[Openmp-commits] [clang] [openmp] [OpenMP][SIMD][FIX] Use conservative "omp simd ordered" lowering (PR #126172)
via Openmp-commits
openmp-commits at lists.llvm.org
Thu Feb 6 20:09:04 PST 2025
https://github.com/MattPD updated https://github.com/llvm/llvm-project/pull/126172
>From 81b6b3138a18bceabefd0ecd0e219bda3af44b84 Mon Sep 17 00:00:00 2001
From: "Matt P. Dziubinski" <matt-p.dziubinski at hpe.com>
Date: Tue, 21 Jan 2025 19:42:41 -0600
Subject: [PATCH] [OpenMP][SIMD][FIX] Use conservative "omp simd ordered"
lowering
A proposed fix for #95611 [OpenMP][SIMD] ordered has no effect in a loop SIMD region as of LLVM 18.1.0
Changes:
- Implement new lowering behavior: Conservatively serialize "omp simd" loops that have `omp simd ordered` directive to prevent incorrect vectorization (which results in incorrect execution behavior of the miscompiled program).
Implementation outline:
- We start with the optimistic default initial value of `LoopStack.setParallel(/Enable=/true);` in `CodeGenFunction::EmitOMPSimdInit(const OMPLoopDirective &D)`.
- We only disable the loop parallel memory access assumption with `if (HasOrderedDirective) LoopStack.setParallel(/Enable=/false);` using the `HasOrderedDirective` (which tests for the presence of an `OMPOrderedDirective`).
- This results in no longer incorrectly vectorizing the loop when the `omp simd ordered` directive is present.
Motivation: We'd like to prevent incorrect vectorization of the loops marked with the `#pragma omp ordered simd` directive which has previously resulted in miscompiled code.
At the same time, we'd like the usage outside of the `#pragma omp ordered simd` context to remain unaffected: Note that in the test "clang/test/OpenMP/ordered_codegen.cpp" we only "lose" the `!llvm.access.group` metadata in `foo_simd` alone.
This is a conservative, in that it's possible some of the loops would be possible to vectorize, but we prefer to avoid miscompilation of the loops that are currently illegal to vectorize.
A concrete example follows:
```cpp
// "test.c"
#include <float.h>
#include <math.h>
#include <omp.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
int compare_float(float x1, float x2, float scalar) {
const float diff = fabsf(x1 - x2);
x1 = fabsf(x1);
x2 = fabsf(x2);
const float l = (x2 > x1) ? x2 : x1;
if (diff <= l * scalar * FLT_EPSILON)
return 1;
else
return 0;
}
#define ARRAY_SIZE 256
__attribute__((noinline)) void initialization_loop(
float X[ARRAY_SIZE][ARRAY_SIZE], float Y[ARRAY_SIZE][ARRAY_SIZE]) {
const float max = 1000.0;
srand(time(NULL));
for (int r = 0; r < ARRAY_SIZE; r++) {
for (int c = 0; c < ARRAY_SIZE; c++) {
X[r][c] = ((float)rand() / (float)(RAND_MAX)) * max;
Y[r][c] = X[r][c];
}
}
}
__attribute__((noinline)) void omp_simd_loop(float X[ARRAY_SIZE][ARRAY_SIZE]) {
for (int r = 1; r < ARRAY_SIZE; ++r) {
for (int c = 1; c < ARRAY_SIZE; ++c) {
#pragma omp simd
for (int k = 2; k < ARRAY_SIZE; ++k) {
#pragma omp ordered simd
X[r][k] = X[r][k - 2] + sinf((float)(r / c));
}
}
}
}
__attribute__((noinline)) int comparison_loop(float X[ARRAY_SIZE][ARRAY_SIZE],
float Y[ARRAY_SIZE][ARRAY_SIZE]) {
int totalErrors_simd = 0;
const float scalar = 1.0;
for (int r = 1; r < ARRAY_SIZE; ++r) {
for (int c = 1; c < ARRAY_SIZE; ++c) {
for (int k = 2; k < ARRAY_SIZE; ++k) {
Y[r][k] = Y[r][k - 2] + sinf((float)(r / c));
}
}
// check row for simd update
for (int k = 0; k < ARRAY_SIZE; ++k) {
if (!compare_float(X[r][k], Y[r][k], scalar)) {
++totalErrors_simd;
}
}
}
return totalErrors_simd;
}
int main(void) {
float X[ARRAY_SIZE][ARRAY_SIZE];
float Y[ARRAY_SIZE][ARRAY_SIZE];
initialization_loop(X, Y);
omp_simd_loop(X);
const int totalErrors_simd = comparison_loop(X, Y);
if (totalErrors_simd) {
fprintf(stdout, "totalErrors_simd: %d \n", totalErrors_simd);
fprintf(stdout, "%s : %d - FAIL: error in ordered simd computation.\n",
__FILE__, __LINE__);
} else {
fprintf(stdout, "Success!\n");
}
return totalErrors_simd;
}
```
Before:
```
$ clang -fopenmp-simd -O3 -ffast-math -lm test.c -o test && ./test
totalErrors_simd: 15408
test.c : 76 - FAIL: error in ordered simd computation.
```
clang 19.1.0: https://godbolt.org/z/6EvhxqEhe
After:
```
$ clang -fopenmp-simd -O3 -ffast-math test.c -o test && ./test
Success!
```
---
clang/lib/CodeGen/CGStmtOpenMP.cpp | 76 ++++++
clang/test/OpenMP/ordered_codegen.cpp | 232 +++++++++---------
.../misc_bugs/simd_conservative_ordered.c | 84 +++++++
3 files changed, 276 insertions(+), 116 deletions(-)
create mode 100644 openmp/runtime/test/misc_bugs/simd_conservative_ordered.c
diff --git a/clang/lib/CodeGen/CGStmtOpenMP.cpp b/clang/lib/CodeGen/CGStmtOpenMP.cpp
index 8e694b95dc7e73b..3542e939678cf42 100644
--- a/clang/lib/CodeGen/CGStmtOpenMP.cpp
+++ b/clang/lib/CodeGen/CGStmtOpenMP.cpp
@@ -2457,10 +2457,86 @@ static void emitSimdlenSafelenClause(CodeGenFunction &CGF,
}
}
+// Check for the presence of an `OMPOrderedDirective`,
+// i.e., `ordered` in `#pragma omp ordered simd`.
