intel/llvm
1
0
Fork 0
mirror of https://github.com/intel/llvm.git synced 2026-01-15 12:25:46 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Flang] Add -ffast-real-mod back for further control of MOD optimizations (#167118)

Browse Source 
It turns out that having `-ffast-math` as the only option to control
optimizations for MOD for REAL kinds (PR #160660) is too coarse-grained
for some applications. Thus, this PR adds back `-ffast-real-mod` to have
more control over the optimization. The `-ffast-math` flag will still
enable the optimization, and `-fno-fast-real-mod` allows one to disable
it.
This commit is contained in:
Michael Klemm
2025-11-22 12:55:32 +01:00
committed by GitHub
parent 7305b6eb54
commit 9daf4345ec
8 changed files with 127 additions and 60 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
clang/include/clang/Options/Options.td
View File

@@ -2788,6 +2788,9 @@ def fno_unsafe_math_optimizations : Flag<["-"], "fno-unsafe-math-optimizations">
Group<f_Group>;
def fassociative_math : Flag<["-"], "fassociative-math">, Visibility<[ClangOption, FlangOption]>, Group<f_Group>;
def fno_associative_math : Flag<["-"], "fno-associative-math">, Visibility<[ClangOption, FlangOption]>, Group<f_Group>;
def ffast_real_mod : Flag<["-"], "ffast-real-mod">,
Group<f_Group>, Visibility<[FlangOption, FC1Option]>,
HelpText<"Enable optimization of MOD for REAL types">;
def fno_fast_real_mod : Flag<["-"], "fno-fast-real-mod">,
Group<f_Group>, Visibility<[FlangOption, FC1Option]>,
HelpText<"Disable optimization of MOD for REAL types in presence of -ffast-math">;

10
clang/lib/Driver/ToolChains/Flang.cpp
View File

@@ -822,8 +822,14 @@ static void addFloatingPointOptions(const Driver &D, const ArgList &Args,
complexRangeKindToStr(Range)));
}
if (Args.hasArg(options::OPT_fno_fast_real_mod))
CmdArgs.push_back("-fno-fast-real-mod");
if (llvm::opt::Arg *A =
Args.getLastArg(clang::options::OPT_ffast_real_mod,
clang::options::OPT_fno_fast_real_mod)) {
if (A->getOption().matches(clang::options::OPT_ffast_real_mod))
CmdArgs.push_back("-ffast-real-mod");
else if (A->getOption().matches(clang::options::OPT_fno_fast_real_mod))
CmdArgs.push_back("-fno-fast-real-mod");
}
if (!HonorINFs && !HonorNaNs && AssociativeMath && ReciprocalMath &&
ApproxFunc && !SignedZeros &&

2
flang/include/flang/Support/LangOptions.def
View File

@@ -61,7 +61,7 @@ LANGOPT(OpenMPNoNestedParallelism, 1, 0)
/// Use SIMD only OpenMP support.
LANGOPT(OpenMPSimd, 1, false)
/// Enable fast MOD operations for REAL
LANGOPT(NoFastRealMod, 1, false)
LANGOPT(FastRealMod, 1, false)
LANGOPT(VScaleMin, 32, 0) ///< Minimum vscale range value
LANGOPT(VScaleMax, 32, 0) ///< Maximum vscale range value

11
flang/lib/Frontend/CompilerInvocation.cpp
View File

@@ -1403,11 +1403,18 @@ static bool parseFloatingPointArgs(CompilerInvocation &invoc,
opts.ReciprocalMath = true;
opts.ApproxFunc = true;
opts.NoSignedZeros = true;
opts.FastRealMod = true;
opts.setFPContractMode(Fortran::common::LangOptions::FPM_Fast);
}
if (args.hasArg(clang::options::OPT_fno_fast_real_mod))
opts.NoFastRealMod = true;
if (llvm::opt::Arg *arg =
args.getLastArg(clang::options::OPT_ffast_real_mod,
clang::options::OPT_fno_fast_real_mod)) {
if (arg->getOption().matches(clang::options::OPT_ffast_real_mod))
opts.FastRealMod = true;
if (arg->getOption().matches(clang::options::OPT_fno_fast_real_mod))
opts.FastRealMod = false;
}
return true;
}

