32 [[TMP13]], [[TMP14]]
+// CHECK2-NEXT: store i32 [[TMP10]], ptr [[J]], align 4
+// CHECK2-NEXT: [[TMP11:%.*]] = load i32, ptr [[START2_ADDR]], align 4
+// CHECK2-NEXT: store i32 [[TMP11]], ptr [[DOTCAPTURE_EXPR_6]], align 4
+// CHECK2-NEXT: [[TMP12:%.*]] = load i32, ptr [[END2_ADDR]], align 4
+// CHECK2-NEXT: store i32 [[TMP12]], ptr [[DOTCAPTURE_EXPR_7]], align 4
+// CHECK2-NEXT: [[TMP13:%.*]] = load i32, ptr [[STEP2_ADDR]], align 4
+// CHECK2-NEXT: store i32 [[TMP13]], ptr [[DOTNEW_STEP8]], align 4
+// CHECK2-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_7]], align 4
+// CHECK2-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
+// CHECK2-NEXT: [[SUB10:%.*]] = sub i32 [[TMP14]], [[TMP15]]
// CHECK2-NEXT: [[SUB11:%.*]] = sub i32 [[SUB10]], 1
-// CHECK2-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
-// CHECK2-NEXT: [[ADD12:%.*]] = add i32 [[SUB11]], [[TMP15]]
// CHECK2-NEXT: [[TMP16:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
-// CHECK2-NEXT: [[DIV13:%.*]] = udiv i32 [[ADD12]], [[TMP16]]
+// CHECK2-NEXT: [[ADD12:%.*]] = add i32 [[SUB11]], [[TMP16]]
+// CHECK2-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
+// CHECK2-NEXT: [[DIV13:%.*]] = udiv i32 [[ADD12]], [[TMP17]]
// CHECK2-NEXT: [[SUB14:%.*]] = sub i32 [[DIV13]], 1
// CHECK2-NEXT: store i32 [[SUB14]], ptr [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK2-NEXT: [[TMP18:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK2-NEXT: store i32 [[TMP18]], ptr [[DOTOMP_UB1]], align 4
// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_LB1]], align 4
// CHECK2-NEXT: store i32 1, ptr [[DOTOMP_ST1]], align 4
-// CHECK2-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK2-NEXT: [[ADD15:%.*]] = add i32 [[TMP17]], 1
+// CHECK2-NEXT: [[TMP19:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK2-NEXT: [[ADD15:%.*]] = add i32 [[TMP19]], 1
// CHECK2-NEXT: store i32 [[ADD15]], ptr [[DOTOMP_NI1]], align 4
-// CHECK2-NEXT: [[TMP18:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
-// CHECK2-NEXT: store i32 [[TMP18]], ptr [[DOTOMP_TEMP_1]], align 4
-// CHECK2-NEXT: [[TMP19:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
-// CHECK2-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
-// CHECK2-NEXT: [[CMP:%.*]] = icmp ugt i32 [[TMP19]], [[TMP20]]
+// CHECK2-NEXT: [[TMP20:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK2-NEXT: store i32 [[TMP20]], ptr [[DOTOMP_TEMP_1]], align 4
+// CHECK2-NEXT: [[TMP21:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
+// CHECK2-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK2-NEXT: [[CMP:%.*]] = icmp ugt i32 [[TMP21]], [[TMP22]]
// CHECK2-NEXT: br i1 [[CMP]], label %[[COND_TRUE:.*]], label %[[COND_FALSE:.*]]
// CHECK2: [[COND_TRUE]]:
-// CHECK2-NEXT: [[TMP21:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
+// CHECK2-NEXT: [[TMP23:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
// CHECK2-NEXT: br label %[[COND_END:.*]]
// CHECK2: [[COND_FALSE]]:
-// CHECK2-NEXT: [[TMP22:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK2-NEXT: [[TMP24:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
// CHECK2-NEXT: br label %[[COND_END]]
// CHECK2: [[COND_END]]:
-// CHECK2-NEXT: [[COND:%.*]] = phi i32 [ [[TMP21]], %[[COND_TRUE]] ], [ [[TMP22]], %[[COND_FALSE]] ]
+// CHECK2-NEXT: [[COND:%.*]] = phi i32 [ [[TMP23]], %[[COND_TRUE]] ], [ [[TMP24]], %[[COND_FALSE]] ]
// CHECK2-NEXT: store i32 [[COND]], ptr [[DOTOMP_FUSE_MAX]], align 4
// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_FUSE_INDEX]], align 4
// CHECK2-NEXT: br label %[[FOR_COND:.*]]
// CHECK2: [[FOR_COND]]:
-// CHECK2-NEXT: [[TMP23:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[TMP24:%.*]] = load i32, ptr [[DOTOMP_FUSE_MAX]], align 4
-// CHECK2-NEXT: [[CMP16:%.*]] = icmp ult i32 [[TMP23]], [[TMP24]]
+// CHECK2-NEXT: [[TMP25:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[TMP26:%.*]] = load i32, ptr [[DOTOMP_FUSE_MAX]], align 4
+// CHECK2-NEXT: [[CMP16:%.*]] = icmp ult i32 [[TMP25]], [[TMP26]]
// CHECK2-NEXT: br i1 [[CMP16]], label %[[FOR_BODY:.*]], label %[[FOR_END:.*]]
// CHECK2: [[FOR_BODY]]:
-// CHECK2-NEXT: [[TMP25:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[TMP26:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
-// CHECK2-NEXT: [[CMP17:%.*]] = icmp ult i32 [[TMP25]], [[TMP26]]
+// CHECK2-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK2-NEXT: [[CMP17:%.*]] = icmp ult i32 [[TMP27]], [[TMP28]]
// CHECK2-NEXT: br i1 [[CMP17]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
// CHECK2: [[IF_THEN]]:
-// CHECK2-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
-// CHECK2-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
-// CHECK2-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[MUL:%.*]] = mul i32 [[TMP28]], [[TMP29]]
-// CHECK2-NEXT: [[ADD18:%.*]] = add i32 [[TMP27]], [[MUL]]
+// CHECK2-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
+// CHECK2-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
+// CHECK2-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[MUL:%.*]] = mul i32 [[TMP30]], [[TMP31]]
+// CHECK2-NEXT: [[ADD18:%.*]] = add i32 [[TMP29]], [[MUL]]
// CHECK2-NEXT: store i32 [[ADD18]], ptr [[DOTOMP_IV0]], align 4
-// CHECK2-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
-// CHECK2-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
-// CHECK2-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
-// CHECK2-NEXT: [[MUL19:%.*]] = mul i32 [[TMP31]], [[TMP32]]
-// CHECK2-NEXT: [[ADD20:%.*]] = add i32 [[TMP30]], [[MUL19]]
+// CHECK2-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
+// CHECK2-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK2-NEXT: [[MUL19:%.*]] = mul i32 [[TMP33]], [[TMP34]]
+// CHECK2-NEXT: [[ADD20:%.*]] = add i32 [[TMP32]], [[MUL19]]
// CHECK2-NEXT: store i32 [[ADD20]], ptr [[I]], align 4
-// CHECK2-NEXT: [[TMP33:%.*]] = load i32, ptr [[I]], align 4
-// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP33]])
+// CHECK2-NEXT: [[TMP35:%.*]] = load i32, ptr [[I]], align 4
+// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP35]])
// CHECK2-NEXT: br label %[[IF_END]]
// CHECK2: [[IF_END]]:
-// CHECK2-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
-// CHECK2-NEXT: [[CMP21:%.*]] = icmp ult i32 [[TMP34]], [[TMP35]]
+// CHECK2-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[TMP37:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK2-NEXT: [[CMP21:%.*]] = icmp ult i32 [[TMP36]], [[TMP37]]
// CHECK2-NEXT: br i1 [[CMP21]], label %[[IF_THEN22:.*]], label %[[IF_END27:.*]]
// CHECK2: [[IF_THEN22]]:
-// CHECK2-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
-// CHECK2-NEXT: [[TMP37:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
-// CHECK2-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[MUL23:%.*]] = mul i32 [[TMP37]], [[TMP38]]
-// CHECK2-NEXT: [[ADD24:%.*]] = add i32 [[TMP36]], [[MUL23]]
+// CHECK2-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
+// CHECK2-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
+// CHECK2-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[MUL23:%.*]] = mul i32 [[TMP39]], [[TMP40]]
+// CHECK2-NEXT: [[ADD24:%.*]] = add i32 [[TMP38]], [[MUL23]]
// CHECK2-NEXT: store i32 [[ADD24]], ptr [[DOTOMP_IV1]], align 4
-// CHECK2-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
-// CHECK2-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
-// CHECK2-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
-// CHECK2-NEXT: [[MUL25:%.*]] = mul i32 [[TMP40]], [[TMP41]]
-// CHECK2-NEXT: [[ADD26:%.*]] = add i32 [[TMP39]], [[MUL25]]
+// CHECK2-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
+// CHECK2-NEXT: [[TMP42:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
+// CHECK2-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
+// CHECK2-NEXT: [[MUL25:%.*]] = mul i32 [[TMP42]], [[TMP43]]
+// CHECK2-NEXT: [[ADD26:%.*]] = add i32 [[TMP41]], [[MUL25]]
// CHECK2-NEXT: store i32 [[ADD26]], ptr [[J]], align 4
-// CHECK2-NEXT: [[TMP42:%.*]] = load i32, ptr [[J]], align 4
-// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP42]])
+// CHECK2-NEXT: [[TMP44:%.*]] = load i32, ptr [[J]], align 4
+// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP44]])
// CHECK2-NEXT: br label %[[IF_END27]]
// CHECK2: [[IF_END27]]:
// CHECK2-NEXT: br label %[[FOR_INC:.*]]
// CHECK2: [[FOR_INC]]:
-// CHECK2-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[INC:%.*]] = add i32 [[TMP43]], 1
+// CHECK2-NEXT: [[TMP45:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[INC:%.*]] = add i32 [[TMP45]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP3:![0-9]+]]
+// CHECK2-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP2:![0-9]+]]
// CHECK2: [[FOR_END]]:
+// CHECK2-NEXT: [[TMP46:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
+// CHECK2-NEXT: [[TMP47:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK2-NEXT: [[TMP48:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
+// CHECK2-NEXT: [[MUL28:%.*]] = mul i32 [[TMP47]], [[TMP48]]
+// CHECK2-NEXT: [[ADD29:%.*]] = add i32 [[TMP46]], [[MUL28]]
+// CHECK2-NEXT: store i32 [[ADD29]], ptr [[I]], align 4
+// CHECK2-NEXT: [[TMP49:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
+// CHECK2-NEXT: [[TMP50:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK2-NEXT: [[TMP51:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
+// CHECK2-NEXT: [[MUL30:%.*]] = mul i32 [[TMP50]], [[TMP51]]
+// CHECK2-NEXT: [[ADD31:%.*]] = add i32 [[TMP49]], [[MUL30]]
+// CHECK2-NEXT: store i32 [[ADD31]], ptr [[J]], align 4
// CHECK2-NEXT: ret void
//
//
@@ -1364,11 +1466,13 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[ENTRY:.*:]]
// CHECK2-NEXT: [[ARR:%.*]] = alloca [256 x double], align 16
// CHECK2-NEXT: [[I:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_UB0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_LB0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_ST0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_NI0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_IV0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[J:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_UB1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_LB1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_ST1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_NI1:%.*]] = alloca i32, align 4
@@ -1378,43 +1482,48 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[DOTOMP_FUSE_INDEX:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTCAPTURE_EXPR_1:%.*]] = alloca i32, align 4
-// CHECK2-NEXT: [[DOTOMP_LB03:%.*]] = alloca i32, align 4
-// CHECK2-NEXT: [[DOTOMP_ST04:%.*]] = alloca i32, align 4
-// CHECK2-NEXT: [[DOTOMP_NI05:%.*]] = alloca i64, align 8
-// CHECK2-NEXT: [[DOTOMP_IV06:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_UB03:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_LB04:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_ST05:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_NI06:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_IV07:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[C:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[__RANGE2:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: [[__END2:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: [[__BEGIN2:%.*]] = alloca ptr, align 8
-// CHECK2-NEXT: [[DOTCAPTURE_EXPR_8:%.*]] = alloca ptr, align 8
-// CHECK2-NEXT: [[DOTCAPTURE_EXPR_10:%.*]] = alloca ptr, align 8
-// CHECK2-NEXT: [[DOTCAPTURE_EXPR_11:%.*]] = alloca i64, align 8
-// CHECK2-NEXT: [[DOTOMP_LB116:%.*]] = alloca i64, align 8
-// CHECK2-NEXT: [[DOTOMP_ST117:%.*]] = alloca i64, align 8
-// CHECK2-NEXT: [[DOTOMP_NI118:%.*]] = alloca i64, align 8
-// CHECK2-NEXT: [[DOTOMP_IV120:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTCAPTURE_EXPR_9:%.*]] = alloca ptr, align 8
+// CHECK2-NEXT: [[DOTCAPTURE_EXPR_11:%.*]] = alloca ptr, align 8
+// CHECK2-NEXT: [[DOTCAPTURE_EXPR_12:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_UB117:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_LB118:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_ST119:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_NI120:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_IV122:%.*]] = alloca i64, align 8
// CHECK2-NEXT: [[CC:%.*]] = alloca i32, align 4
-// CHECK2-NEXT: [[__RANGE221:%.*]] = alloca ptr, align 8
-// CHECK2-NEXT: [[__END222:%.*]] = alloca ptr, align 8
-// CHECK2-NEXT: [[__BEGIN225:%.*]] = alloca ptr, align 8
-// CHECK2-NEXT: [[DOTCAPTURE_EXPR_27:%.*]] = alloca ptr, align 8
+// CHECK2-NEXT: [[__RANGE223:%.*]] = alloca ptr, align 8
+// CHECK2-NEXT: [[__END224:%.*]] = alloca ptr, align 8
+// CHECK2-NEXT: [[__BEGIN227:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: [[DOTCAPTURE_EXPR_29:%.*]] = alloca ptr, align 8
-// CHECK2-NEXT: [[DOTCAPTURE_EXPR_30:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTCAPTURE_EXPR_31:%.*]] = alloca ptr, align 8
+// CHECK2-NEXT: [[DOTCAPTURE_EXPR_32:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_UB2:%.*]] = alloca i64, align 8