+//
+// Consider the following source code:
+// ```
+// __attribute__((noinline)) void omp_simd_loop(float X[ARRAY_SIZE][ARRAY_SIZE])
+// {
+// for (int r = 1; r < ARRAY_SIZE; ++r) {
+// for (int c = 1; c < ARRAY_SIZE; ++c) {
+// #pragma omp simd
+// for (int k = 2; k < ARRAY_SIZE; ++k) {
+// #pragma omp ordered simd
+// X[r][k] = X[r][k - 2] + sinf((float)(r / c));
+// }
+// }
+// }
+// }
+// ```
+//
+// Suppose we are in `CodeGenFunction::EmitOMPSimdInit(const OMPLoopDirective
+// &D)`. By examining `D.dump()` we have the following AST containing
+// `OMPOrderedDirective`:
+//
+// ```
+// OMPSimdDirective 0x1c32950
+// `-CapturedStmt 0x1c32028
+// |-CapturedDecl 0x1c310e8
+// | |-ForStmt 0x1c31e30
+// | | |-DeclStmt 0x1c31298
+// | | | `-VarDecl 0x1c31208 used k 'int' cinit
+// | | | `-IntegerLiteral 0x1c31278 'int' 2
+// | | |-<<<NULL>>>
+// | | |-BinaryOperator 0x1c31308 'int' '<'
+// | | | |-ImplicitCastExpr 0x1c312f0 'int' <LValueToRValue>
+// | | | | `-DeclRefExpr 0x1c312b0 'int' lvalue Var 0x1c31208 'k' 'int'
+// | | | `-IntegerLiteral 0x1c312d0 'int' 256
+// | | |-UnaryOperator 0x1c31348 'int' prefix '++'
+// | | | `-DeclRefExpr 0x1c31328 'int' lvalue Var 0x1c31208 'k' 'int'
+// | | `-CompoundStmt 0x1c31e18
+// | | `-OMPOrderedDirective 0x1c31dd8
+// | | |-OMPSimdClause 0x1c31380
+// | | `-CapturedStmt 0x1c31cd0
+// ```
+//
+// Note the presence of `OMPOrderedDirective` above:
+// It's (transitively) nested in a `CapturedStmt` representing the pragma
+// annotated compound statement. Thus, we need to consider this nesting and
+// include checking the `getCapturedStmt` in this case.
+static bool hasOrderedDirective(const Stmt *S) {
+ if (isa<OMPOrderedDirective>(S))
+ return true;
+
+ if (const auto *CS = dyn_cast<CapturedStmt>(S))
+ return hasOrderedDirective(CS->getCapturedStmt());
+
+ for (const Stmt *Child : S->children()) {
+ if (Child && hasOrderedDirective(Child))
+ return true;
+ }
+
+ return false;
+}
+
+static void applyConservativeSimdOrderedDirective(const Stmt &AssociatedStmt,
+ LoopInfoStack &LoopStack) {
+ // Check for the presence of an `OMPOrderedDirective`
+ // i.e., `ordered` in `#pragma omp ordered simd`
+ bool HasOrderedDirective = hasOrderedDirective(&AssociatedStmt);
+ // If present then conservatively disable loop vectorization
+ // analogously to how `emitSimdlenSafelenClause` does.
+ if (HasOrderedDirective)
+ LoopStack.setParallel(/*Enable=*/false);
+}
+
void CodeGenFunction::EmitOMPSimdInit(const OMPLoopDirective &D) {
// Walk clauses and process safelen/lastprivate.
LoopStack.setParallel(/*Enable=*/true);
LoopStack.setVectorizeEnable();
+ const Stmt *AssociatedStmt = D.getAssociatedStmt();
+ applyConservativeSimdOrderedDirective(*AssociatedStmt, LoopStack);
emitSimdlenSafelenClause(*this, D);
if (const auto *C = D.getSingleClause<OMPOrderClause>())
if (C->getKind() == OMPC_ORDER_concurrent)
diff --git a/clang/test/OpenMP/ordered_codegen.cpp b/clang/test/OpenMP/ordered_codegen.cpp
index 67285cfaef34d54..5cd95f1927e5ced 100644
--- a/clang/test/OpenMP/ordered_codegen.cpp
+++ b/clang/test/OpenMP/ordered_codegen.cpp
@@ -572,30 +572,30 @@ void foo_simd(int low, int up) {
// CHECK1-NEXT: store i32 0, ptr [[DOTOMP_IV]], align 4
// CHECK1-NEXT: br label [[OMP_INNER_FOR_COND:%.*]]
// CHECK1: omp.inner.for.cond:
-// CHECK1-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3:![0-9]+]]
-// CHECK1-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK1-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4
+// CHECK1-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
// CHECK1-NEXT: [[ADD6:%.*]] = add i32 [[TMP9]], 1
// CHECK1-NEXT: [[CMP7:%.*]] = icmp ult i32 [[TMP8]], [[ADD6]]
// CHECK1-NEXT: br i1 [[CMP7]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
// CHECK1: omp.inner.for.body:
-// CHECK1-NEXT: [[TMP10:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK1-NEXT: [[TMP11:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK1-NEXT: [[TMP10:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-NEXT: [[TMP11:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4
// CHECK1-NEXT: [[MUL:%.*]] = mul i32 [[TMP11]], 1
// CHECK1-NEXT: [[ADD8:%.*]] = add i32 [[TMP10]], [[MUL]]
-// CHECK1-NEXT: store i32 [[ADD8]], ptr [[I5]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK1-NEXT: [[TMP12:%.*]] = load i32, ptr [[I5]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK1-NEXT: store i32 [[ADD8]], ptr [[I5]], align 4
+// CHECK1-NEXT: [[TMP12:%.*]] = load i32, ptr [[I5]], align 4
// CHECK1-NEXT: [[IDXPROM:%.*]] = sext i32 [[TMP12]] to i64
// CHECK1-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], ptr @f, i64 0, i64 [[IDXPROM]]
-// CHECK1-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK1-NEXT: call void @__captured_stmt(ptr [[I5]]), !llvm.access.group [[ACC_GRP3]]
+// CHECK1-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX]], align 4
+// CHECK1-NEXT: call void @__captured_stmt(ptr [[I5]])
// CHECK1-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK1: omp.body.continue:
// CHECK1-NEXT: br label [[OMP_INNER_FOR_INC:%.*]]
// CHECK1: omp.inner.for.inc:
-// CHECK1-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK1-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4
// CHECK1-NEXT: [[ADD9:%.*]] = add i32 [[TMP13]], 1
-// CHECK1-NEXT: store i32 [[ADD9]], ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK1-NEXT: br label [[OMP_INNER_FOR_COND]], !llvm.loop [[LOOP4:![0-9]+]]
