[tarantool-patches] [PATCH v2 0/3] Add LTO support

[tarantool-patches] [PATCH v2 0/3] Add LTO support

[Alexander Turenko]

Changes since the last Alex's patch:

* Discarded submodules update in order to use
  `set(CMAKE_POLICY_DEFAULT_CMP0069 NEW)` (deleted the branches).
* Splitted ld / gold version parsing.
* Always give an error when LTO is requested, but cannot be enabled for
  some reason.
* Added more comments to cmake/lto.cmake.
* Suppressed -Wlto-type-mismatch wargings, filed #3742.
* Noted that LuaJIT will not be subject for LTO, filed #3743.
* Rebased on top of fresh 1.10, rewrote CI targets definitions.

https://github.com/tarantool/tarantool/issues/3117
https://github.com/tarantool/tarantool/tree/kh/gh-3117-lto-3
https://travis-ci.org/tarantool/tarantool/builds/441418628

AKhatskevich (3):
  test: prevent guard-breaker optimization with LTO
  Add LTO support
  test: add LTO targets into CI

 .travis.mk         |  4 +-
 .travis.yml        |  8 ++++
 CMakeLists.txt     |  3 ++
 cmake/lto.cmake    | 95 ++++++++++++++++++++++++++++++++++++++++++++++
 test/unit/guard.cc |  6 ++-
 5 files changed, 113 insertions(+), 3 deletions(-)
 create mode 100644 cmake/lto.cmake

-- 
2.19.1

[tarantool-patches] [PATCH v2 1/3] test: prevent guard-breaker optimization with LTO

[Alexander Turenko]

From: AKhatskevich <avkhatskevich@tarantool.org>

In case of very aggressive optimizations the compiler can
optimize guard-breaker function away and the `unit/guard`
test would fail.
---
 test/unit/guard.cc | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/test/unit/guard.cc b/test/unit/guard.cc
index 231b44c7d..3d42fee31 100644
--- a/test/unit/guard.cc
+++ b/test/unit/guard.cc
@@ -13,7 +13,11 @@ static int __attribute__((noinline))
 stack_break_f(char *ptr)
 {
 	char block[2048];
-	char sum = 0;
+	/*
+	 * Make sum volatile to prevent a compiler from
+	 * optimizing away call to this function.
+	 */
+	volatile char sum = 0;
 	memset(block, 0xff, 2048);
 	sum += block[block[4]];
 	ptrdiff_t stack_diff = ptr > block ? ptr - block : block - ptr;
-- 
2.19.1

[tarantool-patches] [PATCH v2 2/3] Add LTO support

[Alexander Turenko]

From: AKhatskevich <avkhatskevich@tarantool.org>

Added -DENABLE_LTO=ON/OFF cmake option, OFF by default.

LTO speeds up cpu-intensive workloads by up to 20% (see [1] and [2]).

Requirements for LTO enabling:

- cmake >= 3.9;
- Linux: ld.bfd / ld.gold from binutils >= 2.31 (or later 2.30)
  (gold >= 1.15);
- Mac OS: xcode >= 8 (earlier versions are not tested).

The requirement of the recent ld version is due to bug with exporting
symbols from dynamic list when LTO is enabled, see [3].

Note: -Wlto-type-mismatch on GCC (enabled by default with -flto) gives
namy warnings. Filed [4] to investigate it.

Note: LuaJIT will be compiled w/o LTO despite the option set, filed [5].

[1]: https://github.com/tarantool/tarantool/wiki/performance-research
[2]: https://gist.github.com/Khatskevich/31a2da6ab46ce903120e7a03d65966db
[3]: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84901
[4]: https://github.com/tarantool/tarantool/issues/3742
[5]: https://github.com/tarantool/tarantool/issues/3743

Closes #3117
---
 CMakeLists.txt  |  3 ++
 cmake/lto.cmake | 95 +++++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 98 insertions(+)
 create mode 100644 cmake/lto.cmake