4
flang/lib/Frontend/FrontendActions.cpp
View File

@@ -277,11 +277,11 @@ bool CodeGenAction::beginSourceFileAction() {
ci.getInvocation().getLangOpts().OpenMPVersion);
}
if (ci.getInvocation().getLangOpts().NoFastRealMod) {
if (ci.getInvocation().getLangOpts().FastRealMod) {
mlir::ModuleOp mod = lb.getModule();
mod.getOperation()->setAttr(
mlir::StringAttr::get(mod.getContext(),
llvm::Twine{"fir.no_fast_real_mod"}),
llvm::Twine{"fir.fast_real_mod"}),
mlir::BoolAttr::get(mod.getContext(), true));
}

10
flang/lib/Optimizer/Builder/IntrinsicCall.cpp
View File

@@ -6526,11 +6526,9 @@ static mlir::Value genFastMod(fir::FirOpBuilder &builder, mlir::Location loc,
mlir::Value IntrinsicLibrary::genMod(mlir::Type resultType,
llvm::ArrayRef<mlir::Value> args) {
auto mod = builder.getModule();
bool dontUseFastRealMod = false;
bool canUseApprox = mlir::arith::bitEnumContainsAny(
builder.getFastMathFlags(), mlir::arith::FastMathFlags::afn);
if (auto attr = mod->getAttrOfType<mlir::BoolAttr>("fir.no_fast_real_mod"))
dontUseFastRealMod = attr.getValue();
bool useFastRealMod = false;
if (auto attr = mod->getAttrOfType<mlir::BoolAttr>("fir.fast_real_mod"))
useFastRealMod = attr.getValue();
assert(args.size() == 2);
if (resultType.isUnsignedInteger()) {
@@ -6543,7 +6541,7 @@ mlir::Value IntrinsicLibrary::genMod(mlir::Type resultType,
if (mlir::isa<mlir::IntegerType>(resultType))
return mlir::arith::RemSIOp::create(builder, loc, args[0], args[1]);
if (resultType.isFloat() && canUseApprox && !dontUseFastRealMod) {
if (resultType.isFloat() && useFastRealMod) {
// Treat MOD as an approximate function and code-gen inline code
// instead of calling into the Fortran runtime library.
return builder.createConvert(loc, resultType,

6
flang/test/Driver/fast-real-mod.f90
View File

@@ -1,6 +1,12 @@
! RUN: %flang -ffast-real-mod -### -c %s 2>&1 | FileCheck %s -check-prefix CHECK-FAST-REAL-MOD
! RUN: %flang -fno-fast-real-mod -### -c %s 2>&1 | FileCheck %s -check-prefix CHECK-NO-FAST-REAL-MOD
! RUN: %flang -fno-fast-real-mod -ffast-real-mod -### -c %s 2>&1 | FileCheck %s -check-prefix CHECK-FAST-REAL-MOD
! RUN: %flang -ffast-real-mod -fno-fast-real-mod -### -c %s 2>&1 | FileCheck %s -check-prefix CHECK-NO-FAST-REAL-MOD
! CHECK-FAST-REAL-MOD: "-ffast-real-mod"
! CHECK-FAST-REAL-MOD-NOT: "-fno-fast-real-mod"
! CHECK-NO-FAST-REAL-MOD: "-fno-fast-real-mod"
! CHECK-NO-FAST-REAL-MOD-NOT: "-fast-real-mod"
program test
! nothing to be done in here