+// CHECK1-NEXT: store i32 [[ADD9]], ptr [[DOTOMP_IV]], align 4
+// CHECK1-NEXT: br label [[OMP_INNER_FOR_COND]], !llvm.loop [[LOOP3:![0-9]+]]
// CHECK1: omp.inner.for.end:
// CHECK1-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
// CHECK1-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
@@ -645,31 +645,31 @@ void foo_simd(int low, int up) {
// CHECK1-NEXT: store i32 [[TMP27]], ptr [[DOTOMP_IV16]], align 4
// CHECK1-NEXT: br label [[OMP_INNER_FOR_COND29:%.*]]
// CHECK1: omp.inner.for.cond29:
-// CHECK1-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7:![0-9]+]]
-// CHECK1-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTOMP_UB]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK1-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4
+// CHECK1-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTOMP_UB]], align 4
// CHECK1-NEXT: [[ADD30:%.*]] = add i32 [[TMP29]], 1
// CHECK1-NEXT: [[CMP31:%.*]] = icmp ult i32 [[TMP28]], [[ADD30]]
// CHECK1-NEXT: br i1 [[CMP31]], label [[OMP_INNER_FOR_BODY32:%.*]], label [[OMP_INNER_FOR_END40:%.*]]
// CHECK1: omp.inner.for.body32:
-// CHECK1-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_18]], align 4, !llvm.access.group [[ACC_GRP7]]
-// CHECK1-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK1-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK1-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4
// CHECK1-NEXT: [[MUL33:%.*]] = mul i32 [[TMP31]], 1
// CHECK1-NEXT: [[ADD34:%.*]] = add i32 [[TMP30]], [[MUL33]]
-// CHECK1-NEXT: store i32 [[ADD34]], ptr [[I28]], align 4, !llvm.access.group [[ACC_GRP7]]
-// CHECK1-NEXT: [[TMP32:%.*]] = load i32, ptr [[I28]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK1-NEXT: store i32 [[ADD34]], ptr [[I28]], align 4
+// CHECK1-NEXT: [[TMP32:%.*]] = load i32, ptr [[I28]], align 4
// CHECK1-NEXT: [[IDXPROM35:%.*]] = sext i32 [[TMP32]] to i64
// CHECK1-NEXT: [[ARRAYIDX36:%.*]] = getelementptr inbounds [10 x float], ptr @f, i64 0, i64 [[IDXPROM35]]
-// CHECK1-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX36]], align 4, !llvm.access.group [[ACC_GRP7]]
-// CHECK1-NEXT: call void @__captured_stmt.1(ptr [[I28]]), !llvm.access.group [[ACC_GRP7]]
+// CHECK1-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX36]], align 4
+// CHECK1-NEXT: call void @__captured_stmt.1(ptr [[I28]])
// CHECK1-NEXT: br label [[OMP_BODY_CONTINUE37:%.*]]
// CHECK1: omp.body.continue37:
// CHECK1-NEXT: br label [[OMP_INNER_FOR_INC38:%.*]]
// CHECK1: omp.inner.for.inc38:
-// CHECK1-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK1-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4
// CHECK1-NEXT: [[ADD39:%.*]] = add i32 [[TMP33]], 1
-// CHECK1-NEXT: store i32 [[ADD39]], ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7]]
-// CHECK1-NEXT: call void @__kmpc_dispatch_fini_4u(ptr @[[GLOB1]], i32 [[TMP0]]), !llvm.access.group [[ACC_GRP7]]
-// CHECK1-NEXT: br label [[OMP_INNER_FOR_COND29]], !llvm.loop [[LOOP8:![0-9]+]]
+// CHECK1-NEXT: store i32 [[ADD39]], ptr [[DOTOMP_IV16]], align 4
+// CHECK1-NEXT: call void @__kmpc_dispatch_fini_4u(ptr @[[GLOB1]], i32 [[TMP0]])
+// CHECK1-NEXT: br label [[OMP_INNER_FOR_COND29]], !llvm.loop [[LOOP5:![0-9]+]]
// CHECK1: omp.inner.for.end40:
// CHECK1-NEXT: br label [[OMP_DISPATCH_INC:%.*]]
// CHECK1: omp.dispatch.inc:
@@ -1201,32 +1201,32 @@ void foo_simd(int low, int up) {
// CHECK1-IRBUILDER-NEXT: store i32 0, ptr [[DOTOMP_IV]], align 4
// CHECK1-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_COND:%.*]]
// CHECK1-IRBUILDER: omp.inner.for.cond:
-// CHECK1-IRBUILDER-NEXT: [[TMP7:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3:![0-9]+]]
-// CHECK1-IRBUILDER-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK1-IRBUILDER-NEXT: [[TMP7:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4
+// CHECK1-IRBUILDER-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
// CHECK1-IRBUILDER-NEXT: [[ADD6:%.*]] = add i32 [[TMP8]], 1
// CHECK1-IRBUILDER-NEXT: [[CMP7:%.*]] = icmp ult i32 [[TMP7]], [[ADD6]]
// CHECK1-IRBUILDER-NEXT: br i1 [[CMP7]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
// CHECK1-IRBUILDER: omp.inner.for.body:
-// CHECK1-IRBUILDER-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK1-IRBUILDER-NEXT: [[TMP10:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK1-IRBUILDER-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK1-IRBUILDER-NEXT: [[TMP10:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4
// CHECK1-IRBUILDER-NEXT: [[MUL:%.*]] = mul i32 [[TMP10]], 1
// CHECK1-IRBUILDER-NEXT: [[ADD8:%.*]] = add i32 [[TMP9]], [[MUL]]
-// CHECK1-IRBUILDER-NEXT: store i32 [[ADD8]], ptr [[I5]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK1-IRBUILDER-NEXT: [[TMP11:%.*]] = load i32, ptr [[I5]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK1-IRBUILDER-NEXT: store i32 [[ADD8]], ptr [[I5]], align 4
+// CHECK1-IRBUILDER-NEXT: [[TMP11:%.*]] = load i32, ptr [[I5]], align 4
// CHECK1-IRBUILDER-NEXT: [[IDXPROM:%.*]] = sext i32 [[TMP11]] to i64
// CHECK1-IRBUILDER-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], ptr @f, i64 0, i64 [[IDXPROM]]
-// CHECK1-IRBUILDER-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK1-IRBUILDER-NEXT: call void @__captured_stmt(ptr [[I5]]), !llvm.access.group [[ACC_GRP3]]
+// CHECK1-IRBUILDER-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX]], align 4
