[flang] Support FLUSH as an intrinsic subroutine #165942

Saldivarcher · 2025-10-31T23:59:55Z

Previously FLUSH was only recognized in statement form (e.g. flush(unit)); a
subroutine invocation call flush(unit) was treated as a generic user call with
no special semantics. This change teaches lowering/semantics to handle
CALL FLUSH equivalently.

Fixes #119418

llvmbot · 2025-11-01T00:00:27Z

@llvm/pr-subscribers-flang-fir-hlfir

@llvm/pr-subscribers-flang-semantics

Author: Miguel Saldivar (Saldivarcher)

Changes

Currently FLUSH was only available as a statement, so something like:

program flush_stmt_example
  implicit none

  integer :: unit

  ! Open a file for writing
  open(unit=10, file="flush_stmt.txt", status="replace", action="write")

  write(10,*) "This line is written to the file."
  flush(10)  ! &lt;-- flush statement, not a CALL

  print *, "The file has been flushed. You can read it now before close."

  close(10)
end program flush_stmt_example

If FLUSH was used as subroutine call, like call flush(10) there was no special handling of it. This PR should add that functionality.

Fixes #119418

Full diff: https://github.com/llvm/llvm-project/pull/165942.diff

7 Files Affected:

(modified) flang-rt/lib/runtime/extensions.cpp (+11)
(modified) flang/include/flang/Optimizer/Builder/IntrinsicCall.h (+1)
(modified) flang/include/flang/Optimizer/Builder/Runtime/Intrinsics.h (+2)
(modified) flang/include/flang/Runtime/extensions.h (+1)
(modified) flang/lib/Evaluate/intrinsics.cpp (+4)
(modified) flang/lib/Optimizer/Builder/IntrinsicCall.cpp (+18)
(modified) flang/lib/Optimizer/Builder/Runtime/Intrinsics.cpp (+9)

diff --git a/flang-rt/lib/runtime/extensions.cpp b/flang-rt/lib/runtime/extensions.cpp
index 19e75143705ab..d3a618c1a39ec 100644
--- a/flang-rt/lib/runtime/extensions.cpp
+++ b/flang-rt/lib/runtime/extensions.cpp
@@ -163,6 +163,17 @@ void FORTRAN_PROCEDURE_NAME(flush)(const int &unit) {
   Cookie cookie{IONAME(BeginFlush)(unit, __FILE__, __LINE__)};
   IONAME(EndIoStatement)(cookie);
 }
+
+void RTNAME(Flush)(int unit) {
+  // We set the `unit == -1` on the `flush()` case, so flush all units.
+  if (unit < 0) {
+    Terminator terminator{__FILE__, __LINE__};
+    IoErrorHandler handler{terminator};
+    ExternalFileUnit::FlushAll(handler);
+    return;
+  }
+  FORTRAN_PROCEDURE_NAME(flush)(unit);
+}
 } // namespace io
 
 // CALL FDATE(DATE)
diff --git a/flang/include/flang/Optimizer/Builder/IntrinsicCall.h b/flang/include/flang/Optimizer/Builder/IntrinsicCall.h
index 3407dd01dd504..51a3ae5af7688 100644
--- a/flang/include/flang/Optimizer/Builder/IntrinsicCall.h
+++ b/flang/include/flang/Optimizer/Builder/IntrinsicCall.h
@@ -278,6 +278,7 @@ struct IntrinsicLibrary {
   mlir::Value genExtremum(mlir::Type, llvm::ArrayRef<mlir::Value>);
   void genFenceProxyAsync(llvm::ArrayRef<fir::ExtendedValue>);
   mlir::Value genFloor(mlir::Type, llvm::ArrayRef<mlir::Value>);
+  void genFlush(llvm::ArrayRef<fir::ExtendedValue>);
   mlir::Value genFraction(mlir::Type resultType,
                           mlir::ArrayRef<mlir::Value> args);
   void genFree(mlir::ArrayRef<fir::ExtendedValue> args);
diff --git a/flang/include/flang/Optimizer/Builder/Runtime/Intrinsics.h b/flang/include/flang/Optimizer/Builder/Runtime/Intrinsics.h
index 7a97172cfbb9a..5121ccce921c6 100644
--- a/flang/include/flang/Optimizer/Builder/Runtime/Intrinsics.h
+++ b/flang/include/flang/Optimizer/Builder/Runtime/Intrinsics.h
@@ -51,6 +51,8 @@ mlir::Value genDsecnds(fir::FirOpBuilder &builder, mlir::Location loc,
 void genEtime(fir::FirOpBuilder &builder, mlir::Location loc,
               mlir::Value values, mlir::Value time);
 
