Thread: power() function in Windows: "value out of range: underflow"
Hi, There are some cases that power() function does not work correctly with 'NaN' arguments in Windows environment. Something like, postgres=# select power('NaN',11); ERROR: value out of range: underflow postgres=# select power('NaN','NaN'); ERROR: value out of range: underflow postgres=# select power(11,'NaN'); ERROR: value out of range: underflow In Linux environment, instead of ERROR it returns 'NaN'. The reason here is, When pow() in float.c:dpow() is called with 'NaN' arguments, pow() returns 'NaN' but in Windows environment errno is set to EDOM(invalid floating-point exception). So, PostgreSQL update "result" and cause an ERROR in CHECKFLOATVAL macro. I think it should be return 'NaN' in all of above cases. I have tried to create a patch to fix it. Please confirm the attached file. --- Thanks and best regards, Dang Minh Huong NEC Solution Innovators, Ltd. http://www.nec-solutioninnovators.co.jp/en/
Attachment
2018-04-10 5:30 GMT-03:00 Huong Dangminh <huo-dangminh@ys.jp.nec.com>: > There are some cases that power() function does not work > correctly with 'NaN' arguments in Windows environment. > Something like, > What is your exact OS version? What is your postgres version? I tested with Windows 10 (10.0.16299) x64 and couldn't reproduce the error. > postgres=# select power('NaN',11); > ERROR: value out of range: underflow > postgres=# select power('NaN','NaN'); > ERROR: value out of range: underflow > postgres=# select power(11,'NaN'); > ERROR: value out of range: underflow > > In Linux environment, instead of ERROR it returns 'NaN'. > Could you show us a simple test case? Print arguments, result and errno. -- Euler Taveira Timbira - http://www.timbira.com.br/ PostgreSQL: Consultoria, Desenvolvimento, Suporte 24x7 e Treinamento
2018-04-10 5:30 GMT-03:00 Huong Dangminh <huo-dangminh@ys.jp.nec.com>: > There are some cases that power() function does not work > correctly with 'NaN' arguments in Windows environment. > Something like, > What is your exact OS version? What is your postgres version? I tested with Windows 10 (10.0.16299) x64 and couldn't reproduce the error. > postgres=# select power('NaN',11); > ERROR: value out of range: underflow > postgres=# select power('NaN','NaN'); > ERROR: value out of range: underflow > postgres=# select power(11,'NaN'); > ERROR: value out of range: underflow > > In Linux environment, instead of ERROR it returns 'NaN'. > Could you show us a simple test case? Print arguments, result and errno. -- Euler Taveira Timbira - http://www.timbira.com.br/ PostgreSQL: Consultoria, Desenvolvimento, Suporte 24x7 e Treinamento
Hi, Thanks for response # I will add this thread to current CF soon. > 2018-04-10 5:30 GMT-03:00 Huong Dangminh <huo-dangminh@ys.jp.nec.com>: > > There are some cases that power() function does not work correctly > > with 'NaN' arguments in Windows environment. > > Something like, > > > What is your exact OS version? What is your postgres version? I tested with > Windows 10 (10.0.16299) x64 and couldn't reproduce the error. I think it is not depended on OS version but the visual c++ runtime libraries version (It seem also be included in installers). My environment: ==================================== - PostgreSQL: 9.6.8 # I am using the EDB PostgreSQL Installer # It also can reproduce in PostgreSQL 10.3 or 9.5.12 >postgres -V postgres (PostgreSQL) 9.6.8 >psql -c "select version()" version ------------------------------------------------------------- PostgreSQL 9.6.8, compiled by Visual C++ build 1800, 64-bit (1 row) >psql -c "select 'NaN'^11" ERROR: value out of range: underflow - OS: Microsoft Windows 10 Pro 10.0.14393 build 14393 ============================= > > postgres=# select power('NaN',11); > > ERROR: value out of range: underflow > > postgres=# select power('NaN','NaN'); > > ERROR: value out of range: underflow > > postgres=# select power(11,'NaN'); > > ERROR: value out of range: underflow > > > > In Linux environment, instead of ERROR it returns 'NaN'. > > > Could you show us a simple test case? Print arguments, result and errno. It just is my confirmation when debug PostgreSQL in Windows environment. # I built PostgreSQL with VC++ 2012 in debug mode and confirmed. I don't know how to print those values. --- Thanks and best regards, Dang Minh Huong NEC Solution Innovators, Ltd. http://www.nec-solutioninnovators.co.jp/en/
