Re: RISC-V glibc port, v6

[Joseph Myers <joseph at codesourcery dot com>]

On Thu, 25 Jan 2018, Palmer Dabbelt wrote:

>     RV64IMAC/LP64 results:
>     
>     Summary of test results:
>          26 FAIL
>        5551 PASS
>           4 UNRESOLVED
>          18 UNSUPPORTED
>          18 XFAIL
>     
>     Test output [1].
>     
>     Includes ~9 expected cross test failures, ~5 timed out that should pass
>     with a longer timeout.  There are 3 failures due to rounding mode
>     issues, tests that should be ignored due to soft float but are not.  I
>     will investigate more tomorrow.
>     
>     [1] http://bluespec.com/tmp/glibc-check-lp64.tar.xz

I see some warnings in those logs:

../sysdeps/riscv/math-tests.h:26:5: warning: "__riscv_flen" is not defined, evaluates to 0 [-Wundef]
 #if __riscv_flen == 0
     ^~~~~~~~~~~~

(a) You're using --disable-werror to build glibc.  Don't do that.  It may 
defeat the point of some tests (that operate at compile time and don't 
test anything at runtime) that rely on errors for compilation warnings.  
It may also defeat the rules about typo-proof macro conventions in glibc, 
which only work with -Wundef warnings are errors.

(b) The source line in the error message is not part of the posted patch 
series, i.e. these test logs don't seem to correspond to the posted 
version of the port.  The math-tests.h posted uses "#ifdef __riscv_flen", 
twice (which is clearly inverted, you should have "#ifndef __riscv_flen", 
twice, so see if using that avoids the rounding mode issues you mention), 
not "#if __riscv_flen == 0"

-- 
Joseph S. Myers
joseph@codesourcery.com