Tcl Source Code

Hello, sorry if this isn't the correct place to report this.

While running Tcllib tests with Tcl 8.6.12 on 32-bit x86 Alpine Linux, which uses musl libc, I ran into a problem with some math tests failing.

I proceeded to try running the same tests on previous versions of Tcl 8.6, and found that the tests pass on 8.6.10 but fail on 8.6.11.

I think the problem started with https://core.tcl-lang.org/tcl/info/5cb42f70ad0e65ba , and I managed to suppress the problem by disabling the special i386 TCL_IEEE_DOUBLE_ROUNDING and TCL_DEFAULT_DOUBLE_ROUNDING macros.

However, not being familiar with low-level programming, I'm not sure if this is the correct way of fixing the problem, and indeed it causes Tcl's expr tests to fail instead: https://core.tcl-lang.org/tcl/artifact/5ad80ae6f0ae48d7 .

I have attached the output from Tcllib's math tests, and Tcl's expr-old and format tests.

I would appreciate it if someone could look into this issue.

Thank you.

Fix merged now. Closing.

There is no Tcl release planned now, it could happen somewhere between September and December.

Thanks for the fix, the affected tests are all passing now.

Just wondering, will a release of Tcl 8.6 containing this fix be made available soon? Otherwise, I'll be submitting this patch for tclStrToD.c to Alpine.

Thanks, @kbk, for the fix. Given rubicon's explanation, I think the proposed solution is the way to go. (I already revised a similar fix, but yours is more complete....)

    if (significand == 0) {
	TCL_DEFAULT_DOUBLE_ROUNDING;
	return copysign(0.0, -signum);
    }

    if (significand == 0) {
	return copysign(0.0, -signum);
    }

Here is some more information.

I grepped the build log for 'tclStrToD' to see the gcc command used to build and link it (sorry if this is just noise and not helpful):

gcc -c -I"." -I/builds/rubicon/aports/main/tcl/src/tcl8.6.12/unix -I/builds/rubicon/aports/main/tcl/src/tcl8.6.12/generic -I/builds/rubicon/aports/main/tcl/src/tcl8.6.12/libtommath -O2 -Os -fomit-frame-pointer -pipe -Os -fomit-frame-pointer -Wall -Wpointer-arith -fPIC -DBUILD_tcl -DPACKAGE_NAME=\"tcl\" -DPACKAGE_TARNAME=\"tcl\" -DPACKAGE_VERSION=\"8.6\" -DPACKAGE_STRING=\"tcl\ 8.6\" -DPACKAGE_BUGREPORT=\"\" -DSTDC_HEADERS=1 -DHAVE_SYS_TYPES_H=1 -DHAVE_SYS_STAT_H=1 -DHAVE_STDLIB_H=1 -DHAVE_STRING_H=1 -DHAVE_MEMORY_H=1 -DHAVE_STRINGS_H=1 -DHAVE_INTTYPES_H=1 -DHAVE_STDINT_H=1 -DHAVE_UNISTD_H=1 -DHAVE_SYS_PARAM_H=1 -DUSE_THREAD_ALLOC=1 -D_REENTRANT=1 -D_THREAD_SAFE=1 -DHAVE_PTHREAD_ATTR_SETSTACKSIZE=1 -DHAVE_PTHREAD_ATFORK=1 -DTCL_THREADS=1 -DTCL_CFGVAL_ENCODING=\"iso8859-1\" -DHAVE_ZLIB=1 -DMODULE_SCOPE=extern\ __attribute__\(\(__visibility__\(\"hidden\"\)\)\) -DHAVE_HIDDEN=1 -DHAVE_CAST_TO_UNION=1 -DTCL_SHLIB_EXT=\".so\" -DNDEBUG=1 -DTCL_CFG_OPTIMIZED=1 -DTCL_TOMMATH=1 -DMP_PREC=4 -D_LARGEFILE64_SOURCE=1 -DTCL_WIDE_INT_TYPE=long\ long -DHAVE_STRUCT_DIRENT64=1 -DHAVE_STRUCT_STAT64=1 -DHAVE_OPEN64=1 -DHAVE_LSEEK64=1 -DHAVE_TYPE_OFF64_T=1 -DHAVE_GETCWD=1 -DHAVE_MKSTEMP=1 -DHAVE_OPENDIR=1 -DHAVE_STRTOL=1 -DHAVE_WAITPID=1 -DNO_GETWD=1 -DHAVE_GETNAMEINFO=1 -DHAVE_GETADDRINFO=1 -DHAVE_FREEADDRINFO=1 -DHAVE_GAI_STRERROR=1 -DHAVE_STRUCT_ADDRINFO=1 -DHAVE_STRUCT_IN6_ADDR=1 -DHAVE_STRUCT_SOCKADDR_IN6=1 -DHAVE_STRUCT_SOCKADDR_STORAGE=1 -DHAVE_GETPWUID_R_5=1 -DHAVE_GETPWUID_R=1 -DHAVE_GETPWNAM_R_5=1 -DHAVE_GETPWNAM_R=1 -DHAVE_GETGRGID_R_5=1 -DHAVE_GETGRGID_R=1 -DHAVE_GETGRNAM_R_5=1 -DHAVE_GETGRNAM_R=1 -DHAVE_DECL_GETHOSTBYNAME_R=1 -DHAVE_GETHOSTBYNAME_R_6=1 -DHAVE_GETHOSTBYNAME_R=1 -DHAVE_DECL_GETHOSTBYADDR_R=1 -DHAVE_GETHOSTBYADDR_R_8=1 -DHAVE_GETHOSTBYADDR_R=1 -DHAVE_TERMIOS_H=1 -DHAVE_SYS_IOCTL_H=1 -DHAVE_SYS_TIME_H=1 -DTIME_WITH_SYS_TIME=1 -DHAVE_GMTIME_R=1 -DHAVE_LOCALTIME_R=1 -DHAVE_MKTIME=1 -DHAVE_TM_GMTOFF=1 -DHAVE_TIMEZONE_VAR=1 -DHAVE_STRUCT_STAT_ST_BLOCKS=1 -DHAVE_STRUCT_STAT_ST_BLKSIZE=1 -DHAVE_BLKCNT_T=1 -DHAVE_INTPTR_T=1 -DHAVE_UINTPTR_T=1 -DNO_UNION_WAIT=1 -DHAVE_SIGNED_CHAR=1 -DHAVE_LANGINFO=1 -DHAVE_MKSTEMPS=1 -DHAVE_SYS_IOCTL_H=1 -DTCL_UNLOAD_DLLS=1 -DHAVE_CPUID=1      -DMP_FIXED_CUTOFFS -DMP_NO_STDINT /builds/rubicon/aports/main/tcl/src/tcl8.6.12/generic/tclStrToD.c

