This is a fix for PR 15632. The mips32 and mips64 glibc builds are not
using the sqrt instruction in the sqrt libm functions. Each of the new files is a single line to include mips/fpu/e_sqrt.c (or e_sqrtf.c). I initially tried to create mips64/fpu but that did not work so I had to create both mips64/n32/fpu and mips64/n64/fpu instead. I did not see any sqrt tests in the glibc performance benchmarks, so I just ran a test of my own that did a bunch of sqrt operations and compiled it with -O2 -fno-builtins. It went from 61 seconds to 17 seconds on my 74k board. I also ran the glibc testsuite with the o32, n32, and n64 ABI's to verify there were no regressions in the testsuite. Ok to checkin? 2013-10-03 Steve Ellcey <[hidden email]> * sysdeps/mips/mips32/fpu/e_sqrt.c: New. * sysdeps/mips/mips32/fpu/e_sqrtf.c: New. * sysdeps/mips/mips64/n32/fpu/e_sqrt.c: New. * sysdeps/mips/mips64/n32/fpu/e_sqrtf.c: New. * sysdeps/mips/mips64/n64/fpu/e_sqrt.c: New. * sysdeps/mips/mips64/n64/fpu/e_sqrtf.c: New. diff --git a/ports/sysdeps/mips/mips32/fpu/e_sqrt.c b/ports/sysdeps/mips/mips32/fpu/e_sqrt.c new file mode 100644 index 0000000..81f4e77 --- /dev/null +++ b/ports/sysdeps/mips/mips32/fpu/e_sqrt.c @@ -0,0 +1 @@ +#include <sysdeps/mips/fpu/e_sqrt.c> diff --git a/ports/sysdeps/mips/mips32/fpu/e_sqrtf.c b/ports/sysdeps/mips/mips32/fpu/e_sqrtf.c new file mode 100644 index 0000000..fb0700d --- /dev/null +++ b/ports/sysdeps/mips/mips32/fpu/e_sqrtf.c @@ -0,0 +1 @@ +#include <sysdeps/mips/fpu/e_sqrtf.c> diff --git a/ports/sysdeps/mips/mips64/n32/fpu/e_sqrt.c b/ports/sysdeps/mips/mips64/n32/fpu/e_sqrt.c new file mode 100644 index 0000000..81f4e77 --- /dev/null +++ b/ports/sysdeps/mips/mips64/n32/fpu/e_sqrt.c @@ -0,0 +1 @@ +#include <sysdeps/mips/fpu/e_sqrt.c> diff --git a/ports/sysdeps/mips/mips64/n32/fpu/e_sqrtf.c b/ports/sysdeps/mips/mips64/n32/fpu/e_sqrtf.c new file mode 100644 index 0000000..fb0700d --- /dev/null +++ b/ports/sysdeps/mips/mips64/n32/fpu/e_sqrtf.c @@ -0,0 +1 @@ +#include <sysdeps/mips/fpu/e_sqrtf.c> diff --git a/ports/sysdeps/mips/mips64/n64/fpu/e_sqrt.c b/ports/sysdeps/mips/mips64/n64/fpu/e_sqrt.c new file mode 100644 index 0000000..81f4e77 --- /dev/null +++ b/ports/sysdeps/mips/mips64/n64/fpu/e_sqrt.c @@ -0,0 +1 @@ +#include <sysdeps/mips/fpu/e_sqrt.c> diff --git a/ports/sysdeps/mips/mips64/n64/fpu/e_sqrtf.c b/ports/sysdeps/mips/mips64/n64/fpu/e_sqrtf.c new file mode 100644 index 0000000..fb0700d --- /dev/null +++ b/ports/sysdeps/mips/mips64/n64/fpu/e_sqrtf.c @@ -0,0 +1 @@ +#include <sysdeps/mips/fpu/e_sqrtf.c> |
On Thu, 3 Oct 2013, Steve Ellcey wrote:
> This is a fix for PR 15632. The mips32 and mips64 glibc builds are not > using the sqrt instruction in the sqrt libm functions. Each of the new files > is a single line to include mips/fpu/e_sqrt.c (or e_sqrtf.c). I initially > tried to create mips64/fpu but that did not work so I had to create both > mips64/n32/fpu and mips64/n64/fpu instead. > > I did not see any sqrt tests in the glibc performance benchmarks, so I just > ran a test of my own that did a bunch of sqrt operations and compiled it > with -O2 -fno-builtins. It went from 61 seconds to 17 seconds on my 74k > board. > > I also ran the glibc testsuite with the o32, n32, and n64 ABI's to verify > there were no regressions in the testsuite. > > Ok to checkin? OK (with of course [BZ #15632] in the ChangeLog.mips entry, and that bug number added to the list of fixed bugs in the NEWS file). It would be a good idea to add an architecture-independent sqrt performance test, though. -- Joseph S. Myers [hidden email] |
Free forum by Nabble | Edit this page |