+// CHECK1-IRBUILDER-NEXT: call void @__captured_stmt(ptr [[I5]])
// CHECK1-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_BODY_ORDERED_AFTER:%.*]]
// CHECK1-IRBUILDER: omp.inner.for.body.ordered.after:
// CHECK1-IRBUILDER-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK1-IRBUILDER: omp.body.continue:
// CHECK1-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_INC:%.*]]
// CHECK1-IRBUILDER: omp.inner.for.inc:
-// CHECK1-IRBUILDER-NEXT: [[TMP12:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK1-IRBUILDER-NEXT: [[TMP12:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4
// CHECK1-IRBUILDER-NEXT: [[ADD9:%.*]] = add i32 [[TMP12]], 1
-// CHECK1-IRBUILDER-NEXT: store i32 [[ADD9]], ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK1-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_COND]], !llvm.loop [[LOOP4:![0-9]+]]
+// CHECK1-IRBUILDER-NEXT: store i32 [[ADD9]], ptr [[DOTOMP_IV]], align 4
+// CHECK1-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_COND]], !llvm.loop [[LOOP3:![0-9]+]]
// CHECK1-IRBUILDER: omp.inner.for.end:
// CHECK1-IRBUILDER-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
// CHECK1-IRBUILDER-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
@@ -1278,34 +1278,34 @@ void foo_simd(int low, int up) {
// CHECK1-IRBUILDER-NEXT: store i32 [[TMP26]], ptr [[DOTOMP_IV16]], align 4
// CHECK1-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_COND30:%.*]]
// CHECK1-IRBUILDER: omp.inner.for.cond30:
-// CHECK1-IRBUILDER-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7:![0-9]+]]
-// CHECK1-IRBUILDER-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTOMP_UB]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK1-IRBUILDER-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4
+// CHECK1-IRBUILDER-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTOMP_UB]], align 4
// CHECK1-IRBUILDER-NEXT: [[ADD31:%.*]] = add i32 [[TMP28]], 1
// CHECK1-IRBUILDER-NEXT: [[CMP32:%.*]] = icmp ult i32 [[TMP27]], [[ADD31]]
// CHECK1-IRBUILDER-NEXT: br i1 [[CMP32]], label [[OMP_INNER_FOR_BODY33:%.*]], label [[OMP_INNER_FOR_END42:%.*]]
// CHECK1-IRBUILDER: omp.inner.for.body33:
-// CHECK1-IRBUILDER-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_18]], align 4, !llvm.access.group [[ACC_GRP7]]
-// CHECK1-IRBUILDER-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK1-IRBUILDER-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK1-IRBUILDER-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4
// CHECK1-IRBUILDER-NEXT: [[MUL34:%.*]] = mul i32 [[TMP30]], 1
// CHECK1-IRBUILDER-NEXT: [[ADD35:%.*]] = add i32 [[TMP29]], [[MUL34]]
-// CHECK1-IRBUILDER-NEXT: store i32 [[ADD35]], ptr [[I28]], align 4, !llvm.access.group [[ACC_GRP7]]
-// CHECK1-IRBUILDER-NEXT: [[TMP31:%.*]] = load i32, ptr [[I28]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK1-IRBUILDER-NEXT: store i32 [[ADD35]], ptr [[I28]], align 4
+// CHECK1-IRBUILDER-NEXT: [[TMP31:%.*]] = load i32, ptr [[I28]], align 4
// CHECK1-IRBUILDER-NEXT: [[IDXPROM36:%.*]] = sext i32 [[TMP31]] to i64
// CHECK1-IRBUILDER-NEXT: [[ARRAYIDX37:%.*]] = getelementptr inbounds [10 x float], ptr @f, i64 0, i64 [[IDXPROM36]]
-// CHECK1-IRBUILDER-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX37]], align 4, !llvm.access.group [[ACC_GRP7]]
-// CHECK1-IRBUILDER-NEXT: call void @__captured_stmt.1(ptr [[I28]]), !llvm.access.group [[ACC_GRP7]]
+// CHECK1-IRBUILDER-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX37]], align 4
+// CHECK1-IRBUILDER-NEXT: call void @__captured_stmt.1(ptr [[I28]])
// CHECK1-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_BODY33_ORDERED_AFTER:%.*]]
// CHECK1-IRBUILDER: omp.inner.for.body33.ordered.after:
// CHECK1-IRBUILDER-NEXT: br label [[OMP_BODY_CONTINUE38:%.*]]
// CHECK1-IRBUILDER: omp.body.continue38:
// CHECK1-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_INC39:%.*]]
// CHECK1-IRBUILDER: omp.inner.for.inc39:
-// CHECK1-IRBUILDER-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK1-IRBUILDER-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4
// CHECK1-IRBUILDER-NEXT: [[ADD40:%.*]] = add i32 [[TMP32]], 1
-// CHECK1-IRBUILDER-NEXT: store i32 [[ADD40]], ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK1-IRBUILDER-NEXT: store i32 [[ADD40]], ptr [[DOTOMP_IV16]], align 4
// CHECK1-IRBUILDER-NEXT: [[OMP_GLOBAL_THREAD_NUM41:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB12]])
-// CHECK1-IRBUILDER-NEXT: call void @__kmpc_dispatch_fini_4u(ptr @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM41]]), !llvm.access.group [[ACC_GRP7]]
-// CHECK1-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_COND30]], !llvm.loop [[LOOP8:![0-9]+]]
+// CHECK1-IRBUILDER-NEXT: call void @__kmpc_dispatch_fini_4u(ptr @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM41]])
+// CHECK1-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_COND30]], !llvm.loop [[LOOP5:![0-9]+]]
// CHECK1-IRBUILDER: omp.inner.for.end42:
// CHECK1-IRBUILDER-NEXT: br label [[OMP_DISPATCH_INC:%.*]]
// CHECK1-IRBUILDER: omp.dispatch.inc:
@@ -1812,30 +1812,30 @@ void foo_simd(int low, int up) {
// CHECK3-NEXT: store i32 0, ptr [[DOTOMP_IV]], align 4
// CHECK3-NEXT: br label [[OMP_INNER_FOR_COND:%.*]]
// CHECK3: omp.inner.for.cond:
-// CHECK3-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3:![0-9]+]]
-// CHECK3-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK3-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4