+void genFlush(fir::FirOpBuilder &builder, mlir::Location loc, mlir::Value unit);
+
 void genFree(fir::FirOpBuilder &builder, mlir::Location loc, mlir::Value ptr);
 
 mlir::Value genFseek(fir::FirOpBuilder &builder, mlir::Location loc,
diff --git a/flang/include/flang/Runtime/extensions.h b/flang/include/flang/Runtime/extensions.h
index 9fd3e118a0f22..8db68eb9c245c 100644
--- a/flang/include/flang/Runtime/extensions.h
+++ b/flang/include/flang/Runtime/extensions.h
@@ -34,6 +34,7 @@ double RTNAME(Dsecnds)(double *refTime, const char *sourceFile, int line);
 
 // CALL FLUSH(n) antedates the Fortran 2003 FLUSH statement.
 void FORTRAN_PROCEDURE_NAME(flush)(const int &unit);
+void RTNAME(Flush)(int unit);
 
 // GNU extension subroutine FDATE
 void FORTRAN_PROCEDURE_NAME(fdate)(char *string, std::int64_t length);
diff --git a/flang/lib/Evaluate/intrinsics.cpp b/flang/lib/Evaluate/intrinsics.cpp
index 1de5e6b53ba71..d403afe9de307 100644
--- a/flang/lib/Evaluate/intrinsics.cpp
+++ b/flang/lib/Evaluate/intrinsics.cpp
@@ -1597,6 +1597,10 @@ static const IntrinsicInterface intrinsicSubroutine[]{
     {"exit", {{"status", DefaultInt, Rank::scalar, Optionality::optional}}, {},
         Rank::elemental, IntrinsicClass::impureSubroutine},
     {"free", {{"ptr", Addressable}}, {}},
+    {"flush",
+        {{"unit", AnyInt, Rank::scalar, Optionality::optional,
+            common::Intent::In}},
+        {}, Rank::elemental, IntrinsicClass::impureSubroutine},
     {"fseek",
         {{"unit", AnyInt, Rank::scalar}, {"offset", AnyInt, Rank::scalar},
             {"whence", AnyInt, Rank::scalar},
diff --git a/flang/lib/Optimizer/Builder/IntrinsicCall.cpp b/flang/lib/Optimizer/Builder/IntrinsicCall.cpp
index 15ea84565dd75..317414ef0fec6 100644
--- a/flang/lib/Optimizer/Builder/IntrinsicCall.cpp
+++ b/flang/lib/Optimizer/Builder/IntrinsicCall.cpp
@@ -525,6 +525,10 @@ static constexpr IntrinsicHandler handlers[]{
        {"back", asValue, handleDynamicOptional}}},
      /*isElemental=*/false},
     {"floor", &I::genFloor},
+    {"flush",
+     &I::genFlush,
+     {{{"unit", asValue, handleDynamicOptional}}},
+     /*isElemental=*/false},
     {"fraction", &I::genFraction},
     {"free", &I::genFree},
     {"fseek",
@@ -4601,6 +4605,20 @@ mlir::Value IntrinsicLibrary::genFloor(mlir::Type resultType,
   return builder.createConvert(loc, resultType, floor);
 }
 
