MySQL Bugs: #20803: ICC 9.1.039 produces bad code compiling "mysql_client

Bug #20803	ICC 9.1.039 produces bad code compiling "mysql_client_test.c"
Submitted:	1 Jul 2006 0:36	Modified:	30 Oct 2007 0:30
Reporter:	Kent Boortz	Email Updates:
Status:	Closed	Impact on me:	None
Category:	MySQL Server: Compiling	Severity:	S2 (Serious)
Version:	5.0.23-pre	OS:	Linux (IA64)
Assigned to:	Kent Boortz	CPU Architecture:	Any
Tags:	icc

Description:
The test 'test_basic_cursors' inside "mysql_client_test.c" gives
a segmentation fault, if compiled with icc and -O2 or above. You
can test this with (in a built source tree)

 % cd mysql-test
 % ./mysql-test-run.pl --start-and-exit mysql_client_test
 % ../tests/mysql_client_test --no-defaults --testcase --user=root \ 
   --port=9306 --socket=var/tmp/master.sock test_basic_cursors

Inlining migh not be the reason, but a -fno-inline makes the code
work again. Examining in gdb shows that the execution stops in
function stmt_fetch_fetch_row() and the head of the "for" loop.
Unfortunately optimization and "-g" doesn't work well, and I could
not get any real information on what is going on.

How to repeat:
Compile the 5.0.23 sources with icc 9.1.039, and configure like

 CC=icc
 CXX=icpc
 CFLAGS="-O3 -unroll2 -ip -mp -restrict"
 CXXFLAGS="-O3 -unroll2 -ip -mp -restrict"
 LDFLAGS="-static-libgcc -static-libcxa -i-static"
 export CC CXX CFLAGS CXXFLAGS LDFLAGS

 ./configure --disable-shared \  
 --without-embedded-server --with-innodb \ 
 --with-archive-storage-engine --without-bench \ 
 --with-zlib-dir=bundled --with-big-tables \ 
 --enable-assembler --enable-local-infile \ 
 --with-mysqld-user=mysql --with-pic --prefix=/ \ 
 --with-extra-charsets=complex \ 
 --enable-thread-safe-client --with-readline

 make

 cd mysql-test
 ./mysql-test-run.pl mysql_client_test

Until proof of the contrary, I assume that this failure is related:

=====
mysql_client_test              [ fail ]

Errors are (from /PATH/mysqltest-time) :
mysqltest: At line NNN: command "$MYSQL_CLIENT_TEST --getopt-ll-test=25600M" failed
(the last lines may be the most important ones)
=====

Occurred in the 5.0.24 release build in these tests:
Do-rpm-rhas4-ia64-5.0-pro-gpl.rhel4.log   normal
Do-rpm-rhas4-ia64-5.0-pro.rhel4.log   normal
Do-rpm-rhas4-ia64-5.0-standard.rhel4.log   normal
Do-rpm-rx2620b-5.0-pro-gpl.rhel3.log   normal
Do-rpm-rx2620b-5.0-pro.rhel3.log   normal
Do-rpm-rx2620b-5.0-standard.rhel3.log   normal
Do-rpm-sles9-ia64-5.0-pro-gpl.sles9.log   normal
Do-rpm-sles9-ia64-5.0-pro.sles9.log   normal
Do-rpm-sles9-ia64-5.0-standard.sles9.log   normal
rx2620b-icc-glibc23-5.0-cluster.log   normal
rx2620b-icc-glibc23-5.0-cluster.log   ps
rx2620b-icc-glibc23-5.0-cluster.log   NDB
rx2620b-icc-glibc23-5.0-max.log   normal
rx2620b-icc-glibc23-5.0-max.log   ps
rx2620b-icc-glibc23-5.0-max.log   NDB
rx2620b-icc-glibc23-5.0-pro.log   normal
rx2620b-icc-glibc23-5.0-pro.log   ps
rx2620b-icc-glibc23-5.0-standard.log   normal
rx2620b-icc-glibc23-5.0-standard.log   ps
rx2620b-icc-glibc23-5.0-pro-gpl.log   normal
rx2620b-icc-glibc23-5.0-pro-gpl.log   ps

All the above are platforms with the IA64 CPU, where we use the Intel "icc" compiler.
The test *passed* on "production", where we use "icc" on a "x86" CPU.

Reverified with icc 9.1.0.45 and MySQL 5.0.30,
the problem is still there.

Small test file for icc on IA-64

Attachment: bug20803.c (text/x-csrc), 1.57 KiB.

Uploaded small C program that will cause segfault when compiled as "icc file.c && ./a.out" with:

$ gcc -V
Intel(R) C Itanium(R) Compiler for Itanium(R)-based applications
Version 9.1    Build 20061105 Package ID: l_cc_c_9.1.045
Copyright (C) 1985-2006 Intel Corporation.  All rights reserved.
$ cat /proc/cpuinfo
processor  : 0
vendor     : GenuineIntel
arch       : IA-64
family     : Itanium 2
model      : 2
revision   : 2
archrev    : 0
features   : branchlong
cpu number : 0
cpu regs   : 4
cpu MHz    : 1299.969209
itc MHz    : 1299.969209
BogoMIPS   : 1946.15

processor  : 1
vendor     : GenuineIntel
arch       : IA-64
family     : Itanium 2
model      : 2
revision   : 2
archrev    : 0
features   : branchlong
cpu number : 0
cpu regs   : 4
cpu MHz    : 1299.969209
itc MHz    : 1299.969209
BogoMIPS   : 1946.15

$ uname -a
Linux rx2620b.mysql.com 2.4.21-37.EL #1 SMP Wed Sep 7 13:24:42 EDT 2005 ia64 ia64 ia64 GNU/Linu

Bug has been reported and escalated to Intel Itanium Compiler Engineering. Issue number Intel Premier Support: Q430699

The cause has been identified as an incorrect machine instruction being generated in the context of software pre-fetching after a subroutine got in-lined.

As a workaround, the bug can be avoided by disabling in-lining ( -ip--no-inlining) or lowering optimziation to O1 or less ( default is -O2)

I will update the case as soon as a compiler with a bug fix will be available

The issue has been fixed by Intel Compiler engineering. It is triggered by pre-fetching data in the context of procedure in-lining. The fix will be available in the next update of the 9.1 and 10.0 release.

The recommended workaround until availability of the fix: Add option -prefetch- to the compilation to prevent pre-fetching. Pre-fetching doesn't help very much on a code like MySQL DB thus the performance impact should be minimal to invisble. 

Heinz

Intel Compiler 10.025 has a fix for the issue; ran test case and verified fix on 2007-07-23

This problem was solved by upgrading to icc 10.0.026,
was used in MySQL server version 5.0.44sp1, and will be
used in upcoming 5.0.50 and 5.1.23.

Noted in 5.0.50, 5.1.23 changelogs.

Worked around an icc problem with an incorrect machine instruction
being generated in the context of software pre-fetching after a
subroutine got in-lined. (Upgrading to icc 10.0.026 makes the
workaround unnecessary.)