diff options
author | Jan Leyonberg <jan_sjodin@yahoo.com> | 2024-04-19 15:50:55 -0400 |
---|---|---|
committer | GitHub <noreply@github.com> | 2024-04-19 15:50:55 -0400 |
commit | 0c455ee34823cb991a35e33ff020bb7cc4e44c8a (patch) | |
tree | 762b24400af91ab653aa1e1ee9e31a1c634e1b7f | |
parent | c32712d1763d74329b42c1cd68a24d4c0075b596 (diff) | |
download | llvm-0c455ee34823cb991a35e33ff020bb7cc4e44c8a.zip llvm-0c455ee34823cb991a35e33ff020bb7cc4e44c8a.tar.gz llvm-0c455ee34823cb991a35e33ff020bb7cc4e44c8a.tar.bz2 |
[flang][OpenMP] Use maxnum/minnum for lowering of max/min reduction operators (#89258)
This patch changes the lowering of max and min to be lowered to
arith::MaxNumFop and arith::MinNumFOp instead of using arith::MaximumFOp
and arith::MinimumFOp. The arith::MaximumFOp and arith::MinimumFOp map
to the corresponding intrinsics llvm.maximum.* and llvm.minimum.*
intrinsics which conform to the semantics specified in the draft of IEEE
754-2019, which is not supported by all hardware. Instead using
arith::MaximumFOp and arith::MinimumFOp will allow code generation for
more targets and match the code generated by clang OpenMP.
fixes #87955
9 files changed, 10 insertions, 10 deletions
diff --git a/flang/lib/Lower/OpenMP/ReductionProcessor.cpp b/flang/lib/Lower/OpenMP/ReductionProcessor.cpp index 6a91ee4..9f8352a 100644 --- a/flang/lib/Lower/OpenMP/ReductionProcessor.cpp +++ b/flang/lib/Lower/OpenMP/ReductionProcessor.cpp @@ -220,12 +220,12 @@ mlir::Value ReductionProcessor::createScalarCombiner( switch (redId) { case ReductionIdentifier::MAX: reductionOp = - getReductionOperation<mlir::arith::MaximumFOp, mlir::arith::MaxSIOp>( + getReductionOperation<mlir::arith::MaxNumFOp, mlir::arith::MaxSIOp>( builder, type, loc, op1, op2); break; case ReductionIdentifier::MIN: reductionOp = - getReductionOperation<mlir::arith::MinimumFOp, mlir::arith::MinSIOp>( + getReductionOperation<mlir::arith::MinNumFOp, mlir::arith::MinSIOp>( builder, type, loc, op1, op2); break; case ReductionIdentifier::IOR: diff --git a/flang/test/Lower/OpenMP/FIR/wsloop-reduction-max-byref.f90 b/flang/test/Lower/OpenMP/FIR/wsloop-reduction-max-byref.f90 index f0979ab..80b720e 100644 --- a/flang/test/Lower/OpenMP/FIR/wsloop-reduction-max-byref.f90 +++ b/flang/test/Lower/OpenMP/FIR/wsloop-reduction-max-byref.f90 @@ -11,7 +11,7 @@ !CHECK: ^bb0(%[[ARG0:.*]]: !fir.ref<f32>, %[[ARG1:.*]]: !fir.ref<f32>): !CHECK: %[[LD0:.*]] = fir.load %[[ARG0]] : !fir.ref<f32> !CHECK: %[[LD1:.