linux

Commit Graph

Author	SHA1	Message	Date
Chris Metcalf	c753bf34c9	word-at-a-time.h: support zero_bytemask() on alpha and tile Both alpha and tile needed implementations of zero_bytemask. The alpha version is untested. Signed-off-by: Chris Metcalf <cmetcalf@ezchip.com>	2015-10-06 14:53:16 -04:00
Michael Cree	f2db633d30	alpha: Use new generic strncpy_from_user() and strnlen_user() Similar to x86/sparc/powerpc implementations except: 1) we implement an extremely efficient has_zero()/find_zero() sequence with both prep_zero_mask() and create_zero_mask() no-operations. 2) Our output from prep_zero_mask() differs in that only the lowest eight bits are used to represent the zero bytes nevertheless it can be safely ORed with other similar masks from prep_zero_mask() and forms input to create_zero_mask(), the two fundamental properties prep_zero_mask() must satisfy. Tests on EV67 and EV68 CPUs revealed that the generic code is essentially as fast (to within 0.5% of CPU cycles) of the old Alpha specific code for large quadword-aligned strings, despite the 30% extra CPU instructions executed. In contrast, the generic code for unaligned strings is substantially slower (by more than a factor of 3) than the old Alpha specific code. Signed-off-by: Michael Cree <mcree@orcon.net.nz> Acked-by: Matt Turner <mattst88@gmail.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>	2012-08-19 08:41:18 -07:00

Author

SHA1

Message

Date

Chris Metcalf

c753bf34c9

word-at-a-time.h: support zero_bytemask() on alpha and tile

Both alpha and tile needed implementations of zero_bytemask.

The alpha version is untested.

Signed-off-by: Chris Metcalf <cmetcalf@ezchip.com>

2015-10-06 14:53:16 -04:00

Michael Cree

f2db633d30

alpha: Use new generic strncpy_from_user() and strnlen_user()

Similar to x86/sparc/powerpc implementations except:
1) we implement an extremely efficient has_zero()/find_zero()
   sequence with both prep_zero_mask() and create_zero_mask()
   no-operations.
2) Our output from prep_zero_mask() differs in that only the
   lowest eight bits are used to represent the zero bytes
   nevertheless it can be safely ORed with other similar masks
   from prep_zero_mask() and forms input to create_zero_mask(),
   the two fundamental properties prep_zero_mask() must satisfy.

Tests on EV67 and EV68 CPUs revealed that the generic code is
essentially as fast (to within 0.5% of CPU cycles) of the old
Alpha specific code for large quadword-aligned strings, despite
the 30% extra CPU instructions executed.  In contrast, the
generic code for unaligned strings is substantially slower (by
more than a factor of 3) than the old Alpha specific code.

Signed-off-by: Michael Cree <mcree@orcon.net.nz>
Acked-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

2012-08-19 08:41:18 -07:00

2 Commits