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 439a2750a..491ccaf0b 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -63,6 +63,9 @@ include(cmake/pod2man.cmake)
 include(cmake/arch.cmake)
 include(cmake/os.cmake)
 include(cmake/compiler.cmake)
+# NO_POLICY_SCOPE is to suppress CMP0069 warnings on the unset
+# policy.
+include(cmake/lto.cmake NO_POLICY_SCOPE)
 include(cmake/simd.cmake)
 include(cmake/atomic.cmake)
 include(cmake/profile.cmake)
diff --git a/cmake/lto.cmake b/cmake/lto.cmake
new file mode 100644
index 000000000..44a6f5771
--- /dev/null
+++ b/cmake/lto.cmake
@@ -0,0 +1,95 @@
+#
+# Manage LTO (Link-Time-Optimization) and IPO
+# (Inter-Procedural-Optimization)
+#
+
+# Tarantool uses both dynamic-list and lto link options, which
+# works only since binutils:
+#
+# - 2.30 for linking with ld.gold (gold version is 1.15);
+# - last 2.30 or 2.31 in case of ld.bfd.
+
+# This cmake module exports CMP0069 policy and should be included
+# with NO_POLICY_SCOPE option.
+
+# The file gives an error if LTO is requested, but cannot be
+# enabled for some reason.
+
+if (NOT DEFINED ENABLE_LTO)
+    set(ENABLE_LTO OFF)
+endif()
+
+# Disable LTO if not requested.
+if (NOT ENABLE_LTO)
+    message(STATUS "Enabling LTO: FALSE")
+    return()
+endif()
+
+if(CMAKE_VERSION VERSION_LESS 3.9)
+    message(FATAL_ERROR "cmake >= 3.9 is needed for LTO enabling")
+endif()
+
+# Set 'CMP0069 NEW' behaviour for this project to support
+# compilers other than Intel Compiler and suppress cmake
+# warnings on the unset policy.
+cmake_policy(SET CMP0069 NEW)
+
+# Retain 'CMP0069 NEW' behaviour after
+# 'cmake_minimum_required(VERSION ...) in subprojects to
+# avoid cmake warnings on the unset policy.
+set(CMAKE_POLICY_DEFAULT_CMP0069 NEW)
+
+# Check whether LTO is supported by the compiler / toolchain and
+# give an error otherwise.
+include(CheckIPOSupported)
+check_ipo_supported(RESULT CMAKE_IPO_AVAILABLE)
+if (NOT CMAKE_IPO_AVAILABLE)
+    message(FATAL_ERROR "LTO is not supported by the compiler / toolchain")
+endif()
+
+# Extra checks on Linux whether all needed LTO features are
+# supported. Mac OS seems to work correctly with xcode >= 8.
+if (NOT TARGET_OS_DARWIN)
+    execute_process(
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        COMMAND ld -v OUTPUT_VARIABLE linker_version_str)
+    message(STATUS "ld version string: ${linker_version_str}")
+
+    # GNU ld (Gentoo 2.31.1 p3) 2.31.1
+    # GNU ld (GNU Binutils for Ubuntu) 2.30
+    # GNU ld version 2.27-10.el7
+    string(REGEX MATCH "^GNU ld.* (2\\.[0-9]+)[^ ]*$" matched_bfd
+        ${linker_version_str})
+
+    # GNU gold (Gentoo 2.31.1 p3 2.31.1) 1.16
+    # GNU gold (GNU Binutils for Ubuntu 2.30) 1.15
+    # GNU gold (version 2.27-10.el7) 1.12
+    if (NOT matched_bfd)
+        string(REGEX MATCH "^GNU gold.* (1\\.[0-9]+)[^ ]*$" matched_gold
+            ${linker_version_str})
+    endif()
+
+    if(matched_bfd)
+        set(linker_version ${CMAKE_MATCH_1})
+        message(STATUS "Found ld.bfd version: ${linker_version}")
+
+        if (linker_version VERSION_LESS "2.31")
+            message(FATAL_ERROR "ld.bfd >= 2.31 is needed for LTO")
+        endif()
+    elseif(matched_gold)
+        set(linker_version ${CMAKE_MATCH_1})
+        message(STATUS "Found ld.gold version: ${linker_version}")
+
+        if (linker_version VERSION_LESS "1.15")
+            message(FATAL_ERROR "ld.gold >= 1.15 is needed for LTO")
+        endif()
+    else()
+        message(FATAL_ERROR "Unsupported ld version format")
+    endif()
+endif()
+
+# gh-3742: investigate LTO warnings.
+set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -Wno-lto-type-mismatch")
+
+set(CMAKE_INTERPROCEDURAL_OPTIMIZATION TRUE)
+message(STATUS "Enabling LTO: TRUE")
-- 
2.19.1