*]] = fir.load %[[ARG1]] : !fir.ref<f32> -!CHECK: %[[RES:.*]] = arith.maximumf %[[LD0]], %[[LD1]] {{.*}}: f32 +!CHECK: %[[RES:.*]] = arith.maxnumf %[[LD0]], %[[LD1]] {{.*}}: f32 !CHECK: fir.store %[[RES]] to %[[ARG0]] : !fir.ref<f32> !CHECK: omp.yield(%[[ARG0]] : !fir.ref<f32>) diff --git a/flang/test/Lower/OpenMP/FIR/wsloop-reduction-max.f90 b/flang/test/Lower/OpenMP/FIR/wsloop-reduction-max.f90 index 996296c..c3b821e 100644 --- a/flang/test/Lower/OpenMP/FIR/wsloop-reduction-max.f90 +++ b/flang/test/Lower/OpenMP/FIR/wsloop-reduction-max.f90 @@ -6,7 +6,7 @@ !CHECK: omp.yield(%[[MINIMUM_VAL_F]] : f32) !CHECK: combiner !CHECK: ^bb0(%[[ARG0_F:.*]]: f32, %[[ARG1_F:.*]]: f32): -!CHECK: %[[COMB_VAL_F:.*]] = arith.maximumf %[[ARG0_F]], %[[ARG1_F]] {{.*}}: f32 +!CHECK: %[[COMB_VAL_F:.*]] = arith.maxnumf %[[ARG0_F]], %[[ARG1_F]] {{.*}}: f32 !CHECK: omp.yield(%[[COMB_VAL_F]] : f32) !CHECK: omp.declare_reduction @[[MAX_DECLARE_I:.*]] : i32 init { diff --git a/flang/test/Lower/OpenMP/FIR/wsloop-reduction-min-byref.f90 b/flang/test/Lower/OpenMP/FIR/wsloop-reduction-min-byref.f90 index 24aa8e4..b284f8e5 100644 --- a/flang/test/Lower/OpenMP/FIR/wsloop-reduction-min-byref.f90 +++ b/flang/test/Lower/OpenMP/FIR/wsloop-reduction-min-byref.f90 @@ -11,7 +11,7 @@ !CHECK: ^bb0(%[[ARG0:.*]]: !fir.ref<f32>, %[[ARG1:.*]]: !fir.ref<f32>): !CHECK: %[[LD0:.*]] = fir.load %[[ARG0]] : !fir.ref<f32> !CHECK: %[[LD1:.*]] = fir.load %[[ARG1]] : !fir.ref<f32> -!CHECK: %[[RES:.*]] = arith.minimumf %[[LD0]], %[[LD1]] {{.*}}: f32 +!CHECK: %[[RES:.*]] = arith.minnumf %[[LD0]], %[[LD1]] {{.*}}: f32 !CHECK: fir.store %[[RES]] to %[[ARG0]] : !fir.ref<f32> !CHECK: omp.yield(%[[ARG0]] : !fir.ref<f32>) diff --git a/flang/test/Lower/OpenMP/FIR/wsloop-reduction-min.f90 b/flang/test/Lower/OpenMP/FIR/wsloop-reduction-min.f90 index 268f51c..ab33e18 100644 --- a/flang/test/Lower/OpenMP/FIR/wsloop-reduction-min.f90 +++ b/flang/test/Lower/OpenMP/FIR/wsloop-reduction-min.f90 @@ -6,7 +6,7 @@ !CHECK: omp.yield(%[[MAXIMUM_VAL_F]] : f32) !CHECK: combiner !CHECK: ^bb0(%[[ARG0_F:.*]]: f32, %[[ARG1_F:.*]]: f32): -!CHECK: %[[COMB_VAL_F:.*]] = arith.minimumf %[[ARG0_F]], %[[ARG1_F]] {{.*}}: f32 +!CHECK: %[[COMB_VAL_F:.*]] = arith.minnumf %[[ARG0_F]], %[[ARG1_F]] {{.*}}: f32 !CHECK: omp.yield(%[[COMB_VAL_F]] : f32) !CHECK: omp.declare_reduction @[[MIN_DECLARE_I:.