Message ID | CAFULd4aJVUjr57x+_AO8k81w8Dsf1gshdsL-qbMUX72s+ZoVFA@mail.gmail.com |
---|---|
State | New |
Headers | show |
Dear Uros, Thanks very much for the patch. I have a few questions: > the patch removes -O0 from dg-additiona-options in IEEE testsuite, as this always override default optimization flag. That was my purpose: this test can fail with optimization (on x86_64, IIRC), hence the -O0 should override the default optimization flags in the testsuite. > * gfortran.dg/ieee/ieee_rounding_1.f90 (dg-additional-options): Add. If -mfp-rounding-mode=d is required on alpha for IEEE conformance, it should be added to add-ieee-options in lib/fortran-torture.exp, shouldn’t it? Rather than enabled on a case-by-case basis. Also, we might want to document these target-specific options here: https://gcc.gnu.org/onlinedocs/gfortran/IEEE-modules.html > BTW: On alpha, it is possible to enable underflow handling: > > #ifdef __USE_GNU > /* On later hardware, and later kernels for earlier hardware, we can forcibly > underflow denormal inputs and outputs. This can speed up certain programs > significantly, usually without affecting accuracy. */ > enum > { > FE_MAP_DMZ = 1UL << 12, /* Map denorm inputs to zero */ > #define FE_MAP_DMZ FE_MAP_DMZ > > FE_MAP_UMZ = 1UL << 13, /* Map underflowed outputs to zero */ > #define FE_MAP_UMZ FE_MAP_UMZ > }; > #endif > > FX, if you care for this option, I can help test the patch and > corresponding testcases. Yes, that’s interesting. What libc function do you call with those FE_MAP_{D,U}MZ values? I would also like to enable FTZ for i386/x86_64 at some point, but my issue there is that it’s not “universal”, ie it’s only for SSE math if I understand correctly. One point of view is that it’s better than nothing (especially now that SSE math is probably the most used mode), but another point of view is that it wouldn’t be standard conforming. FX
On Wed, Jul 2, 2014 at 11:16 AM, FX <fxcoudert@gmail.com> wrote: > Thanks very much for the patch. I have a few questions: > >> the patch removes -O0 from dg-additiona-options in IEEE testsuite, as this always override default optimization flag. > > That was my purpose: this test can fail with optimization (on x86_64, IIRC), hence the -O0 should override the default optimization flags in the testsuite. I did test this on x86_64-linux-gnu (with and without -m32) and there were no problems with the testrun. >> * gfortran.dg/ieee/ieee_rounding_1.f90 (dg-additional-options): Add. > > If -mfp-rounding-mode=d is required on alpha for IEEE conformance, it should be added to add-ieee-options in lib/fortran-torture.exp, shouldn’t it? Rather than enabled on a case-by-case basis. Also, we might want to document these target-specific options here: https://gcc.gnu.org/onlinedocs/gfortran/IEEE-modules.html I'm not sure if this is good idea ATM, since very few testcases (one exactly ;) ) require strict rounding modes. Due to this, I don't think dynamic rounding mode should be enabled for all testcases, as it comes with a price. We can review this in future, if many testcases depend on this option. >> BTW: On alpha, it is possible to enable underflow handling: >> >> #ifdef __USE_GNU >> /* On later hardware, and later kernels for earlier hardware, we can forcibly >> underflow denormal inputs and outputs. This can speed up certain programs >> significantly, usually without affecting accuracy. */ >> enum >> { >> FE_MAP_DMZ = 1UL << 12, /* Map denorm inputs to zero */ >> #define FE_MAP_DMZ FE_MAP_DMZ >> >> FE_MAP_UMZ = 1UL << 13, /* Map underflowed outputs to zero */ >> #define FE_MAP_UMZ FE_MAP_UMZ >> }; >> #endif >> >> FX, if you care for this option, I can help test the patch and >> corresponding testcases. > > Yes, that’s interesting. What libc function do you call with those FE_MAP_{D,U}MZ values? Hm, it looks that this functionality is handled through fesetenv, but I'm not 100% sure. > I would also like to enable FTZ for i386/x86_64 at some point, but my issue there is that it’s not “universal”, ie it’s only for SSE math if I understand correctly. One point of view is that it’s better than nothing (especially now that SSE math is probably the most used mode), but another point of view is that it wouldn’t be standard conforming. It is possible to test _SSE_MATH_ and _SSE2_MATH_ here? Uros.
Index: libgfortran/configure.host =================================================================== --- libgfortran/configure.host (revision 212221) +++ libgfortran/configure.host (working copy) @@ -48,6 +48,8 @@ esac # Some targets require additional compiler options for NaN/Inf. ieee_flags= case "${host_cpu}" in + alpha*) + ieee_flags="-mieee" ;; sh*) ieee_flags="-mieee" ;; esac Index: libgfortran/config/fpu-aix.h =================================================================== --- libgfortran/config/fpu-aix.h (revision 212221) +++ libgfortran/config/fpu-aix.h (working copy) @@ -372,18 +372,21 @@ support_fpu_rounding_mode (int mode) return 0; #endif + case GFC_FPE_UPWARD: #ifdef FE_UPWARD return 1; #else return 0; #endif + case GFC_FPE_DOWNWARD: #ifdef FE_DOWNWARD return 1; #else return 0; #endif + case GFC_FPE_TOWARDZERO: #ifdef FE_TOWARDZERO return 1; #else Index: libgfortran/config/fpu-glibc.h =================================================================== --- libgfortran/config/fpu-glibc.h (revision 212221) +++ libgfortran/config/fpu-glibc.h (working copy) @@ -387,18 +387,21 @@ support_fpu_rounding_mode (int mode) return 0; #endif + case GFC_FPE_UPWARD: #ifdef FE_UPWARD return 1; #else return 0; #endif + case GFC_FPE_DOWNWARD: #ifdef FE_DOWNWARD return 1; #else return 0; #endif + case GFC_FPE_TOWARDZERO: #ifdef FE_TOWARDZERO return 1; #else Index: gcc/testsuite/gfortran.dg/ieee/ieee_1.F90 =================================================================== --- gcc/testsuite/gfortran.dg/ieee/ieee_1.F90 (revision 212221) +++ gcc/testsuite/gfortran.dg/ieee/ieee_1.F90 (working copy) @@ -1,5 +1,5 @@ ! { dg-do run } -! { dg-additional-options "-ffree-line-length-none -O0" } +! { dg-additional-options "-ffree-line-length-none" } ! ! Use dg-additional-options rather than dg-options to avoid overwriting the ! default IEEE options which are passed by ieee.exp and necessary. Index: gcc/testsuite/gfortran.dg/ieee/ieee_rounding_1.f90 =================================================================== --- gcc/testsuite/gfortran.dg/ieee/ieee_rounding_1.f90 (revision 212221) +++ gcc/testsuite/gfortran.dg/ieee/ieee_rounding_1.f90 (working copy) @@ -1,4 +1,5 @@ ! { dg-do run } +! { dg-additional-options "-mfp-rounding-mode=d" { target alpha*-*-* } } use, intrinsic :: ieee_features, only : ieee_rounding use, intrinsic :: ieee_arithmetic