141
flang/test/Lower/Intrinsics/fast-real-mod.f90
View File

@@ -1,24 +1,38 @@
! RUN: %flang_fc1 -ffast-math -emit-mlir -o - %s | FileCheck %s --check-prefixes=CHECK%if target=x86_64{{.*}} %{,CHECK-KIND10%}%if flang-supports-f128-math %{,CHECK-KIND16%}
! RUN: %flang_fc1 -ffast-real-mod -emit-mlir -o - %s | FileCheck %s --check-prefixes=CHECK-FRM%if target=x86_64{{.*}} %{,CHECK-FRM-KIND10%}%if flang-supports-f128-math %{,CHECK-FRM-KIND16%}
! RUN: %flang_fc1 -ffast-real-mod -fno-fast-real-mod -emit-mlir -o - %s | FileCheck %s --check-prefixes=CHECK-NFRM%if target=x86_64{{.*}} %{,CHECK-NFRM-KIND10%}%if flang-supports-f128-math %{,CHECK-NFRM-KIND16%}
! RUN: %flang_fc1 -fno-fast-real-mod -ffast-real-mod -emit-mlir -o - %s | FileCheck %s --check-prefixes=CHECK-FRM%if target=x86_64{{.*}} %{,CHECK-FRM-KIND10%}%if flang-supports-f128-math %{,CHECK-FRM-KIND16%}
! RUN: %flang_fc1 -ffast-math -emit-mlir -o - %s | FileCheck %s --check-prefixes=CHECK-FM%if target=x86_64{{.*}} %{,CHECK-FM-KIND10%}%if flang-supports-f128-math %{,CHECK-FM-KIND16%}
! RUN: %flang_fc1 -ffast-math -fno-fast-real-mod -emit-mlir -o - %s | FileCheck %s --check-prefixes=CHECK-NFRM%if target=x86_64{{.*}} %{,CHECK-NFRM-KIND10%}%if flang-supports-f128-math %{,CHECK-NFRM-KIND16%}
! TODO: check line that fir.fast_real_mod is not there
! CHECK-NFRM: module attributes {{{.*}}fir.no_fast_real_mod = true{{.*}}}
! CHECK-FM: module attributes {{{.*}}fir.fast_real_mod = true{{.*}}}
! CHECK-FRM: module attributes {{{.*}}fir.fast_real_mod = true{{.*}}}
! CHECK-LABEL: @_QPmod_real4
subroutine mod_real4(r, a, p)
implicit none
real(kind=4) :: r, a, p
! CHECK: %[[A:.*]] = fir.declare{{.*}}a"
! CHECK: %[[P:.*]] = fir.declare{{.*}}p"
! CHECK: %[[R:.*]] = fir.declare{{.*}}r"
! CHECK: %[[A_LOAD:.*]] = fir.load %[[A]]
! CHECK: %[[P_LOAD:.*]] = fir.load %[[P]]
! CHECK: %[[DIV:.*]] = arith.divf %[[A_LOAD]], %[[P_LOAD]] fastmath<fast> : f32
! CHECK: %[[CV1:.*]] = fir.convert %[[DIV]] : (f32) -> si32
! CHECK: %[[CV2:.*]] = fir.convert %[[CV1]] : (si32) -> f32
! CHECK: %[[MUL:.*]] = arith.mulf %[[CV2]], %[[P_LOAD]] fastmath<fast> : f32
! CHECK: %[[SUB:.*]] = arith.subf %[[A_LOAD]], %[[MUL]] fastmath<fast> : f32
! CHECK: fir.store %[[SUB]] to %[[R]] : !fir.ref<f32>
! CHECK-FRM: %[[A:.*]] = fir.declare{{.*}}a"
! CHECK-FRM: %[[P:.*]] = fir.declare{{.*}}p"
! CHECK-FRM: %[[R:.*]] = fir.declare{{.*}}r"
! CHECK-FRM: %[[A_LOAD:.*]] = fir.load %[[A]]
! CHECK-FRM: %[[P_LOAD:.*]] = fir.load %[[P]]
! CHECK-FRM: %[[DIV:.*]] = arith.divf %[[A_LOAD]], %[[P_LOAD]] fastmath<contract> : f32
! CHECK-FRM: %[[CV1:.*]] = fir.convert %[[DIV]] : (f32) -> si32
! CHECK-FRM: %[[CV2:.*]] = fir.convert %[[CV1]] : (si32) -> f32
! CHECK-FRM: %[[MUL:.*]] = arith.mulf %[[CV2]], %[[P_LOAD]] fastmath<contract> : f32
