| ! Check that loop trip-count versioning fires on real lowering output and is | |
| ! controlled by -f[no-]version-loops-for-trip-count. The pass matches the | |
| ! canonical form of array-section loops (arith.maxsi bound, shared slice and | |
| ! loop bound), so this test guards against lowering or canonicalization | |
| ! changes that would silently stop it from matching. | |
| ! | |
| ! RUN: %flang_fc1 -emit-llvm -O3 -fversion-loops-for-trip-count \ | |
| ! RUN: -mmlir --mlir-disable-threading \ | |
| ! RUN: -mmlir --mlir-print-ir-after=loop-tripcount-versioning \ | |
| ! RUN: %s -o /dev/null 2>&1 | FileCheck %s | |
| ! RUN: %flang_fc1 -emit-llvm -O3 \ | |
| ! RUN: -mmlir --mlir-disable-threading \ | |
| ! RUN: -mmlir --mlir-print-ir-after=loop-tripcount-versioning \ | |
| ! RUN: %s -o /dev/null 2>&1 | FileCheck %s --check-prefix=OFF --allow-empty | |
| ! | |
| ! The -mmlir overrides reach the pass. `zero` has extent 4 and a three- | |
| ! operation body (12 unrolled operations), so a limit one below each is the | |
| ! boundary: 3 and 11 leave it unversioned, 4 and 12 version it. | |
| ! RUN: %flang_fc1 -emit-llvm -O3 -fversion-loops-for-trip-count \ | |
| ! RUN: -mmlir --mlir-disable-threading \ | |
| ! RUN: -mmlir --mlir-print-ir-after=loop-tripcount-versioning \ | |
| ! RUN: -mmlir -loop-tripcount-versioning-max-section-extent=3 \ | |
| ! RUN: %s -o /dev/null 2>&1 | FileCheck %s --check-prefix=LIMIT | |
| ! RUN: %flang_fc1 -emit-llvm -O3 -fversion-loops-for-trip-count \ | |
| ! RUN: -mmlir --mlir-disable-threading \ | |
| ! RUN: -mmlir --mlir-print-ir-after=loop-tripcount-versioning \ | |
| ! RUN: -mmlir -loop-tripcount-versioning-max-section-extent=4 \ | |
| ! RUN: %s -o /dev/null 2>&1 | FileCheck %s | |
| ! RUN: %flang_fc1 -emit-llvm -O3 -fversion-loops-for-trip-count \ | |
| ! RUN: -mmlir --mlir-disable-threading \ | |
| ! RUN: -mmlir --mlir-print-ir-after=loop-tripcount-versioning \ | |
| ! RUN: -mmlir -loop-tripcount-versioning-max-unrolled-ops=11 \ | |
| ! RUN: %s -o /dev/null 2>&1 | FileCheck %s --check-prefix=LIMIT | |
| ! RUN: %flang_fc1 -emit-llvm -O3 -fversion-loops-for-trip-count \ | |
| ! RUN: -mmlir --mlir-disable-threading \ | |
| ! RUN: -mmlir --mlir-print-ir-after=loop-tripcount-versioning \ | |
| ! RUN: -mmlir -loop-tripcount-versioning-max-unrolled-ops=12 \ | |
| ! RUN: %s -o /dev/null 2>&1 | FileCheck %s | |
| ! The section bound is a runtime value; the declared leading extent is 4. | |
| subroutine zero(tau, ncol) | |
| real(8), intent(out) :: tau(4, 26) | |
| integer, intent(in) :: ncol | |
| tau(1:ncol, :) = 0.0d0 | |
| end subroutine | |
| ! CHECK-LABEL: IR Dump After LoopTripCountVersioning | |
| ! CHECK-LABEL: func.func @_QPzero( | |
| ! CHECK: %[[UB:.*]] = arith.maxsi %{{.*}}, %{{.*}} : index | |
| ! CHECK: fir.do_loop | |
| ! CHECK: %[[C4:.*]] = arith.constant 4 : index | |
| ! CHECK: %[[CMP:.*]] = arith.cmpi eq, %[[UB]], %[[C4]] : index | |
| ! CHECK: fir.if %[[CMP]] { | |
| ! CHECK: fir.do_loop %{{.*}} = %{{.*}} to %[[C4]] step | |
| ! CHECK: } else { | |
| ! CHECK: fir.do_loop %{{.*}} = %{{.*}} to %[[UB]] step | |
| ! CHECK: } | |
| ! Without the flag the pass is not in the pipeline, so there is no IR dump. | |
| ! OFF-NOT: IR Dump After LoopTripCountVersioning | |
| ! With a binding limit the pass runs but leaves the loop alone. | |
| ! LIMIT-LABEL: IR Dump After LoopTripCountVersioning | |
| ! LIMIT-LABEL: func.func @_QPzero( | |
| ! LIMIT-NOT: fir.if | |
| ! LIMIT: return |