+// FLUSH
+void IntrinsicLibrary::genFlush(llvm::ArrayRef<fir::ExtendedValue> args) {
+  assert(args.size() == 1);
+
+  mlir::Value unit;
+  if (isStaticallyAbsent(args[0]))
+    // Give a sentinal value of `-1` on the `()` case.
+    unit = builder.createIntegerConstant(loc, builder.getI32Type(), -1);
+  else
+    unit = fir::getBase(args[0]);
+
+  fir::runtime::genFlush(builder, loc, unit);
+}
+
 // FRACTION
 mlir::Value IntrinsicLibrary::genFraction(mlir::Type resultType,
                                           llvm::ArrayRef<mlir::Value> args) {
diff --git a/flang/lib/Optimizer/Builder/Runtime/Intrinsics.cpp b/flang/lib/Optimizer/Builder/Runtime/Intrinsics.cpp
index 110b1b20898c7..9fa3b18a255bd 100644
--- a/flang/lib/Optimizer/Builder/Runtime/Intrinsics.cpp
+++ b/flang/lib/Optimizer/Builder/Runtime/Intrinsics.cpp
@@ -137,6 +137,15 @@ void fir::runtime::genEtime(fir::FirOpBuilder &builder, mlir::Location loc,
   fir::CallOp::create(builder, loc, runtimeFunc, args);
 }
 
+void fir::runtime::genFlush(fir::FirOpBuilder &builder, mlir::Location loc,
+                            mlir::Value unit) {
+  auto runtimeFunc = fir::runtime::getRuntimeFunc<mkRTKey(Flush)>(loc, builder);
+  llvm::SmallVector<mlir::Value> args = fir::runtime::createArguments(
+      builder, loc, runtimeFunc.getFunctionType(), unit);
+
+  fir::CallOp::create(builder, loc, runtimeFunc, args);
+}
+
 void fir::runtime::genFree(fir::FirOpBuilder &builder, mlir::Location loc,
                            mlir::Value ptr) {
   auto runtimeFunc = fir::runtime::getRuntimeFunc<mkRTKey(Free)>(loc, builder);

tblah

Thank you for contributing this.

Please could you add tests showing the correct lowering for flush with an argument, without an argument, and with a dynamically optional argument. There are examples in flang/test/Lower/Intrinsics.

Saldivarcher · 2025-11-11T17:13:40Z

Thank you for contributing this.

Please could you add tests showing the correct lowering for flush with an argument, without an argument, and with a dynamically optional argument. There are examples in flang/test/Lower/Intrinsics.

Test added, thank you!

tblah

What about something dynamically optional like

subroutine flush_unit(unit)
  integer, optional :: unit
  call flush(unit)
end subroutine

Saldivarcher · 2025-11-11T19:58:15Z

What about something dynamically optional like
subroutine flush_unit(unit)
  integer, optional :: unit
  call flush(unit)
end subroutine

flang and gfortran both don't like the optional. I'm guessing it isn't a part of the standard: https://godbolt.org/z/WTnafdK87

tblah · 2025-11-12T11:15:58Z

gfortran doesn't support it. classic-flang (armflang 24.04) does support it.

llvm-flang with your patch does not prevent the usage of a dynamically optional argument. I built your branch and tried it and it generates

  func.func @_QPflush_unit(%arg0: !fir.ref<i32> {fir.bindc_name = "unit", fir.optional}) {
    %0 = fir.dummy_scope : !fir.dscope
    %1:2 = hlfir.declare %arg0 dummy_scope %0 {fortran_attrs = #fir.var_attrs<optional>, uniq_name = "_QFflush_unitEunit"} : (!fir.ref<i32>, !fir.dscope) -> (!fir.ref<i32>, !fir.ref<i32>)
    %2 = fir.is_present %1#0 : (!fir.ref<i32>) -> i1
    %3 = fir.if %2 -> (i32) {
      %4 = fir.load %1#0 : !fir.ref<i32>
      fir.result %4 : i32
    } else {
      %c0_i32 = arith.constant 0 : i32
      fir.result %c0_i32 : i32
    }
    fir.call @_FortranAFlush(%3) fastmath<contract> : (i32) -> ()
    return
  }
  func.func private @_FortranAFlush(i32) attributes {fir.runtime}

From this we can see that whilst the dynamic optional is handled, it defaults to 0. I think you will need -1 for how you defined flush.

