[MDEV-14318] cmake updates to build on arm64 Created: 2017-11-07  Updated: 2018-02-19  Resolved: 2018-02-19

Status: Closed
Project: MariaDB Server
Component/s: Compiling, Storage Engine - InnoDB
Affects Version/s: 10.2
Fix Version/s: 10.2.14, 10.3.5

Type: Bug Priority: Major
Reporter: David Thompson (Inactive) Assignee: Sergey Vojtovich
Resolution: Fixed Votes: 0
Labels: centos7, gcc

Epic Link: arm64 optimization

 Description   

In order to build 10.2 and 10.3 on arm64, additional innodb files need to be built with -O0 when on 4.x gcc compiler, for example on centos7. The following patch includes the necessary fixes to the innobase CMakeLists.txt

index d291f6f..299efe3 100644
 
--- a/storage/innobase/CMakeLists.txt
 
+++ b/storage/innobase/CMakeLists.txt
 
@@ -175,7 +175,10 @@ IF(CMAKE_COMPILER_IS_GNUCXX AND CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64")
 
       btr/btr0btr.cc
 
       btr/btr0cur.cc
 
       buf/buf0buf.cc
 
+      fts/fts0fts.cc
 
       gis/gis0sea.cc
 
+      handler/handler0alter.cc
 
+      row/row0mysql.cc
 
       COMPILE_FLAGS "-O0"
 
       )
 
   ENDIF()



 Comments   
Comment by Sergei Golubchik [ 2017-11-08 ]

What happens if they're not compiled with -O0 ?

Comment by David Thompson (Inactive) [ 2017-11-09 ]

Doesn't compile, so this will block buildbot once we get some additional servers for build / development. The patch simply expands the list of files that are compiled with no optimization for gcc < 5.2. It's a known compiler bug for arm64 and you get something like the following error:

/root/server/storage/innobase/fts/fts0fts.cc: In function 'dberr_t fts_sync_index(fts_sync_t*, fts_index_cache_t*)':
 
/root/server/storage/innobase/fts/fts0fts.cc:4035:1: error: could not split insn
 
 }
 
 ^
 
(insn 627 1909 1426 (parallel [
 
            (set (reg:DI 1 x1 [orig:138 D.112421 ] [138])
 
                (mem/v:DI (reg/f:DI 19 x19 [orig:98 D.112432 ] [98]) [-1  S8 A64]))
 
            (set (mem/v:DI (reg/f:DI 19 x19 [orig:98 D.112432 ] [98]) [-1  S8 A64])
 
                (unspec_volatile:DI [
 
                        (plus:DI (mem/v:DI (reg/f:DI 19 x19 [orig:98 D.112432 ] [98]) [-1  S8 A64])
 
                            (const_int 536870912 [0x20000000]))
 
                        (const_int 5 [0x5])
 
                    ] UNSPECV_ATOMIC_OP))
 
            (clobber (reg:CC 66 cc))
 
            (clobber (reg:DI 2 x2))
 
            (clobber (reg:SI 0 x0))
 
        ]) /root/server/include/my_atomic.h:274 1831 {atomic_fetch_adddi}
 
     (expr_list:REG_UNUSED (reg:CC 66 cc)
 
        (expr_list:REG_UNUSED (reg:DI 2 x2)
 
            (expr_list:REG_UNUSED (reg:SI 0 x0)
 
                (nil)))))
 
/root/server/storage/innobase/fts/fts0fts.cc:4035:1: internal compiler error: in final_scan_insn, at final.c:2897
 
Please submit a full bug report,
 
with preprocessed source if appropriate.
 
See <http://bugzilla.redhat.com/bugzilla> for instructions.

Comment by Daniel Black [ 2017-11-22 ]

dthompson as RHEL supports Arm64 I suggest lodging a bug at http://bugzilla.redhat.com/bugzilla and getting redhat to backport the appropriate compiler fix (it is their business model after all). make VERBOSE=1 and then take the failing compile line and add "-save-temps". Attach the .i* file to the bug report along with the compile line.
From: https://gcc.gnu.org/bugs/

Comment by Sergey Vojtovich [ 2018-02-19 ]

I attempted to fine tune gcc flags:

diff --git a/storage/innobase/CMakeLists.txt b/storage/innobase/CMakeLists.txt
 
index c2d6144..3b3ec37 100644
 
--- a/storage/innobase/CMakeLists.txt
 
+++ b/storage/innobase/CMakeLists.txt
 
@@ -164,19 +164,19 @@ ENDIF()
 
 
 
 INCLUDE(${MYSQL_CMAKE_SCRIPT_DIR}/compile_flags.cmake)
 
 # A GCC bug causes crash when compiling these files on ARM64 with -O1+
 
-# Compile them with -O0 as a workaround.
 
+# See also https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67143
 
 IF(CMAKE_COMPILER_IS_GNUCXX AND CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64")
 
   # Bug was fixed in GCC 5.2, so workaround only needed < 5.2
 
   EXECUTE_PROCESS(COMMAND ${CMAKE_C_COMPILER} -dumpversion
 
                   OUTPUT_VARIABLE GCC_VERSION)
 
   IF(GCC_VERSION VERSION_LESS 5.2)
 
-    ADD_COMPILE_FLAGS(
 
-      btr/btr0btr.cc
 
-      btr/btr0cur.cc
 
-      buf/buf0buf.cc
 
-      gis/gis0sea.cc
 
-      COMPILE_FLAGS "-O0"
 
-      )
 
+    ADD_COMPILE_FLAGS(btr/btr0btr.cc row/row0merge.cc srv/srv0srv.cc COMPILE_FLAGS "-fno-gcse")
 
+    ADD_COMPILE_FLAGS(buf/buf0buf.cc COMPILE_FLAGS "-fno-gcse -fno-cse-follow-jumps")
 
+    ADD_COMPILE_FLAGS(fts/fts0fts.cc COMPILE_FLAGS "-fno-gcse -fno-cse-follow-jumps -fno-ipa-cp-clone")
 
+    ADD_COMPILE_FLAGS(gis/gis0sea.cc COMPILE_FLAGS "-O0")
 
+    ADD_COMPILE_FLAGS(handler/handler0alter.cc COMPILE_FLAGS "-fno-gcse -fno-tree-dominator-opts")
 
+    ADD_COMPILE_FLAGS(mtr/mtr0mtr.cc COMPILE_FLAGS "-fno-inline-functions")
 
+    ADD_COMPILE_FLAGS(row/row0mysql.cc COMPILE_FLAGS "-fno-tree-dominator-opts -fno-tree-ter")
 
   ENDIF()
 
 ENDIF()
 
 IF(MSVC)

But it is quite fragile: every minor change to affected files may require adjusting these flags. Instead I've just added affected files as suggested in description.

Generated at Thu Feb 08 08:12:37 UTC 2024 using Jira 8.20.16#820016-sha1:9d11dbea5f4be3d4cc21f03a88dd11d8c8687422.