// CHECK2-NEXT: [[DOTOMP_LB2:%.*]] = alloca i64, align 8
// CHECK2-NEXT: [[DOTOMP_ST2:%.*]] = alloca i64, align 8
// CHECK2-NEXT: [[DOTOMP_NI2:%.*]] = alloca i64, align 8
// CHECK2-NEXT: [[DOTOMP_IV2:%.*]] = alloca i64, align 8
-// CHECK2-NEXT: [[DOTOMP_TEMP_140:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_TEMP_142:%.*]] = alloca i64, align 8
// CHECK2-NEXT: [[DOTOMP_TEMP_2:%.*]] = alloca i64, align 8
-// CHECK2-NEXT: [[DOTOMP_FUSE_MAX46:%.*]] = alloca i64, align 8
-// CHECK2-NEXT: [[DOTOMP_FUSE_INDEX52:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_FUSE_MAX48:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_FUSE_INDEX54:%.*]] = alloca i64, align 8
// CHECK2-NEXT: [[V:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: [[VV:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: store i32 0, ptr [[I]], align 4
+// CHECK2-NEXT: store i32 127, ptr [[DOTOMP_UB0]], align 4
// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_LB0]], align 4
// CHECK2-NEXT: store i32 1, ptr [[DOTOMP_ST0]], align 4
// CHECK2-NEXT: store i32 128, ptr [[DOTOMP_NI0]], align 4
// CHECK2-NEXT: store i32 0, ptr [[J]], align 4
+// CHECK2-NEXT: store i32 127, ptr [[DOTOMP_UB1]], align 4
// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_LB1]], align 4
// CHECK2-NEXT: store i32 1, ptr [[DOTOMP_ST1]], align 4
// CHECK2-NEXT: store i32 128, ptr [[DOTOMP_NI1]], align 4
@@ -1441,220 +1550,226 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[DIV:%.*]] = sdiv i32 [[SUB]], 1
// CHECK2-NEXT: [[SUB2:%.*]] = sub nsw i32 [[DIV]], 1
// CHECK2-NEXT: store i32 [[SUB2]], ptr [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_LB03]], align 4
-// CHECK2-NEXT: store i32 1, ptr [[DOTOMP_ST04]], align 4
// CHECK2-NEXT: [[TMP7:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK2-NEXT: [[ADD:%.*]] = add nsw i32 [[TMP7]], 1
+// CHECK2-NEXT: store i32 [[TMP7]], ptr [[DOTOMP_UB03]], align 4
+// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_LB04]], align 4
+// CHECK2-NEXT: store i32 1, ptr [[DOTOMP_ST05]], align 4
+// CHECK2-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK2-NEXT: [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
// CHECK2-NEXT: [[CONV:%.*]] = sext i32 [[ADD]] to i64
-// CHECK2-NEXT: store i64 [[CONV]], ptr [[DOTOMP_NI05]], align 8
+// CHECK2-NEXT: store i64 [[CONV]], ptr [[DOTOMP_NI06]], align 8
// CHECK2-NEXT: store i32 42, ptr [[C]], align 4
// CHECK2-NEXT: store ptr [[ARR]], ptr [[__RANGE2]], align 8
-// CHECK2-NEXT: [[TMP8:%.*]] = load ptr, ptr [[__RANGE2]], align 8
-// CHECK2-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP8]], i64 0, i64 0
+// CHECK2-NEXT: [[TMP9:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META4:![0-9]+]], !align [[META5:![0-9]+]]
+// CHECK2-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP9]], i64 0, i64 0
// CHECK2-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 256
// CHECK2-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK2-NEXT: [[TMP9:%.*]] = load ptr, ptr [[__RANGE2]], align 8
-// CHECK2-NEXT: [[ARRAYDECAY7:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP9]], i64 0, i64 0
-// CHECK2-NEXT: store ptr [[ARRAYDECAY7]], ptr [[__BEGIN2]], align 8
-// CHECK2-NEXT: [[TMP10:%.*]] = load ptr, ptr [[__RANGE2]], align 8
-// CHECK2-NEXT: [[ARRAYDECAY9:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP10]], i64 0, i64 0
-// CHECK2-NEXT: store ptr [[ARRAYDECAY9]], ptr [[DOTCAPTURE_EXPR_8]], align 8
-// CHECK2-NEXT: [[TMP11:%.*]] = load ptr, ptr [[__END2]], align 8
-// CHECK2-NEXT: store ptr [[TMP11]], ptr [[DOTCAPTURE_EXPR_10]], align 8
-// CHECK2-NEXT: [[TMP12:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_10]], align 8
-// CHECK2-NEXT: [[TMP13:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_8]], align 8
-// CHECK2-NEXT: [[SUB_PTR_LHS_CAST:%.*]] = ptrtoint ptr [[TMP12]] to i64
-// CHECK2-NEXT: [[SUB_PTR_RHS_CAST:%.*]] = ptrtoint ptr [[TMP13]] to i64
+// CHECK2-NEXT: [[TMP10:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META4]], !align [[META5]]
+// CHECK2-NEXT: [[ARRAYDECAY8:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP10]], i64 0, i64 0
+// CHECK2-NEXT: store ptr [[ARRAYDECAY8]], ptr [[__BEGIN2]], align 8
+// CHECK2-NEXT: [[TMP11:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META4]], !align [[META5]]
+// CHECK2-NEXT: [[ARRAYDECAY10:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP11]], i64 0, i64 0
+// CHECK2-NEXT: store ptr [[ARRAYDECAY10]], ptr [[DOTCAPTURE_EXPR_9]], align 8
+// CHECK2-NEXT: [[TMP12:%.*]] = load ptr, ptr [[__END2]], align 8
+// CHECK2-NEXT: store ptr [[TMP12]], ptr [[DOTCAPTURE_EXPR_11]], align 8
+// CHECK2-NEXT: [[TMP13:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_11]], align 8
+// CHECK2-NEXT: [[TMP14:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_9]], align 8
+// CHECK2-NEXT: [[SUB_PTR_LHS_CAST:%.*]] = ptrtoint ptr [[TMP13]] to i64
+// CHECK2-NEXT: [[SUB_PTR_RHS_CAST:%.*]] = ptrtoint ptr [[TMP14]] to i64
// CHECK2-NEXT: [[SUB_PTR_SUB:%.*]] = sub i64 [[SUB_PTR_LHS_CAST]], [[SUB_PTR_RHS_CAST]]
// CHECK2-NEXT: [[SUB_PTR_DIV:%.*]] = sdiv exact i64 [[SUB_PTR_SUB]], 8
-// CHECK2-NEXT: [[SUB12:%.*]] = sub nsw i64 [[SUB_PTR_DIV]], 1
-// CHECK2-NEXT: [[ADD13:%.*]] = add nsw i64 [[SUB12]], 1
-// CHECK2-NEXT: [[DIV14:%.*]] = sdiv i64 [[ADD13]], 1
-// CHECK2-NEXT: [[SUB15:%.*]] = sub nsw i64 [[DIV14]], 1
-// CHECK2-NEXT: store i64 [[SUB15]], ptr [[DOTCAPTURE_EXPR_11]], align 8
-// CHECK2-NEXT: store i64 0, ptr [[DOTOMP_LB116]], align 8
-// CHECK2-NEXT: store i64 1, ptr [[DOTOMP_ST117]], align 8
-// CHECK2-NEXT: [[TMP14:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_11]], align 8
-// CHECK2-NEXT: [[ADD19:%.*]] = add nsw i64 [[TMP14]], 1
-// CHECK2-NEXT: store i64 [[ADD19]], ptr [[DOTOMP_NI118]], align 8
+// CHECK2-NEXT: [[SUB13:%.*]] = sub nsw i64 [[SUB_PTR_DIV]], 1
+// CHECK2-NEXT: [[ADD14:%.*]] = add nsw i64 [[SUB13]], 1
+// CHECK2-NEXT: [[DIV15:%.*]] = sdiv i64 [[ADD14]], 1
+// CHECK2-NEXT: [[SUB16:%.*]] = sub nsw i64 [[DIV15]], 1
+// CHECK2-NEXT: store i64 [[SUB16]], ptr [[DOTCAPTURE_EXPR_12]], align 8
+// CHECK2-NEXT: [[TMP15:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_12]], align 8
+// CHECK2-NEXT: store i64 [[TMP15]], ptr [[DOTOMP_UB117]], align 8
+// CHECK2-NEXT: store i64 0, ptr [[DOTOMP_LB118]], align 8
+// CHECK2-NEXT: store i64 1, ptr [[DOTOMP_ST119]], align 8
+// CHECK2-NEXT: [[TMP16:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_12]], align 8
+// CHECK2-NEXT: [[ADD21:%.*]] = add nsw i64 [[TMP16]], 1
+// CHECK2-NEXT: store i64 [[ADD21]], ptr [[DOTOMP_NI120]], align 8
// CHECK2-NEXT: store i32 37, ptr [[CC]], align 4
-// CHECK2-NEXT: store ptr [[ARR]], ptr [[__RANGE221]], align 8
-// CHECK2-NEXT: [[TMP15:%.*]] = load ptr, ptr [[__RANGE221]], align 8
-// CHECK2-NEXT: [[ARRAYDECAY23:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP15]], i64 0, i64 0
-// CHECK2-NEXT: [[ADD_PTR24:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY23]], i64 256
-// CHECK2-NEXT: store ptr [[ADD_PTR24]], ptr [[__END222]], align 8
-// CHECK2-NEXT: [[TMP16:%.*]] = load ptr, ptr [[__RANGE221]], align 8
-// CHECK2-NEXT: [[ARRAYDECAY26:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP16]], i64 0, i64 0
-// CHECK2-NEXT: store ptr [[ARRAYDECAY26]], ptr [[__BEGIN225]], align 8
-// CHECK2-NEXT: [[TMP17:%.*]] = load ptr, ptr [[__RANGE221]], align 8
-// CHECK2-NEXT: [[ARRAYDECAY28:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP17]], i64 0, i64 0
-// CHECK2-NEXT: store ptr [[ARRAYDECAY28]], ptr [[DOTCAPTURE_EXPR_27]], align 8
-// CHECK2-NEXT: [[TMP18:%.*]] = load ptr, ptr [[__END222]], align 8
-// CHECK2-NEXT: store ptr [[TMP18]], ptr [[DOTCAPTURE_EXPR_29]], align 8
-// CHECK2-NEXT: [[TMP19:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_29]], align 8
-// CHECK2-NEXT: [[TMP20:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_27]], align 8
-// CHECK2-NEXT: [[SUB_PTR_LHS_CAST31:%.*]] = ptrtoint ptr [[TMP19]] to i64
-// CHECK2-NEXT: [[SUB_PTR_RHS_CAST32:%.*]] = ptrtoint ptr [[TMP20]] to i64
-// CHECK2-NEXT: [[SUB_PTR_SUB33:%.*]] = sub i64 [[SUB_PTR_LHS_CAST31]], [[SUB_PTR_RHS_CAST32]]
-// CHECK2-NEXT: [[SUB_PTR_DIV34:%.*]] = sdiv exact i64 [[SUB_PTR_SUB33]], 8
-// CHECK2-NEXT: [[SUB35:%.*]] = sub nsw i64 [[SUB_PTR_DIV34]], 1
-// CHECK2-NEXT: [[ADD36:%.*]] = add nsw i64 [[SUB35]], 1
-// CHECK2-NEXT: [[DIV37:%.*]] = sdiv i64 [[ADD36]], 1
-// CHECK2-NEXT: [[SUB38:%.*]] = sub nsw i64 [[DIV37]], 1
-// CHECK2-NEXT: store i64 [[SUB38]], ptr [[DOTCAPTURE_EXPR_30]], align 8
+// CHECK2-NEXT: store ptr [[ARR]], ptr [[__RANGE223]], align 8
+// CHECK2-NEXT: [[TMP17:%.*]] = load ptr, ptr [[__RANGE223]], align 8, !nonnull [[META4]], !align [[META5]]
+// CHECK2-NEXT: [[ARRAYDECAY25:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP17]], i64 0, i64 0
+// CHECK2-NEXT: [[ADD_PTR26:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY25]], i64 256
+// CHECK2-NEXT: store ptr [[ADD_PTR26]], ptr [[__END224]], align 8
+// CHECK2-NEXT: [[TMP18:%.*]] = load ptr, ptr [[__RANGE223]], align 8, !nonnull [[META4]], !align [[META5]]
+// CHECK2-NEXT: [[ARRAYDECAY28:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP18]], i64 0, i64 0
+// CHECK2-NEXT: store ptr [[ARRAYDECAY28]], ptr [[__BEGIN227]], align 8
+// CHECK2-NEXT: [[TMP19:%.*]] = load ptr, ptr [[__RANGE223]], align 8, !nonnull [[META4]], !align [[META5]]
+// CHECK2-NEXT: [[ARRAYDECAY30:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP19]], i64 0, i64 0
+// CHECK2-NEXT: store ptr [[ARRAYDECAY30]], ptr [[DOTCAPTURE_EXPR_29]], align 8
+// CHECK2-NEXT: [[TMP20:%.*]] = load ptr, ptr [[__END224]], align 8
+// CHECK2-NEXT: store ptr [[TMP20]], ptr [[DOTCAPTURE_EXPR_31]], align 8
+// CHECK2-NEXT: [[TMP21:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_31]], align 8
+// CHECK2-NEXT: [[TMP22:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_29]], align 8
+// CHECK2-NEXT: [[SUB_PTR_LHS_CAST33:%.*]] = ptrtoint ptr [[TMP21]] to i64
+// CHECK2-NEXT: [[SUB_PTR_RHS_CAST34:%.*]] = ptrtoint ptr [[TMP22]] to i64
+// CHECK2-NEXT: [[SUB_PTR_SUB35:%.*]] = sub i64 [[SUB_PTR_LHS_CAST33]], [[SUB_PTR_RHS_CAST34]]
+// CHECK2-NEXT: [[SUB_PTR_DIV36:%.*]] = sdiv exact i64 [[SUB_PTR_SUB35]], 8
+// CHECK2-NEXT: [[SUB37:%.*]] = sub nsw i64 [[SUB_PTR_DIV36]], 1
+// CHECK2-NEXT: [[ADD38:%.*]] = add nsw i64 [[SUB37]], 1
+// CHECK2-NEXT: [[DIV39:%.*]] = sdiv i64 [[ADD38]], 1
+// CHECK2-NEXT: [[SUB40:%.*]] = sub nsw i64 [[DIV39]], 1
+// CHECK2-NEXT: store i64 [[SUB40]], ptr [[DOTCAPTURE_EXPR_32]], align 8
+// CHECK2-NEXT: [[TMP23:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_32]], align 8
+// CHECK2-NEXT: store i64 [[TMP23]], ptr [[DOTOMP_UB2]], align 8
// CHECK2-NEXT: store i64 0, ptr [[DOTOMP_LB2]], align 8
// CHECK2-NEXT: store i64 1, ptr [[DOTOMP_ST2]], align 8
-// CHECK2-NEXT: [[TMP21:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_30]], align 8
-// CHECK2-NEXT: [[ADD39:%.*]] = add nsw i64 [[TMP21]], 1
-// CHECK2-NEXT: store i64 [[ADD39]], ptr [[DOTOMP_NI2]], align 8
-// CHECK2-NEXT: [[TMP22:%.*]] = load i64, ptr [[DOTOMP_NI05]], align 8
-// CHECK2-NEXT: store i64 [[TMP22]], ptr [[DOTOMP_TEMP_140]], align 8
-// CHECK2-NEXT: [[TMP23:%.*]] = load i64, ptr [[DOTOMP_TEMP_140]], align 8
-// CHECK2-NEXT: [[TMP24:%.*]] = load i64, ptr [[DOTOMP_NI118]], align 8
-// CHECK2-NEXT: [[CMP41:%.*]] = icmp sgt i64 [[TMP23]], [[TMP24]]
-// CHECK2-NEXT: br i1 [[CMP41]], label %[[COND_TRUE42:.*]], label %[[COND_FALSE43:.*]]
-// CHECK2: [[COND_TRUE42]]:
-// CHECK2-NEXT: [[TMP25:%.*]] = load i64, ptr [[DOTOMP_TEMP_140]], align 8
-// CHECK2-NEXT: br label %[[COND_END44:.*]]
-// CHECK2: [[COND_FALSE43]]:
-// CHECK2-NEXT: [[TMP26:%.*]] = load i64, ptr [[DOTOMP_NI118]], align 8
-// CHECK2-NEXT: br label %[[COND_END44]]
-// CHECK2: [[COND_END44]]:
-// CHECK2-NEXT: [[COND45:%.*]] = phi i64 [ [[TMP25]], %[[COND_TRUE42]] ], [ [[TMP26]], %[[COND_FALSE43]] ]
-// CHECK2-NEXT: store i64 [[COND45]], ptr [[DOTOMP_TEMP_2]], align 8
-// CHECK2-NEXT: [[TMP27:%.*]] = load i64, ptr [[DOTOMP_TEMP_2]], align 8
-// CHECK2-NEXT: [[TMP28:%.*]] = load i64, ptr [[DOTOMP_NI2]], align 8
-// CHECK2-NEXT: [[CMP47:%.*]] = icmp sgt i64 [[TMP27]], [[TMP28]]
-// CHECK2-NEXT: br i1 [[CMP47]], label %[[COND_TRUE48:.*]], label %[[COND_FALSE49:.*]]
-// CHECK2: [[COND_TRUE48]]:
-// CHECK2-NEXT: [[TMP29:%.*]] = load i64, ptr [[DOTOMP_TEMP_2]], align 8
-// CHECK2-NEXT: br label %[[COND_END50:.*]]
-// CHECK2: [[COND_FALSE49]]:
-// CHECK2-NEXT: [[TMP30:%.*]] = load i64, ptr [[DOTOMP_NI2]], align 8
-// CHECK2-NEXT: br label %[[COND_END50]]
-// CHECK2: [[COND_END50]]:
-// CHECK2-NEXT: [[COND51:%.*]] = phi i64 [ [[TMP29]], %[[COND_TRUE48]] ], [ [[TMP30]], %[[COND_FALSE49]] ]
-// CHECK2-NEXT: store i64 [[COND51]], ptr [[DOTOMP_FUSE_MAX46]], align 8
-// CHECK2-NEXT: store i64 0, ptr [[DOTOMP_FUSE_INDEX52]], align 8
+// CHECK2-NEXT: [[TMP24:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_32]], align 8
+// CHECK2-NEXT: [[ADD41:%.*]] = add nsw i64 [[TMP24]], 1
+// CHECK2-NEXT: store i64 [[ADD41]], ptr [[DOTOMP_NI2]], align 8
+// CHECK2-NEXT: [[TMP25:%.*]] = load i64, ptr [[DOTOMP_NI06]], align 8
