[flang] Lower F08 mask intrinsics

Lower F08 maskl and maskr intrinsics.

Differential Revision: https://reviews.llvm.org/D129296

3 files changed, 197 insertions, 0 deletions

diff --git a/flang/lib/Lower/IntrinsicCall.cpp b/flang/lib/Lower/IntrinsicCall.cpp
index e5036a3..2db221f 100644
--- a/flang/lib/Lower/IntrinsicCall.cpp
+++ b/flang/lib/Lower/IntrinsicCall.cpp
@@ -515,6 +515,8 @@ struct IntrinsicLibrary {
   mlir::Value genLeadz(mlir::Type, llvm::ArrayRef<mlir::Value>);
   fir::ExtendedValue genLen(mlir::Type, llvm::ArrayRef<fir::ExtendedValue>);
   fir::ExtendedValue genLenTrim(mlir::Type, llvm::ArrayRef<fir::ExtendedValue>);
+  template <typename Shift>
+  mlir::Value genMask(mlir::Type, llvm::ArrayRef<mlir::Value>);
   fir::ExtendedValue genMatmul(mlir::Type, llvm::ArrayRef<fir::ExtendedValue>);
   fir::ExtendedValue genMaxloc(mlir::Type, llvm::ArrayRef<fir::ExtendedValue>);
   fir::ExtendedValue genMaxval(mlir::Type, llvm::ArrayRef<fir::ExtendedValue>);
@@ -813,6 +815,8 @@ static constexpr IntrinsicHandler handlers[]{
     {"lgt", &I::genCharacterCompare<mlir::arith::CmpIPredicate::sgt>},
     {"lle", &I::genCharacterCompare<mlir::arith::CmpIPredicate::sle>},
     {"llt", &I::genCharacterCompare<mlir::arith::CmpIPredicate::slt>},
+    {"maskl", &I::genMask<mlir::arith::ShLIOp>},
+    {"maskr", &I::genMask<mlir::arith::ShRUIOp>},
     {"matmul",
      &I::genMatmul,
      {{{"matrix_a", asAddr}, {"matrix_b", asAddr}}},
@@ -3314,6 +3318,27 @@ IntrinsicLibrary::genCharacterCompare(mlir::Type resultType,
       fir::getBase(args[1]), fir::getLen(args[1]));
 }
 
+// MASKL, MASKR
+template <typename Shift>
+mlir::Value IntrinsicLibrary::genMask(mlir::Type resultType,
+                                      llvm::ArrayRef<mlir::Value> args) {
+  assert(args.size() == 2);
+
+  mlir::Value ones = builder.createIntegerConstant(loc, resultType, -1);
+  mlir::Value bitSize = builder.createIntegerConstant(
+      loc, resultType, resultType.getIntOrFloatBitWidth());
+  mlir::Value bitsToSet = builder.createConvert(loc, resultType, args[0]);
+
+  // The standard does not specify what to return if the number of bits to be
+  // set, I < 0 or I >= BIT_SIZE(KIND). The shift instruction used below will
+  // produce a poison value which may return a possibly platform-specific and/or
+  // non-deterministic result. Other compilers don't produce a consistent result
+  // in this case either, so we choose the most efficient implementation.