gcc -O2 -Os -fomit-frame-pointer -pipe -Os -fomit-frame-pointer  -Wl,--as-needed,-O1,--sort-common -Wl,--export-dynamic  -shared -o libtcl8.6.so regcomp.o regexec.o regfree.o regerror.o tclAlloc.o tclAssembly.o tclAsync.o tclBasic.o tclBinary.o tclCkalloc.o tclClock.o tclCmdAH.o tclCmdIL.o tclCmdMZ.o tclCompCmds.o tclCompCmdsGR.o tclCompCmdsSZ.o tclCompExpr.o tclCompile.o tclConfig.o tclDate.o tclDictObj.o tclDisassemble.o tclEncoding.o tclEnsemble.o tclEnv.o tclEvent.o tclExecute.o tclFCmd.o tclFileName.o tclGet.o tclHash.o tclHistory.o tclIndexObj.o tclInterp.o tclIO.o tclIOCmd.o tclIORChan.o tclIORTrans.o tclIOGT.o tclIOSock.o tclIOUtil.o tclLink.o tclListObj.o tclLiteral.o tclLoad.o tclMain.o tclNamesp.o tclNotify.o tclObj.o tclOptimize.o tclPanic.o tclParse.o tclPathObj.o tclPipe.o tclPkg.o tclPkgConfig.o tclPosixStr.o tclPreserve.o tclProc.o tclRegexp.o tclResolve.o tclResult.o tclScan.o tclStringObj.o tclStrToD.o tclThread.o tclThreadAlloc.o tclThreadJoin.o tclThreadStorage.o tclStubInit.o tclTimer.o tclTrace.o tclUtf.o tclUtil.o tclVar.o tclZlib.o tclTomMathInterface.o tclUnixChan.o tclUnixEvent.o tclUnixFCmd.o tclUnixFile.o tclUnixPipe.o tclUnixSock.o tclUnixTime.o tclUnixInit.o tclUnixThrd.o tclUnixCompat.o tclUnixNotfy.o  tclOO.o tclOOBasic.o tclOOCall.o tclOODefineCmds.o tclOOInfo.o tclOOMethod.o tclOOStubInit.o tclLoadDl.o  bn_s_mp_reverse.o bn_s_mp_mul_digs_fast.o bn_s_mp_sqr_fast.o bn_mp_add.o bn_mp_and.o bn_mp_add_d.o bn_mp_clamp.o bn_mp_clear.o bn_mp_clear_multi.o bn_mp_cmp.o bn_mp_cmp_d.o bn_mp_cmp_mag.o bn_mp_cnt_lsb.o bn_mp_copy.o bn_mp_count_bits.o bn_mp_div.o bn_mp_div_d.o bn_mp_div_2.o bn_mp_div_2d.o bn_mp_div_3.o bn_mp_exch.o bn_mp_expt_u32.o bn_mp_grow.o bn_mp_init.o bn_mp_init_copy.o bn_mp_init_multi.o bn_mp_init_set.o bn_mp_init_size.o bn_s_mp_karatsuba_mul.o bn_s_mp_karatsuba_sqr.o bn_s_mp_balance_mul.o bn_mp_lshd.o bn_mp_mod.o bn_mp_mod_2d.o bn_mp_mul.o bn_mp_mul_2.o bn_mp_mul_2d.o bn_mp_mul_d.o bn_mp_neg.o bn_mp_or.o bn_mp_radix_size.o bn_mp_radix_smap.o bn_mp_read_radix.o bn_mp_rshd.o bn_mp_set.o bn_mp_shrink.o bn_mp_sqr.o bn_mp_sqrt.o bn_mp_sub.o bn_mp_sub_d.o bn_mp_signed_rsh.o bn_mp_to_ubin.o bn_s_mp_toom_mul.o bn_s_mp_toom_sqr.o bn_mp_to_radix.o bn_mp_ubin_size.o bn_mp_xor.o bn_mp_zero.o bn_s_mp_add.o bn_s_mp_mul_digs.o bn_s_mp_sqr.o bn_s_mp_sub.o    -Wl,--as-needed,-O1,--sort-common -Wl,--export-dynamic  -ldl -lz  -lpthread    "-Wl,-rpath,/usr/lib"