*]] : i32 init { diff --git a/flang/test/Lower/OpenMP/wsloop-reduction-max-byref.f90 b/flang/test/Lower/OpenMP/wsloop-reduction-max-byref.f90 index ee562bb..2f6921e 100644 --- a/flang/test/Lower/OpenMP/wsloop-reduction-max-byref.f90 +++ b/flang/test/Lower/OpenMP/wsloop-reduction-max-byref.f90 @@ -13,7 +13,7 @@ !CHECK: ^bb0(%[[ARG0:.*]]: !fir.ref<f32>, %[[ARG1:.*]]: !fir.ref<f32>): !CHECK: %[[LD0:.*]] = fir.load %[[ARG0]] : !fir.ref<f32> !CHECK: %[[LD1:.*]] = fir.load %[[ARG1]] : !fir.ref<f32> -!CHECK: %[[RES:.*]] = arith.maximumf %[[LD0]], %[[LD1]] {{.*}}: f32 +!CHECK: %[[RES:.*]] = arith.maxnumf %[[LD0]], %[[LD1]] {{.*}}: f32 !CHECK: fir.store %[[RES]] to %[[ARG0]] : !fir.ref<f32> !CHECK: omp.yield(%[[ARG0]] : !fir.ref<f32>) diff --git a/flang/test/Lower/OpenMP/wsloop-reduction-max.f90 b/flang/test/Lower/OpenMP/wsloop-reduction-max.f90 index 6f11f0e..c9cf5cb 100644 --- a/flang/test/Lower/OpenMP/wsloop-reduction-max.f90 +++ b/flang/test/Lower/OpenMP/wsloop-reduction-max.f90 @@ -10,7 +10,7 @@ ! CHECK-LABEL: } combiner { ! CHECK: ^bb0(%[[VAL_0:.*]]: f32, %[[VAL_1:.*]]: f32): -! CHECK: %[[VAL_2:.*]] = arith.maximumf %[[VAL_0]], %[[VAL_1]] fastmath<contract> : f32 +! CHECK: %[[VAL_2:.*]] = arith.maxnumf %[[VAL_0]], %[[VAL_1]] fastmath<contract> : f32 ! CHECK: omp.yield(%[[VAL_2]] : f32) ! CHECK: } diff --git a/flang/test/Lower/OpenMP/wsloop-reduction-min-byref.f90 b/flang/test/Lower/OpenMP/wsloop-reduction-min-byref.f90 index c037211..84a376b 100644 --- a/flang/test/Lower/OpenMP/wsloop-reduction-min-byref.f90 +++ b/flang/test/Lower/OpenMP/wsloop-reduction-min-byref.f90 @@ -13,7 +13,7 @@ !CHECK: ^bb0(%[[ARG0:.*]]: !fir.ref<f32>, %[[ARG1:.*]]: !fir.ref<f32>): !CHECK: %[[LD0:.*]] = fir.load %[[ARG0]] : !fir.ref<f32> !CHECK: %[[LD1:.*]] = fir.load %[[ARG1]] : !fir.ref<f32> -!CHECK: %[[RES:.*]] = arith.minimumf %[[LD0]], %[[LD1]] {{.*}}: f32 +!CHECK: %[[RES:.*]] = arith.minnumf %[[LD0]], %[[LD1]] {{.*}}: f32 !CHECK: fir.store %[[RES]] to %[[ARG0]] : !fir.ref<f32> !CHECK: omp.yield(%[[ARG0]] : !fir.ref<f32>) diff --git a/flang/test/Lower/OpenMP/wsloop-reduction-min.f90 b/flang/test/Lower/OpenMP/wsloop-reduction-min.f90 index 2c694f8..3ba279a 100644 --- a/flang/test/Lower/OpenMP/wsloop-reduction-min.f90 +++ b/flang/test/Lower/OpenMP/wsloop-reduction-min.f90 @@ -10,7 +10,7 @@ ! CHECK-LABEL: } combiner { ! CHECK: ^bb0(%[[VAL_0:.*]]: f32, %[[VAL_1:.*]]: f32): -! CHECK: %[[VAL_2:.*]] = arith.minimumf %[[VAL_0]], %[[VAL_1]] fastmath<contract> : f32 +! CHECK: %[[VAL_2:.*]] = arith.minnumf %[[VAL_0]], %[[VAL_1]] fastmath<contract> : f32 ! CHECK: omp.yield(%[[VAL_2]] : f32) ! CHECK: } |