[tarantool-patches] [PATCH v2 3/3] test: add LTO targets into CI

[Alexander Turenko]

From: AKhatskevich <avkhatskevich@tarantool.org>

---
 .travis.mk  | 4 ++--
 .travis.yml | 8 ++++++++
 2 files changed, 10 insertions(+), 2 deletions(-)

diff --git a/.travis.mk b/.travis.mk
index 88335474c..1b7d05149 100644
--- a/.travis.mk
+++ b/.travis.mk
@@ -40,7 +40,7 @@ deps_ubuntu:
 		lcov ruby
 
 test_ubuntu: deps_ubuntu
-	cmake . -DCMAKE_BUILD_TYPE=RelWithDebInfoWError
+	cmake . -DCMAKE_BUILD_TYPE=RelWithDebInfoWError ${CMAKE_EXTRA_PARAMS}
 	make -j8
 	cd test && /usr/bin/python test-run.py -j 1
 
@@ -49,7 +49,7 @@ deps_osx:
 	brew install openssl readline curl icu4c --force
 
 test_osx: deps_osx
-	cmake . -DCMAKE_BUILD_TYPE=RelWithDebInfoWError
+	cmake . -DCMAKE_BUILD_TYPE=RelWithDebInfoWError ${CMAKE_EXTRA_PARAMS}
 	# Increase the maximum number of open file descriptors on macOS
 	sudo sysctl -w kern.maxfiles=20480 || :
 	sudo sysctl -w kern.maxfilesperproc=20480 || :
diff --git a/.travis.yml b/.travis.yml
index a9d1036da..dbd8c91c0 100644
--- a/.travis.yml
+++ b/.travis.yml
@@ -33,6 +33,14 @@ matrix:
         os: osx
       - name: Debug build + test + coverage (Linux, gcc)
         env: TARGET=coverage
+      - name: LTO build + test (Linux, gcc)
+        env: TARGET=test CMAKE_EXTRA_PARAMS=-DENABLE_LTO=ON
+      - name: LTO build + test (Linux, clang)
+        env: TARGET=test CMAKE_EXTRA_PARAMS=-DENABLE_LTO=ON
+        compiler: clang
+      - name: LTO build + test (OS X)
+        os: osx
+        env: TARGET=test CMAKE_EXTRA_PARAMS=-DENABLE_LTO=ON
       - name: CentOS 6 build + test + deploy RPM
         env: OS=el DIST=6
       - name: CentOS 7 build + test + deploy RPM
-- 
2.19.1

[tarantool-patches] Re: [PATCH v2 2/3] Add LTO support

[Konstantin Osipov]

* Alexander Turenko <alexander.turenko@tarantool.org> [18/10/15 10:15]:
> From: AKhatskevich <avkhatskevich@tarantool.org>
> 
> Added -DENABLE_LTO=ON/OFF cmake option, OFF by default.
> 
> LTO speeds up cpu-intensive workloads by up to 20% (see [1] and [2]).
> 
> Requirements for LTO enabling:

to enable LTO, not "for LTO enabling".

The same applies to cmake message:
> +if(CMAKE_VERSION VERSION_LESS 3.9)
> +    message(FATAL_ERROR "cmake >= 3.9 is needed for LTO enabling")
> +endif()


> +
> +# Set 'CMP0069 NEW' behaviour for this project to support
> +# compilers other than Intel Compiler and suppress cmake
> +# warnings on the unset policy.
> +cmake_policy(SET CMP0069 NEW)

Please describe what this policy is about, not only just what it
is for.