+// CHECK2-NEXT: store i64 [[TMP25]], ptr [[DOTOMP_TEMP_142]], align 8
+// CHECK2-NEXT: [[TMP26:%.*]] = load i64, ptr [[DOTOMP_TEMP_142]], align 8
+// CHECK2-NEXT: [[TMP27:%.*]] = load i64, ptr [[DOTOMP_NI120]], align 8
+// CHECK2-NEXT: [[CMP43:%.*]] = icmp sgt i64 [[TMP26]], [[TMP27]]
+// CHECK2-NEXT: br i1 [[CMP43]], label %[[COND_TRUE44:.*]], label %[[COND_FALSE45:.*]]
+// CHECK2: [[COND_TRUE44]]:
+// CHECK2-NEXT: [[TMP28:%.*]] = load i64, ptr [[DOTOMP_TEMP_142]], align 8
+// CHECK2-NEXT: br label %[[COND_END46:.*]]
+// CHECK2: [[COND_FALSE45]]:
+// CHECK2-NEXT: [[TMP29:%.*]] = load i64, ptr [[DOTOMP_NI120]], align 8
+// CHECK2-NEXT: br label %[[COND_END46]]
+// CHECK2: [[COND_END46]]:
+// CHECK2-NEXT: [[COND47:%.*]] = phi i64 [ [[TMP28]], %[[COND_TRUE44]] ], [ [[TMP29]], %[[COND_FALSE45]] ]
+// CHECK2-NEXT: store i64 [[COND47]], ptr [[DOTOMP_TEMP_2]], align 8
+// CHECK2-NEXT: [[TMP30:%.*]] = load i64, ptr [[DOTOMP_TEMP_2]], align 8
+// CHECK2-NEXT: [[TMP31:%.*]] = load i64, ptr [[DOTOMP_NI2]], align 8
+// CHECK2-NEXT: [[CMP49:%.*]] = icmp sgt i64 [[TMP30]], [[TMP31]]
+// CHECK2-NEXT: br i1 [[CMP49]], label %[[COND_TRUE50:.*]], label %[[COND_FALSE51:.*]]
+// CHECK2: [[COND_TRUE50]]:
+// CHECK2-NEXT: [[TMP32:%.*]] = load i64, ptr [[DOTOMP_TEMP_2]], align 8
+// CHECK2-NEXT: br label %[[COND_END52:.*]]
+// CHECK2: [[COND_FALSE51]]:
+// CHECK2-NEXT: [[TMP33:%.*]] = load i64, ptr [[DOTOMP_NI2]], align 8
+// CHECK2-NEXT: br label %[[COND_END52]]
+// CHECK2: [[COND_END52]]:
+// CHECK2-NEXT: [[COND53:%.*]] = phi i64 [ [[TMP32]], %[[COND_TRUE50]] ], [ [[TMP33]], %[[COND_FALSE51]] ]
+// CHECK2-NEXT: store i64 [[COND53]], ptr [[DOTOMP_FUSE_MAX48]], align 8
+// CHECK2-NEXT: store i64 0, ptr [[DOTOMP_FUSE_INDEX54]], align 8
// CHECK2-NEXT: br label %[[FOR_COND:.*]]
// CHECK2: [[FOR_COND]]:
-// CHECK2-NEXT: [[TMP31:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK2-NEXT: [[TMP32:%.*]] = load i64, ptr [[DOTOMP_FUSE_MAX46]], align 8
-// CHECK2-NEXT: [[CMP53:%.*]] = icmp slt i64 [[TMP31]], [[TMP32]]
-// CHECK2-NEXT: br i1 [[CMP53]], label %[[FOR_BODY:.*]], label %[[FOR_END:.*]]
+// CHECK2-NEXT: [[TMP34:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK2-NEXT: [[TMP35:%.*]] = load i64, ptr [[DOTOMP_FUSE_MAX48]], align 8
+// CHECK2-NEXT: [[CMP55:%.*]] = icmp slt i64 [[TMP34]], [[TMP35]]
+// CHECK2-NEXT: br i1 [[CMP55]], label %[[FOR_BODY:.*]], label %[[FOR_END:.*]]
// CHECK2: [[FOR_BODY]]:
-// CHECK2-NEXT: [[TMP33:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK2-NEXT: [[TMP34:%.*]] = load i64, ptr [[DOTOMP_NI05]], align 8
-// CHECK2-NEXT: [[CMP54:%.*]] = icmp slt i64 [[TMP33]], [[TMP34]]
-// CHECK2-NEXT: br i1 [[CMP54]], label %[[IF_THEN:.*]], label %[[IF_END74:.*]]
+// CHECK2-NEXT: [[TMP36:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK2-NEXT: [[TMP37:%.*]] = load i64, ptr [[DOTOMP_NI06]], align 8
+// CHECK2-NEXT: [[CMP56:%.*]] = icmp slt i64 [[TMP36]], [[TMP37]]
+// CHECK2-NEXT: br i1 [[CMP56]], label %[[IF_THEN:.*]], label %[[IF_END76:.*]]
// CHECK2: [[IF_THEN]]:
-// CHECK2-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTOMP_LB03]], align 4
-// CHECK2-NEXT: [[CONV55:%.*]] = sext i32 [[TMP35]] to i64
-// CHECK2-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_ST04]], align 4
-// CHECK2-NEXT: [[CONV56:%.*]] = sext i32 [[TMP36]] to i64
-// CHECK2-NEXT: [[TMP37:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK2-NEXT: [[MUL:%.*]] = mul nsw i64 [[CONV56]], [[TMP37]]
-// CHECK2-NEXT: [[ADD57:%.*]] = add nsw i64 [[CONV55]], [[MUL]]
-// CHECK2-NEXT: [[CONV58:%.*]] = trunc i64 [[ADD57]] to i32
-// CHECK2-NEXT: store i32 [[CONV58]], ptr [[DOTOMP_IV06]], align 4
-// CHECK2-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_IV06]], align 4
-// CHECK2-NEXT: [[MUL59:%.*]] = mul nsw i32 [[TMP38]], 1
-// CHECK2-NEXT: [[ADD60:%.*]] = add nsw i32 0, [[MUL59]]
-// CHECK2-NEXT: store i32 [[ADD60]], ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
-// CHECK2-NEXT: [[CMP61:%.*]] = icmp slt i32 [[TMP39]], [[TMP40]]
-// CHECK2-NEXT: br i1 [[CMP61]], label %[[IF_THEN62:.*]], label %[[IF_END:.*]]
-// CHECK2: [[IF_THEN62]]:
-// CHECK2-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
-// CHECK2-NEXT: [[TMP42:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
-// CHECK2-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[MUL63:%.*]] = mul nsw i32 [[TMP42]], [[TMP43]]
-// CHECK2-NEXT: [[ADD64:%.*]] = add nsw i32 [[TMP41]], [[MUL63]]
-// CHECK2-NEXT: store i32 [[ADD64]], ptr [[DOTOMP_IV0]], align 4
-// CHECK2-NEXT: [[TMP44:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
-// CHECK2-NEXT: [[MUL65:%.*]] = mul nsw i32 [[TMP44]], 1
-// CHECK2-NEXT: [[ADD66:%.*]] = add nsw i32 0, [[MUL65]]
-// CHECK2-NEXT: store i32 [[ADD66]], ptr [[I]], align 4
-// CHECK2-NEXT: [[TMP45:%.*]] = load i32, ptr [[I]], align 4
-// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP45]])
+// CHECK2-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_LB04]], align 4
+// CHECK2-NEXT: [[CONV57:%.*]] = sext i32 [[TMP38]] to i64
+// CHECK2-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTOMP_ST05]], align 4
+// CHECK2-NEXT: [[CONV58:%.*]] = sext i32 [[TMP39]] to i64
+// CHECK2-NEXT: [[TMP40:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK2-NEXT: [[MUL:%.*]] = mul nsw i64 [[CONV58]], [[TMP40]]
+// CHECK2-NEXT: [[ADD59:%.*]] = add nsw i64 [[CONV57]], [[MUL]]
+// CHECK2-NEXT: [[CONV60:%.*]] = trunc i64 [[ADD59]] to i32
+// CHECK2-NEXT: store i32 [[CONV60]], ptr [[DOTOMP_IV07]], align 4
+// CHECK2-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTOMP_IV07]], align 4
+// CHECK2-NEXT: [[MUL61:%.*]] = mul nsw i32 [[TMP41]], 1
+// CHECK2-NEXT: [[ADD62:%.*]] = add nsw i32 0, [[MUL61]]
+// CHECK2-NEXT: store i32 [[ADD62]], ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[TMP42:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK2-NEXT: [[CMP63:%.*]] = icmp slt i32 [[TMP42]], [[TMP43]]
+// CHECK2-NEXT: br i1 [[CMP63]], label %[[IF_THEN64:.*]], label %[[IF_END:.*]]
+// CHECK2: [[IF_THEN64]]:
+// CHECK2-NEXT: [[TMP44:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
+// CHECK2-NEXT: [[TMP45:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
+// CHECK2-NEXT: [[TMP46:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[MUL65:%.*]] = mul nsw i32 [[TMP45]], [[TMP46]]
+// CHECK2-NEXT: [[ADD66:%.*]] = add nsw i32 [[TMP44]], [[MUL65]]
+// CHECK2-NEXT: store i32 [[ADD66]], ptr [[DOTOMP_IV0]], align 4
+// CHECK2-NEXT: [[TMP47:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
+// CHECK2-NEXT: [[MUL67:%.*]] = mul nsw i32 [[TMP47]], 1
+// CHECK2-NEXT: [[ADD68:%.*]] = add nsw i32 0, [[MUL67]]
+// CHECK2-NEXT: store i32 [[ADD68]], ptr [[I]], align 4
+// CHECK2-NEXT: [[TMP48:%.*]] = load i32, ptr [[I]], align 4
+// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP48]])
// CHECK2-NEXT: br label %[[IF_END]]
// CHECK2: [[IF_END]]:
-// CHECK2-NEXT: [[TMP46:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[TMP47:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
-// CHECK2-NEXT: [[CMP67:%.*]] = icmp slt i32 [[TMP46]], [[TMP47]]
-// CHECK2-NEXT: br i1 [[CMP67]], label %[[IF_THEN68:.*]], label %[[IF_END73:.*]]
-// CHECK2: [[IF_THEN68]]:
-// CHECK2-NEXT: [[TMP48:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
-// CHECK2-NEXT: [[TMP49:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
-// CHECK2-NEXT: [[TMP50:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[MUL69:%.*]] = mul nsw i32 [[TMP49]], [[TMP50]]
-// CHECK2-NEXT: [[ADD70:%.*]] = add nsw i32 [[TMP48]], [[MUL69]]
-// CHECK2-NEXT: store i32 [[ADD70]], ptr [[DOTOMP_IV1]], align 4
-// CHECK2-NEXT: [[TMP51:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
-// CHECK2-NEXT: [[MUL71:%.*]] = mul nsw i32 [[TMP51]], 2
-// CHECK2-NEXT: [[ADD72:%.*]] = add nsw i32 0, [[MUL71]]
-// CHECK2-NEXT: store i32 [[ADD72]], ptr [[J]], align 4
-// CHECK2-NEXT: [[TMP52:%.*]] = load i32, ptr [[J]], align 4
-// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP52]])
-// CHECK2-NEXT: br label %[[IF_END73]]
-// CHECK2: [[IF_END73]]:
-// CHECK2-NEXT: br label %[[IF_END74]]
-// CHECK2: [[IF_END74]]:
-// CHECK2-NEXT: [[TMP53:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK2-NEXT: [[TMP54:%.*]] = load i64, ptr [[DOTOMP_NI118]], align 8
-// CHECK2-NEXT: [[CMP75:%.*]] = icmp slt i64 [[TMP53]], [[TMP54]]
-// CHECK2-NEXT: br i1 [[CMP75]], label %[[IF_THEN76:.*]], label %[[IF_END81:.*]]
-// CHECK2: [[IF_THEN76]]:
-// CHECK2-NEXT: [[TMP55:%.*]] = load i64, ptr [[DOTOMP_LB116]], align 8
-// CHECK2-NEXT: [[TMP56:%.*]] = load i64, ptr [[DOTOMP_ST117]], align 8
-// CHECK2-NEXT: [[TMP57:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK2-NEXT: [[MUL77:%.*]] = mul nsw i64 [[TMP56]], [[TMP57]]
-// CHECK2-NEXT: [[ADD78:%.*]] = add nsw i64 [[TMP55]], [[MUL77]]
-// CHECK2-NEXT: store i64 [[ADD78]], ptr [[DOTOMP_IV120]], align 8
-// CHECK2-NEXT: [[TMP58:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_8]], align 8
-// CHECK2-NEXT: [[TMP59:%.*]] = load i64, ptr [[DOTOMP_IV120]], align 8
-// CHECK2-NEXT: [[MUL79:%.*]] = mul nsw i64 [[TMP59]], 1
-// CHECK2-NEXT: [[ADD_PTR80:%.*]] = getelementptr inbounds double, ptr [[TMP58]], i64 [[MUL79]]
-// CHECK2-NEXT: store ptr [[ADD_PTR80]], ptr [[__BEGIN2]], align 8
-// CHECK2-NEXT: [[TMP60:%.*]] = load ptr, ptr [[__BEGIN2]], align 8
-// CHECK2-NEXT: store ptr [[TMP60]], ptr [[V]], align 8
-// CHECK2-NEXT: [[TMP61:%.*]] = load i32, ptr [[C]], align 4
-// CHECK2-NEXT: [[TMP62:%.*]] = load ptr, ptr [[V]], align 8
-// CHECK2-NEXT: [[TMP63:%.*]] = load double, ptr [[TMP62]], align 8
-// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP61]], double noundef [[TMP63]])
-// CHECK2-NEXT: br label %[[IF_END81]]
-// CHECK2: [[IF_END81]]:
-// CHECK2-NEXT: [[TMP64:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK2-NEXT: [[TMP65:%.*]] = load i64, ptr [[DOTOMP_NI2]], align 8
-// CHECK2-NEXT: [[CMP82:%.*]] = icmp slt i64 [[TMP64]], [[TMP65]]
-// CHECK2-NEXT: br i1 [[CMP82]], label %[[IF_THEN83:.*]], label %[[IF_END88:.*]]
-// CHECK2: [[IF_THEN83]]:
-// CHECK2-NEXT: [[TMP66:%.*]] = load i64, ptr [[DOTOMP_LB2]], align 8
-// CHECK2-NEXT: [[TMP67:%.*]] = load i64, ptr [[DOTOMP_ST2]], align 8
-// CHECK2-NEXT: [[TMP68:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK2-NEXT: [[MUL84:%.*]] = mul nsw i64 [[TMP67]], [[TMP68]]
-// CHECK2-NEXT: [[ADD85:%.*]] = add nsw i64 [[TMP66]], [[MUL84]]
-// CHECK2-NEXT: store i64 [[ADD85]], ptr [[DOTOMP_IV2]], align 8
-// CHECK2-NEXT: [[TMP69:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_27]], align 8
-// CHECK2-NEXT: [[TMP70:%.*]] = load i64, ptr [[DOTOMP_IV2]], align 8
-// CHECK2-NEXT: [[MUL86:%.*]] = mul nsw i64 [[TMP70]], 1
-// CHECK2-NEXT: [[ADD_PTR87:%.*]] = getelementptr inbounds double, ptr [[TMP69]], i64 [[MUL86]]
-// CHECK2-NEXT: store ptr [[ADD_PTR87]], ptr [[__BEGIN225]], align 8
-// CHECK2-NEXT: [[TMP71:%.*]] = load ptr, ptr [[__BEGIN225]], align 8
-// CHECK2-NEXT: store ptr [[TMP71]], ptr [[VV]], align 8
-// CHECK2-NEXT: [[TMP72:%.*]] = load i32, ptr [[CC]], align 4
-// CHECK2-NEXT: [[TMP73:%.*]] = load ptr, ptr [[VV]], align 8
-// CHECK2-NEXT: [[TMP74:%.*]] = load double, ptr [[TMP73]], align 8
-// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP72]], double noundef [[TMP74]])
-// CHECK2-NEXT: br label %[[IF_END88]]
-// CHECK2: [[IF_END88]]:
+// CHECK2-NEXT: [[TMP49:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[TMP50:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK2-NEXT: [[CMP69:%.*]] = icmp slt i32 [[TMP49]], [[TMP50]]
+// CHECK2-NEXT: br i1 [[CMP69]], label %[[IF_THEN70:.*]], label %[[IF_END75:.*]]
+// CHECK2: [[IF_THEN70]]:
+// CHECK2-NEXT: [[TMP51:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
+// CHECK2-NEXT: [[TMP52:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
+// CHECK2-NEXT: [[TMP53:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[MUL71:%.*]] = mul nsw i32 [[TMP52]], [[TMP53]]
+// CHECK2-NEXT: [[ADD72:%.*]] = add nsw i32 [[TMP51]], [[MUL71]]
+// CHECK2-NEXT: store i32 [[ADD72]], ptr [[DOTOMP_IV1]], align 4
+// CHECK2-NEXT: [[TMP54:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
+// CHECK2-NEXT: [[MUL73:%.*]] = mul nsw i32 [[TMP54]], 2
+// CHECK2-NEXT: [[ADD74:%.*]] = add nsw i32 0, [[MUL73]]
+// CHECK2-NEXT: store i32 [[ADD74]], ptr [[J]], align 4
+// CHECK2-NEXT: [[TMP55:%.*]] = load i32, ptr [[J]], align 4
+// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP55]])
+// CHECK2-NEXT: br label %[[IF_END75]]
+// CHECK2: [[IF_END75]]:
+// CHECK2-NEXT: br label %[[IF_END76]]
+// CHECK2: [[IF_END76]]:
+// CHECK2-NEXT: [[TMP56:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK2-NEXT: [[TMP57:%.*]] = load i64, ptr [[DOTOMP_NI120]], align 8
+// CHECK2-NEXT: [[CMP77:%.*]] = icmp slt i64 [[TMP56]], [[TMP57]]
+// CHECK2-NEXT: br i1 [[CMP77]], label %[[IF_THEN78:.*]], label %[[IF_END83:.*]]
+// CHECK2: [[IF_THEN78]]:
+// CHECK2-NEXT: [[TMP58:%.*]] = load i64, ptr [[DOTOMP_LB118]], align 8
+// CHECK2-NEXT: [[TMP59:%.*]] = load i64, ptr [[DOTOMP_ST119]], align 8