+  mlir::Value shift =
+      builder.create<mlir::arith::SubIOp>(loc, bitSize, bitsToSet);
+  return builder.create<Shift>(loc, ones, shift);
+}
+
 // MATMUL
 fir::ExtendedValue
 IntrinsicLibrary::genMatmul(mlir::Type resultType,
diff --git a/flang/test/Lower/Intrinsics/maskl.f90 b/flang/test/Lower/Intrinsics/maskl.f90
new file mode 100644
index 0000000..ffcaf53
--- /dev/null
+++ b/flang/test/Lower/Intrinsics/maskl.f90
@@ -0,0 +1,86 @@
+! RUN: bbc -emit-fir %s -o - | FileCheck %s
+! RUN: %flang_fc1 -emit-fir %s -o - | FileCheck %s
+
+! CHECK-LABEL: maskl_test
+! CHECK-SAME: %[[A:.*]]: !fir.ref<i32>{{.*}}, %[[B:.*]]: !fir.ref<i32>{{.*}}
+subroutine maskl_test(a, b)
+  integer :: a
+  integer :: b
+
+  ! CHECK: %[[A_VAL:.*]] = fir.load %[[A]] : !fir.ref<i32>
+  b = maskl(a)
+  ! CHECK: %[[C__1:.*]] = arith.constant -1 : i32
+  ! CHECK: %[[BITS:.*]] = arith.constant 32 : i32
+  ! CHECK: %[[LEN:.*]] = arith.subi %[[BITS]], %[[A_VAL]] : i32
+  ! CHECK: %[[SHIFT:.*]] = arith.shli %[[C__1]], %[[LEN]] : i32
+  ! CHECK: fir.store %[[SHIFT]] to %[[B]] : !fir.ref<i32>
+end subroutine maskl_test
+
+! CHECK-LABEL: maskl1_test
+! CHECK-SAME: %[[A:.*]]: !fir.ref<i32>{{.*}}, %[[B:.*]]: !fir.ref<i8>{{.*}}
+subroutine maskl1_test(a, b)
+  integer :: a
+  integer(kind=1) :: b
+
+  ! CHECK: %[[A_VAL:.*]] = fir.load %[[A]] : !fir.ref<i32>
+  b = maskl(a, 1)
+  ! CHECK: %[[C__1:.*]] = arith.constant -1 : i8
+  ! CHECK: %[[BITS:.*]] = arith.constant 8 : i8
+  ! CHECK: %[[A_CONV:.*]] = fir.convert %[[A_VAL]] : (i32) -> i8
+  ! CHECK: %[[LEN:.*]] = arith.subi %[[BITS]], %[[A_CONV]] : i8
+  ! CHECK: %[[SHIFT:.*]] = arith.shli %[[C__1]], %[[LEN]] : i8
+  ! CHECK: fir.store %[[SHIFT]] to %[[B]] : !fir.ref<i8>
+end subroutine maskl1_test
+
+! CHECK-LABEL: maskl2_test
+! CHECK-SAME: %[[A:.*]]: !fir.ref<i32>{{.*}}, %[[B:.*]]: !fir.ref<i16>{{.*}}
+subroutine maskl2_test(a, b)
+  integer :: a
+  integer(kind=2) :: b
+
+  ! CHECK: %[[A_VAL:.*]] = fir.load %[[A]] : !fir.ref<i32>
+  b = maskl(a, 2)
+  ! CHECK: %[[C__1:.*]] = arith.constant -1 : i16
+  ! CHECK: %[[BITS:.*]] = arith.constant 16 : i16
+  ! CHECK: %[[A_CONV:.*]] = fir.convert %[[A_VAL]] : (i32) -> i16
+  ! CHECK: %[[LEN:.*]] = arith.subi %[[BITS]], %[[A_CONV]] : i16
+  ! CHECK: %[[SHIFT:.*]] = arith.shli %[[C__1]], %[[LEN]] : i16
+  ! CHECK: fir.store %[[SHIFT]] to %[[B]] : !fir.ref<i16>
+end subroutine maskl2_test
+
+! CHECK-LABEL: maskl4_test
+! CHECK-SAME: %[[A:.*]]: !fir.ref<i32>{{.*}}, %[[B:.*]]: !fir.ref<i32>{{.*}}