I worked on this earlier this month, and didn't use fossil bisect, but by replacing 8.6.10 with 8.6.11, file-by-file (sort of), I isolated the changes that cause the tests to fail to tclStrToD.c, and by patching the file a little bit at a time, I stumbled upon those lines in MakeLowPrecisionDouble that I mentioned in my previous remark.

Lastly, the math tests are only failing on x86, Alpine is built for a few other architectures (aarch64, armhf, armv7, ppc64le, s390x, and of course x86_64), and they don't fail on those.

I'll try to answer the questions as best as I can, but as I am a little under the weather today, I may be a bit slow.

> what compiler is this?

GCC 11.2.1

> Are you changing optimization flags

Alpine Linux's build system sets the following environment variables ( from https://git.alpinelinux.org/abuild/tree/abuild.conf ):

export CFLAGS="-Os -fomit-frame-pointer"
export CXXFLAGS="$CFLAGS"
export CPPFLAGS="$CFLAGS"
export LDFLAGS="-Wl,--as-needed,-O1,--sort-common"

> You did say that everything in Tcl's own tests passed before you started adjusting the rounding modes, right?

Sorry for not clarifying this better.

The expr tests (that I linked to, but which is no longer attached to this bug report) fail when I adjust the rounding modes.

The expr-old and format tests which are still attached fail together with the Tcllib math tests.

> I've not found that to be necessary at the -O2 setting that Tcl uses by default

Alpine uses -Os, is this likely the source of the problem?

> the next thing that I'd suspect is that something is messed up in the C library

Through trial and error (as I am not a C programmer), after commenting the following lines in MakeLowPrecisionDouble:

    if (significand == 0) {
	return copysign(0.0, -signum);
    }

Both Tcl's expr-old and format tests and Tcllib's math and units tests all pass.

Alpine is now using musl v1.2.3, which has this implementation of copysign: https://git.musl-libc.org/cgit/musl/tree/src/math/copysign.c?h=v1.2.3

> Are we sure that 8.6.10 and 8.6.11 were built with the same compiler, include files, system libraries and optimization flags?

I tested 8.6.10 and 8.6.11 in a Gitlab CI within about an hour of each other, I don't think anything changed within that short period of time.

double retval;		/* Value of the number. */

volatile double retval; /* Value of the number */

Kevin, something for you?