+// CHECK2-NEXT: [[TMP60:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK2-NEXT: [[MUL79:%.*]] = mul nsw i64 [[TMP59]], [[TMP60]]
+// CHECK2-NEXT: [[ADD80:%.*]] = add nsw i64 [[TMP58]], [[MUL79]]
+// CHECK2-NEXT: store i64 [[ADD80]], ptr [[DOTOMP_IV122]], align 8
+// CHECK2-NEXT: [[TMP61:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_9]], align 8
+// CHECK2-NEXT: [[TMP62:%.*]] = load i64, ptr [[DOTOMP_IV122]], align 8
+// CHECK2-NEXT: [[MUL81:%.*]] = mul nsw i64 [[TMP62]], 1
+// CHECK2-NEXT: [[ADD_PTR82:%.*]] = getelementptr inbounds double, ptr [[TMP61]], i64 [[MUL81]]
+// CHECK2-NEXT: store ptr [[ADD_PTR82]], ptr [[__BEGIN2]], align 8
+// CHECK2-NEXT: [[TMP63:%.*]] = load ptr, ptr [[__BEGIN2]], align 8
+// CHECK2-NEXT: store ptr [[TMP63]], ptr [[V]], align 8
+// CHECK2-NEXT: [[TMP64:%.*]] = load i32, ptr [[C]], align 4
+// CHECK2-NEXT: [[TMP65:%.*]] = load ptr, ptr [[V]], align 8, !nonnull [[META4]], !align [[META5]]
+// CHECK2-NEXT: [[TMP66:%.*]] = load double, ptr [[TMP65]], align 8
+// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP64]], double noundef [[TMP66]])
+// CHECK2-NEXT: br label %[[IF_END83]]
+// CHECK2: [[IF_END83]]:
+// CHECK2-NEXT: [[TMP67:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK2-NEXT: [[TMP68:%.*]] = load i64, ptr [[DOTOMP_NI2]], align 8
+// CHECK2-NEXT: [[CMP84:%.*]] = icmp slt i64 [[TMP67]], [[TMP68]]
+// CHECK2-NEXT: br i1 [[CMP84]], label %[[IF_THEN85:.*]], label %[[IF_END90:.*]]
+// CHECK2: [[IF_THEN85]]:
+// CHECK2-NEXT: [[TMP69:%.*]] = load i64, ptr [[DOTOMP_LB2]], align 8
+// CHECK2-NEXT: [[TMP70:%.*]] = load i64, ptr [[DOTOMP_ST2]], align 8
+// CHECK2-NEXT: [[TMP71:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK2-NEXT: [[MUL86:%.*]] = mul nsw i64 [[TMP70]], [[TMP71]]
+// CHECK2-NEXT: [[ADD87:%.*]] = add nsw i64 [[TMP69]], [[MUL86]]
+// CHECK2-NEXT: store i64 [[ADD87]], ptr [[DOTOMP_IV2]], align 8
+// CHECK2-NEXT: [[TMP72:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_29]], align 8
+// CHECK2-NEXT: [[TMP73:%.*]] = load i64, ptr [[DOTOMP_IV2]], align 8
+// CHECK2-NEXT: [[MUL88:%.*]] = mul nsw i64 [[TMP73]], 1
+// CHECK2-NEXT: [[ADD_PTR89:%.*]] = getelementptr inbounds double, ptr [[TMP72]], i64 [[MUL88]]
+// CHECK2-NEXT: store ptr [[ADD_PTR89]], ptr [[__BEGIN227]], align 8
+// CHECK2-NEXT: [[TMP74:%.*]] = load ptr, ptr [[__BEGIN227]], align 8
+// CHECK2-NEXT: store ptr [[TMP74]], ptr [[VV]], align 8
+// CHECK2-NEXT: [[TMP75:%.*]] = load i32, ptr [[CC]], align 4
+// CHECK2-NEXT: [[TMP76:%.*]] = load ptr, ptr [[VV]], align 8, !nonnull [[META4]], !align [[META5]]
+// CHECK2-NEXT: [[TMP77:%.*]] = load double, ptr [[TMP76]], align 8
+// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP75]], double noundef [[TMP77]])
+// CHECK2-NEXT: br label %[[IF_END90]]
+// CHECK2: [[IF_END90]]:
// CHECK2-NEXT: br label %[[FOR_INC:.*]]
// CHECK2: [[FOR_INC]]:
-// CHECK2-NEXT: [[TMP75:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK2-NEXT: [[INC:%.*]] = add nsw i64 [[TMP75]], 1
-// CHECK2-NEXT: store i64 [[INC]], ptr [[DOTOMP_FUSE_INDEX52]], align 8
-// CHECK2-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP5:![0-9]+]]
+// CHECK2-NEXT: [[TMP78:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK2-NEXT: [[INC:%.*]] = add nsw i64 [[TMP78]], 1
+// CHECK2-NEXT: store i64 [[INC]], ptr [[DOTOMP_FUSE_INDEX54]], align 8
+// CHECK2-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
// CHECK2: [[FOR_END]]:
// CHECK2-NEXT: ret void
//
@@ -1664,11 +1779,13 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[ENTRY:.*:]]
// CHECK2-NEXT: [[ARR:%.*]] = alloca [256 x double], align 16
// CHECK2-NEXT: [[J:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_UB0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_LB0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_ST0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_NI0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_IV0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[K:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_UB1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_LB1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_ST1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_NI1:%.*]] = alloca i32, align 4
@@ -1683,10 +1800,12 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[__END2:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: [[V:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: store i32 0, ptr [[J]], align 4
+// CHECK2-NEXT: store i32 127, ptr [[DOTOMP_UB0]], align 4
// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_LB0]], align 4
// CHECK2-NEXT: store i32 1, ptr [[DOTOMP_ST0]], align 4
// CHECK2-NEXT: store i32 128, ptr [[DOTOMP_NI0]], align 4
// CHECK2-NEXT: store i32 0, ptr [[K]], align 4
+// CHECK2-NEXT: store i32 63, ptr [[DOTOMP_UB1]], align 4
// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_LB1]], align 4
// CHECK2-NEXT: store i32 1, ptr [[DOTOMP_ST1]], align 4
// CHECK2-NEXT: store i32 64, ptr [[DOTOMP_NI1]], align 4
@@ -1719,7 +1838,7 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[TMP7:%.*]] = load i32, ptr [[I]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP7]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[I]], align 4
-// CHECK2-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP6:![0-9]+]]
+// CHECK2-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP7:![0-9]+]]
// CHECK2: [[FOR_END]]:
// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_FUSE_INDEX]], align 4
// CHECK2-NEXT: br label %[[FOR_COND2:.*]]
@@ -1772,14 +1891,14 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[TMP24:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
// CHECK2-NEXT: [[INC16:%.*]] = add nsw i32 [[TMP24]], 1
// CHECK2-NEXT: store i32 [[INC16]], ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: br label %[[FOR_COND2]], !llvm.loop [[LOOP7:![0-9]+]]
+// CHECK2-NEXT: br label %[[FOR_COND2]], !llvm.loop [[LOOP8:![0-9]+]]
// CHECK2: [[FOR_END17]]:
// CHECK2-NEXT: store i32 42, ptr [[C]], align 4
// CHECK2-NEXT: store ptr [[ARR]], ptr [[__RANGE2]], align 8
-// CHECK2-NEXT: [[TMP25:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP25:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META4]], !align [[META5]]
// CHECK2-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP25]], i64 0, i64 0
// CHECK2-NEXT: store ptr [[ARRAYDECAY]], ptr [[__BEGIN2]], align 8
-// CHECK2-NEXT: [[TMP26:%.*]] = load ptr, ptr [[__RANGE2]], align 8
+// CHECK2-NEXT: [[TMP26:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META4]], !align [[META5]]
// CHECK2-NEXT: [[ARRAYDECAY18:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP26]], i64 0, i64 0
// CHECK2-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY18]], i64 256
// CHECK2-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
@@ -1793,7 +1912,7 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[TMP29:%.*]] = load ptr, ptr [[__BEGIN2]], align 8
// CHECK2-NEXT: store ptr [[TMP29]], ptr [[V]], align 8
// CHECK2-NEXT: [[TMP30:%.*]] = load i32, ptr [[C]], align 4
-// CHECK2-NEXT: [[TMP31:%.*]] = load ptr, ptr [[V]], align 8
+// CHECK2-NEXT: [[TMP31:%.*]] = load ptr, ptr [[V]], align 8, !nonnull [[META4]], !align [[META5]]
// CHECK2-NEXT: [[TMP32:%.*]] = load double, ptr [[TMP31]], align 8
// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP30]], double noundef [[TMP32]])
// CHECK2-NEXT: br label %[[FOR_INC22:.*]]
@@ -1803,6 +1922,18 @@ extern "C" void foo5() {
// CHECK2-NEXT: store ptr [[INCDEC_PTR]], ptr [[__BEGIN2]], align 8
// CHECK2-NEXT: br label %[[FOR_COND19]]
// CHECK2: [[FOR_END23]]:
+// CHECK2-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
+// CHECK2-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK2-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
+// CHECK2-NEXT: [[MUL24:%.*]] = mul nsw i32 [[TMP35]], [[TMP36]]
+// CHECK2-NEXT: [[ADD25:%.*]] = add nsw i32 [[TMP34]], [[MUL24]]
+// CHECK2-NEXT: store i32 [[ADD25]], ptr [[J]], align 4
+// CHECK2-NEXT: [[TMP37:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
+// CHECK2-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK2-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
+// CHECK2-NEXT: [[MUL26:%.*]] = mul nsw i32 [[TMP38]], [[TMP39]]
+// CHECK2-NEXT: [[ADD27:%.*]] = add nsw i32 [[TMP37]], [[MUL26]]
+// CHECK2-NEXT: store i32 [[ADD27]], ptr [[K]], align 4
// CHECK2-NEXT: ret void
//
//
@@ -1811,11 +1942,13 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[ENTRY:.*:]]
// CHECK2-NEXT: [[ARR:%.*]] = alloca [256 x double], align 16
// CHECK2-NEXT: [[J:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_UB0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_LB0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_ST0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_NI0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_IV0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[K:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_UB1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_LB1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_ST1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_NI1:%.*]] = alloca i32, align 4
@@ -1825,36 +1958,40 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[DOTOMP_FUSE_INDEX:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTCAPTURE_EXPR_1:%.*]] = alloca i32, align 4
-// CHECK2-NEXT: [[DOTOMP_LB03:%.*]] = alloca i32, align 4
-// CHECK2-NEXT: [[DOTOMP_ST04:%.*]] = alloca i32, align 4
-// CHECK2-NEXT: [[DOTOMP_NI05:%.*]] = alloca i64, align 8
-// CHECK2-NEXT: [[DOTOMP_IV06:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_UB03:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_LB04:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_ST05:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_NI06:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_IV07:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[C:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[__RANGE2:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: [[__END2:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: [[__BEGIN2:%.*]] = alloca ptr, align 8
-// CHECK2-NEXT: [[DOTCAPTURE_EXPR_8:%.*]] = alloca ptr, align 8
-// CHECK2-NEXT: [[DOTCAPTURE_EXPR_10:%.*]] = alloca ptr, align 8
-// CHECK2-NEXT: [[DOTCAPTURE_EXPR_11:%.*]] = alloca i64, align 8
-// CHECK2-NEXT: [[DOTOMP_LB116:%.*]] = alloca i64, align 8
-// CHECK2-NEXT: [[DOTOMP_ST117:%.*]] = alloca i64, align 8
-// CHECK2-NEXT: [[DOTOMP_NI118:%.*]] = alloca i64, align 8
-// CHECK2-NEXT: [[DOTOMP_IV120:%.*]] = alloca i64, align 8
-// CHECK2-NEXT: [[DOTOMP_TEMP_121:%.*]] = alloca i64, align 8
-// CHECK2-NEXT: [[DOTOMP_FUSE_MAX22:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTCAPTURE_EXPR_9:%.*]] = alloca ptr, align 8
+// CHECK2-NEXT: [[DOTCAPTURE_EXPR_11:%.*]] = alloca ptr, align 8
+// CHECK2-NEXT: [[DOTCAPTURE_EXPR_12:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_UB117:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_LB118:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_ST119:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_NI120:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_IV122:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_TEMP_123:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_FUSE_MAX24:%.*]] = alloca i64, align 8
// CHECK2-NEXT: [[I:%.*]] = alloca i32, align 4
-// CHECK2-NEXT: [[DOTOMP_FUSE_INDEX29:%.*]] = alloca i64, align 8
+// CHECK2-NEXT: [[DOTOMP_FUSE_INDEX31:%.*]] = alloca i64, align 8
// CHECK2-NEXT: [[V:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: [[CC:%.*]] = alloca i32, align 4
-// CHECK2-NEXT: [[__RANGE264:%.*]] = alloca ptr, align 8
-// CHECK2-NEXT: [[__BEGIN265:%.*]] = alloca ptr, align 8
-// CHECK2-NEXT: [[__END267:%.*]] = alloca ptr, align 8
+// CHECK2-NEXT: [[__RANGE266:%.*]] = alloca ptr, align 8
+// CHECK2-NEXT: [[__BEGIN267:%.*]] = alloca ptr, align 8
+// CHECK2-NEXT: [[__END269:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: [[VV:%.*]] = alloca ptr, align 8
// CHECK2-NEXT: store i32 0, ptr [[J]], align 4
+// CHECK2-NEXT: store i32 127, ptr [[DOTOMP_UB0]], align 4
// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_LB0]], align 4
// CHECK2-NEXT: store i32 1, ptr [[DOTOMP_ST0]], align 4
// CHECK2-NEXT: store i32 128, ptr [[DOTOMP_NI0]], align 4
// CHECK2-NEXT: store i32 0, ptr [[K]], align 4
+// CHECK2-NEXT: store i32 511, ptr [[DOTOMP_UB1]], align 4
// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_LB1]], align 4
// CHECK2-NEXT: store i32 1, ptr [[DOTOMP_ST1]], align 4
// CHECK2-NEXT: store i32 512, ptr [[DOTOMP_NI1]], align 4
@@ -1881,199 +2018,203 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[DIV:%.*]] = sdiv i32 [[SUB]], 1
// CHECK2-NEXT: [[SUB2:%.*]] = sub nsw i32 [[DIV]], 1
// CHECK2-NEXT: store i32 [[SUB2]], ptr [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_LB03]], align 4
-// CHECK2-NEXT: store i32 1, ptr [[DOTOMP_ST04]], align 4
// CHECK2-NEXT: [[TMP7:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
-// CHECK2-NEXT: [[ADD:%.*]] = add nsw i32 [[TMP7]], 1