Hi, Thanks for response # I will add this thread to current CF soon. > 2018-04-10 5:30 GMT-03:00 Huong Dangminh <huo-dangminh@ys.jp.nec.com>: > > There are some cases that power() function does not work correctly > > with 'NaN' arguments in Windows environment. > > Something like, > > > What is your exact OS version? What is your postgres version? I tested with > Windows 10 (10.0.16299) x64 and couldn't reproduce the error. I think it is not depended on OS version but the visual c++ runtime libraries version (It seem also be included in installers). My environment: ==================================== - PostgreSQL: 9.6.8 # I am using the EDB PostgreSQL Installer # It also can reproduce in PostgreSQL 10.3 or 9.5.12 >postgres -V postgres (PostgreSQL) 9.6.8 >psql -c "select version()" version ------------------------------------------------------------- PostgreSQL 9.6.8, compiled by Visual C++ build 1800, 64-bit (1 row) >psql -c "select 'NaN'^11" ERROR: value out of range: underflow - OS: Microsoft Windows 10 Pro 10.0.14393 build 14393 ============================= > > postgres=# select power('NaN',11); > > ERROR: value out of range: underflow > > postgres=# select power('NaN','NaN'); > > ERROR: value out of range: underflow > > postgres=# select power(11,'NaN'); > > ERROR: value out of range: underflow > > > > In Linux environment, instead of ERROR it returns 'NaN'. > > > Could you show us a simple test case? Print arguments, result and errno. It just is my confirmation when debug PostgreSQL in Windows environment. # I built PostgreSQL with VC++ 2012 in debug mode and confirmed. I don't know how to print those values. --- Thanks and best regards, Dang Minh Huong NEC Solution Innovators, Ltd. http://www.nec-solutioninnovators.co.jp/en/
On 10 April 2018 at 20:30, Huong Dangminh <huo-dangminh@ys.jp.nec.com> wrote: > Hi, > > There are some cases that power() function does not work > correctly with 'NaN' arguments in Windows environment. > Something like, > > postgres=# select power('NaN',11); > ERROR: value out of range: underflow > postgres=# select power('NaN','NaN'); > ERROR: value out of range: underflow > postgres=# select power(11,'NaN'); > ERROR: value out of range: underflow > > In Linux environment, instead of ERROR it returns 'NaN'. > > The reason here is, > When pow() in float.c:dpow() is called with 'NaN' arguments, > pow() returns 'NaN' but in Windows environment errno is set to > EDOM(invalid floating-point exception). > So, PostgreSQL update "result" and cause an ERROR in CHECKFLOATVAL macro. I can recreate this when building with MSVC 2012. I confirm that I see the same as you. Microsoft are setting errno to EDOM in the above 3 cases, where in Linux the result is still NaN, just the errno is not set. Looking at [1], it says: * If x or y is a NaN, a NaN shall be returned (unless specified elsewhere in this description). * For any value of y (including NaN), if x is +1, 1.0 shall be returned. * For any value of x (including NaN), if y is ±0, 1.0 shall be returned. Which Microsoft seem to not have broken, they're just also setting the errno to EDOM in the first case, which seems a little strange, but the standard there does not seem to mention that it shouldn't do that. You patch fixes the problem for me, and importantly the two following cases still work: postgres=# select power(1,'NaN'); power ------- 1 (1 row) postgres=# select power('NaN', 0); power ------- 1 (1 row) There's no mention in the SQL standard about NaN handling in the above two cases, but I wonder if it's worth also adding a couple of tests for the above two cases to ensure all platforms are doing the same thing... ? I'd also rather see this line: if (errno == EDOM && isnan(result) && !