+// CHECK3-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
// CHECK3-NEXT: [[ADD6:%.*]] = add i32 [[TMP9]], 1
// CHECK3-NEXT: [[CMP7:%.*]] = icmp ult i32 [[TMP8]], [[ADD6]]
// CHECK3-NEXT: br i1 [[CMP7]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
// CHECK3: omp.inner.for.body:
-// CHECK3-NEXT: [[TMP10:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK3-NEXT: [[TMP11:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK3-NEXT: [[TMP10:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK3-NEXT: [[TMP11:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4
// CHECK3-NEXT: [[MUL:%.*]] = mul i32 [[TMP11]], 1
// CHECK3-NEXT: [[ADD8:%.*]] = add i32 [[TMP10]], [[MUL]]
-// CHECK3-NEXT: store i32 [[ADD8]], ptr [[I5]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK3-NEXT: [[TMP12:%.*]] = load i32, ptr [[I5]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK3-NEXT: store i32 [[ADD8]], ptr [[I5]], align 4
+// CHECK3-NEXT: [[TMP12:%.*]] = load i32, ptr [[I5]], align 4
// CHECK3-NEXT: [[IDXPROM:%.*]] = sext i32 [[TMP12]] to i64
// CHECK3-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], ptr @f, i64 0, i64 [[IDXPROM]]
-// CHECK3-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK3-NEXT: call void @__captured_stmt(ptr [[I5]]), !llvm.access.group [[ACC_GRP3]]
+// CHECK3-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX]], align 4
+// CHECK3-NEXT: call void @__captured_stmt(ptr [[I5]])
// CHECK3-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK3: omp.body.continue:
// CHECK3-NEXT: br label [[OMP_INNER_FOR_INC:%.*]]
// CHECK3: omp.inner.for.inc:
-// CHECK3-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK3-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4
// CHECK3-NEXT: [[ADD9:%.*]] = add i32 [[TMP13]], 1
-// CHECK3-NEXT: store i32 [[ADD9]], ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK3-NEXT: br label [[OMP_INNER_FOR_COND]], !llvm.loop [[LOOP4:![0-9]+]]
+// CHECK3-NEXT: store i32 [[ADD9]], ptr [[DOTOMP_IV]], align 4
+// CHECK3-NEXT: br label [[OMP_INNER_FOR_COND]], !llvm.loop [[LOOP3:![0-9]+]]
// CHECK3: omp.inner.for.end:
// CHECK3-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
// CHECK3-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
@@ -1885,31 +1885,31 @@ void foo_simd(int low, int up) {
// CHECK3-NEXT: store i32 [[TMP27]], ptr [[DOTOMP_IV16]], align 4
// CHECK3-NEXT: br label [[OMP_INNER_FOR_COND29:%.*]]
// CHECK3: omp.inner.for.cond29:
-// CHECK3-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7:![0-9]+]]
-// CHECK3-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTOMP_UB]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK3-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4
+// CHECK3-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTOMP_UB]], align 4
// CHECK3-NEXT: [[ADD30:%.*]] = add i32 [[TMP29]], 1
// CHECK3-NEXT: [[CMP31:%.*]] = icmp ult i32 [[TMP28]], [[ADD30]]
// CHECK3-NEXT: br i1 [[CMP31]], label [[OMP_INNER_FOR_BODY32:%.*]], label [[OMP_INNER_FOR_END40:%.*]]
// CHECK3: omp.inner.for.body32:
-// CHECK3-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_18]], align 4, !llvm.access.group [[ACC_GRP7]]
-// CHECK3-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK3-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK3-NEXT: [[TMP31:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4
// CHECK3-NEXT: [[MUL33:%.*]] = mul i32 [[TMP31]], 1
// CHECK3-NEXT: [[ADD34:%.*]] = add i32 [[TMP30]], [[MUL33]]
-// CHECK3-NEXT: store i32 [[ADD34]], ptr [[I28]], align 4, !llvm.access.group [[ACC_GRP7]]
-// CHECK3-NEXT: [[TMP32:%.*]] = load i32, ptr [[I28]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK3-NEXT: store i32 [[ADD34]], ptr [[I28]], align 4
+// CHECK3-NEXT: [[TMP32:%.*]] = load i32, ptr [[I28]], align 4
// CHECK3-NEXT: [[IDXPROM35:%.*]] = sext i32 [[TMP32]] to i64
// CHECK3-NEXT: [[ARRAYIDX36:%.*]] = getelementptr inbounds [10 x float], ptr @f, i64 0, i64 [[IDXPROM35]]
-// CHECK3-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX36]], align 4, !llvm.access.group [[ACC_GRP7]]
-// CHECK3-NEXT: call void @__captured_stmt.1(ptr [[I28]]), !llvm.access.group [[ACC_GRP7]]
+// CHECK3-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX36]], align 4
+// CHECK3-NEXT: call void @__captured_stmt.1(ptr [[I28]])
// CHECK3-NEXT: br label [[OMP_BODY_CONTINUE37:%.*]]
// CHECK3: omp.body.continue37:
// CHECK3-NEXT: br label [[OMP_INNER_FOR_INC38:%.*]]
// CHECK3: omp.inner.for.inc38:
-// CHECK3-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK3-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4
// CHECK3-NEXT: [[ADD39:%.*]] = add i32 [[TMP33]], 1
-// CHECK3-NEXT: store i32 [[ADD39]], ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7]]
-// CHECK3-NEXT: call void @__kmpc_dispatch_fini_4u(ptr @[[GLOB1]], i32 [[TMP0]]), !llvm.access.group [[ACC_GRP7]]
-// CHECK3-NEXT: br label [[OMP_INNER_FOR_COND29]], !llvm.loop [[LOOP8:![0-9]+]]
+// CHECK3-NEXT: store i32 [[ADD39]], ptr [[DOTOMP_IV16]], align 4
+// CHECK3-NEXT: call void @__kmpc_dispatch_fini_4u(ptr @[[GLOB1]], i32 [[TMP0]])
+// CHECK3-NEXT: br label [[OMP_INNER_FOR_COND29]], !llvm.loop [[LOOP5:![0-9]+]]
// CHECK3: omp.inner.for.end40:
// CHECK3-NEXT: br label [[OMP_DISPATCH_INC:%.*]]
// CHECK3: omp.dispatch.inc:
@@ -2441,32 +2441,32 @@ void foo_simd(int low, int up) {
// CHECK3-IRBUILDER-NEXT: store i32 0, ptr [[DOTOMP_IV]], align 4
