| // RUN: %clang_cc1 -verify -triple x86_64-unknown-linux-gnu -fclang-abi-compat=latest -std=c++17 -fopenmp -O2 -vectorize-loops -vectorize-slp -emit-llvm %s -o - | FileCheck %s |
| // REQUIRES: x86-registered-target |
| // expected-no-diagnostics |
| |
| // Why the intra-tile loop is stored min-bounded: that form carries its trip |
| // count in the loop bound and vectorizes, while the rectangular form a |
| // 'collapse' consumer needs adds a body guard that blocks the vectorizer. Both |
| // sides are checked here so the tradeoff is visible in one place. |
| |
| // CHECK-LABEL: define {{.*}} @vec_tile( |
| // CHECK: fadd <4 x float> |
| extern "C" void vec_tile(float *a, float *b, int n) { |
| #pragma omp tile sizes(64) |
| for (int i = 0; i < n; ++i) |
| a[i] = a[i] + b[i]; |
| } |
| |
| // CHECK-LABEL: define internal void @vec_tile_collapsed.omp_outlined( |
| // CHECK-NOT: fadd <4 x float> |
| // CHECK: br i1 %{{.*}}, label %omp.body.next, label %omp.inner.for.inc |
| // CHECK-NOT: fadd <4 x float> |
| // CHECK: ret void |
| extern "C" void vec_tile_collapsed(float *a, float *b, int n) { |
| #pragma omp parallel for collapse(2) |
| #pragma omp tile sizes(64) |
| for (int i = 0; i < n; ++i) |
| a[i] = a[i] + b[i]; |
| } |
