From f2db633d301b4b50f5f93de0e8314cc81e9bc7de Mon Sep 17 00:00:00 2001 From: Michael Cree Date: Sun, 19 Aug 2012 14:40:58 +1200 Subject: 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 Acked-by: Matt Turner Signed-off-by: Linus Torvalds --- arch/alpha/lib/ev67-strlen_user.S | 107 -------------------------------------- 1 file changed, 107 deletions(-) delete mode 100644 arch/alpha/lib/ev67-strlen_user.S (limited to 'arch/alpha/lib/ev67-strlen_user.S') diff --git a/arch/alpha/lib/ev67-strlen_user.S b/arch/alpha/lib/ev67-strlen_user.S deleted file mode 100644 index 57e0d77..0000000 --- a/arch/alpha/lib/ev67-strlen_user.S +++ /dev/null @@ -1,107 +0,0 @@ -/* - * arch/alpha/lib/ev67-strlen_user.S - * 21264 version contributed by Rick Gorton - * - * Return the length of the string including the NULL terminator - * (strlen+1) or zero if an error occurred. - * - * In places where it is critical to limit the processing time, - * and the data is not trusted, strnlen_user() should be used. - * It will return a value greater than its second argument if - * that limit would be exceeded. This implementation is allowed - * to access memory beyond the limit, but will not cross a page - * boundary when doing so. - * - * Much of the information about 21264 scheduling/coding comes from: - * Compiler Writer's Guide for the Alpha 21264 - * abbreviated as 'CWG' in other comments here - * ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html - * Scheduling notation: - * E - either cluster - * U - upper subcluster; U0 - subcluster U0; U1 - subcluster U1 - * L - lower subcluster; L0 - subcluster L0; L1 - subcluster L1 - * Try not to change the actual algorithm if possible for consistency. - */ - -#include - - -/* Allow an exception for an insn; exit if we get one. */ -#define EX(x,y...) \ - 99: x,##y; \ - .section __ex_table,"a"; \ - .long 99b - .; \ - lda v0, $exception-99b(zero); \ - .previous - - - .set noreorder - .set noat - .text - - .globl __strlen_user - .ent __strlen_user - .frame sp, 0, ra - - .align 4 -__strlen_user: - ldah a1, 32767(zero) # do not use plain strlen_user() for strings - # that might be almost 2 GB long; you should - # be using strnlen_user() instead - nop - nop - nop - - .globl __strnlen_user - - .align 4 -__strnlen_user: - .prologue 0 - EX( ldq_u t0, 0(a0) ) # L : load first quadword (a0 may be misaligned) - lda t1, -1(zero) # E : - - insqh t1, a0, t1 # U : - andnot a0, 7, v0 # E : - or t1, t0, t0 # E : - subq a0, 1, a0 # E : get our +1 for the return - - cmpbge zero, t0, t1 # E : t1 <- bitmask: bit i == 1 <==> i-th byte == 0 - subq a1, 7, t2 # E : - subq a0, v0, t0 # E : - bne t1, $found # U : - - addq t2, t0, t2 # E : - addq a1, 1, a1 # E : - nop # E : - nop # E : - - .align 4 -$loop: ble t2, $limit # U : - EX( ldq t0, 8(v0) ) # L : - nop # E : - nop # E : - - cmpbge zero, t0, t1 # E : - subq t2, 8, t2 # E : - addq v0, 8, v0 # E : addr += 8 - beq t1, $loop # U : - -$found: cttz t1, t2 # U0 : - addq v0, t2, v0 # E : - subq v0, a0, v0 # E : - ret # L0 : - -$exception: - nop - nop - nop - ret - - .align 4 # currently redundant -$limit: - nop - nop - subq a1, t2, v0 - ret - - .end __strlen_user -- cgit v1.1