// CHECK3-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_COND:%.*]]
// CHECK3-IRBUILDER: omp.inner.for.cond:
-// CHECK3-IRBUILDER-NEXT: [[TMP7:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3:![0-9]+]]
-// CHECK3-IRBUILDER-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK3-IRBUILDER-NEXT: [[TMP7:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4
+// CHECK3-IRBUILDER-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
// CHECK3-IRBUILDER-NEXT: [[ADD6:%.*]] = add i32 [[TMP8]], 1
// CHECK3-IRBUILDER-NEXT: [[CMP7:%.*]] = icmp ult i32 [[TMP7]], [[ADD6]]
// CHECK3-IRBUILDER-NEXT: br i1 [[CMP7]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
// CHECK3-IRBUILDER: omp.inner.for.body:
-// CHECK3-IRBUILDER-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK3-IRBUILDER-NEXT: [[TMP10:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK3-IRBUILDER-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK3-IRBUILDER-NEXT: [[TMP10:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4
// CHECK3-IRBUILDER-NEXT: [[MUL:%.*]] = mul i32 [[TMP10]], 1
// CHECK3-IRBUILDER-NEXT: [[ADD8:%.*]] = add i32 [[TMP9]], [[MUL]]
-// CHECK3-IRBUILDER-NEXT: store i32 [[ADD8]], ptr [[I5]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK3-IRBUILDER-NEXT: [[TMP11:%.*]] = load i32, ptr [[I5]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK3-IRBUILDER-NEXT: store i32 [[ADD8]], ptr [[I5]], align 4
+// CHECK3-IRBUILDER-NEXT: [[TMP11:%.*]] = load i32, ptr [[I5]], align 4
// CHECK3-IRBUILDER-NEXT: [[IDXPROM:%.*]] = sext i32 [[TMP11]] to i64
// CHECK3-IRBUILDER-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], ptr @f, i64 0, i64 [[IDXPROM]]
-// CHECK3-IRBUILDER-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK3-IRBUILDER-NEXT: call void @__captured_stmt(ptr [[I5]]), !llvm.access.group [[ACC_GRP3]]
+// CHECK3-IRBUILDER-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX]], align 4
+// CHECK3-IRBUILDER-NEXT: call void @__captured_stmt(ptr [[I5]])
// CHECK3-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_BODY_ORDERED_AFTER:%.*]]
// CHECK3-IRBUILDER: omp.inner.for.body.ordered.after:
// CHECK3-IRBUILDER-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK3-IRBUILDER: omp.body.continue:
// CHECK3-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_INC:%.*]]
// CHECK3-IRBUILDER: omp.inner.for.inc:
-// CHECK3-IRBUILDER-NEXT: [[TMP12:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3]]
+// CHECK3-IRBUILDER-NEXT: [[TMP12:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4
// CHECK3-IRBUILDER-NEXT: [[ADD9:%.*]] = add i32 [[TMP12]], 1
-// CHECK3-IRBUILDER-NEXT: store i32 [[ADD9]], ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP3]]
-// CHECK3-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_COND]], !llvm.loop [[LOOP4:![0-9]+]]
+// CHECK3-IRBUILDER-NEXT: store i32 [[ADD9]], ptr [[DOTOMP_IV]], align 4
+// CHECK3-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_COND]], !llvm.loop [[LOOP3:![0-9]+]]
// CHECK3-IRBUILDER: omp.inner.for.end:
// CHECK3-IRBUILDER-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
// CHECK3-IRBUILDER-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
@@ -2518,34 +2518,34 @@ void foo_simd(int low, int up) {
// CHECK3-IRBUILDER-NEXT: store i32 [[TMP26]], ptr [[DOTOMP_IV16]], align 4
// CHECK3-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_COND30:%.*]]
// CHECK3-IRBUILDER: omp.inner.for.cond30:
-// CHECK3-IRBUILDER-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7:![0-9]+]]
-// CHECK3-IRBUILDER-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTOMP_UB]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK3-IRBUILDER-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4
+// CHECK3-IRBUILDER-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTOMP_UB]], align 4
// CHECK3-IRBUILDER-NEXT: [[ADD31:%.*]] = add i32 [[TMP28]], 1
// CHECK3-IRBUILDER-NEXT: [[CMP32:%.*]] = icmp ult i32 [[TMP27]], [[ADD31]]
// CHECK3-IRBUILDER-NEXT: br i1 [[CMP32]], label [[OMP_INNER_FOR_BODY33:%.*]], label [[OMP_INNER_FOR_END42:%.*]]
// CHECK3-IRBUILDER: omp.inner.for.body33:
-// CHECK3-IRBUILDER-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_18]], align 4, !llvm.access.group [[ACC_GRP7]]
-// CHECK3-IRBUILDER-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK3-IRBUILDER-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_18]], align 4
+// CHECK3-IRBUILDER-NEXT: [[TMP30:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4
// CHECK3-IRBUILDER-NEXT: [[MUL34:%.*]] = mul i32 [[TMP30]], 1
// CHECK3-IRBUILDER-NEXT: [[ADD35:%.*]] = add i32 [[TMP29]], [[MUL34]]
-// CHECK3-IRBUILDER-NEXT: store i32 [[ADD35]], ptr [[I28]], align 4, !llvm.access.group [[ACC_GRP7]]
-// CHECK3-IRBUILDER-NEXT: [[TMP31:%.*]] = load i32, ptr [[I28]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK3-IRBUILDER-NEXT: store i32 [[ADD35]], ptr [[I28]], align 4
+// CHECK3-IRBUILDER-NEXT: [[TMP31:%.*]] = load i32, ptr [[I28]], align 4
// CHECK3-IRBUILDER-NEXT: [[IDXPROM36:%.*]] = sext i32 [[TMP31]] to i64
// CHECK3-IRBUILDER-NEXT: [[ARRAYIDX37:%.*]] = getelementptr inbounds [10 x float], ptr @f, i64 0, i64 [[IDXPROM36]]
-// CHECK3-IRBUILDER-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX37]], align 4, !llvm.access.group [[ACC_GRP7]]
-// CHECK3-IRBUILDER-NEXT: call void @__captured_stmt.1(ptr [[I28]]), !llvm.access.group [[ACC_GRP7]]
+// CHECK3-IRBUILDER-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX37]], align 4