+subroutine maskl4_test(a, b)
+  integer :: a
+  integer(kind=4) :: b
+
+  ! CHECK: %[[A_VAL:.*]] = fir.load %[[A]] : !fir.ref<i32>
+  b = maskl(a, 4)
+  ! CHECK: %[[C__1:.*]] = arith.constant -1 : i32
+  ! CHECK: %[[BITS:.*]] = arith.constant 32 : i32
+  ! CHECK: %[[LEN:.*]] = arith.subi %[[BITS]], %[[A_VAL]] : i32
+  ! CHECK: %[[SHIFT:.*]] = arith.shli %[[C__1]], %[[LEN]] : i32
+  ! CHECK: fir.store %[[SHIFT]] to %[[B]] : !fir.ref<i32>
+end subroutine maskl4_test
+
+! CHECK-LABEL: maskl8_test
+! CHECK-SAME: %[[A:.*]]: !fir.ref<i32>{{.*}}, %[[B:.*]]: !fir.ref<i64>{{.*}}
+subroutine maskl8_test(a, b)
+  integer :: a
+  integer(kind=8) :: b
+
+  ! CHECK: %[[A_VAL:.*]] = fir.load %[[A]] : !fir.ref<i32>
+  b = maskl(a, 8)
+  ! CHECK: %[[C__1:.*]] = arith.constant -1 : i64
+  ! CHECK: %[[BITS:.*]] = arith.constant 64 : i64
+  ! CHECK: %[[A_CONV:.*]] = fir.convert %[[A_VAL]] : (i32) -> i64
+  ! CHECK: %[[LEN:.*]] = arith.subi %[[BITS]], %[[A_CONV]] : i64
+  ! CHECK: %[[SHIFT:.*]] = arith.shli %[[C__1]], %[[LEN]] : i64
+  ! CHECK: fir.store %[[SHIFT]] to %[[B]] : !fir.ref<i64>
+end subroutine maskl8_test
+
+! TODO: Code containing 128-bit integer literals current breaks. This is
+! probably related to the issue linked below. When that is fixed, a test
+! for kind=16 should be added here.
+!
+! https://github.com/llvm/llvm-project/issues/56446
diff --git a/flang/test/Lower/Intrinsics/maskr.f90 b/flang/test/Lower/Intrinsics/maskr.f90
new file mode 100644
index 0000000..e19890d
--- /dev/null
+++ b/flang/test/Lower/Intrinsics/maskr.f90
@@ -0,0 +1,86 @@
+! RUN: bbc -emit-fir %s -o - | FileCheck %s
+! RUN: %flang_fc1 -emit-fir %s -o - | FileCheck %s
+
+! CHECK-LABEL: maskr_test
+! CHECK-SAME: %[[A:.*]]: !fir.ref<i32>{{.*}}, %[[B:.*]]: !fir.ref<i32>{{.*}}
+subroutine maskr_test(a, b)
+  integer :: a
+  integer :: b
+
+  ! CHECK: %[[A_VAL:.*]] = fir.load %[[A]] : !fir.ref<i32>
+  b = maskr(a)
+  ! CHECK: %[[C__1:.*]] = arith.constant -1 : i32
+  ! CHECK: %[[BITS:.*]] = arith.constant 32 : i32
+  ! CHECK: %[[LEN:.*]] = arith.subi %[[BITS]], %[[A_VAL]] : i32
+  ! CHECK: %[[SHIFT:.*]] = arith.shrui %[[C__1]], %[[LEN]] : i32
+  ! CHECK: fir.store %[[SHIFT]] to %[[B]] : !fir.ref<i32>
+end subroutine maskr_test
+
+! CHECK-LABEL: maskr1_test
+! CHECK-SAME: %[[A:.*]]: !fir.ref<i32>{{.*}}, %[[B:.*]]: !fir.ref<i8>{{.*}}
+subroutine maskr1_test(a, b)
+  integer :: a
+  integer(kind=1) :: b
+
+  ! CHECK: %[[A_VAL:.*]] = fir.load %[[A]] : !fir.ref<i32>
+  b = maskr(a, 1)