! CHECK-FRM: %[[SUB:.*]] = arith.subf %[[A_LOAD]], %[[MUL]] fastmath<contract> : f32
! CHECK-FRM: fir.store %[[SUB]] to %[[R]] : !fir.ref<f32>
! CHECK-FM: %[[A:.*]] = fir.declare{{.*}}a"
! CHECK-FM: %[[P:.*]] = fir.declare{{.*}}p"
! CHECK-FM: %[[R:.*]] = fir.declare{{.*}}r"
! CHECK-FM: %[[A_LOAD:.*]] = fir.load %[[A]]
! CHECK-FM: %[[P_LOAD:.*]] = fir.load %[[P]]
! CHECK-FM: %[[DIV:.*]] = arith.divf %[[A_LOAD]], %[[P_LOAD]] fastmath<fast> : f32
! CHECK-FM: %[[CV1:.*]] = fir.convert %[[DIV]] : (f32) -> si32
! CHECK-FM: %[[CV2:.*]] = fir.convert %[[CV1]] : (si32) -> f32
! CHECK-FM: %[[MUL:.*]] = arith.mulf %[[CV2]], %[[P_LOAD]] fastmath<fast> : f32
! CHECK-FM: %[[SUB:.*]] = arith.subf %[[A_LOAD]], %[[MUL]] fastmath<fast> : f32
! CHECK-FM: fir.store %[[SUB]] to %[[R]] : !fir.ref<f32>
! CHECK-NFRM: fir.call @_FortranAModReal4(%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) {{.*}}: (f32, f32, !fir.ref<i8>, i32) -> f32
r = mod(a, p)
end subroutine mod_real4
@@ -27,17 +41,28 @@ end subroutine mod_real4
subroutine mod_real8(r, a, p)
implicit none
real(kind=8) :: r, a, p
! CHECK: %[[A:.*]] = fir.declare{{.*}}a"
! CHECK: %[[P:.*]] = fir.declare{{.*}}p"
! CHECK: %[[R:.*]] = fir.declare{{.*}}r"
! CHECK: %[[A_LOAD:.*]] = fir.load %[[A]]
! CHECK: %[[P_LOAD:.*]] = fir.load %[[P]]
! CHECK: %[[DIV:.*]] = arith.divf %[[A_LOAD]], %[[P_LOAD]] fastmath<fast> : f64
! CHECK: %[[CV1:.*]] = fir.convert %[[DIV]] : (f64) -> si64
! CHECK: %[[CV2:.*]] = fir.convert %[[CV1]] : (si64) -> f64
! CHECK: %[[MUL:.*]] = arith.mulf %[[CV2]], %[[P_LOAD]] fastmath<fast> : f64
! CHECK: %[[SUB:.*]] = arith.subf %[[A_LOAD]], %[[MUL]] fastmath<fast> : f64
! CHECK: fir.store %[[SUB]] to %[[R]] : !fir.ref<f64>
! CHECK-FRM: %[[A:.*]] = fir.declare{{.*}}a"
! CHECK-FRM: %[[P:.*]] = fir.declare{{.*}}p"
! CHECK-FRM: %[[R:.*]] = fir.declare{{.*}}r"
! CHECK-FRM: %[[A_LOAD:.*]] = fir.load %[[A]]
! CHECK-FRM: %[[P_LOAD:.*]] = fir.load %[[P]]
! CHECK-FRM: %[[DIV:.*]] = arith.divf %[[A_LOAD]], %[[P_LOAD]] fastmath<contract> : f64
! CHECK-FRM: %[[CV1:.*]] = fir.convert %[[DIV]] : (f64) -> si64
! CHECK-FRM: %[[CV2:.*]] = fir.convert %[[CV1]] : (si64) -> f64
! CHECK-FRM: %[[MUL:.*]] = arith.mulf %[[CV2]], %[[P_LOAD]] fastmath<contract> : f64
! CHECK-FRM: %[[SUB:.*]] = arith.subf %[[A_LOAD]], %[[MUL]] fastmath<contract> : f64
! CHECK-FRM: fir.store %[[SUB]] to %[[R]] : !fir.ref<f64>
! CHECK-FM: %[[A:.*]] = fir.declare{{.*}}a"
! CHECK-FM: %[[P:.*]] = fir.declare{{.*}}p"
! CHECK-FM: %[[R:.*]] = fir.declare{{.*}}r"
! CHECK-FM: %[[A_LOAD:.*]] = fir.load %[[A]]
! CHECK-FM: %[[P_LOAD:.*]] = fir.load %[[P]]
! CHECK-FM: %[[DIV:.*]] = arith.divf %[[A_LOAD]], %[[P_LOAD]] fastmath<fast> : f64
! CHECK-FM: %[[CV1:.*]] = fir.convert %[[DIV]] : (f64) -> si64