+// CHECK3-IRBUILDER-NEXT: call void @__captured_stmt.1(ptr [[I28]])
// CHECK3-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_BODY33_ORDERED_AFTER:%.*]]
// CHECK3-IRBUILDER: omp.inner.for.body33.ordered.after:
// CHECK3-IRBUILDER-NEXT: br label [[OMP_BODY_CONTINUE38:%.*]]
// CHECK3-IRBUILDER: omp.body.continue38:
// CHECK3-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_INC39:%.*]]
// CHECK3-IRBUILDER: omp.inner.for.inc39:
-// CHECK3-IRBUILDER-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK3-IRBUILDER-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTOMP_IV16]], align 4
// CHECK3-IRBUILDER-NEXT: [[ADD40:%.*]] = add i32 [[TMP32]], 1
-// CHECK3-IRBUILDER-NEXT: store i32 [[ADD40]], ptr [[DOTOMP_IV16]], align 4, !llvm.access.group [[ACC_GRP7]]
+// CHECK3-IRBUILDER-NEXT: store i32 [[ADD40]], ptr [[DOTOMP_IV16]], align 4
// CHECK3-IRBUILDER-NEXT: [[OMP_GLOBAL_THREAD_NUM41:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB12]])
-// CHECK3-IRBUILDER-NEXT: call void @__kmpc_dispatch_fini_4u(ptr @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM41]]), !llvm.access.group [[ACC_GRP7]]
-// CHECK3-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_COND30]], !llvm.loop [[LOOP8:![0-9]+]]
+// CHECK3-IRBUILDER-NEXT: call void @__kmpc_dispatch_fini_4u(ptr @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM41]])
+// CHECK3-IRBUILDER-NEXT: br label [[OMP_INNER_FOR_COND30]], !llvm.loop [[LOOP5:![0-9]+]]
// CHECK3-IRBUILDER: omp.inner.for.end42:
// CHECK3-IRBUILDER-NEXT: br label [[OMP_DISPATCH_INC:%.*]]
// CHECK3-IRBUILDER: omp.dispatch.inc:
@@ -2885,33 +2885,33 @@ void foo_simd(int low, int up) {
// CHECK5-NEXT: store i32 0, ptr [[DOTOMP_IV]], align 4
// CHECK5-NEXT: br label [[OMP_INNER_FOR_COND:%.*]]
// CHECK5: omp.inner.for.cond:
-// CHECK5-NEXT: [[TMP7:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP9:![0-9]+]]
-// CHECK5-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4, !llvm.access.group [[ACC_GRP9]]
+// CHECK5-NEXT: [[TMP7:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4
+// CHECK5-NEXT: [[TMP8:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_2]], align 4
// CHECK5-NEXT: [[ADD6:%.*]] = add i32 [[TMP8]], 1
// CHECK5-NEXT: [[CMP7:%.*]] = icmp ult i32 [[TMP7]], [[ADD6]]
// CHECK5-NEXT: br i1 [[CMP7]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
// CHECK5: omp.inner.for.body:
-// CHECK5-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4, !llvm.access.group [[ACC_GRP9]]
-// CHECK5-NEXT: [[TMP10:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP9]]
+// CHECK5-NEXT: [[TMP9:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
+// CHECK5-NEXT: [[TMP10:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4
// CHECK5-NEXT: [[MUL:%.*]] = mul i32 [[TMP10]], 1
// CHECK5-NEXT: [[ADD8:%.*]] = add i32 [[TMP9]], [[MUL]]
-// CHECK5-NEXT: store i32 [[ADD8]], ptr [[I5]], align 4, !llvm.access.group [[ACC_GRP9]]
-// CHECK5-NEXT: [[TMP11:%.*]] = load i32, ptr [[I5]], align 4, !llvm.access.group [[ACC_GRP9]]
+// CHECK5-NEXT: store i32 [[ADD8]], ptr [[I5]], align 4
+// CHECK5-NEXT: [[TMP11:%.*]] = load i32, ptr [[I5]], align 4
// CHECK5-NEXT: [[IDXPROM:%.*]] = sext i32 [[TMP11]] to i64
// CHECK5-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x float], ptr @f, i64 0, i64 [[IDXPROM]]
-// CHECK5-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX]], align 4, !llvm.access.group [[ACC_GRP9]]
-// CHECK5-NEXT: [[TMP12:%.*]] = load i32, ptr [[I5]], align 4, !llvm.access.group [[ACC_GRP9]]
+// CHECK5-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX]], align 4
+// CHECK5-NEXT: [[TMP12:%.*]] = load i32, ptr [[I5]], align 4
// CHECK5-NEXT: [[IDXPROM9:%.*]] = sext i32 [[TMP12]] to i64
// CHECK5-NEXT: [[ARRAYIDX10:%.*]] = getelementptr inbounds [10 x float], ptr @f, i64 0, i64 [[IDXPROM9]]
-// CHECK5-NEXT: store float 1.000000e+00, ptr [[ARRAYIDX10]], align 4, !llvm.access.group [[ACC_GRP9]]
+// CHECK5-NEXT: store float 1.000000e+00, ptr [[ARRAYIDX10]], align 4
// CHECK5-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK5: omp.body.continue:
// CHECK5-NEXT: br label [[OMP_INNER_FOR_INC:%.*]]
// CHECK5: omp.inner.for.inc:
-// CHECK5-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP9]]
+// CHECK5-NEXT: [[TMP13:%.*]] = load i32, ptr [[DOTOMP_IV]], align 4
// CHECK5-NEXT: [[ADD11:%.*]] = add i32 [[TMP13]], 1
-// CHECK5-NEXT: store i32 [[ADD11]], ptr [[DOTOMP_IV]], align 4, !llvm.access.group [[ACC_GRP9]]
-// CHECK5-NEXT: br label [[OMP_INNER_FOR_COND]], !llvm.loop [[LOOP10:![0-9]+]]
+// CHECK5-NEXT: store i32 [[ADD11]], ptr [[DOTOMP_IV]], align 4
+// CHECK5-NEXT: br label [[OMP_INNER_FOR_COND]], !llvm.loop [[LOOP9:![0-9]+]]
// CHECK5: omp.inner.for.end:
// CHECK5-NEXT: [[TMP14:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]], align 4
// CHECK5-NEXT: [[TMP15:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_1]], align 4
@@ -2951,33 +2951,33 @@ void foo_simd(int low, int up) {
// CHECK5-NEXT: store i32 [[TMP25]], ptr [[DOTOMP_IV30]], align 4
// CHECK5-NEXT: br label [[OMP_INNER_FOR_COND32:%.*]]
// CHECK5: omp.inner.for.cond32:
-// CHECK5-NEXT: [[TMP26:%.*]] = load i32, ptr [[DOTOMP_IV30]], align 4, !llvm.access.group [[ACC_GRP13:![0-9]+]]
-// CHECK5-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTOMP_UB]], align 4, !llvm.access.group [[ACC_GRP13]]
+// CHECK5-NEXT: [[TMP26:%.*]] = load i32, ptr [[DOTOMP_IV30]], align 4
