[flang] Use precompiled headers in Frontend, Lower, Parser, Semantics and Evaluate #131137

mrkajetanp · 2025-03-13T13:11:13Z

Precompiling larger headers can save a lot of compile time across various compilation units.

Selected compile time & memory improvements are as follows:

flang/lib/Parser/Fortran-parsers.cpp:
Elapsed (wall clock) time (h:mm:ss or m:ss): 0:47.31 -> 0:41.68
Maximum resident set size (kbytes): 2062140 -> 1745584

flang/lib/Lower/Bridge.cpp:
Elapsed (wall clock) time (h:mm:ss or m:ss): 1:19.16 -> 0:45.86
Maximum resident set size (kbytes): 3849144 -> 2443476

flang/lib/Lower/PFTBuilder.cpp
Elapsed (wall clock) time (h:mm:ss or m:ss): 1:29.24 -> 1:00.99
Maximum resident set size (kbytes): 4218368 -> 2923128

flang/lib/Lower/Allocatable.cpp
Elapsed (wall clock) time (h:mm:ss or m:ss): 0:53.03 -> 0:22.50
Maximum resident set size (kbytes): 3092840 -> 2116908

flang/lib/Semantics/Semantics.cpp
Elapsed (wall clock) time (h:mm:ss or m:ss): 1:18.75 -> 1:00.20
Maximum resident set size (kbytes): 3527744 -> 2545308

While the newly added precompiled headers are as follows:

Parser:
Elapsed (wall clock) time (h:mm:ss or m:ss): 0:09.62
Maximum resident set size (kbytes): 1034608

Lower:
Elapsed (wall clock) time (h:mm:ss or m:ss): 0:41.33
Maximum resident set size (kbytes): 3615240

Semantics:
Elapsed (wall clock) time (h:mm:ss or m:ss): 0:26.69
Maximum resident set size (kbytes): 2403776

llvmbot · 2025-03-13T13:11:49Z

@llvm/pr-subscribers-flang-driver
@llvm/pr-subscribers-flang-semantics

@llvm/pr-subscribers-flang-parser

Author: Kajetan Puchalski (mrkajetanp)

Changes

Precompiling larger headers can save a lot of compile time across various compilation units.

Selected compile time & memory improvements are as follows:

flang/lib/Parser/Fortran-parsers.cpp:
Elapsed (wall clock) time (h:mm:ss or m:ss): 0:47.31 -> 0:41.68 Maximum resident set size (kbytes): 2062140 -> 1745584

flang/lib/Lower/Bridge.cpp:
Elapsed (wall clock) time (h:mm:ss or m:ss): 1:19.16 -> 0:45.86 Maximum resident set size (kbytes): 3849144 -> 2443476

flang/lib/Lower/PFTBuilder.cpp
Elapsed (wall clock) time (h:mm:ss or m:ss): 1:29.24 -> 1:00.99 Maximum resident set size (kbytes): 4218368 -> 2923128

flang/lib/Lower/Allocatable.cpp
Elapsed (wall clock) time (h:mm:ss or m:ss): 0:53.03 -> 0:22.50 Maximum resident set size (kbytes): 3092840 -> 2116908

flang/lib/Semantics/Semantics.cpp
Elapsed (wall clock) time (h:mm:ss or m:ss): 1:18.75 -> 1:00.20 Maximum resident set size (kbytes): 3527744 -> 2545308

While the newly added precompiled headers are as follows:

Parser:
Elapsed (wall clock) time (h:mm:ss or m:ss): 0:09.62 Maximum resident set size (kbytes): 1034608

Lower:
Elapsed (wall clock) time (h:mm:ss or m:ss): 0:41.33 Maximum resident set size (kbytes): 3615240

Semantics:
Elapsed (wall clock) time (h:mm:ss or m:ss): 0:26.69 Maximum resident set size (kbytes): 2403776