(isnan(arg1) || isnan(arg2))) written as: if (errno == EDOM && isnan(result) && !isnan(arg1) && !isnan(arg2)) [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/pow.html -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On 10 April 2018 at 20:30, Huong Dangminh <huo-dangminh@ys.jp.nec.com> wrote: > Hi, > > There are some cases that power() function does not work > correctly with 'NaN' arguments in Windows environment. > Something like, > > postgres=# select power('NaN',11); > ERROR: value out of range: underflow > postgres=# select power('NaN','NaN'); > ERROR: value out of range: underflow > postgres=# select power(11,'NaN'); > ERROR: value out of range: underflow > > In Linux environment, instead of ERROR it returns 'NaN'. > > The reason here is, > When pow() in float.c:dpow() is called with 'NaN' arguments, > pow() returns 'NaN' but in Windows environment errno is set to > EDOM(invalid floating-point exception). > So, PostgreSQL update "result" and cause an ERROR in CHECKFLOATVAL macro. I can recreate this when building with MSVC 2012. I confirm that I see the same as you. Microsoft are setting errno to EDOM in the above 3 cases, where in Linux the result is still NaN, just the errno is not set. Looking at [1], it says: * If x or y is a NaN, a NaN shall be returned (unless specified elsewhere in this description). * For any value of y (including NaN), if x is +1, 1.0 shall be returned. * For any value of x (including NaN), if y is ±0, 1.0 shall be returned. Which Microsoft seem to not have broken, they're just also setting the errno to EDOM in the first case, which seems a little strange, but the standard there does not seem to mention that it shouldn't do that. You patch fixes the problem for me, and importantly the two following cases still work: postgres=# select power(1,'NaN'); power ------- 1 (1 row) postgres=# select power('NaN', 0); power ------- 1 (1 row) There's no mention in the SQL standard about NaN handling in the above two cases, but I wonder if it's worth also adding a couple of tests for the above two cases to ensure all platforms are doing the same thing... ? I'd also rather see this line: if (errno == EDOM && isnan(result) && !(isnan(arg1) || isnan(arg2))) written as: if (errno == EDOM && isnan(result) && !isnan(arg1) && !isnan(arg2)) [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/pow.html -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
2018-04-11 0:13 GMT-03:00 David Rowley <david.rowley@2ndquadrant.com>: > I can recreate this when building with MSVC 2012. I confirm that I see > the same as you. Microsoft are setting errno to EDOM in the above 3 > cases, where in Linux the result is still NaN, just the errno is not > set. > FWIW, I tested in MSVC 2017 (15.6.4) and it worked like expected. Looking at [1], it seems there could be nasty bugs when using math functions in MSVC < 2013 (or 2015). I don't have some older MSVC here to try /fp:OPTIONHERE [2] to see if it makes any difference. [1] https://docs.microsoft.com/en-us/cpp/porting/floating-point-migration-issues [2] https://msdn.microsoft.com/pt-br/library/e7s85ffb.aspx -- Euler Taveira Timbira - http://www.timbira.com.br/ PostgreSQL: Consultoria, Desenvolvimento, Suporte 24x7 e Treinamento
2018-04-11 0:13 GMT-03:00 David Rowley <david.rowley@2ndquadrant.com>: > I can recreate this when building with MSVC 2012. I confirm that I see > the same as you. Microsoft are setting errno to EDOM in the above 3 > cases, where in Linux the result is still NaN, just the errno is not > set. > FWIW, I tested in MSVC 2017 (15.6.4) and it worked like expected. Looking at [1], it seems there could be nasty bugs when using math functions in MSVC < 2013 (or 2015). I don't have some older MSVC here to try /fp:OPTIONHERE [2] to see if it makes any difference. [1] https://docs.microsoft.com/en-us/cpp/porting/floating-point-migration-issues [2] https://msdn.microsoft.com/pt-br/library/e7s85ffb.aspx -- Euler Taveira Timbira - http://www.timbira.com.br/ PostgreSQL: Consultoria, Desenvolvimento, Suporte 24x7 e Treinamento