> +
> +# Retain 'CMP0069 NEW' behaviour after
> +# 'cmake_minimum_required(VERSION ...) in subprojects to
> +# avoid cmake warnings on the unset policy.
> +set(CMAKE_POLICY_DEFAULT_CMP0069 NEW)


-- 
Konstantin Osipov, Moscow, Russia, +7 903 626 22 32
http://tarantool.io - www.twitter.com/kostja_osipov

[tarantool-patches] Re: [PATCH v2 2/3] Add LTO support

[Alexander Turenko]

On Tue, Oct 16, 2018 at 09:09:02PM +0300, Konstantin Osipov wrote:
> * Alexander Turenko <alexander.turenko@tarantool.org> [18/10/15 10:15]:
> > From: AKhatskevich <avkhatskevich@tarantool.org>
> > 
> > Added -DENABLE_LTO=ON/OFF cmake option, OFF by default.
> > 
> > LTO speeds up cpu-intensive workloads by up to 20% (see [1] and [2]).
> > 
> > Requirements for LTO enabling:
> 
> to enable LTO, not "for LTO enabling".
> 
> The same applies to cmake message:
> > +if(CMAKE_VERSION VERSION_LESS 3.9)
> > +    message(FATAL_ERROR "cmake >= 3.9 is needed for LTO enabling")
> > +endif()
> 

Fixed.

> 
> > +
> > +# Set 'CMP0069 NEW' behaviour for this project to support
> > +# compilers other than Intel Compiler and suppress cmake
> > +# warnings on the unset policy.
> > +cmake_policy(SET CMP0069 NEW)
> 
> Please describe what this policy is about, not only just what it
> is for.
> 

-# Set 'CMP0069 NEW' behaviour for this project to support
-# compilers other than Intel Compiler and suppress cmake
-# warnings on the unset policy.
+# 'CMP0069 NEW' behaviour enables LTO for compilers other then
+# Intel Compiler when CMAKE_INTERPROCEDURAL_OPTIMIZATION is
+# enabled and provides CheckIPOSupported module. We set the policy
+# to support LTO with GCC / Clang and to suppress cmake warnings
+# on the unset policy.

> > +
> > +# Retain 'CMP0069 NEW' behaviour after
> > +# 'cmake_minimum_required(VERSION ...) in subprojects to
> > +# avoid cmake warnings on the unset policy.
> > +set(CMAKE_POLICY_DEFAULT_CMP0069 NEW)
> 
> 
> -- 
> Konstantin Osipov, Moscow, Russia, +7 903 626 22 32
> http://tarantool.io - www.twitter.com/kostja_osipov
> 

[tarantool-patches] Re: [PATCH v2 3/3] test: add LTO targets into CI

[Georgy Kirichenko]

[-- Attachment #1: Type: text/plain, Size: 1788 bytes --]

Seems Ok

On Monday, October 15, 2018 2:17:39 AM MSK Alexander Turenko wrote:
> From: AKhatskevich <avkhatskevich@tarantool.org>
> 
> ---
>  .travis.mk  | 4 ++--
>  .travis.yml | 8 ++++++++
>  2 files changed, 10 insertions(+), 2 deletions(-)
> 
> diff --git a/.travis.mk b/.travis.mk
> index 88335474c..1b7d05149 100644
> --- a/.travis.mk
> +++ b/.travis.mk
> @@ -40,7 +40,7 @@ deps_ubuntu:
>  		lcov ruby
> 
>  test_ubuntu: deps_ubuntu
> -	cmake . -DCMAKE_BUILD_TYPE=RelWithDebInfoWError
> +	cmake . -DCMAKE_BUILD_TYPE=RelWithDebInfoWError ${CMAKE_EXTRA_PARAMS}
>  	make -j8
>  	cd test && /usr/bin/python test-run.py -j 1
> 
> @@ -49,7 +49,7 @@ deps_osx:
>  	brew install openssl readline curl icu4c --force
> 
>  test_osx: deps_osx
> -	cmake . -DCMAKE_BUILD_TYPE=RelWithDebInfoWError
> +	cmake . -DCMAKE_BUILD_TYPE=RelWithDebInfoWError ${CMAKE_EXTRA_PARAMS}
>  	# Increase the maximum number of open file descriptors on macOS
>  	sudo sysctl -w kern.maxfiles=20480 || :
>  	sudo sysctl -w kern.maxfilesperproc=20480 || :
> diff --git a/.travis.yml b/.travis.yml
> index a9d1036da..dbd8c91c0 100644
> --- a/.travis.yml
> +++ b/.travis.yml
> @@ -33,6 +33,14 @@ matrix:
>          os: osx
>        - name: Debug build + test + coverage (Linux, gcc)
>          env: TARGET=coverage
> +      - name: LTO build + test (Linux, gcc)
> +        env: TARGET=test CMAKE_EXTRA_PARAMS=-DENABLE_LTO=ON
> +      - name: LTO build + test (Linux, clang)
> +        env: TARGET=test CMAKE_EXTRA_PARAMS=-DENABLE_LTO=ON
> +        compiler: clang
> +      - name: LTO build + test (OS X)
> +        os: osx
> +        env: TARGET=test CMAKE_EXTRA_PARAMS=-DENABLE_LTO=ON
>        - name: CentOS 6 build + test + deploy RPM
>          env: OS=el DIST=6
>        - name: CentOS 7 build + test + deploy RPM