I don't mind if you choose to support this or not. If you do support it, you need to make sure the default value is -1 and have a test for the lowering of a dynamically optional argument. If you choose not to support this, please make it clear in the documentation and add a test to make sure that this is correctly diagnosed by the compiler (if you just remove handleDynamicOptional, there could be segfaults at runtime because the program will unconditionally try to load the argument): e.g.

  func.func @_QPflush_unit(%arg0: !fir.ref<i32> {fir.bindc_name = "unit", fir.optional}) {
    %0 = fir.dummy_scope : !fir.dscope
    %1:2 = hlfir.declare %arg0 dummy_scope %0 {fortran_attrs = #fir.var_attrs<optional>, uniq_name = "_QFflush_unitEunit"} : (!fir.ref<i32>, !fir.dscope) -> (!fir.ref<i32>, !fir.ref<i32>)
    %2 = fir.load %1#0 : !fir.ref<i32>
    fir.call @_FortranAFlush(%2) fastmath<contract> : (i32) -> ()
    return
  }
  func.func private @_FortranAFlush(i32) attributes {fir.runtime}

Saldivarcher · 2025-11-17T06:10:03Z

gfortran doesn't support it. classic-flang (armflang 24.04) does support it.

llvm-flang with your patch does not prevent the usage of a dynamically optional argument. I built your branch and tried it and it generates
  func.func @_QPflush_unit(%arg0: !fir.ref<i32> {fir.bindc_name = "unit", fir.optional}) {
    %0 = fir.dummy_scope : !fir.dscope
    %1:2 = hlfir.declare %arg0 dummy_scope %0 {fortran_attrs = #fir.var_attrs<optional>, uniq_name = "_QFflush_unitEunit"} : (!fir.ref<i32>, !fir.dscope) -> (!fir.ref<i32>, !fir.ref<i32>)
    %2 = fir.is_present %1#0 : (!fir.ref<i32>) -> i1
    %3 = fir.if %2 -> (i32) {
      %4 = fir.load %1#0 : !fir.ref<i32>
      fir.result %4 : i32
    } else {
      %c0_i32 = arith.constant 0 : i32
      fir.result %c0_i32 : i32
    }
    fir.call @_FortranAFlush(%3) fastmath<contract> : (i32) -> ()
    return
  }
  func.func private @_FortranAFlush(i32) attributes {fir.runtime}
From this we can see that whilst the dynamic optional is handled, it defaults to 0. I think you will need -1 for how you defined flush.

I don't mind if you choose to support this or not. If you do support it, you need to make sure the default value is -1 and have a test for the lowering of a dynamically optional argument. If you choose not to support this, please make it clear in the documentation and add a test to make sure that this is correctly diagnosed by the compiler (if you just remove handleDynamicOptional, there could be segfaults at runtime because the program will unconditionally try to load the argument): e.g.
  func.func @_QPflush_unit(%arg0: !fir.ref<i32> {fir.bindc_name = "unit", fir.optional}) {
    %0 = fir.dummy_scope : !fir.dscope
    %1:2 = hlfir.declare %arg0 dummy_scope %0 {fortran_attrs = #fir.var_attrs<optional>, uniq_name = "_QFflush_unitEunit"} : (!fir.ref<i32>, !fir.dscope) -> (!fir.ref<i32>, !fir.ref<i32>)
    %2 = fir.load %1#0 : !fir.ref<i32>
    fir.call @_FortranAFlush(%2) fastmath<contract> : (i32) -> ()
    return
  }
  func.func private @_FortranAFlush(i32) attributes {fir.runtime}

Ended up adding support for it! I also added the testcase to the flush.f90 test. Let me know how that looks.