Full diff: https://github.com/llvm/llvm-project/pull/131137.diff

3 Files Affected:

(modified) flang/lib/Lower/CMakeLists.txt (+11)
(modified) flang/lib/Parser/CMakeLists.txt (+8)
(modified) flang/lib/Semantics/CMakeLists.txt (+9)

diff --git a/flang/lib/Lower/CMakeLists.txt b/flang/lib/Lower/CMakeLists.txt
index 0bd9a47cd040f..bc817ff8f1f3e 100644
--- a/flang/lib/Lower/CMakeLists.txt
+++ b/flang/lib/Lower/CMakeLists.txt
@@ -73,3 +73,14 @@ add_flang_library(FortranLower
   MLIRLLVMDialect
   MLIRSCFToControlFlow
 )
+
+target_precompile_headers(FortranLower PRIVATE
+  [["flang/Lower/ConvertExpr.h"]]
+  [["flang/Lower/SymbolMap.h"]]
+  [["flang/Lower/AbstractConverter.h"]]
+  [["flang/Lower/IterationSpace.h"]]
+  [["flang/Lower/CallInterface.h"]]
+  [["flang/Lower/BoxAnalyzer.h"]]
+  [["flang/Lower/PFTBuilder.h"]]
+  [["flang/Lower/DirectivesCommon.h"]]
+)
diff --git a/flang/lib/Parser/CMakeLists.txt b/flang/lib/Parser/CMakeLists.txt
index 76fe3d7ce6ba4..1855b8a841ba7 100644
--- a/flang/lib/Parser/CMakeLists.txt
+++ b/flang/lib/Parser/CMakeLists.txt
@@ -36,3 +36,11 @@ add_flang_library(FortranParser
   omp_gen
   acc_gen
 )
+
+target_precompile_headers(FortranParser PRIVATE
+  [["flang/Parser/parsing.h"]]
+  [["flang/Parser/parse-tree.h"]]
+  [["flang/Parser/provenance.h"]]
+  [["flang/Parser/message.h"]]
+  [["flang/Parser/parse-tree-visitor.h"]]
+)
diff --git a/flang/lib/Semantics/CMakeLists.txt b/flang/lib/Semantics/CMakeLists.txt
index 93bf0c7c5facd..bd8cc47365f06 100644
--- a/flang/lib/Semantics/CMakeLists.txt
+++ b/flang/lib/Semantics/CMakeLists.txt
@@ -64,3 +64,12 @@ add_flang_library(FortranSemantics
   FrontendOpenACC
   TargetParser
 )
+
+target_precompile_headers(FortranSemantics PRIVATE
+  [["flang/Semantics/semantics.h"]]
+  [["flang/Semantics/type.h"]]
+  [["flang/Semantics/openmp-modifiers.h"]]
+  [["flang/Semantics/expression.h"]]
+  [["flang/Semantics/tools.h"]]
+  [["flang/Semantics/symbol.h"]]
+)

jeanPerier · 2025-03-13T15:06:21Z

Lowering depends on Semantics and the Parser, and Semantics depends on Parser, have you tried using REUSE_FROM https://cmake.org/cmake/help/latest/command/target_precompile_headers.html#reusing-precompile-headers so that lowering can reuse the precompiled headers too?

Just curious, I am not familiar with that cmake feature at all.

mrkajetanp · 2025-03-13T15:14:59Z

Lowering depends on Semantics and the Parser, and Semantics depends on Parser, have you tried using REUSE_FROM

We shortly discussed this in a subthread under #130600.
In short, REUSE_FROM is an all-or-nothing kind of situation. With the feature it's not possible to reuse a different precompiled header unit and also add extra ones at the same time, you either re-use an existing one or have to build a new one completely from scratch.

Since the main goal is to reduce the memory usage per compilation thread, I didn't want to put too much into the precompiled headers either. If I tried adding the same Semantics headers on top of the Lower ones then the precompiled header for Lower/ would become probably the biggest TU in all of flang.

klausler · 2025-03-13T15:42:16Z

Why not Evaluate? It might benefit the most of all of these directories.

mrkajetanp · 2025-03-13T15:46:07Z

Why not Evaluate? It might benefit the most of all of these directories.

Thanks for the suggestion! No reason not to indeed, I'll add it.

jeanPerier · 2025-03-13T16:31:26Z

What is the impact on the size of the build directories? (I am asking because this blog post is pointing out that the build directory size can increase).

mrkajetanp · 2025-03-13T16:39:43Z

What is the impact on the size of the build directories?