! CHECK-FM: %[[CV2:.*]] = fir.convert %[[CV1]] : (si64) -> f64
! CHECK-FM: %[[MUL:.*]] = arith.mulf %[[CV2]], %[[P_LOAD]] fastmath<fast> : f64
! CHECK-FM: %[[SUB:.*]] = arith.subf %[[A_LOAD]], %[[MUL]] fastmath<fast> : f64
! CHECK-FM: fir.store %[[SUB]] to %[[R]] : !fir.ref<f64>
! CHECK-NFRM: fir.call @_FortranAModReal8(%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) {{.*}}: (f64, f64, !fir.ref<i8>, i32) -> f64
r = mod(a, p)
end subroutine mod_real8
@@ -47,17 +72,28 @@ subroutine mod_real10(r, a, p)
implicit none
integer, parameter :: kind10 = merge(10, 4, selected_real_kind(p=18).eq.10)
real(kind=kind10) :: r, a, p
! CHECK-KIND10: %[[A:.*]] = fir.declare{{.*}}a"
! CHECK-KIND10: %[[P:.*]] = fir.declare{{.*}}p"
! CHECK-KIND10: %[[R:.*]] = fir.declare{{.*}}r"
! CHECK-KIND10: %[[A_LOAD:.*]] = fir.load %[[A]]
! CHECK-KIND10: %[[P_LOAD:.*]] = fir.load %[[P]]
! CHECK-KIND10: %[[DIV:.*]] = arith.divf %[[A_LOAD]], %[[P_LOAD]] fastmath<fast> : f80
! CHECK-KIND10: %[[CV1:.*]] = fir.convert %[[DIV]] : (f80) -> si80
! CHECK-KIND10: %[[CV2:.*]] = fir.convert %[[CV1]] : (si80) -> f80
! CHECK-KIND10: %[[MUL:.*]] = arith.mulf %[[CV2]], %[[P_LOAD]] fastmath<fast> : f80
! CHECK-KIND10: %[[SUB:.*]] = arith.subf %[[A_LOAD]], %[[MUL]] fastmath<fast> : f80
! CHECK-KIND10: fir.store %[[SUB]] to %[[R]] : !fir.ref<f80>
! CHECK-FRM-KIND10: %[[A:.*]] = fir.declare{{.*}}a"
! CHECK-FRM-KIND10: %[[P:.*]] = fir.declare{{.*}}p"
! CHECK-FRM-KIND10: %[[R:.*]] = fir.declare{{.*}}r"
! CHECK-FRM-KIND10: %[[A_LOAD:.*]] = fir.load %[[A]]
! CHECK-FRM-KIND10: %[[P_LOAD:.*]] = fir.load %[[P]]
! CHECK-FRM-KIND10: %[[DIV:.*]] = arith.divf %[[A_LOAD]], %[[P_LOAD]] fastmath<contract> : f80
! CHECK-FRM-KIND10: %[[CV1:.*]] = fir.convert %[[DIV]] : (f80) -> si80
! CHECK-FRM-KIND10: %[[CV2:.*]] = fir.convert %[[CV1]] : (si80) -> f80
! CHECK-FRM-KIND10: %[[MUL:.*]] = arith.mulf %[[CV2]], %[[P_LOAD]] fastmath<contract> : f80
! CHECK-FRM-KIND10: %[[SUB:.*]] = arith.subf %[[A_LOAD]], %[[MUL]] fastmath<contract> : f80
! CHECK-FRM-KIND10: fir.store %[[SUB]] to %[[R]] : !fir.ref<f80>
! CHECK-FM-KIND10: %[[A:.*]] = fir.declare{{.*}}a"
! CHECK-FM-KIND10: %[[P:.*]] = fir.declare{{.*}}p"
! CHECK-FM-KIND10: %[[R:.*]] = fir.declare{{.*}}r"
! CHECK-FM-KIND10: %[[A_LOAD:.*]] = fir.load %[[A]]
! CHECK-FM-KIND10: %[[P_LOAD:.*]] = fir.load %[[P]]
! CHECK-FM-KIND10: %[[DIV:.*]] = arith.divf %[[A_LOAD]], %[[P_LOAD]] fastmath<fast> : f80
! CHECK-FM-KIND10: %[[CV1:.*]] = fir.convert %[[DIV]] : (f80) -> si80
! CHECK-FM-KIND10: %[[CV2:.*]] = fir.convert %[[CV1]] : (si80) -> f80
! CHECK-FM-KIND10: %[[MUL:.*]] = arith.mulf %[[CV2]], %[[P_LOAD]] fastmath<fast> : f80
! CHECK-FM-KIND10: %[[SUB:.*]] = arith.subf %[[A_LOAD]], %[[MUL]] fastmath<fast> : f80
! CHECK-FM-KIND10: fir.store %[[SUB]] to %[[R]] : !fir.ref<f80>