Thanks for the review btw, I appreciate it 🙏

kkwli · 2025-11-17T20:16:06Z

I think we need add this non-standard subroutine in https://github.com/llvm/llvm-project/blob/main/flang/docs/Intrinsics.md.

Previously `FLUSH` was only recognized in statement form (e.g. `flush(unit)`); a subroutine invocation `call flush(unit)` was treated as a generic user call with no special semantics. This change teaches lowering/semantics to handle `CALL FLUSH` equivalently. Fixes llvm#119418

Saldivarcher · 2025-11-17T21:51:38Z

I think we need add this non-standard subroutine in https://github.com/llvm/llvm-project/blob/main/flang/docs/Intrinsics.md.

Done ✔️

github-actions · 2025-11-17T22:05:22Z

🐧 Linux x64 Test Results

4059 tests passed
202 tests skipped

kkwli · 2025-11-18T16:10:37Z

I think we need add this non-standard subroutine in https://github.com/llvm/llvm-project/blob/main/flang/docs/Intrinsics.md.

Done ✔️

Thanks. LG

llvmbot added flang Flang issues not falling into any other category flang:fir-hlfir flang:semantics labels Nov 1, 2025

Saldivarcher force-pushed the pr119418 branch 2 times, most recently from 9d3543e to a40172f Compare November 2, 2025 15:51

Saldivarcher changed the title ~~[flang] Add functionality for calling FLUSH~~ [flang] Support FLUSH as an intrinsic subroutine Nov 2, 2025

tblah reviewed Nov 11, 2025

View reviewed changes

tblah requested a review from klausler November 11, 2025 10:55

Saldivarcher force-pushed the pr119418 branch from a40172f to cecc18b Compare November 11, 2025 17:13

klausler removed their request for review November 11, 2025 17:31

tblah reviewed Nov 11, 2025

View reviewed changes

Saldivarcher force-pushed the pr119418 branch from cecc18b to 9deab44 Compare November 17, 2025 06:08

Saldivarcher force-pushed the pr119418 branch from 9deab44 to a8ed9f5 Compare November 17, 2025 21:48

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[flang] Support FLUSH as an intrinsic subroutine #165942

[flang] Support FLUSH as an intrinsic subroutine #165942

Uh oh!

Saldivarcher commented Oct 31, 2025 •

edited

Loading

Uh oh!

llvmbot commented Nov 1, 2025 •

edited

Loading

Uh oh!

tblah left a comment

Uh oh!

Saldivarcher commented Nov 11, 2025

Uh oh!

tblah left a comment

Uh oh!

Saldivarcher commented Nov 11, 2025

Uh oh!

tblah commented Nov 12, 2025

Uh oh!

Saldivarcher commented Nov 17, 2025

Uh oh!

kkwli commented Nov 17, 2025

Uh oh!

Saldivarcher commented Nov 17, 2025

Uh oh!

github-actions bot commented Nov 17, 2025

Uh oh!

kkwli commented Nov 18, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

[flang] Support FLUSH as an intrinsic subroutine #165942

Are you sure you want to change the base?

[flang] Support FLUSH as an intrinsic subroutine #165942

Uh oh!

Conversation

Saldivarcher commented Oct 31, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Nov 1, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

tblah left a comment

Choose a reason for hiding this comment

Uh oh!

Saldivarcher commented Nov 11, 2025

Uh oh!

tblah left a comment

Choose a reason for hiding this comment

Uh oh!

Saldivarcher commented Nov 11, 2025

Uh oh!

tblah commented Nov 12, 2025

Uh oh!

Saldivarcher commented Nov 17, 2025

Uh oh!

kkwli commented Nov 17, 2025

Uh oh!

Saldivarcher commented Nov 17, 2025

Uh oh!

github-actions bot commented Nov 17, 2025

🐧 Linux x64 Test Results

Uh oh!

kkwli commented Nov 18, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Saldivarcher commented Oct 31, 2025 •

edited

Loading

llvmbot commented Nov 1, 2025 •

edited

Loading