From du -h:
440M lib/Lower/CMakeFiles/FortranLower.dir/cmake_pch.hxx.pch
122M lib/Parser/CMakeFiles/FortranParser.dir/cmake_pch.hxx.pch
278M lib/Semantics/CMakeFiles/FortranSemantics.dir/cmake_pch.hxx.pch
164M lib/Evaluate/CMakeFiles/FortranEvaluate.dir/cmake_pch.hxx.pch

But then at the same time I think some space might be saved in the specific compilation units that now no longer need to include them, so it might not be the biggest deal.

Meinersbur · 2025-03-14T20:53:19Z

Consider that this does not work together with ccache: #130600 (comment)

mgorny

Thanks for doing this! I can confirm, that (combined with #131397) I can build Flang with -j6 — and basically it'd go with -j12 if not for a few libraries.

DavidTruby

LGTM thanks!

Reverts #130600 Reverting on account of Windows issues with ccache, will bring it back along with #131137 once those are resolved.

mrkajetanp · 2025-04-08T14:43:04Z

I reverted #130600 as it was causing some buildbot issues in the absence of the ccache fixes.
I've now added the changes in that PR here and rebased this so all the PCH can be merged together once the ccache issues are resolved.

Precompiling larger headers can save a lot of compile time across various compilation units. Selected compile time & memory improvements are as follows: flang/lib/Parser/Fortran-parsers.cpp: Elapsed (wall clock) time (h:mm:ss or m:ss): 0:47.31 -> 0:41.68 Maximum resident set size (kbytes): 2062140 -> 1745584 flang/lib/Lower/Bridge.cpp: Elapsed (wall clock) time (h:mm:ss or m:ss): 1:19.16 -> 0:45.86 Maximum resident set size (kbytes): 3849144 -> 2443476 flang/lib/Lower/PFTBuilder.cpp Elapsed (wall clock) time (h:mm:ss or m:ss): 1:29.24 -> 1:00.99 Maximum resident set size (kbytes): 4218368 -> 2923128 flang/lib/Lower/Allocatable.cpp Elapsed (wall clock) time (h:mm:ss or m:ss): 0:53.03 -> 0:22.50 Maximum resident set size (kbytes): 3092840 -> 2116908 flang/lib/Semantics/Semantics.cpp Elapsed (wall clock) time (h:mm:ss or m:ss): 1:18.75 -> 1:00.20 Maximum resident set size (kbytes): 3527744 -> 2545308 While the newly added precompiled headers are as follows: Parser: Elapsed (wall clock) time (h:mm:ss or m:ss): 0:09.62 Maximum resident set size (kbytes): 1034608 Lower: Elapsed (wall clock) time (h:mm:ss or m:ss): 0:41.33 Maximum resident set size (kbytes): 3615240 Semantics: Elapsed (wall clock) time (h:mm:ss or m:ss): 0:26.69 Maximum resident set size (kbytes): 2403776 Signed-off-by: Kajetan Puchalski <kajetan.puchalski@arm.com>

flang/lib/Evaluate/check-expression.cpp: Elapsed (wall clock) time (h:mm:ss or m:ss): 1:11.91 -> 1:02.29 Maximum resident set size (kbytes): 2710788 -> 2414740 Similar improvements for other compilation units under Evaluate. cmake_pch.hxx.cxx compilation time: Elapsed (wall clock) time (h:mm:ss or m:ss): 0:13.93 Maximum resident set size (kbytes): 1492744 Signed-off-by: Kajetan Puchalski <kajetan.puchalski@arm.com>

Reland of llvm#130600 which was reverted on account of waiting for required ccache compatibility fixes. Signed-off-by: Kajetan Puchalski <kajetan.puchalski@arm.com>

Signed-off-by: Kajetan Puchalski <kajetan.puchalski@arm.com>

mrkajetanp · 2025-04-25T09:36:23Z

With #136856 merged there is now proper ccache support for this on Linux. I added a line here to disable pch on Windows for the time being, meaning this should now be ready to go.