+// CHECK5-NEXT: [[TMP27:%.*]] = load i32, ptr [[DOTOMP_UB]], align 4
// CHECK5-NEXT: [[ADD33:%.*]] = add i32 [[TMP27]], 1
// CHECK5-NEXT: [[CMP34:%.*]] = icmp ult i32 [[TMP26]], [[ADD33]]
// CHECK5-NEXT: br i1 [[CMP34]], label [[OMP_INNER_FOR_BODY35:%.*]], label [[OMP_INNER_FOR_END45:%.*]]
// CHECK5: omp.inner.for.body35:
-// CHECK5-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_19]], align 4, !llvm.access.group [[ACC_GRP13]]
-// CHECK5-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTOMP_IV30]], align 4, !llvm.access.group [[ACC_GRP13]]
+// CHECK5-NEXT: [[TMP28:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_19]], align 4
+// CHECK5-NEXT: [[TMP29:%.*]] = load i32, ptr [[DOTOMP_IV30]], align 4
// CHECK5-NEXT: [[MUL36:%.*]] = mul i32 [[TMP29]], 1
// CHECK5-NEXT: [[ADD37:%.*]] = add i32 [[TMP28]], [[MUL36]]
-// CHECK5-NEXT: store i32 [[ADD37]], ptr [[I31]], align 4, !llvm.access.group [[ACC_GRP13]]
-// CHECK5-NEXT: [[TMP30:%.*]] = load i32, ptr [[I31]], align 4, !llvm.access.group [[ACC_GRP13]]
+// CHECK5-NEXT: store i32 [[ADD37]], ptr [[I31]], align 4
+// CHECK5-NEXT: [[TMP30:%.*]] = load i32, ptr [[I31]], align 4
// CHECK5-NEXT: [[IDXPROM38:%.*]] = sext i32 [[TMP30]] to i64
// CHECK5-NEXT: [[ARRAYIDX39:%.*]] = getelementptr inbounds [10 x float], ptr @f, i64 0, i64 [[IDXPROM38]]
-// CHECK5-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX39]], align 4, !llvm.access.group [[ACC_GRP13]]
-// CHECK5-NEXT: [[TMP31:%.*]] = load i32, ptr [[I31]], align 4, !llvm.access.group [[ACC_GRP13]]
+// CHECK5-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX39]], align 4
+// CHECK5-NEXT: [[TMP31:%.*]] = load i32, ptr [[I31]], align 4
// CHECK5-NEXT: [[IDXPROM40:%.*]] = sext i32 [[TMP31]] to i64
// CHECK5-NEXT: [[ARRAYIDX41:%.*]] = getelementptr inbounds [10 x float], ptr @f, i64 0, i64 [[IDXPROM40]]
-// CHECK5-NEXT: store float 1.000000e+00, ptr [[ARRAYIDX41]], align 4, !llvm.access.group [[ACC_GRP13]]
+// CHECK5-NEXT: store float 1.000000e+00, ptr [[ARRAYIDX41]], align 4
// CHECK5-NEXT: br label [[OMP_BODY_CONTINUE42:%.*]]
// CHECK5: omp.body.continue42:
// CHECK5-NEXT: br label [[OMP_INNER_FOR_INC43:%.*]]
// CHECK5: omp.inner.for.inc43:
-// CHECK5-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTOMP_IV30]], align 4, !llvm.access.group [[ACC_GRP13]]
+// CHECK5-NEXT: [[TMP32:%.*]] = load i32, ptr [[DOTOMP_IV30]], align 4
// CHECK5-NEXT: [[ADD44:%.*]] = add i32 [[TMP32]], 1
-// CHECK5-NEXT: store i32 [[ADD44]], ptr [[DOTOMP_IV30]], align 4, !llvm.access.group [[ACC_GRP13]]
-// CHECK5-NEXT: br label [[OMP_INNER_FOR_COND32]], !llvm.loop [[LOOP14:![0-9]+]]
+// CHECK5-NEXT: store i32 [[ADD44]], ptr [[DOTOMP_IV30]], align 4
+// CHECK5-NEXT: br label [[OMP_INNER_FOR_COND32]], !llvm.loop [[LOOP11:![0-9]+]]
// CHECK5: omp.inner.for.end45:
// CHECK5-NEXT: [[TMP33:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_19]], align 4
// CHECK5-NEXT: [[TMP34:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_20]], align 4
diff --git a/openmp/runtime/test/misc_bugs/simd_conservative_ordered.c b/openmp/runtime/test/misc_bugs/simd_conservative_ordered.c
new file mode 100644
index 000000000000000..af0fa66eee1d7c3
--- /dev/null
+++ b/openmp/runtime/test/misc_bugs/simd_conservative_ordered.c
@@ -0,0 +1,84 @@
+// RUN: %libomp-compile -O3 -ffast-math
+// RUN: %libomp-run
+#include <float.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <time.h>
+
+int compare_float(float x1, float x2, float scalar) {
+ const float diff = fabsf(x1 - x2);
+ x1 = fabsf(x1);
+ x2 = fabsf(x2);
+ const float l = (x2 > x1) ? x2 : x1;
+ if (diff <= l * scalar * FLT_EPSILON)
+ return 1;
+ else
+ return 0;
+}
+
+#define ARRAY_SIZE 256
+
+__attribute__((noinline)) void
+initialization_loop(float X[ARRAY_SIZE][ARRAY_SIZE],
+ float Y[ARRAY_SIZE][ARRAY_SIZE]) {
+ const float max = 1000.0;
+ srand(time(NULL));
+ for (int r = 0; r < ARRAY_SIZE; r++) {
+ for (int c = 0; c < ARRAY_SIZE; c++) {
+ X[r][c] = ((float)rand() / (float)(RAND_MAX)) * max;
+ Y[r][c] = X[r][c];
+ }
+ }
+}
+
+__attribute__((noinline)) void omp_simd_loop(float X[ARRAY_SIZE][ARRAY_SIZE]) {
+ for (int r = 1; r < ARRAY_SIZE; ++r) {
+ for (int c = 1; c < ARRAY_SIZE; ++c) {
+#pragma omp simd
+ for (int k = 2; k < ARRAY_SIZE; ++k) {
+#pragma omp ordered simd
+ X[r][k] = X[r][k - 2] + sinf((float)(r / c));
+ }
+ }
+ }
+}
+
+__attribute__((noinline)) int comparison_loop(float X[ARRAY_SIZE][ARRAY_SIZE],
+ float Y[ARRAY_SIZE][ARRAY_SIZE]) {
+ int totalErrors_simd = 0;
+ const float scalar = 1.0;
+ for (int r = 1; r < ARRAY_SIZE; ++r) {
+ for (int c = 1; c < ARRAY_SIZE; ++c) {
+ for (int k = 2; k < ARRAY_SIZE; ++k) {
+ Y[r][k] = Y[r][k - 2] + sinf((float)(r / c));
+ }
+ }
+ // check row for simd update
+ for (int k = 0; k < ARRAY_SIZE; ++k) {
+ if (!compare_float(X[r][k], Y[r][k], scalar)) {
+ ++totalErrors_simd;
+ }
+ }
+ }
+ return totalErrors_simd;
+}
+
+int main(void) {
+ float X[ARRAY_SIZE][ARRAY_SIZE];
+ float Y[ARRAY_SIZE][ARRAY_SIZE];
+
+ initialization_loop(X, Y);
+ omp_simd_loop(X);
+ const int totalErrors_simd = comparison_loop(X, Y);
+
+ if (totalErrors_simd) {
+ fprintf(stdout, "totalErrors_simd: %d \n", totalErrors_simd);
+ fprintf(stdout, "%s : %d - FAIL: error in ordered simd computation.\n",
+ __FILE__, __LINE__);
+ } else {
+ fprintf(stdout, "Success!\n");
+ }
+
+ return totalErrors_simd;
+}
More information about the Openmp-commits
mailing list