+// CHECK2-NEXT: store i32 [[TMP7]], ptr [[DOTOMP_UB03]], align 4
+// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_LB04]], align 4
+// CHECK2-NEXT: store i32 1, ptr [[DOTOMP_ST05]], align 4
+// CHECK2-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
+// CHECK2-NEXT: [[ADD:%.*]] = add nsw i32 [[TMP8]], 1
// CHECK2-NEXT: [[CONV:%.*]] = sext i32 [[ADD]] to i64
-// CHECK2-NEXT: store i64 [[CONV]], ptr [[DOTOMP_NI05]], align 8
+// CHECK2-NEXT: store i64 [[CONV]], ptr [[DOTOMP_NI06]], align 8
// CHECK2-NEXT: store i32 42, ptr [[C]], align 4
// CHECK2-NEXT: store ptr [[ARR]], ptr [[__RANGE2]], align 8
-// CHECK2-NEXT: [[TMP8:%.*]] = load ptr, ptr [[__RANGE2]], align 8
-// CHECK2-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP8]], i64 0, i64 0
+// CHECK2-NEXT: [[TMP9:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META4]], !align [[META5]]
+// CHECK2-NEXT: [[ARRAYDECAY:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP9]], i64 0, i64 0
// CHECK2-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY]], i64 256
// CHECK2-NEXT: store ptr [[ADD_PTR]], ptr [[__END2]], align 8
-// CHECK2-NEXT: [[TMP9:%.*]] = load ptr, ptr [[__RANGE2]], align 8
-// CHECK2-NEXT: [[ARRAYDECAY7:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP9]], i64 0, i64 0
-// CHECK2-NEXT: store ptr [[ARRAYDECAY7]], ptr [[__BEGIN2]], align 8
-// CHECK2-NEXT: [[TMP10:%.*]] = load ptr, ptr [[__RANGE2]], align 8
-// CHECK2-NEXT: [[ARRAYDECAY9:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP10]], i64 0, i64 0
-// CHECK2-NEXT: store ptr [[ARRAYDECAY9]], ptr [[DOTCAPTURE_EXPR_8]], align 8
-// CHECK2-NEXT: [[TMP11:%.*]] = load ptr, ptr [[__END2]], align 8
-// CHECK2-NEXT: store ptr [[TMP11]], ptr [[DOTCAPTURE_EXPR_10]], align 8
-// CHECK2-NEXT: [[TMP12:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_10]], align 8
-// CHECK2-NEXT: [[TMP13:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_8]], align 8
-// CHECK2-NEXT: [[SUB_PTR_LHS_CAST:%.*]] = ptrtoint ptr [[TMP12]] to i64
-// CHECK2-NEXT: [[SUB_PTR_RHS_CAST:%.*]] = ptrtoint ptr [[TMP13]] to i64
+// CHECK2-NEXT: [[TMP10:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META4]], !align [[META5]]
+// CHECK2-NEXT: [[ARRAYDECAY8:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP10]], i64 0, i64 0
+// CHECK2-NEXT: store ptr [[ARRAYDECAY8]], ptr [[__BEGIN2]], align 8
+// CHECK2-NEXT: [[TMP11:%.*]] = load ptr, ptr [[__RANGE2]], align 8, !nonnull [[META4]], !align [[META5]]
+// CHECK2-NEXT: [[ARRAYDECAY10:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP11]], i64 0, i64 0
+// CHECK2-NEXT: store ptr [[ARRAYDECAY10]], ptr [[DOTCAPTURE_EXPR_9]], align 8
+// CHECK2-NEXT: [[TMP12:%.*]] = load ptr, ptr [[__END2]], align 8
+// CHECK2-NEXT: store ptr [[TMP12]], ptr [[DOTCAPTURE_EXPR_11]], align 8
+// CHECK2-NEXT: [[TMP13:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_11]], align 8
+// CHECK2-NEXT: [[TMP14:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_9]], align 8
+// CHECK2-NEXT: [[SUB_PTR_LHS_CAST:%.*]] = ptrtoint ptr [[TMP13]] to i64
+// CHECK2-NEXT: [[SUB_PTR_RHS_CAST:%.*]] = ptrtoint ptr [[TMP14]] to i64
// CHECK2-NEXT: [[SUB_PTR_SUB:%.*]] = sub i64 [[SUB_PTR_LHS_CAST]], [[SUB_PTR_RHS_CAST]]
// CHECK2-NEXT: [[SUB_PTR_DIV:%.*]] = sdiv exact i64 [[SUB_PTR_SUB]], 8
-// CHECK2-NEXT: [[SUB12:%.*]] = sub nsw i64 [[SUB_PTR_DIV]], 1
-// CHECK2-NEXT: [[ADD13:%.*]] = add nsw i64 [[SUB12]], 1
-// CHECK2-NEXT: [[DIV14:%.*]] = sdiv i64 [[ADD13]], 1
-// CHECK2-NEXT: [[SUB15:%.*]] = sub nsw i64 [[DIV14]], 1
-// CHECK2-NEXT: store i64 [[SUB15]], ptr [[DOTCAPTURE_EXPR_11]], align 8
-// CHECK2-NEXT: store i64 0, ptr [[DOTOMP_LB116]], align 8
-// CHECK2-NEXT: store i64 1, ptr [[DOTOMP_ST117]], align 8
-// CHECK2-NEXT: [[TMP14:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_11]], align 8
-// CHECK2-NEXT: [[ADD19:%.*]] = add nsw i64 [[TMP14]], 1
-// CHECK2-NEXT: store i64 [[ADD19]], ptr [[DOTOMP_NI118]], align 8
-// CHECK2-NEXT: [[TMP15:%.*]] = load i64, ptr [[DOTOMP_NI05]], align 8
-// CHECK2-NEXT: store i64 [[TMP15]], ptr [[DOTOMP_TEMP_121]], align 8
-// CHECK2-NEXT: [[TMP16:%.*]] = load i64, ptr [[DOTOMP_TEMP_121]], align 8
-// CHECK2-NEXT: [[TMP17:%.*]] = load i64, ptr [[DOTOMP_NI118]], align 8
-// CHECK2-NEXT: [[CMP23:%.*]] = icmp sgt i64 [[TMP16]], [[TMP17]]
-// CHECK2-NEXT: br i1 [[CMP23]], label %[[COND_TRUE24:.*]], label %[[COND_FALSE25:.*]]
-// CHECK2: [[COND_TRUE24]]:
-// CHECK2-NEXT: [[TMP18:%.*]] = load i64, ptr [[DOTOMP_TEMP_121]], align 8
-// CHECK2-NEXT: br label %[[COND_END26:.*]]
-// CHECK2: [[COND_FALSE25]]:
-// CHECK2-NEXT: [[TMP19:%.*]] = load i64, ptr [[DOTOMP_NI118]], align 8
-// CHECK2-NEXT: br label %[[COND_END26]]
-// CHECK2: [[COND_END26]]:
-// CHECK2-NEXT: [[COND27:%.*]] = phi i64 [ [[TMP18]], %[[COND_TRUE24]] ], [ [[TMP19]], %[[COND_FALSE25]] ]
-// CHECK2-NEXT: store i64 [[COND27]], ptr [[DOTOMP_FUSE_MAX22]], align 8
+// CHECK2-NEXT: [[SUB13:%.*]] = sub nsw i64 [[SUB_PTR_DIV]], 1
+// CHECK2-NEXT: [[ADD14:%.*]] = add nsw i64 [[SUB13]], 1
+// CHECK2-NEXT: [[DIV15:%.*]] = sdiv i64 [[ADD14]], 1
+// CHECK2-NEXT: [[SUB16:%.*]] = sub nsw i64 [[DIV15]], 1
+// CHECK2-NEXT: store i64 [[SUB16]], ptr [[DOTCAPTURE_EXPR_12]], align 8
+// CHECK2-NEXT: [[TMP15:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_12]], align 8
+// CHECK2-NEXT: store i64 [[TMP15]], ptr [[DOTOMP_UB117]], align 8
+// CHECK2-NEXT: store i64 0, ptr [[DOTOMP_LB118]], align 8
+// CHECK2-NEXT: store i64 1, ptr [[DOTOMP_ST119]], align 8
+// CHECK2-NEXT: [[TMP16:%.*]] = load i64, ptr [[DOTCAPTURE_EXPR_12]], align 8
+// CHECK2-NEXT: [[ADD21:%.*]] = add nsw i64 [[TMP16]], 1
+// CHECK2-NEXT: store i64 [[ADD21]], ptr [[DOTOMP_NI120]], align 8
+// CHECK2-NEXT: [[TMP17:%.*]] = load i64, ptr [[DOTOMP_NI06]], align 8
+// CHECK2-NEXT: store i64 [[TMP17]], ptr [[DOTOMP_TEMP_123]], align 8
+// CHECK2-NEXT: [[TMP18:%.*]] = load i64, ptr [[DOTOMP_TEMP_123]], align 8
+// CHECK2-NEXT: [[TMP19:%.*]] = load i64, ptr [[DOTOMP_NI120]], align 8
+// CHECK2-NEXT: [[CMP25:%.*]] = icmp sgt i64 [[TMP18]], [[TMP19]]
+// CHECK2-NEXT: br i1 [[CMP25]], label %[[COND_TRUE26:.*]], label %[[COND_FALSE27:.*]]
+// CHECK2: [[COND_TRUE26]]:
+// CHECK2-NEXT: [[TMP20:%.*]] = load i64, ptr [[DOTOMP_TEMP_123]], align 8
+// CHECK2-NEXT: br label %[[COND_END28:.*]]
+// CHECK2: [[COND_FALSE27]]:
+// CHECK2-NEXT: [[TMP21:%.*]] = load i64, ptr [[DOTOMP_NI120]], align 8
+// CHECK2-NEXT: br label %[[COND_END28]]
+// CHECK2: [[COND_END28]]:
+// CHECK2-NEXT: [[COND29:%.*]] = phi i64 [ [[TMP20]], %[[COND_TRUE26]] ], [ [[TMP21]], %[[COND_FALSE27]] ]
+// CHECK2-NEXT: store i64 [[COND29]], ptr [[DOTOMP_FUSE_MAX24]], align 8
// CHECK2-NEXT: store i32 0, ptr [[I]], align 4
// CHECK2-NEXT: br label %[[FOR_COND:.*]]
// CHECK2: [[FOR_COND]]:
-// CHECK2-NEXT: [[TMP20:%.*]] = load i32, ptr [[I]], align 4
-// CHECK2-NEXT: [[CMP28:%.*]] = icmp slt i32 [[TMP20]], 128
-// CHECK2-NEXT: br i1 [[CMP28]], label %[[FOR_BODY:.*]], label %[[FOR_END:.*]]
+// CHECK2-NEXT: [[TMP22:%.*]] = load i32, ptr [[I]], align 4
+// CHECK2-NEXT: [[CMP30:%.*]] = icmp slt i32 [[TMP22]], 128
+// CHECK2-NEXT: br i1 [[CMP30]], label %[[FOR_BODY:.*]], label %[[FOR_END:.*]]
// CHECK2: [[FOR_BODY]]:
-// CHECK2-NEXT: [[TMP21:%.*]] = load i32, ptr [[I]], align 4
-// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP21]])
+// CHECK2-NEXT: [[TMP23:%.*]] = load i32, ptr [[I]], align 4
+// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP23]])
// CHECK2-NEXT: br label %[[FOR_INC:.*]]
// CHECK2: [[FOR_INC]]:
-// CHECK2-NEXT: [[TMP22:%.*]] = load i32, ptr [[I]], align 4
-// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP22]], 1
+// CHECK2-NEXT: [[TMP24:%.*]] = load i32, ptr [[I]], align 4
+// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP24]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[I]], align 4
-// CHECK2-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
+// CHECK2-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP9:![0-9]+]]
// CHECK2: [[FOR_END]]:
-// CHECK2-NEXT: store i64 0, ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK2-NEXT: br label %[[FOR_COND30:.*]]
-// CHECK2: [[FOR_COND30]]:
-// CHECK2-NEXT: [[TMP23:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK2-NEXT: [[TMP24:%.*]] = load i64, ptr [[DOTOMP_FUSE_MAX22]], align 8
-// CHECK2-NEXT: [[CMP31:%.*]] = icmp slt i64 [[TMP23]], [[TMP24]]
-// CHECK2-NEXT: br i1 [[CMP31]], label %[[FOR_BODY32:.*]], label %[[FOR_END63:.*]]
-// CHECK2: [[FOR_BODY32]]:
-// CHECK2-NEXT: [[TMP25:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK2-NEXT: [[TMP26:%.*]] = load i64, ptr [[DOTOMP_NI05]], align 8
+// CHECK2-NEXT: store i64 0, ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK2-NEXT: br label %[[FOR_COND32:.*]]
+// CHECK2: [[FOR_COND32]]:
+// CHECK2-NEXT: [[TMP25:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK2-NEXT: [[TMP26:%.*]] = load i64, ptr [[DOTOMP_FUSE_MAX24]], align 8
// CHECK2-NEXT: [[CMP33:%.*]] = icmp slt i64 [[TMP25]], [[TMP26]]
-// CHECK2-NEXT: br i1 [[CMP33]], label %[[IF_THEN:.*]], label %[[IF_END53:.*]]
+// CHECK2-NEXT: br i1 [[CMP33]], label %[[FOR_BODY34:.*]], label %[[FOR_END65:.*]]
+// CHECK2: [[FOR_BODY34]]:
+// CHECK2-NEXT: [[TMP27:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK2-NEXT: [[TMP28:%.*]] = load i64, ptr [[DOTOMP_NI06]], align 8
+// CHECK2-NEXT: [[CMP35:%.*]] = icmp slt i64 [[TMP27]], [[TMP28]]
+// CHECK2-NEXT: br i1 [[CMP35]], label %[[IF_THEN:.*]], label %[[IF_END55:.*]]
// CHECK2: [[IF_THEN]]:
-// CHECK2-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTOMP_LB03]], align 4
-// CHECK2-NEXT: [[CONV34:%.*]] = sext i32 [[TMP27]] to i64
-// CHECK2-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTOMP_ST04]], align 4
-// CHECK2-NEXT: [[CONV35:%.*]] = sext i32 [[TMP28]] to i64
-// CHECK2-NEXT: [[TMP29:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK2-NEXT: [[MUL:%.*]] = mul nsw i64 [[CONV35]], [[TMP29]]
-// CHECK2-NEXT: [[ADD36:%.*]] = add nsw i64 [[CONV34]], [[MUL]]
-// CHECK2-NEXT: [[CONV37:%.*]] = trunc i64 [[ADD36]] to i32
-// CHECK2-NEXT: store i32 [[CONV37]], ptr [[DOTOMP_IV06]], align 4
-// CHECK2-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTOMP_IV06]], align 4
-// CHECK2-NEXT: [[MUL38:%.*]] = mul nsw i32 [[TMP30]], 1
-// CHECK2-NEXT: [[ADD39:%.*]] = add nsw i32 0, [[MUL38]]
-// CHECK2-NEXT: store i32 [[ADD39]], ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
-// CHECK2-NEXT: [[CMP40:%.*]] = icmp slt i32 [[TMP31]], [[TMP32]]
-// CHECK2-NEXT: br i1 [[CMP40]], label %[[IF_THEN41:.*]], label %[[IF_END:.*]]
-// CHECK2: [[IF_THEN41]]:
-// CHECK2-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
-// CHECK2-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
-// CHECK2-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[MUL42:%.*]] = mul nsw i32 [[TMP34]], [[TMP35]]
-// CHECK2-NEXT: [[ADD43:%.*]] = add nsw i32 [[TMP33]], [[MUL42]]
-// CHECK2-NEXT: store i32 [[ADD43]], ptr [[DOTOMP_IV0]], align 4
-// CHECK2-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
-// CHECK2-NEXT: [[MUL44:%.*]] = mul nsw i32 [[TMP36]], 2
-// CHECK2-NEXT: [[ADD45:%.*]] = add nsw i32 0, [[MUL44]]
-// CHECK2-NEXT: store i32 [[ADD45]], ptr [[J]], align 4
-// CHECK2-NEXT: [[TMP37:%.*]] = load i32, ptr [[J]], align 4
-// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP37]])
+// CHECK2-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTOMP_LB04]], align 4
+// CHECK2-NEXT: [[CONV36:%.*]] = sext i32 [[TMP29]] to i64
+// CHECK2-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTOMP_ST05]], align 4
+// CHECK2-NEXT: [[CONV37:%.*]] = sext i32 [[TMP30]] to i64
+// CHECK2-NEXT: [[TMP31:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK2-NEXT: [[MUL:%.*]] = mul nsw i64 [[CONV37]], [[TMP31]]
+// CHECK2-NEXT: [[ADD38:%.*]] = add nsw i64 [[CONV36]], [[MUL]]
+// CHECK2-NEXT: [[CONV39:%.*]] = trunc i64 [[ADD38]] to i32
+// CHECK2-NEXT: store i32 [[CONV39]], ptr [[DOTOMP_IV07]], align 4
+// CHECK2-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTOMP_IV07]], align 4
+// CHECK2-NEXT: [[MUL40:%.*]] = mul nsw i32 [[TMP32]], 1
+// CHECK2-NEXT: [[ADD41:%.*]] = add nsw i32 0, [[MUL40]]
+// CHECK2-NEXT: store i32 [[ADD41]], ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK2-NEXT: [[CMP42:%.*]] = icmp slt i32 [[TMP33]], [[TMP34]]
+// CHECK2-NEXT: br i1 [[CMP42]], label %[[IF_THEN43:.*]], label %[[IF_END:.*]]
+// CHECK2: [[IF_THEN43]]:
+// CHECK2-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
+// CHECK2-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
+// CHECK2-NEXT: [[TMP37:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[MUL44:%.*]] = mul nsw i32 [[TMP36]], [[TMP37]]
+// CHECK2-NEXT: [[ADD45:%.*]] = add nsw i32 [[TMP35]], [[MUL44]]
+// CHECK2-NEXT: store i32 [[ADD45]], ptr [[DOTOMP_IV0]], align 4
+// CHECK2-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
+// CHECK2-NEXT: [[MUL46:%.*]] = mul nsw i32 [[TMP38]], 2
+// CHECK2-NEXT: [[ADD47:%.*]] = add nsw i32 0, [[MUL46]]
+// CHECK2-NEXT: store i32 [[ADD47]], ptr [[J]], align 4
+// CHECK2-NEXT: [[TMP39:%.*]] = load i32, ptr [[J]], align 4
+// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP39]])
// CHECK2-NEXT: br label %[[IF_END]]
// CHECK2: [[IF_END]]:
-// CHECK2-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
-// CHECK2-NEXT: [[CMP46:%.*]] = icmp slt i32 [[TMP38]], [[TMP39]]
-// CHECK2-NEXT: br i1 [[CMP46]], label %[[IF_THEN47:.*]], label %[[IF_END52:.*]]
-// CHECK2: [[IF_THEN47]]:
-// CHECK2-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
-// CHECK2-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
-// CHECK2-NEXT: [[TMP42:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[MUL48:%.*]] = mul nsw i32 [[TMP41]], [[TMP42]]
-// CHECK2-NEXT: [[ADD49:%.*]] = add nsw i32 [[TMP40]], [[MUL48]]
-// CHECK2-NEXT: store i32 [[ADD49]], ptr [[DOTOMP_IV1]], align 4
-// CHECK2-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
-// CHECK2-NEXT: [[MUL50:%.*]] = mul nsw i32 [[TMP43]], 1
-// CHECK2-NEXT: [[ADD51:%.*]] = add nsw i32 0, [[MUL50]]
-// CHECK2-NEXT: store i32 [[ADD51]], ptr [[K]], align 4
-// CHECK2-NEXT: [[TMP44:%.*]] = load i32, ptr [[K]], align 4
-// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP44]])
-// CHECK2-NEXT: br label %[[IF_END52]]
-// CHECK2: [[IF_END52]]:
-// CHECK2-NEXT: br label %[[IF_END53]]
-// CHECK2: [[IF_END53]]:
-// CHECK2-NEXT: [[TMP45:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK2-NEXT: [[TMP46:%.*]] = load i64, ptr [[DOTOMP_NI118]], align 8
-// CHECK2-NEXT: [[CMP54:%.*]] = icmp slt i64 [[TMP45]], [[TMP46]]
-// CHECK2-NEXT: br i1 [[CMP54]], label %[[IF_THEN55:.*]], label %[[IF_END60:.*]]
-// CHECK2: [[IF_THEN55]]:
-// CHECK2-NEXT: [[TMP47:%.*]] = load i64, ptr [[DOTOMP_LB116]], align 8
-// CHECK2-NEXT: [[TMP48:%.*]] = load i64, ptr [[DOTOMP_ST117]], align 8
-// CHECK2-NEXT: [[TMP49:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK2-NEXT: [[MUL56:%.*]] = mul nsw i64 [[TMP48]], [[TMP49]]
-// CHECK2-NEXT: [[ADD57:%.*]] = add nsw i64 [[TMP47]], [[MUL56]]
-// CHECK2-NEXT: store i64 [[ADD57]], ptr [[DOTOMP_IV120]], align 8
-// CHECK2-NEXT: [[TMP50:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_8]], align 8
-// CHECK2-NEXT: [[TMP51:%.*]] = load i64, ptr [[DOTOMP_IV120]], align 8
-// CHECK2-NEXT: [[MUL58:%.*]] = mul nsw i64 [[TMP51]], 1
-// CHECK2-NEXT: [[ADD_PTR59:%.*]] = getelementptr inbounds double, ptr [[TMP50]], i64 [[MUL58]]
-// CHECK2-NEXT: store ptr [[ADD_PTR59]], ptr [[__BEGIN2]], align 8
-// CHECK2-NEXT: [[TMP52:%.*]] = load ptr, ptr [[__BEGIN2]], align 8
-// CHECK2-NEXT: store ptr [[TMP52]], ptr [[V]], align 8
-// CHECK2-NEXT: [[TMP53:%.*]] = load i32, ptr [[C]], align 4
-// CHECK2-NEXT: [[TMP54:%.*]] = load ptr, ptr [[V]], align 8
-// CHECK2-NEXT: [[TMP55:%.*]] = load double, ptr [[TMP54]], align 8
-// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP53]], double noundef [[TMP55]])
-// CHECK2-NEXT: br label %[[IF_END60]]
-// CHECK2: [[IF_END60]]:
-// CHECK2-NEXT: br label %[[FOR_INC61:.*]]
-// CHECK2: [[FOR_INC61]]:
-// CHECK2-NEXT: [[TMP56:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK2-NEXT: [[INC62:%.*]] = add nsw i64 [[TMP56]], 1
-// CHECK2-NEXT: store i64 [[INC62]], ptr [[DOTOMP_FUSE_INDEX29]], align 8
-// CHECK2-NEXT: br label %[[FOR_COND30]], !llvm.loop [[LOOP9:![0-9]+]]
-// CHECK2: [[FOR_END63]]:
+// CHECK2-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK2-NEXT: [[CMP48:%.*]] = icmp slt i32 [[TMP40]], [[TMP41]]
+// CHECK2-NEXT: br i1 [[CMP48]], label %[[IF_THEN49:.*]], label %[[IF_END54:.*]]
+// CHECK2: [[IF_THEN49]]:
+// CHECK2-NEXT: [[TMP42:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
+// CHECK2-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
+// CHECK2-NEXT: [[TMP44:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[MUL50:%.*]] = mul nsw i32 [[TMP43]], [[TMP44]]
+// CHECK2-NEXT: [[ADD51:%.*]] = add nsw i32 [[TMP42]], [[MUL50]]
+// CHECK2-NEXT: store i32 [[ADD51]], ptr [[DOTOMP_IV1]], align 4
+// CHECK2-NEXT: [[TMP45:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
+// CHECK2-NEXT: [[MUL52:%.*]] = mul nsw i32 [[TMP45]], 1
+// CHECK2-NEXT: [[ADD53:%.*]] = add nsw i32 0, [[MUL52]]
+// CHECK2-NEXT: store i32 [[ADD53]], ptr [[K]], align 4
+// CHECK2-NEXT: [[TMP46:%.*]] = load i32, ptr [[K]], align 4
+// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP46]])
+// CHECK2-NEXT: br label %[[IF_END54]]
+// CHECK2: [[IF_END54]]:
+// CHECK2-NEXT: br label %[[IF_END55]]
+// CHECK2: [[IF_END55]]:
+// CHECK2-NEXT: [[TMP47:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK2-NEXT: [[TMP48:%.*]] = load i64, ptr [[DOTOMP_NI120]], align 8
+// CHECK2-NEXT: [[CMP56:%.*]] = icmp slt i64 [[TMP47]], [[TMP48]]
+// CHECK2-NEXT: br i1 [[CMP56]], label %[[IF_THEN57:.*]], label %[[IF_END62:.*]]
+// CHECK2: [[IF_THEN57]]:
+// CHECK2-NEXT: [[TMP49:%.*]] = load i64, ptr [[DOTOMP_LB118]], align 8
+// CHECK2-NEXT: [[TMP50:%.*]] = load i64, ptr [[DOTOMP_ST119]], align 8
+// CHECK2-NEXT: [[TMP51:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK2-NEXT: [[MUL58:%.*]] = mul nsw i64 [[TMP50]], [[TMP51]]
+// CHECK2-NEXT: [[ADD59:%.*]] = add nsw i64 [[TMP49]], [[MUL58]]
+// CHECK2-NEXT: store i64 [[ADD59]], ptr [[DOTOMP_IV122]], align 8
+// CHECK2-NEXT: [[TMP52:%.*]] = load ptr, ptr [[DOTCAPTURE_EXPR_9]], align 8
+// CHECK2-NEXT: [[TMP53:%.*]] = load i64, ptr [[DOTOMP_IV122]], align 8
+// CHECK2-NEXT: [[MUL60:%.*]] = mul nsw i64 [[TMP53]], 1
+// CHECK2-NEXT: [[ADD_PTR61:%.*]] = getelementptr inbounds double, ptr [[TMP52]], i64 [[MUL60]]
+// CHECK2-NEXT: store ptr [[ADD_PTR61]], ptr [[__BEGIN2]], align 8
+// CHECK2-NEXT: [[TMP54:%.*]] = load ptr, ptr [[__BEGIN2]], align 8
+// CHECK2-NEXT: store ptr [[TMP54]], ptr [[V]], align 8
+// CHECK2-NEXT: [[TMP55:%.*]] = load i32, ptr [[C]], align 4
+// CHECK2-NEXT: [[TMP56:%.*]] = load ptr, ptr [[V]], align 8, !nonnull [[META4]], !align [[META5]]
+// CHECK2-NEXT: [[TMP57:%.*]] = load double, ptr [[TMP56]], align 8
+// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP55]], double noundef [[TMP57]])
+// CHECK2-NEXT: br label %[[IF_END62]]
+// CHECK2: [[IF_END62]]:
+// CHECK2-NEXT: br label %[[FOR_INC63:.*]]
+// CHECK2: [[FOR_INC63]]:
+// CHECK2-NEXT: [[TMP58:%.*]] = load i64, ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK2-NEXT: [[INC64:%.*]] = add nsw i64 [[TMP58]], 1
+// CHECK2-NEXT: store i64 [[INC64]], ptr [[DOTOMP_FUSE_INDEX31]], align 8
+// CHECK2-NEXT: br label %[[FOR_COND32]], !llvm.loop [[LOOP10:![0-9]+]]
+// CHECK2: [[FOR_END65]]:
// CHECK2-NEXT: store i32 37, ptr [[CC]], align 4
-// CHECK2-NEXT: store ptr [[ARR]], ptr [[__RANGE264]], align 8
-// CHECK2-NEXT: [[TMP57:%.*]] = load ptr, ptr [[__RANGE264]], align 8
-// CHECK2-NEXT: [[ARRAYDECAY66:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP57]], i64 0, i64 0
-// CHECK2-NEXT: store ptr [[ARRAYDECAY66]], ptr [[__BEGIN265]], align 8
-// CHECK2-NEXT: [[TMP58:%.*]] = load ptr, ptr [[__RANGE264]], align 8
-// CHECK2-NEXT: [[ARRAYDECAY68:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP58]], i64 0, i64 0
-// CHECK2-NEXT: [[ADD_PTR69:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY68]], i64 256
-// CHECK2-NEXT: store ptr [[ADD_PTR69]], ptr [[__END267]], align 8
-// CHECK2-NEXT: br label %[[FOR_COND70:.*]]
-// CHECK2: [[FOR_COND70]]:
-// CHECK2-NEXT: [[TMP59:%.*]] = load ptr, ptr [[__BEGIN265]], align 8
-// CHECK2-NEXT: [[TMP60:%.*]] = load ptr, ptr [[__END267]], align 8
-// CHECK2-NEXT: [[CMP71:%.*]] = icmp ne ptr [[TMP59]], [[TMP60]]
-// CHECK2-NEXT: br i1 [[CMP71]], label %[[FOR_BODY72:.*]], label %[[FOR_END74:.*]]
-// CHECK2: [[FOR_BODY72]]:
-// CHECK2-NEXT: [[TMP61:%.*]] = load ptr, ptr [[__BEGIN265]], align 8
-// CHECK2-NEXT: store ptr [[TMP61]], ptr [[VV]], align 8
-// CHECK2-NEXT: [[TMP62:%.*]] = load i32, ptr [[CC]], align 4
-// CHECK2-NEXT: [[TMP63:%.*]] = load ptr, ptr [[VV]], align 8
-// CHECK2-NEXT: [[TMP64:%.*]] = load double, ptr [[TMP63]], align 8
-// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP62]], double noundef [[TMP64]])
-// CHECK2-NEXT: br label %[[FOR_INC73:.*]]
-// CHECK2: [[FOR_INC73]]:
-// CHECK2-NEXT: [[TMP65:%.*]] = load ptr, ptr [[__BEGIN265]], align 8
-// CHECK2-NEXT: [[INCDEC_PTR:%.*]] = getelementptr inbounds nuw double, ptr [[TMP65]], i32 1
-// CHECK2-NEXT: store ptr [[INCDEC_PTR]], ptr [[__BEGIN265]], align 8
-// CHECK2-NEXT: br label %[[FOR_COND70]]
-// CHECK2: [[FOR_END74]]:
+// CHECK2-NEXT: store ptr [[ARR]], ptr [[__RANGE266]], align 8
+// CHECK2-NEXT: [[TMP59:%.*]] = load ptr, ptr [[__RANGE266]], align 8, !nonnull [[META4]], !align [[META5]]
+// CHECK2-NEXT: [[ARRAYDECAY68:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP59]], i64 0, i64 0
+// CHECK2-NEXT: store ptr [[ARRAYDECAY68]], ptr [[__BEGIN267]], align 8
+// CHECK2-NEXT: [[TMP60:%.*]] = load ptr, ptr [[__RANGE266]], align 8, !nonnull [[META4]], !align [[META5]]
+// CHECK2-NEXT: [[ARRAYDECAY70:%.*]] = getelementptr inbounds [256 x double], ptr [[TMP60]], i64 0, i64 0
+// CHECK2-NEXT: [[ADD_PTR71:%.*]] = getelementptr inbounds double, ptr [[ARRAYDECAY70]], i64 256
+// CHECK2-NEXT: store ptr [[ADD_PTR71]], ptr [[__END269]], align 8
+// CHECK2-NEXT: br label %[[FOR_COND72:.*]]
+// CHECK2: [[FOR_COND72]]:
+// CHECK2-NEXT: [[TMP61:%.*]] = load ptr, ptr [[__BEGIN267]], align 8
+// CHECK2-NEXT: [[TMP62:%.*]] = load ptr, ptr [[__END269]], align 8
+// CHECK2-NEXT: [[CMP73:%.*]] = icmp ne ptr [[TMP61]], [[TMP62]]
+// CHECK2-NEXT: br i1 [[CMP73]], label %[[FOR_BODY74:.*]], label %[[FOR_END76:.*]]
+// CHECK2: [[FOR_BODY74]]:
+// CHECK2-NEXT: [[TMP63:%.*]] = load ptr, ptr [[__BEGIN267]], align 8
+// CHECK2-NEXT: store ptr [[TMP63]], ptr [[VV]], align 8
+// CHECK2-NEXT: [[TMP64:%.*]] = load i32, ptr [[CC]], align 4
+// CHECK2-NEXT: [[TMP65:%.*]] = load ptr, ptr [[VV]], align 8, !nonnull [[META4]], !align [[META5]]
+// CHECK2-NEXT: [[TMP66:%.*]] = load double, ptr [[TMP65]], align 8
+// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP64]], double noundef [[TMP66]])
+// CHECK2-NEXT: br label %[[FOR_INC75:.*]]
+// CHECK2: [[FOR_INC75]]:
+// CHECK2-NEXT: [[TMP67:%.*]] = load ptr, ptr [[__BEGIN267]], align 8
+// CHECK2-NEXT: [[INCDEC_PTR:%.*]] = getelementptr inbounds nuw double, ptr [[TMP67]], i32 1
+// CHECK2-NEXT: store ptr [[INCDEC_PTR]], ptr [[__BEGIN267]], align 8
+// CHECK2-NEXT: br label %[[FOR_COND72]]
+// CHECK2: [[FOR_END76]]:
// CHECK2-NEXT: ret void
//
//
@@ -2097,6 +2238,7 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[DOTCAPTURE_EXPR_1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTNEW_STEP:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_UB0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_LB0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_ST0:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_NI0:%.*]] = alloca i32, align 4
@@ -2105,6 +2247,7 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[DOTCAPTURE_EXPR_7:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTNEW_STEP8:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_UB1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_LB1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_ST1:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_NI1:%.*]] = alloca i32, align 4
@@ -2113,6 +2256,7 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[DOTCAPTURE_EXPR_19:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTNEW_STEP21:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTCAPTURE_EXPR_22:%.*]] = alloca i32, align 4
+// CHECK2-NEXT: [[DOTOMP_UB2:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_LB2:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_ST2:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_NI2:%.*]] = alloca i32, align 4
@@ -2142,187 +2286,215 @@ extern "C" void foo5() {
// CHECK2-NEXT: [[DIV:%.*]] = udiv i32 [[ADD]], [[TMP7]]
// CHECK2-NEXT: [[SUB4:%.*]] = sub i32 [[DIV]], 1
// CHECK2-NEXT: store i32 [[SUB4]], ptr [[DOTCAPTURE_EXPR_2]], align 4
+// CHECK2-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
+// CHECK2-NEXT: store i32 [[TMP8]], ptr [[DOTOMP_UB0]], align 4
// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_LB0]], align 4
// CHECK2-NEXT: store i32 1, ptr [[DOTOMP_ST0]], align 4
-// CHECK2-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
-// CHECK2-NEXT: [[ADD5:%.*]] = add i32 [[TMP8]], 1
+// CHECK2-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
+// CHECK2-NEXT: [[ADD5:%.*]] = add i32 [[TMP9]], 1
// CHECK2-NEXT: store i32 [[ADD5]], ptr [[DOTOMP_NI0]], align 4
-// CHECK2-NEXT: [[TMP9:%.*]] = load i32, ptr [[END_ADDR]], align 4
-// CHECK2-NEXT: store i32 [[TMP9]], ptr [[J]], align 4
// CHECK2-NEXT: [[TMP10:%.*]] = load i32, ptr [[END_ADDR]], align 4
-// CHECK2-NEXT: store i32 [[TMP10]], ptr [[DOTCAPTURE_EXPR_6]], align 4
-// CHECK2-NEXT: [[TMP11:%.*]] = load i32, ptr [[START_ADDR]], align 4
-// CHECK2-NEXT: store i32 [[TMP11]], ptr [[DOTCAPTURE_EXPR_7]], align 4
-// CHECK2-NEXT: [[TMP12:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
-// CHECK2-NEXT: store i32 [[TMP12]], ptr [[DOTNEW_STEP8]], align 4
-// CHECK2-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
-// CHECK2-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_7]], align 4