Thanks for confirming. > 2018-04-11 0:13 GMT-03:00 David Rowley <david.rowley@2ndquadrant.com>: > > I can recreate this when building with MSVC 2012. I confirm that I see > > the same as you. Microsoft are setting errno to EDOM in the above 3 > > cases, where in Linux the result is still NaN, just the errno is not > > set. > > > FWIW, I tested in MSVC 2017 (15.6.4) and it worked like expected. > Looking at [1], it seems there could be nasty bugs when using math functions > in MSVC < 2013 (or 2015). I don't have some older MSVC here to try > /fp:OPTIONHERE [2] to see if it makes any difference. Thanks. Anyway currently PostgreSQL installers seem not be compiled with MSVC 2017, so it should be fix for current users? I updated the patch as David Rowley mentioned. --- Thanks and best regards, Dang Minh Huong NEC Solution Innovators, Ltd. http://www.nec-solutioninnovators.co.jp/en/
Attachment
Thanks for confirming. > 2018-04-11 0:13 GMT-03:00 David Rowley <david.rowley@2ndquadrant.com>: > > I can recreate this when building with MSVC 2012. I confirm that I see > > the same as you. Microsoft are setting errno to EDOM in the above 3 > > cases, where in Linux the result is still NaN, just the errno is not > > set. > > > FWIW, I tested in MSVC 2017 (15.6.4) and it worked like expected. > Looking at [1], it seems there could be nasty bugs when using math functions > in MSVC < 2013 (or 2015). I don't have some older MSVC here to try > /fp:OPTIONHERE [2] to see if it makes any difference. Thanks. Anyway currently PostgreSQL installers seem not be compiled with MSVC 2017, so it should be fix for current users? I updated the patch as David Rowley mentioned. --- Thanks and best regards, Dang Minh Huong NEC Solution Innovators, Ltd. http://www.nec-solutioninnovators.co.jp/en/
Attachment
On 11 April 2018 at 16:42, Huong Dangminh <huo-dangminh@ys.jp.nec.com> wrote: > I updated the patch as David Rowley mentioned. Looks fine to me. Please add to the next commitfest. -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On 11 April 2018 at 16:42, Huong Dangminh <huo-dangminh@ys.jp.nec.com> wrote: > I updated the patch as David Rowley mentioned. Looks fine to me. Please add to the next commitfest. -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
> > I updated the patch as David Rowley mentioned. > > Looks fine to me. Please add to the next commitfest. Thanks. Added. > -- > David Rowley http://www.2ndQuadrant.com/ > PostgreSQL Development, 24x7 Support, Training & Services > --- Thanks and best regards, Dang Minh Huong NEC Solution Innovators, Ltd. http://www.nec-solutioninnovators.co.jp/en/
> > I updated the patch as David Rowley mentioned. > > Looks fine to me. Please add to the next commitfest. Thanks. Added. > -- > David Rowley http://www.2ndQuadrant.com/ > PostgreSQL Development, 24x7 Support, Training & Services > --- Thanks and best regards, Dang Minh Huong NEC Solution Innovators, Ltd. http://www.nec-solutioninnovators.co.jp/en/
Huong Dangminh <huo-dangminh@ys.jp.nec.com> writes: >> 2018-04-11 0:13 GMT-03:00 David Rowley <david.rowley@2ndquadrant.com>: >>> I can recreate this when building with MSVC 2012. I confirm that I see >>> the same as you. Microsoft are setting errno to EDOM in the above 3 >>> cases, where in Linux the result is still NaN, just the errno is not >>> set. > I updated the patch as David Rowley mentioned. Pushed. I'd mainly note that you need to update all the variant float8 expected-files, not just the primary one. (Sometimes this requires a bit of guesswork, but here we're expecting all platforms to produce the same result. The buildfarm should tell us if I got it wrong.) regards, tom lane