llvmbot added flang Flang issues not falling into any other category flang:fir-hlfir flang:semantics flang:parser labels Mar 13, 2025

tblah requested review from DavidTruby, kiranchandramohan and kparzysz March 13, 2025 15:06

tblah requested review from mgorny and jeanPerier March 13, 2025 15:06

jeanPerier requested review from sscalpone and klausler March 13, 2025 15:20

mrkajetanp changed the title ~~[flang] Use precompiled headers in Lower, Parser and Semantics~~ [flang] Use precompiled headers in Lower, Parser, Semantics and Evaluate Mar 18, 2025

mgorny approved these changes Apr 6, 2025

View reviewed changes

DavidTruby approved these changes Apr 7, 2025

View reviewed changes

mrkajetanp mentioned this pull request Apr 8, 2025

Revert "[flang] Use precompiled parsing headers" #134851

Merged

mrkajetanp added a commit that referenced this pull request Apr 8, 2025

Revert "[flang] Use precompiled parsing headers" (#134851)

7e1b76c

Reverts #130600 Reverting on account of Windows issues with ccache, will bring it back along with #131137 once those are resolved.

mrkajetanp force-pushed the precompiled-headers-2 branch from ae83e1f to 5c706ec Compare April 8, 2025 14:40

llvmbot added the flang:driver label Apr 8, 2025

mrkajetanp changed the title ~~[flang] Use precompiled headers in Lower, Parser, Semantics and Evaluate~~ [flang] Use precompiled headers in Frontend, Lower, Parser, Semantics and Evaluate Apr 8, 2025

mrkajetanp added 2 commits April 24, 2025 16:51

[flang] Use precompiled headers in Frontend

cd8c8cc

Reland of llvm#130600 which was reverted on account of waiting for required ccache compatibility fixes. Signed-off-by: Kajetan Puchalski <kajetan.puchalski@arm.com>

[CMake] Disable precompiled headers for ccache builds on Windows

6edd33b

Signed-off-by: Kajetan Puchalski <kajetan.puchalski@arm.com>

mrkajetanp force-pushed the precompiled-headers-2 branch from 5c706ec to 6edd33b Compare April 24, 2025 20:00

mrkajetanp requested review from Meinersbur and petrhosek April 25, 2025 09:37

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[flang] Use precompiled headers in Frontend, Lower, Parser, Semantics and Evaluate #131137

[flang] Use precompiled headers in Frontend, Lower, Parser, Semantics and Evaluate #131137

mrkajetanp commented Mar 13, 2025 •

edited

Loading

llvmbot commented Mar 13, 2025 •

edited

Loading

jeanPerier commented Mar 13, 2025

mrkajetanp commented Mar 13, 2025 •

edited

Loading

klausler commented Mar 13, 2025

mrkajetanp commented Mar 13, 2025

jeanPerier commented Mar 13, 2025

mrkajetanp commented Mar 13, 2025

Meinersbur commented Mar 14, 2025 •

edited

Loading

mgorny left a comment

DavidTruby left a comment

mrkajetanp commented Apr 8, 2025

mrkajetanp commented Apr 25, 2025

[flang] Use precompiled headers in Frontend, Lower, Parser, Semantics and Evaluate #131137

Are you sure you want to change the base?

[flang] Use precompiled headers in Frontend, Lower, Parser, Semantics and Evaluate #131137

Conversation

mrkajetanp commented Mar 13, 2025 • edited Loading

llvmbot commented Mar 13, 2025 • edited Loading

jeanPerier commented Mar 13, 2025

mrkajetanp commented Mar 13, 2025 • edited Loading

klausler commented Mar 13, 2025

mrkajetanp commented Mar 13, 2025

jeanPerier commented Mar 13, 2025

mrkajetanp commented Mar 13, 2025

Meinersbur commented Mar 14, 2025 • edited Loading

mgorny left a comment

Choose a reason for hiding this comment

DavidTruby left a comment

Choose a reason for hiding this comment

mrkajetanp commented Apr 8, 2025

mrkajetanp commented Apr 25, 2025

mrkajetanp commented Mar 13, 2025 •

edited

Loading

llvmbot commented Mar 13, 2025 •

edited

Loading

mrkajetanp commented Mar 13, 2025 •

edited

Loading

Meinersbur commented Mar 14, 2025 •

edited

Loading