-// CHECK2-NEXT: [[SUB10:%.*]] = sub i32 [[TMP13]], [[TMP14]]
+// CHECK2-NEXT: store i32 [[TMP10]], ptr [[J]], align 4
+// CHECK2-NEXT: [[TMP11:%.*]] = load i32, ptr [[END_ADDR]], align 4
+// CHECK2-NEXT: store i32 [[TMP11]], ptr [[DOTCAPTURE_EXPR_6]], align 4
+// CHECK2-NEXT: [[TMP12:%.*]] = load i32, ptr [[START_ADDR]], align 4
+// CHECK2-NEXT: store i32 [[TMP12]], ptr [[DOTCAPTURE_EXPR_7]], align 4
+// CHECK2-NEXT: [[TMP13:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
+// CHECK2-NEXT: store i32 [[TMP13]], ptr [[DOTNEW_STEP8]], align 4
+// CHECK2-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
+// CHECK2-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_7]], align 4
+// CHECK2-NEXT: [[SUB10:%.*]] = sub i32 [[TMP14]], [[TMP15]]
// CHECK2-NEXT: [[SUB11:%.*]] = sub i32 [[SUB10]], 1
-// CHECK2-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
-// CHECK2-NEXT: [[ADD12:%.*]] = add i32 [[SUB11]], [[TMP15]]
// CHECK2-NEXT: [[TMP16:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
-// CHECK2-NEXT: [[DIV13:%.*]] = udiv i32 [[ADD12]], [[TMP16]]
+// CHECK2-NEXT: [[ADD12:%.*]] = add i32 [[SUB11]], [[TMP16]]
+// CHECK2-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
+// CHECK2-NEXT: [[DIV13:%.*]] = udiv i32 [[ADD12]], [[TMP17]]
// CHECK2-NEXT: [[SUB14:%.*]] = sub i32 [[DIV13]], 1
// CHECK2-NEXT: store i32 [[SUB14]], ptr [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK2-NEXT: [[TMP18:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK2-NEXT: store i32 [[TMP18]], ptr [[DOTOMP_UB1]], align 4
// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_LB1]], align 4
// CHECK2-NEXT: store i32 1, ptr [[DOTOMP_ST1]], align 4
-// CHECK2-NEXT: [[TMP17:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_9]], align 4
-// CHECK2-NEXT: [[ADD15:%.*]] = add i32 [[TMP17]], 1
+// CHECK2-NEXT: [[TMP19:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_9]], align 4
+// CHECK2-NEXT: [[ADD15:%.*]] = add i32 [[TMP19]], 1
// CHECK2-NEXT: store i32 [[ADD15]], ptr [[DOTOMP_NI1]], align 4
-// CHECK2-NEXT: [[TMP18:%.*]] = load i32, ptr [[START_ADDR]], align 4
-// CHECK2-NEXT: [[TMP19:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
-// CHECK2-NEXT: [[ADD16:%.*]] = add nsw i32 [[TMP18]], [[TMP19]]
-// CHECK2-NEXT: store i32 [[ADD16]], ptr [[K]], align 4
// CHECK2-NEXT: [[TMP20:%.*]] = load i32, ptr [[START_ADDR]], align 4
// CHECK2-NEXT: [[TMP21:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
-// CHECK2-NEXT: [[ADD18:%.*]] = add nsw i32 [[TMP20]], [[TMP21]]
-// CHECK2-NEXT: store i32 [[ADD18]], ptr [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK2-NEXT: [[TMP22:%.*]] = load i32, ptr [[END_ADDR]], align 4
+// CHECK2-NEXT: [[ADD16:%.*]] = add nsw i32 [[TMP20]], [[TMP21]]
+// CHECK2-NEXT: store i32 [[ADD16]], ptr [[K]], align 4
+// CHECK2-NEXT: [[TMP22:%.*]] = load i32, ptr [[START_ADDR]], align 4
// CHECK2-NEXT: [[TMP23:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
-// CHECK2-NEXT: [[ADD20:%.*]] = add nsw i32 [[TMP22]], [[TMP23]]
+// CHECK2-NEXT: [[ADD18:%.*]] = add nsw i32 [[TMP22]], [[TMP23]]
+// CHECK2-NEXT: store i32 [[ADD18]], ptr [[DOTCAPTURE_EXPR_17]], align 4
+// CHECK2-NEXT: [[TMP24:%.*]] = load i32, ptr [[END_ADDR]], align 4
+// CHECK2-NEXT: [[TMP25:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
+// CHECK2-NEXT: [[ADD20:%.*]] = add nsw i32 [[TMP24]], [[TMP25]]
// CHECK2-NEXT: store i32 [[ADD20]], ptr [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK2-NEXT: [[TMP24:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
-// CHECK2-NEXT: store i32 [[TMP24]], ptr [[DOTNEW_STEP21]], align 4
-// CHECK2-NEXT: [[TMP25:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_19]], align 4
-// CHECK2-NEXT: [[TMP26:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK2-NEXT: [[SUB23:%.*]] = sub i32 [[TMP25]], [[TMP26]]
+// CHECK2-NEXT: [[TMP26:%.*]] = load i32, ptr [[STEP_ADDR]], align 4
+// CHECK2-NEXT: store i32 [[TMP26]], ptr [[DOTNEW_STEP21]], align 4
+// CHECK2-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK2-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_17]], align 4
+// CHECK2-NEXT: [[SUB23:%.*]] = sub i32 [[TMP27]], [[TMP28]]
// CHECK2-NEXT: [[SUB24:%.*]] = sub i32 [[SUB23]], 1
-// CHECK2-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTNEW_STEP21]], align 4
-// CHECK2-NEXT: [[ADD25:%.*]] = add i32 [[SUB24]], [[TMP27]]
-// CHECK2-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTNEW_STEP21]], align 4
-// CHECK2-NEXT: [[DIV26:%.*]] = udiv i32 [[ADD25]], [[TMP28]]
+// CHECK2-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTNEW_STEP21]], align 4
+// CHECK2-NEXT: [[ADD25:%.*]] = add i32 [[SUB24]], [[TMP29]]
+// CHECK2-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTNEW_STEP21]], align 4
+// CHECK2-NEXT: [[DIV26:%.*]] = udiv i32 [[ADD25]], [[TMP30]]
// CHECK2-NEXT: [[SUB27:%.*]] = sub i32 [[DIV26]], 1
// CHECK2-NEXT: store i32 [[SUB27]], ptr [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK2-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK2-NEXT: store i32 [[TMP31]], ptr [[DOTOMP_UB2]], align 4
// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_LB2]], align 4
// CHECK2-NEXT: store i32 1, ptr [[DOTOMP_ST2]], align 4
-// CHECK2-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_22]], align 4
-// CHECK2-NEXT: [[ADD28:%.*]] = add i32 [[TMP29]], 1
+// CHECK2-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_22]], align 4
+// CHECK2-NEXT: [[ADD28:%.*]] = add i32 [[TMP32]], 1
// CHECK2-NEXT: store i32 [[ADD28]], ptr [[DOTOMP_NI2]], align 4
-// CHECK2-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
-// CHECK2-NEXT: store i32 [[TMP30]], ptr [[DOTOMP_TEMP_1]], align 4
-// CHECK2-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
-// CHECK2-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
-// CHECK2-NEXT: [[CMP:%.*]] = icmp ugt i32 [[TMP31]], [[TMP32]]
+// CHECK2-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK2-NEXT: store i32 [[TMP33]], ptr [[DOTOMP_TEMP_1]], align 4
+// CHECK2-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
+// CHECK2-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK2-NEXT: [[CMP:%.*]] = icmp ugt i32 [[TMP34]], [[TMP35]]
// CHECK2-NEXT: br i1 [[CMP]], label %[[COND_TRUE:.*]], label %[[COND_FALSE:.*]]
// CHECK2: [[COND_TRUE]]:
-// CHECK2-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
+// CHECK2-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_TEMP_1]], align 4
// CHECK2-NEXT: br label %[[COND_END:.*]]
// CHECK2: [[COND_FALSE]]:
-// CHECK2-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK2-NEXT: [[TMP37:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
// CHECK2-NEXT: br label %[[COND_END]]
// CHECK2: [[COND_END]]:
-// CHECK2-NEXT: [[COND:%.*]] = phi i32 [ [[TMP33]], %[[COND_TRUE]] ], [ [[TMP34]], %[[COND_FALSE]] ]
+// CHECK2-NEXT: [[COND:%.*]] = phi i32 [ [[TMP36]], %[[COND_TRUE]] ], [ [[TMP37]], %[[COND_FALSE]] ]
// CHECK2-NEXT: store i32 [[COND]], ptr [[DOTOMP_TEMP_2]], align 4
-// CHECK2-NEXT: [[TMP35:%.*]] = load i32, ptr [[DOTOMP_TEMP_2]], align 4
-// CHECK2-NEXT: [[TMP36:%.*]] = load i32, ptr [[DOTOMP_NI2]], align 4
-// CHECK2-NEXT: [[CMP29:%.*]] = icmp ugt i32 [[TMP35]], [[TMP36]]
+// CHECK2-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_TEMP_2]], align 4
+// CHECK2-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTOMP_NI2]], align 4
+// CHECK2-NEXT: [[CMP29:%.*]] = icmp ugt i32 [[TMP38]], [[TMP39]]
// CHECK2-NEXT: br i1 [[CMP29]], label %[[COND_TRUE30:.*]], label %[[COND_FALSE31:.*]]
// CHECK2: [[COND_TRUE30]]:
-// CHECK2-NEXT: [[TMP37:%.*]] = load i32, ptr [[DOTOMP_TEMP_2]], align 4
+// CHECK2-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTOMP_TEMP_2]], align 4
// CHECK2-NEXT: br label %[[COND_END32:.*]]
// CHECK2: [[COND_FALSE31]]:
-// CHECK2-NEXT: [[TMP38:%.*]] = load i32, ptr [[DOTOMP_NI2]], align 4
+// CHECK2-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTOMP_NI2]], align 4
// CHECK2-NEXT: br label %[[COND_END32]]
// CHECK2: [[COND_END32]]:
-// CHECK2-NEXT: [[COND33:%.*]] = phi i32 [ [[TMP37]], %[[COND_TRUE30]] ], [ [[TMP38]], %[[COND_FALSE31]] ]
+// CHECK2-NEXT: [[COND33:%.*]] = phi i32 [ [[TMP40]], %[[COND_TRUE30]] ], [ [[TMP41]], %[[COND_FALSE31]] ]
// CHECK2-NEXT: store i32 [[COND33]], ptr [[DOTOMP_FUSE_MAX]], align 4
// CHECK2-NEXT: store i32 0, ptr [[DOTOMP_FUSE_INDEX]], align 4
// CHECK2-NEXT: br label %[[FOR_COND:.*]]
// CHECK2: [[FOR_COND]]:
-// CHECK2-NEXT: [[TMP39:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[TMP40:%.*]] = load i32, ptr [[DOTOMP_FUSE_MAX]], align 4
-// CHECK2-NEXT: [[CMP34:%.*]] = icmp ult i32 [[TMP39]], [[TMP40]]
+// CHECK2-NEXT: [[TMP42:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTOMP_FUSE_MAX]], align 4
+// CHECK2-NEXT: [[CMP34:%.*]] = icmp ult i32 [[TMP42]], [[TMP43]]
// CHECK2-NEXT: br i1 [[CMP34]], label %[[FOR_BODY:.*]], label %[[FOR_END:.*]]
// CHECK2: [[FOR_BODY]]:
-// CHECK2-NEXT: [[TMP41:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[TMP42:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
-// CHECK2-NEXT: [[CMP35:%.*]] = icmp ult i32 [[TMP41]], [[TMP42]]
+// CHECK2-NEXT: [[TMP44:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[TMP45:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK2-NEXT: [[CMP35:%.*]] = icmp ult i32 [[TMP44]], [[TMP45]]
// CHECK2-NEXT: br i1 [[CMP35]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
// CHECK2: [[IF_THEN]]:
-// CHECK2-NEXT: [[TMP43:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
-// CHECK2-NEXT: [[TMP44:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
-// CHECK2-NEXT: [[TMP45:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[MUL:%.*]] = mul i32 [[TMP44]], [[TMP45]]
-// CHECK2-NEXT: [[ADD36:%.*]] = add i32 [[TMP43]], [[MUL]]
+// CHECK2-NEXT: [[TMP46:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
+// CHECK2-NEXT: [[TMP47:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
+// CHECK2-NEXT: [[TMP48:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[MUL:%.*]] = mul i32 [[TMP47]], [[TMP48]]
+// CHECK2-NEXT: [[ADD36:%.*]] = add i32 [[TMP46]], [[MUL]]
// CHECK2-NEXT: store i32 [[ADD36]], ptr [[DOTOMP_IV0]], align 4
-// CHECK2-NEXT: [[TMP46:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
-// CHECK2-NEXT: [[TMP47:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
-// CHECK2-NEXT: [[TMP48:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
-// CHECK2-NEXT: [[MUL37:%.*]] = mul i32 [[TMP47]], [[TMP48]]
-// CHECK2-NEXT: [[ADD38:%.*]] = add i32 [[TMP46]], [[MUL37]]
+// CHECK2-NEXT: [[TMP49:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK2-NEXT: [[TMP50:%.*]] = load i32, ptr [[DOTOMP_IV0]], align 4
+// CHECK2-NEXT: [[TMP51:%.*]] = load i32, ptr [[DOTNEW_STEP]], align 4
+// CHECK2-NEXT: [[MUL37:%.*]] = mul i32 [[TMP50]], [[TMP51]]
+// CHECK2-NEXT: [[ADD38:%.*]] = add i32 [[TMP49]], [[MUL37]]
// CHECK2-NEXT: store i32 [[ADD38]], ptr [[I]], align 4
-// CHECK2-NEXT: [[TMP49:%.*]] = load i32, ptr [[I]], align 4
-// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP49]])
+// CHECK2-NEXT: [[TMP52:%.*]] = load i32, ptr [[I]], align 4
+// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP52]])
// CHECK2-NEXT: br label %[[IF_END]]
// CHECK2: [[IF_END]]:
-// CHECK2-NEXT: [[TMP50:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[TMP51:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
-// CHECK2-NEXT: [[CMP39:%.*]] = icmp ult i32 [[TMP50]], [[TMP51]]
+// CHECK2-NEXT: [[TMP53:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[TMP54:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK2-NEXT: [[CMP39:%.*]] = icmp ult i32 [[TMP53]], [[TMP54]]
// CHECK2-NEXT: br i1 [[CMP39]], label %[[IF_THEN40:.*]], label %[[IF_END45:.*]]
// CHECK2: [[IF_THEN40]]:
-// CHECK2-NEXT: [[TMP52:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
-// CHECK2-NEXT: [[TMP53:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
-// CHECK2-NEXT: [[TMP54:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[MUL41:%.*]] = mul i32 [[TMP53]], [[TMP54]]
-// CHECK2-NEXT: [[ADD42:%.*]] = add i32 [[TMP52]], [[MUL41]]
+// CHECK2-NEXT: [[TMP55:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
+// CHECK2-NEXT: [[TMP56:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
+// CHECK2-NEXT: [[TMP57:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[MUL41:%.*]] = mul i32 [[TMP56]], [[TMP57]]
+// CHECK2-NEXT: [[ADD42:%.*]] = add i32 [[TMP55]], [[MUL41]]
// CHECK2-NEXT: store i32 [[ADD42]], ptr [[DOTOMP_IV1]], align 4
-// CHECK2-NEXT: [[TMP55:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
-// CHECK2-NEXT: [[TMP56:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
-// CHECK2-NEXT: [[TMP57:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
-// CHECK2-NEXT: [[MUL43:%.*]] = mul i32 [[TMP56]], [[TMP57]]
-// CHECK2-NEXT: [[SUB44:%.*]] = sub i32 [[TMP55]], [[MUL43]]
+// CHECK2-NEXT: [[TMP58:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_6]], align 4
+// CHECK2-NEXT: [[TMP59:%.*]] = load i32, ptr [[DOTOMP_IV1]], align 4