[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

[tarantool-patches] Re: [PATCH v2 2/3] Add LTO support

[Georgy Kirichenko]

[-- Attachment #1: Type: text/plain, Size: 1677 bytes --]

Seems OK
On Wednesday, October 17, 2018 10:13:25 AM MSK Alexander Turenko wrote:
> On Tue, Oct 16, 2018 at 09:09:02PM +0300, Konstantin Osipov wrote:
> > * Alexander Turenko <alexander.turenko@tarantool.org> [18/10/15 10:15]:
> > > From: AKhatskevich <avkhatskevich@tarantool.org>
> > > 
> > > Added -DENABLE_LTO=ON/OFF cmake option, OFF by default.
> > > 
> > > LTO speeds up cpu-intensive workloads by up to 20% (see [1] and [2]).
> > 
> > > Requirements for LTO enabling:
> > to enable LTO, not "for LTO enabling".
> > 
> > The same applies to cmake message:
> > > +if(CMAKE_VERSION VERSION_LESS 3.9)
> > > +    message(FATAL_ERROR "cmake >= 3.9 is needed for LTO enabling")
> > > +endif()
> 
> Fixed.
> 
> > > +
> > > +# Set 'CMP0069 NEW' behaviour for this project to support
> > > +# compilers other than Intel Compiler and suppress cmake
> > > +# warnings on the unset policy.
> > > +cmake_policy(SET CMP0069 NEW)
> > 
> > Please describe what this policy is about, not only just what it
> > is for.
> 
> -# Set 'CMP0069 NEW' behaviour for this project to support
> -# compilers other than Intel Compiler and suppress cmake
> -# warnings on the unset policy.
> +# 'CMP0069 NEW' behaviour enables LTO for compilers other then
> +# Intel Compiler when CMAKE_INTERPROCEDURAL_OPTIMIZATION is
> +# enabled and provides CheckIPOSupported module. We set the policy
> +# to support LTO with GCC / Clang and to suppress cmake warnings
> +# on the unset policy.
> 
> > > +
> > > +# Retain 'CMP0069 NEW' behaviour after
> > > +# 'cmake_minimum_required(VERSION ...) in subprojects to
> > > +# avoid cmake warnings on the unset policy.
> > > +set(CMAKE_POLICY_DEFAULT_CMP0069 NEW)


[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

[tarantool-patches] Re: [PATCH v2 0/3] Add LTO support

[Kirill Yukhin]

Hello,
On 15 Oct 02:17, Alexander Turenko wrote:
> Changes since the last Alex's patch:
> 
> * Discarded submodules update in order to use
>   `set(CMAKE_POLICY_DEFAULT_CMP0069 NEW)` (deleted the branches).
> * Splitted ld / gold version parsing.
> * Always give an error when LTO is requested, but cannot be enabled for
>   some reason.
> * Added more comments to cmake/lto.cmake.
> * Suppressed -Wlto-type-mismatch wargings, filed #3742.
> * Noted that LuaJIT will not be subject for LTO, filed #3743.
> * Rebased on top of fresh 1.10, rewrote CI targets definitions.
> 
> https://github.com/tarantool/tarantool/issues/3117
> https://github.com/tarantool/tarantool/tree/kh/gh-3117-lto-3
> https://travis-ci.org/tarantool/tarantool/builds/441418628
I've checked the patchset into 2.1 branch.

--
Regards, Kirill Yukhin