! CHECK-NFRM-KIND10: fir.call @_FortranAModReal10(%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) {{.*}}: (f80, f80, !fir.ref<i8>, i32) -> f80
r = mod(a, p)
end subroutine mod_real10
@@ -67,17 +103,28 @@ subroutine mod_real16(r, a, p)
implicit none
integer, parameter :: kind16 = merge(16, 4, selected_real_kind(p=33).eq.16)
real(kind=kind16) :: r, a, p
! CHECK-KIND16: %[[A:.*]] = fir.declare{{.*}}a"
! CHECK-KIND16: %[[P:.*]] = fir.declare{{.*}}p"
! CHECK-KIND16: %[[R:.*]] = fir.declare{{.*}}r"
! CHECK-KIND16: %[[A_LOAD:.*]] = fir.load %[[A]]
! CHECK-KIND16: %[[P_LOAD:.*]] = fir.load %[[P]]
! CHECK-KIND16: %[[DIV:.*]] = arith.divf %[[A_LOAD]], %[[P_LOAD]] fastmath<fast> : f128
! CHECK-KIND16: %[[CV1:.*]] = fir.convert %[[DIV]] : (f128) -> si128
! CHECK-KIND16: %[[CV2:.*]] = fir.convert %[[CV1]] : (si128) -> f128
! CHECK-KIND16: %[[MUL:.*]] = arith.mulf %[[CV2]], %[[P_LOAD]] fastmath<fast> : f128
! CHECK-KIND16: %[[SUB:.*]] = arith.subf %[[A_LOAD]], %[[MUL]] fastmath<fast> : f128
! CHECK-KIND16: fir.store %[[SUB]] to %[[R]] : !fir.ref<f128>
! CHECK-FRM-KIND16: %[[A:.*]] = fir.declare{{.*}}a"
! CHECK-FRM-KIND16: %[[P:.*]] = fir.declare{{.*}}p"
! CHECK-FRM-KIND16: %[[R:.*]] = fir.declare{{.*}}r"
! CHECK-FRM-KIND16: %[[A_LOAD:.*]] = fir.load %[[A]]
! CHECK-FRM-KIND16: %[[P_LOAD:.*]] = fir.load %[[P]]
! CHECK-FRM-KIND16: %[[DIV:.*]] = arith.divf %[[A_LOAD]], %[[P_LOAD]] fastmath<contract> : f128
! CHECK-FRM-KIND16: %[[CV1:.*]] = fir.convert %[[DIV]] : (f128) -> si128
! CHECK-FRM-KIND16: %[[CV2:.*]] = fir.convert %[[CV1]] : (si128) -> f128
! CHECK-FRM-KIND16: %[[MUL:.*]] = arith.mulf %[[CV2]], %[[P_LOAD]] fastmath<contract> : f128
! CHECK-FRM-KIND16: %[[SUB:.*]] = arith.subf %[[A_LOAD]], %[[MUL]] fastmath<contract> : f128
! CHECK-FRM-KIND16: fir.store %[[SUB]] to %[[R]] : !fir.ref<f128>
! CHECK-FM-KIND16: %[[A:.*]] = fir.declare{{.*}}a"
! CHECK-FM-KIND16: %[[P:.*]] = fir.declare{{.*}}p"
! CHECK-FM-KIND16: %[[R:.*]] = fir.declare{{.*}}r"
! CHECK-FM-KIND16: %[[A_LOAD:.*]] = fir.load %[[A]]
! CHECK-FM-KIND16: %[[P_LOAD:.*]] = fir.load %[[P]]
! CHECK-FM-KIND16: %[[DIV:.*]] = arith.divf %[[A_LOAD]], %[[P_LOAD]] fastmath<fast> : f128
! CHECK-FM-KIND16: %[[CV1:.*]] = fir.convert %[[DIV]] : (f128) -> si128
! CHECK-FM-KIND16: %[[CV2:.*]] = fir.convert %[[CV1]] : (si128) -> f128
! CHECK-FM-KIND16: %[[MUL:.*]] = arith.mulf %[[CV2]], %[[P_LOAD]] fastmath<fast> : f128
! CHECK-FM-KIND16: %[[SUB:.*]] = arith.subf %[[A_LOAD]], %[[MUL]] fastmath<fast> : f128
! CHECK-FM-KIND16: fir.store %[[SUB]] to %[[R]] : !fir.ref<f128>
! CHECK-NFRM-KIND16: fir.call @_FortranAModReal16(%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) {{.*}}: (f128, f128, !fir.ref<i8>, i32) -> f128
r = mod(a, p)
end subroutine mod_real16