Huong Dangminh <huo-dangminh@ys.jp.nec.com> writes: >> 2018-04-11 0:13 GMT-03:00 David Rowley <david.rowley@2ndquadrant.com>: >>> I can recreate this when building with MSVC 2012. I confirm that I see >>> the same as you. Microsoft are setting errno to EDOM in the above 3 >>> cases, where in Linux the result is still NaN, just the errno is not >>> set. > I updated the patch as David Rowley mentioned. Pushed. I'd mainly note that you need to update all the variant float8 expected-files, not just the primary one. (Sometimes this requires a bit of guesswork, but here we're expecting all platforms to produce the same result. The buildfarm should tell us if I got it wrong.) regards, tom lane
On 30 April 2018 at 07:24, Tom Lane <tgl@sss.pgh.pa.us> wrote: > Pushed. I'd mainly note that you need to update all the variant float8 > expected-files, not just the primary one. (Sometimes this requires a > bit of guesswork, but here we're expecting all platforms to produce > the same result. The buildfarm should tell us if I got it wrong.) gaur does not seem happy with this. I get the impression that pow(1, NaN) and pow(NaN, 0) on that machine must be returning NaN and setting errno to EDOM, and now that we're only using that code path when both are are non-NaN we no longer hit the special case which does result = 1; I wonder if it's better just to hard code these two cases before even calling the pow() function. -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On 30 April 2018 at 07:24, Tom Lane <tgl@sss.pgh.pa.us> wrote: > Pushed. I'd mainly note that you need to update all the variant float8 > expected-files, not just the primary one. (Sometimes this requires a > bit of guesswork, but here we're expecting all platforms to produce > the same result. The buildfarm should tell us if I got it wrong.) gaur does not seem happy with this. I get the impression that pow(1, NaN) and pow(NaN, 0) on that machine must be returning NaN and setting errno to EDOM, and now that we're only using that code path when both are are non-NaN we no longer hit the special case which does result = 1; I wonder if it's better just to hard code these two cases before even calling the pow() function. -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
David Rowley <david.rowley@2ndquadrant.com> writes: > I wonder if it's better just to hard code these two cases before even > calling the pow() function. Yeah, see my followup --- I also found out that SUSv2 (POSIX 1997) doesn't require either of these special cases, which helps explain why the inconsistency on older platforms. Hacking on a patch now. regards, tom lane
David Rowley <david.rowley@2ndquadrant.com> writes: > I wonder if it's better just to hard code these two cases before even > calling the pow() function. Yeah, see my followup --- I also found out that SUSv2 (POSIX 1997) doesn't require either of these special cases, which helps explain why the inconsistency on older platforms. Hacking on a patch now. regards, tom lane
> From: Tom Lane [mailto:tgl@sss.pgh.pa.us] > > I updated the patch as David Rowley mentioned. > > Pushed. I'd mainly note that you need to update all the variant float8 > expected-files, not just the primary one. (Sometimes this requires a bit Thank you. I will be careful next time. > of guesswork, but here we're expecting all platforms to produce the same > result. The buildfarm should tell us if I got it wrong.) Also thanks a lot of fixing failure in BSD platform as the result from buildfarm. --- Thanks and best regards, Dang Minh Huong NEC Solution Innovators, Ltd. http://www.nec-solutioninnovators.co.jp/en/
> From: Tom Lane [mailto:tgl@sss.pgh.pa.us] > > I updated the patch as David Rowley mentioned. > > Pushed. I'd mainly note that you need to update all the variant float8 > expected-files, not just the primary one. (Sometimes this requires a bit Thank you. I will be careful next time. > of guesswork, but here we're expecting all platforms to produce the same > result. The buildfarm should tell us if I got it wrong.) Also thanks a lot of fixing failure in BSD platform as the result from buildfarm. --- Thanks and best regards, Dang Minh Huong NEC Solution Innovators, Ltd. http://www.nec-solutioninnovators.co.jp/en/