math tests failed when the argument is out of domain

math tests failed when the argument is out of domain

[Yu-Chien Peter Lin]

Hi all, 

We run the glibc testsuite on RISC-V targets and got some 
common failure of following test in math subdir. 

FAIL: math/test-double-acos
FAIL: math/test-double-asin
FAIL: math/test-float64-acos
FAIL: math/test-float64-asin
FAIL: math/test-float32x-acos
FAIL: math/test-float32x-asin

The return values and errno of acos and asin are correct, but
they don't raise "Invalid operation" exception when the argument
is inf, -inf, 1.125, -1.125, max_value and -max_value.

These failed cases are listed test results [1] since glibc 2.33,
but it seems not a reported bug on sourceware bugzilla. so I want
to know the status on this issue.

[1] https://sourceware.org/glibc/wiki/Release/2.35#RISC-V_.28rv64imafdc.2Flp64d.29

These tests run at below environments.
- glibc 2.35
- gcc 10.2.1
- binutils 2.38
- Linux 5.4

Best Regards,
Peter Lin

Re: math tests failed when the argument is out of domain

[Szabolcs Nagy]

The 10/11/2022 20:25, Yu-Chien Peter Lin wrote:
> Hi all, 
> 
> We run the glibc testsuite on RISC-V targets and got some 
> common failure of following test in math subdir. 
> 
> FAIL: math/test-double-acos
> FAIL: math/test-double-asin
> FAIL: math/test-float64-acos
> FAIL: math/test-float64-asin
> FAIL: math/test-float32x-acos
> FAIL: math/test-float32x-asin
> 
> The return values and errno of acos and asin are correct, but
> they don't raise "Invalid operation" exception when the argument
> is inf, -inf, 1.125, -1.125, max_value and -max_value.


fwiw i saw this on aarch64 with some version of gcc-10 when
'return INFINITY/INFINITY;' is optimized to 'return NAN;'
so the fenv exception is not set.

this could be fixed via a more robust arithmetic to set nan
e.g. (x-x)/(x-x) or ensuring gcc takes exceptions into account
when const folding float operations.

> 
> These failed cases are listed test results [1] since glibc 2.33,
> but it seems not a reported bug on sourceware bugzilla. so I want
> to know the status on this issue.
> 
> [1] https://sourceware.org/glibc/wiki/Release/2.35#RISC-V_.28rv64imafdc.2Flp64d.29
> 
> These tests run at below environments.
> - glibc 2.35
> - gcc 10.2.1
> - binutils 2.38
> - Linux 5.4
> 
> Best Regards,
> Peter Lin

Re: math tests failed when the argument is out of domain

[Yu-Chien Peter Lin]

On Thu, Oct 13, 2022 at 10:24:59AM +0100, Szabolcs Nagy wrote:
> The 10/11/2022 20:25, Yu-Chien Peter Lin wrote:
> > Hi all, 
> > 
> > We run the glibc testsuite on RISC-V targets and got some 
> > common failure of following test in math subdir. 
> > 
> > FAIL: math/test-double-acos
> > FAIL: math/test-double-asin
> > FAIL: math/test-float64-acos
> > FAIL: math/test-float64-asin
> > FAIL: math/test-float32x-acos
> > FAIL: math/test-float32x-asin
> > 
> > The return values and errno of acos and asin are correct, but
> > they don't raise "Invalid operation" exception when the argument
> > is inf, -inf, 1.125, -1.125, max_value and -max_value.
> 
> 
> fwiw i saw this on aarch64 with some version of gcc-10 when
> 'return INFINITY/INFINITY;' is optimized to 'return NAN;'
> so the fenv exception is not set.
> 
> this could be fixed via a more robust arithmetic to set nan
> e.g. (x-x)/(x-x) or ensuring gcc takes exceptions into account
> when const folding float operations.

Hi Szabolcs,

Thanks, these tests are pass with your patch applied.

Best Regards,
Peter Lin

> > 
> > These failed cases are listed test results [1] since glibc 2.33,
> > but it seems not a reported bug on sourceware bugzilla. so I want
> > to know the status on this issue.
> > 
> > [1] https://sourceware.org/glibc/wiki/Release/2.35#RISC-V_.28rv64imafdc.2Flp64d.29
> > 
> > These tests run at below environments.
> > - glibc 2.35
> > - gcc 10.2.1
> > - binutils 2.38
> > - Linux 5.4
> > 
> > Best Regards,
> > Peter Lin