+  ! CHECK: %[[C__1:.*]] = arith.constant -1 : i8
+  ! CHECK: %[[BITS:.*]] = arith.constant 8 : i8
+  ! CHECK: %[[A_CONV:.*]] = fir.convert %[[A_VAL]] : (i32) -> i8
+  ! CHECK: %[[LEN:.*]] = arith.subi %[[BITS]], %[[A_CONV]] : i8
+  ! CHECK: %[[SHIFT:.*]] = arith.shrui %[[C__1]], %[[LEN]] : i8
+  ! CHECK: fir.store %[[SHIFT]] to %[[B]] : !fir.ref<i8>
+end subroutine maskr1_test
+
+! CHECK-LABEL: maskr2_test
+! CHECK-SAME: %[[A:.*]]: !fir.ref<i32>{{.*}}, %[[B:.*]]: !fir.ref<i16>{{.*}}
+subroutine maskr2_test(a, b)
+  integer :: a
+  integer(kind=2) :: b
+
+  ! CHECK: %[[A_VAL:.*]] = fir.load %[[A]] : !fir.ref<i32>
+  b = maskr(a, 2)
+  ! CHECK: %[[C__1:.*]] = arith.constant -1 : i16
+  ! CHECK: %[[BITS:.*]] = arith.constant 16 : i16
+  ! CHECK: %[[A_CONV:.*]] = fir.convert %[[A_VAL]] : (i32) -> i16
+  ! CHECK: %[[LEN:.*]] = arith.subi %[[BITS]], %[[A_CONV]] : i16
+  ! CHECK: %[[SHIFT:.*]] = arith.shrui %[[C__1]], %[[LEN]] : i16
+  ! CHECK: fir.store %[[SHIFT]] to %[[B]] : !fir.ref<i16>
+end subroutine maskr2_test
+
+! CHECK-LABEL: maskr4_test
+! CHECK-SAME: %[[A:.*]]: !fir.ref<i32>{{.*}}, %[[B:.*]]: !fir.ref<i32>{{.*}}
+subroutine maskr4_test(a, b)
+  integer :: a
+  integer(kind=4) :: b
+
+  ! CHECK: %[[A_VAL:.*]] = fir.load %[[A]] : !fir.ref<i32>
+  b = maskr(a, 4)
+  ! CHECK: %[[C__1:.*]] = arith.constant -1 : i32
+  ! CHECK: %[[BITS:.*]] = arith.constant 32 : i32
+  ! CHECK: %[[LEN:.*]] = arith.subi %[[BITS]], %[[A_VAL]] : i32
+  ! CHECK: %[[SHIFT:.*]] = arith.shrui %[[C__1]], %[[LEN]] : i32
+  ! CHECK: fir.store %[[SHIFT]] to %[[B]] : !fir.ref<i32>
+end subroutine maskr4_test
+
+! CHECK-LABEL: maskr8_test
+! CHECK-SAME: %[[A:.*]]: !fir.ref<i32>{{.*}}, %[[B:.*]]: !fir.ref<i64>{{.*}}
+subroutine maskr8_test(a, b)
+  integer :: a
+  integer(kind=8) :: b
+
+  ! CHECK: %[[A_VAL:.*]] = fir.load %[[A]] : !fir.ref<i32>
+  b = maskr(a, 8)
+  ! CHECK: %[[C__1:.*]] = arith.constant -1 : i64
+  ! CHECK: %[[BITS:.*]] = arith.constant 64 : i64
+  ! CHECK: %[[A_CONV:.*]] = fir.convert %[[A_VAL]] : (i32) -> i64
+  ! CHECK: %[[LEN:.*]] = arith.subi %[[BITS]], %[[A_CONV]] : i64
+  ! CHECK: %[[SHIFT:.*]] = arith.shrui %[[C__1]], %[[LEN]] : i64
+  ! CHECK: fir.store %[[SHIFT]] to %[[B]] : !fir.ref<i64>
+end subroutine maskr8_test
+
+! TODO: Code containing 128-bit integer literals current breaks. This is
+! probably related to the issue linked below. When that is fixed, a test
+! for kind=16 should be added here.
+!
+! https://github.com/llvm/llvm-project/issues/56446