+// CHECK2-NEXT: [[TMP60:%.*]] = load i32, ptr [[DOTNEW_STEP8]], align 4
+// CHECK2-NEXT: [[MUL43:%.*]] = mul i32 [[TMP59]], [[TMP60]]
+// CHECK2-NEXT: [[SUB44:%.*]] = sub i32 [[TMP58]], [[MUL43]]
// CHECK2-NEXT: store i32 [[SUB44]], ptr [[J]], align 4
-// CHECK2-NEXT: [[TMP58:%.*]] = load i32, ptr [[J]], align 4
-// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP58]])
+// CHECK2-NEXT: [[TMP61:%.*]] = load i32, ptr [[J]], align 4
+// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP61]])
// CHECK2-NEXT: br label %[[IF_END45]]
// CHECK2: [[IF_END45]]:
-// CHECK2-NEXT: [[TMP59:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[TMP60:%.*]] = load i32, ptr [[DOTOMP_NI2]], align 4
-// CHECK2-NEXT: [[CMP46:%.*]] = icmp ult i32 [[TMP59]], [[TMP60]]
+// CHECK2-NEXT: [[TMP62:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[TMP63:%.*]] = load i32, ptr [[DOTOMP_NI2]], align 4
+// CHECK2-NEXT: [[CMP46:%.*]] = icmp ult i32 [[TMP62]], [[TMP63]]
// CHECK2-NEXT: br i1 [[CMP46]], label %[[IF_THEN47:.*]], label %[[IF_END52:.*]]
// CHECK2: [[IF_THEN47]]:
-// CHECK2-NEXT: [[TMP61:%.*]] = load i32, ptr [[DOTOMP_LB2]], align 4
-// CHECK2-NEXT: [[TMP62:%.*]] = load i32, ptr [[DOTOMP_ST2]], align 4
-// CHECK2-NEXT: [[TMP63:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[MUL48:%.*]] = mul i32 [[TMP62]], [[TMP63]]
-// CHECK2-NEXT: [[ADD49:%.*]] = add i32 [[TMP61]], [[MUL48]]
+// CHECK2-NEXT: [[TMP64:%.*]] = load i32, ptr [[DOTOMP_LB2]], align 4
+// CHECK2-NEXT: [[TMP65:%.*]] = load i32, ptr [[DOTOMP_ST2]], align 4
+// CHECK2-NEXT: [[TMP66:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[MUL48:%.*]] = mul i32 [[TMP65]], [[TMP66]]
+// CHECK2-NEXT: [[ADD49:%.*]] = add i32 [[TMP64]], [[MUL48]]
// CHECK2-NEXT: store i32 [[ADD49]], ptr [[DOTOMP_IV2]], align 4
-// CHECK2-NEXT: [[TMP64:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_17]], align 4
-// CHECK2-NEXT: [[TMP65:%.*]] = load i32, ptr [[DOTOMP_IV2]], align 4
-// CHECK2-NEXT: [[TMP66:%.*]] = load i32, ptr [[DOTNEW_STEP21]], align 4
-// CHECK2-NEXT: [[MUL50:%.*]] = mul i32 [[TMP65]], [[TMP66]]
-// CHECK2-NEXT: [[ADD51:%.*]] = add i32 [[TMP64]], [[MUL50]]
+// CHECK2-NEXT: [[TMP67:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_17]], align 4
+// CHECK2-NEXT: [[TMP68:%.*]] = load i32, ptr [[DOTOMP_IV2]], align 4
+// CHECK2-NEXT: [[TMP69:%.*]] = load i32, ptr [[DOTNEW_STEP21]], align 4
+// CHECK2-NEXT: [[MUL50:%.*]] = mul i32 [[TMP68]], [[TMP69]]
+// CHECK2-NEXT: [[ADD51:%.*]] = add i32 [[TMP67]], [[MUL50]]
// CHECK2-NEXT: store i32 [[ADD51]], ptr [[K]], align 4
-// CHECK2-NEXT: [[TMP67:%.*]] = load i32, ptr [[K]], align 4
-// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP67]])
+// CHECK2-NEXT: [[TMP70:%.*]] = load i32, ptr [[K]], align 4
+// CHECK2-NEXT: call void (...) @body(i32 noundef [[TMP70]])
// CHECK2-NEXT: br label %[[IF_END52]]
// CHECK2: [[IF_END52]]:
// CHECK2-NEXT: br label %[[FOR_INC:.*]]
// CHECK2: [[FOR_INC]]:
-// CHECK2-NEXT: [[TMP68:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: [[INC:%.*]] = add i32 [[TMP68]], 1
+// CHECK2-NEXT: [[TMP71:%.*]] = load i32, ptr [[DOTOMP_FUSE_INDEX]], align 4
+// CHECK2-NEXT: [[INC:%.*]] = add i32 [[TMP71]], 1
// CHECK2-NEXT: store i32 [[INC]], ptr [[DOTOMP_FUSE_INDEX]], align 4
-// CHECK2-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP10:![0-9]+]]
+// CHECK2-NEXT: br label %[[FOR_COND]], !llvm.loop [[LOOP11:![0-9]+]]
// CHECK2: [[FOR_END]]:
+// CHECK2-NEXT: [[TMP72:%.*]] = load i32, ptr [[DOTOMP_LB0]], align 4
+// CHECK2-NEXT: [[TMP73:%.*]] = load i32, ptr [[DOTOMP_NI0]], align 4
+// CHECK2-NEXT: [[TMP74:%.*]] = load i32, ptr [[DOTOMP_ST0]], align 4
+// CHECK2-NEXT: [[MUL53:%.*]] = mul i32 [[TMP73]], [[TMP74]]
+// CHECK2-NEXT: [[ADD54:%.*]] = add i32 [[TMP72]], [[MUL53]]
+// CHECK2-NEXT: store i32 [[ADD54]], ptr [[I]], align 4
+// CHECK2-NEXT: [[TMP75:%.*]] = load i32, ptr [[DOTOMP_LB1]], align 4
+// CHECK2-NEXT: [[TMP76:%.*]] = load i32, ptr [[DOTOMP_NI1]], align 4
+// CHECK2-NEXT: [[TMP77:%.*]] = load i32, ptr [[DOTOMP_ST1]], align 4
+// CHECK2-NEXT: [[MUL55:%.*]] = mul i32 [[TMP76]], [[TMP77]]
+// CHECK2-NEXT: [[ADD56:%.*]] = add i32 [[TMP75]], [[MUL55]]
+// CHECK2-NEXT: store i32 [[ADD56]], ptr [[J]], align 4
+// CHECK2-NEXT: [[TMP78:%.*]] = load i32, ptr [[DOTOMP_LB2]], align 4
+// CHECK2-NEXT: [[TMP79:%.*]] = load i32, ptr [[DOTOMP_NI2]], align 4
+// CHECK2-NEXT: [[TMP80:%.*]] = load i32, ptr [[DOTOMP_ST2]], align 4
+// CHECK2-NEXT: [[MUL57:%.*]] = mul i32 [[TMP79]], [[TMP80]]
+// CHECK2-NEXT: [[ADD58:%.*]] = add i32 [[TMP78]], [[MUL57]]
+// CHECK2-NEXT: store i32 [[ADD58]], ptr [[K]], align 4
// CHECK2-NEXT: ret void
//
//.
-// CHECK1: [[LOOP3]] = distinct !{[[LOOP3]], [[META4:![0-9]+]]}
-// CHECK1: [[META4]] = !{!"llvm.loop.mustprogress"}
-// CHECK1: [[LOOP5]] = distinct !{[[LOOP5]], [[META4]]}
-// CHECK1: [[LOOP6]] = distinct !{[[LOOP6]], [[META4]]}
-// CHECK1: [[LOOP7]] = distinct !{[[LOOP7]], [[META4]]}
-// CHECK1: [[LOOP8]] = distinct !{[[LOOP8]], [[META4]]}
-// CHECK1: [[LOOP9]] = distinct !{[[LOOP9]], [[META4]]}
-// CHECK1: [[LOOP10]] = distinct !{[[LOOP10]], [[META4]]}
+// CHECK1: [[LOOP2]] = distinct !{[[LOOP2]], [[META3:![0-9]+]]}
+// CHECK1: [[META3]] = !{!"llvm.loop.mustprogress"}
+// CHECK1: [[LOOP4]] = distinct !{[[LOOP4]], [[META3]]}
+// CHECK1: [[META5]] = !{}
+// CHECK1: [[META6]] = !{i64 8}
+// CHECK1: [[LOOP7]] = distinct !{[[LOOP7]], [[META3]]}
+// CHECK1: [[LOOP8]] = distinct !{[[LOOP8]], [[META3]]}
+// CHECK1: [[LOOP9]] = distinct !{[[LOOP9]], [[META3]]}
+// CHECK1: [[LOOP10]] = distinct !{[[LOOP10]], [[META3]]}
+// CHECK1: [[LOOP11]] = distinct !{[[LOOP11]], [[META3]]}
//.
-// CHECK2: [[LOOP3]] = distinct !{[[LOOP3]], [[META4:![0-9]+]]}
-// CHECK2: [[META4]] = !{!"llvm.loop.mustprogress"}
-// CHECK2: [[LOOP5]] = distinct !{[[LOOP5]], [[META4]]}
-// CHECK2: [[LOOP6]] = distinct !{[[LOOP6]], [[META4]]}
-// CHECK2: [[LOOP7]] = distinct !{[[LOOP7]], [[META4]]}
-// CHECK2: [[LOOP8]] = distinct !{[[LOOP8]], [[META4]]}
-// CHECK2: [[LOOP9]] = distinct !{[[LOOP9]], [[META4]]}
-// CHECK2: [[LOOP10]] = distinct !{[[LOOP10]], [[META4]]}
+// CHECK2: [[LOOP2]] = distinct !{[[LOOP2]], [[META3:![0-9]+]]}
+// CHECK2: [[META3]] = !{!"llvm.loop.mustprogress"}
+// CHECK2: [[META4]] = !{}
+// CHECK2: [[META5]] = !{i64 8}
+// CHECK2: [[LOOP6]] = distinct !{[[LOOP6]], [[META3]]}
+// CHECK2: [[LOOP7]] = distinct !{[[LOOP7]], [[META3]]}
+// CHECK2: [[LOOP8]] = distinct !{[[LOOP8]], [[META3]]}
+// CHECK2: [[LOOP9]] = distinct !{[[LOOP9]], [[META3]]}
+// CHECK2: [[LOOP10]] = distinct !{[[LOOP10]], [[META3]]}
+// CHECK2: [[LOOP11]] = distinct !{[[LOOP11]], [[META3]]}
//.
>From 34ee25632331a16cecea7126374a401f7f3d2964 Mon Sep 17 00:00:00 2001
From: Zahira Ammarguellat <zahira.ammarguellat at intel.com>
Date: Thu, 23 Jul 2026 10:54:33 -0700
Subject: [PATCH 5/5] Fix format
---
clang/include/clang/AST/StmtOpenMP.h | 21 ++++++++-------
clang/lib/Sema/SemaOpenMP.cpp | 39 +++++++++++++++-------------
2 files changed, 32 insertions(+), 28 deletions(-)
diff --git a/clang/include/clang/AST/StmtOpenMP.h b/clang/include/clang/AST/StmtOpenMP.h
index 2b1a4b4fa72af..78c4d42cd5cf3 100644
--- a/clang/include/clang/AST/StmtOpenMP.h
+++ b/clang/include/clang/AST/StmtOpenMP.h
@@ -5877,10 +5877,11 @@ class OMPReverseDirective final
/// dependent contexts.
/// \param PreInits Helper preinits statements for the loop nest.
/// \param Finals Loop variable finalization statements.
- static OMPReverseDirective
- Create(const ASTContext &C, SourceLocation StartLoc, SourceLocation EndLoc,
- Stmt *AssociatedStmt, unsigned NumLoops, Stmt *TransformedStmt,
- Stmt *PreInits, Stmt *Finals);
+ static OMPReverseDirective *Create(const ASTContext &C,
+ SourceLocation StartLoc,
+ SourceLocation EndLoc, Stmt *AssociatedStmt,
+ unsigned NumLoops, Stmt *TransformedStmt,
+ Stmt *PreInits, Stmt *Finals);
/// Build an empty '#pragma omp reverse' AST node for deserialization.
///
@@ -6069,12 +6070,12 @@ class OMPFuseDirective final
/// dependent
/// \param PreInits Helper preinits statements for the loop nest
/// \param Finals Loop variable finalization statements
- static OMPFuseDirective * Create(const ASTContext &C, SourceLocation StartLoc,
- SourceLocation EndLoc,
- ArrayRef<OMPClause *> Clauses,
- unsigned NumGeneratedTopLevelLoops,
- Stmt *AssociatedStmt, Stmt *TransformedStmt,
- Stmt *PreInits, Stmt *Finals);
+ static OMPFuseDirective *Create(const ASTContext &C, SourceLocation StartLoc,
+ SourceLocation EndLoc,
+ ArrayRef<OMPClause *> Clauses,
+ unsigned NumGeneratedTopLevelLoops,
+ Stmt *AssociatedStmt, Stmt *TransformedStmt,
+ Stmt *PreInits, Stmt *Finals);
/// Build an empty '#pragma omp fuse' AST node for deserialization
///
diff --git a/clang/lib/Sema/SemaOpenMP.cpp b/clang/lib/Sema/SemaOpenMP.cpp
index 1ddf8db56ada6..a7a54fe49c626 100644
--- a/clang/lib/Sema/SemaOpenMP.cpp
+++ b/clang/lib/Sema/SemaOpenMP.cpp
@@ -15006,8 +15006,8 @@ StmtResult SemaOpenMP::ActOnOpenMPTileDirective(ArrayRef<OMPClause *> Clauses,
// Delay tiling to when template is completely instantiated.
if (SemaRef.CurContext->isDependentContext())
- return OMPTileDirective::Create(Context, StartLoc, EndLoc, Clauses, NumLoops,
- AStmt, nullptr, nullptr, nullptr);
+ return OMPTileDirective::Create(Context, StartLoc, EndLoc, Clauses,
+ NumLoops, AStmt, nullptr, nullptr, nullptr);
assert(LoopHelpers.size() == NumLoops &&
"Expecting loop iteration space dimensionality to match number of "
@@ -15249,7 +15249,8 @@ StmtResult SemaOpenMP::ActOnOpenMPTileDirective(ArrayRef<OMPClause *> Clauses,
}
return OMPTileDirective::Create(Context, StartLoc, EndLoc, Clauses, NumLoops,
- AStmt, Inner, buildPreInits(Context, PreInits),
+ AStmt, Inner,
+ buildPreInits(Context, PreInits),
buildLoopFinalization(Context, LoopHelpers));
}
@@ -16294,7 +16295,8 @@ StmtResult SemaOpenMP::ActOnOpenMPInterchangeDirective(
return Idx == Arg;
}))
return OMPInterchangeDirective::Create(Context, StartLoc, EndLoc, Clauses,
- NumLoops, AStmt, AStmt, nullptr, nullptr);
+ NumLoops, AStmt, AStmt, nullptr,
+ nullptr);
// Find the affected loops.
SmallVector<Stmt *> LoopStmts(NumLoops, nullptr);
@@ -16404,10 +16406,10 @@ StmtResult SemaOpenMP::ActOnOpenMPInterchangeDirective(
SourceHelper.Inc->getEndLoc());
}
- return OMPInterchangeDirective::Create(Context, StartLoc, EndLoc, Clauses,
- NumLoops, AStmt, Inner,
- buildPreInits(Context, PreInits),
- buildLoopFinalization(Context, LoopHelpers));
+ return OMPInterchangeDirective::Create(
+ Context, StartLoc, EndLoc, Clauses, NumLoops, AStmt, Inner,
+ buildPreInits(Context, PreInits),
+ buildLoopFinalization(Context, LoopHelpers));
}
StmtResult SemaOpenMP::ActOnOpenMPFuseDirective(ArrayRef<OMPClause *> Clauses,
@@ -16430,7 +16432,8 @@ StmtResult SemaOpenMP::ActOnOpenMPFuseDirective(ArrayRef<OMPClause *> Clauses,
// because a dependent context could prevent determining its true value
if (CurrContext->isDependentContext())
return OMPFuseDirective::Create(Context, StartLoc, EndLoc, Clauses,
- /* NumLoops */ 1, AStmt, nullptr, nullptr, nullptr);
+ /* NumLoops */ 1, AStmt, nullptr, nullptr,
+ nullptr);
// Validate that the potential loop sequence is transformable for fusion
// Also collect the HelperExprs, Loop Stmts, Inits, and Number of loops
@@ -16853,8 +16856,8 @@ StmtResult SemaOpenMP::ActOnOpenMPFuseDirective(ArrayRef<OMPClause *> Clauses,
// Get the user's actual loop variable.
assert(!SeqAnalysis.Loops[I].HelperExprs.Counters.empty() &&
"Expected at least one counter for the loop");
- DeclRefExpr *UserLoopVarRef = cast<DeclRefExpr>(
- SeqAnalysis.Loops[I].HelperExprs.Counters[0]);
+ DeclRefExpr *UserLoopVarRef =
+ cast<DeclRefExpr>(SeqAnalysis.Loops[I].HelperExprs.Counters[0]);
VarDecl *UserLoopVarDecl = cast<VarDecl>(UserLoopVarRef->getDecl());
// Skip internal OpenMP variables (they start with .omp.).
@@ -16864,15 +16867,15 @@ StmtResult SemaOpenMP::ActOnOpenMPFuseDirective(ArrayRef<OMPClause *> Clauses,
// Build: user_var = lb + ni * st.
// ni * st.
ExprResult FinalValue = SemaRef.BuildBinOp(
- CurScope, SourceLocation(), BO_Mul,
- MakeVarDeclRef(NIVarDecls[J]), MakeVarDeclRef(STVarDecls[J]));
+ CurScope, SourceLocation(), BO_Mul, MakeVarDeclRef(NIVarDecls[J]),
+ MakeVarDeclRef(STVarDecls[J]));
if (!FinalValue.isUsable())
return StmtError();
// lb + (ni * st).
- FinalValue = SemaRef.BuildBinOp(
- CurScope, SourceLocation(), BO_Add,
- MakeVarDeclRef(LBVarDecls[J]), FinalValue.get());
+ FinalValue =
+ SemaRef.BuildBinOp(CurScope, SourceLocation(), BO_Add,
+ MakeVarDeclRef(LBVarDecls[J]), FinalValue.get());
if (!FinalValue.isUsable())
return StmtError();
@@ -16941,8 +16944,8 @@ StmtResult SemaOpenMP::ActOnOpenMPFuseDirective(ArrayRef<OMPClause *> Clauses,
Stmt *Finals = nullptr;
if (!FusedLoopFinalization.empty()) {
Finals = CompoundStmt::Create(Context, FusedLoopFinalization,
- FPOptionsOverride(),
- SourceLocation(), SourceLocation());
+ FPOptionsOverride(), SourceLocation(),
+ SourceLocation());
}
return OMPFuseDirective::Create(Context, StartLoc, EndLoc, Clauses,
More information about the cfe-commits
mailing list