1/*
  2 * Optimized version of the strlen_user() function
  3 *
  4 * Inputs:
  5 *	in0	address of buffer
  6 *
  7 * Outputs:
  8 *	ret0	0 in case of fault, strlen(buffer)+1 otherwise
  9 *
 10 * Copyright (C) 1998, 1999, 2001 Hewlett-Packard Co
 11 *	David Mosberger-Tang <davidm@hpl.hp.com>
 12 *	Stephane Eranian <eranian@hpl.hp.com>
 13 *
 14 * 01/19/99 S.Eranian heavily enhanced version (see details below)
 15 * 09/24/99 S.Eranian added speculation recovery code
 16 */
 17
 18#include <asm/asmmacro.h>
 19
 20//
 21// int strlen_user(char *)
 22// ------------------------
 23// Returns:
 24//	- length of string + 1
 25//	- 0 in case an exception is raised
 26//
 27// This is an enhanced version of the basic strlen_user. it includes a
 28// combination of compute zero index (czx), parallel comparisons, speculative
 29// loads and loop unroll using rotating registers.
 30//
 31// General Ideas about the algorithm:
 32//	  The goal is to look at the string in chunks of 8 bytes.
 33//	  so we need to do a few extra checks at the beginning because the
 34//	  string may not be 8-byte aligned. In this case we load the 8byte
 35//	  quantity which includes the start of the string and mask the unused
 36//	  bytes with 0xff to avoid confusing czx.
 37//	  We use speculative loads and software pipelining to hide memory
 38//	  latency and do read ahead safely. This way we defer any exception.
 39//
 40//	  Because we don't want the kernel to be relying on particular
 41//	  settings of the DCR register, we provide recovery code in case
 42//	  speculation fails. The recovery code is going to "redo" the work using
 43//	  only normal loads. If we still get a fault then we return an
 44//	  error (ret0=0). Otherwise we return the strlen+1 as usual.
 45//	  The fact that speculation may fail can be caused, for instance, by
 46//	  the DCR.dm bit being set. In this case TLB misses are deferred, i.e.,
 47//	  a NaT bit will be set if the translation is not present. The normal
 48//	  load, on the other hand, will cause the translation to be inserted
 49//	  if the mapping exists.
 50//
 51//	  It should be noted that we execute recovery code only when we need
 52//	  to use the data that has been speculatively loaded: we don't execute
 53//	  recovery code on pure read ahead data.
 54//
 55// Remarks:
 56//	- the cmp r0,r0 is used as a fast way to initialize a predicate
 57//	  register to 1. This is required to make sure that we get the parallel
 58//	  compare correct.
 59//
 60//	- we don't use the epilogue counter to exit the loop but we need to set
 61//	  it to zero beforehand.
 62//
 63//	- after the loop we must test for Nat values because neither the
 64//	  czx nor cmp instruction raise a NaT consumption fault. We must be
 65//	  careful not to look too far for a Nat for which we don't care.
 66//	  For instance we don't need to look at a NaT in val2 if the zero byte
 67//	  was in val1.
 68//
 69//	- Clearly performance tuning is required.
 70//
 71
 72#define saved_pfs	r11
 73#define	tmp		r10
 74#define base		r16
 75#define orig		r17
 76#define saved_pr	r18
 77#define src		r19
 78#define mask		r20
 79#define val		r21
 80#define val1		r22
 81#define val2		r23
 82
 83GLOBAL_ENTRY(__strlen_user)
 84	.prologue
 85	.save ar.pfs, saved_pfs
 86	alloc saved_pfs=ar.pfs,11,0,0,8
 87
 88	.rotr v[2], w[2]	// declares our 4 aliases
 89
 90	extr.u tmp=in0,0,3	// tmp=least significant 3 bits
 91	mov orig=in0		// keep trackof initial byte address
 92	dep src=0,in0,0,3	// src=8byte-aligned in0 address
 93	.save pr, saved_pr
 94	mov saved_pr=pr		// preserve predicates (rotation)
 95	;;
 96
 97	.body
 98
 99	ld8.s v[1]=[src],8	// load the initial 8bytes (must speculate)
100	shl tmp=tmp,3		// multiply by 8bits/byte
101	mov mask=-1		// our mask
102	;;
103	ld8.s w[1]=[src],8	// load next 8 bytes in 2nd pipeline
104	cmp.eq p6,p0=r0,r0	// sets p6 (required because of // cmp.and)
105	sub tmp=64,tmp		// how many bits to shift our mask on the right
106	;;
107	shr.u	mask=mask,tmp	// zero enough bits to hold v[1] valuable part
108	mov ar.ec=r0		// clear epilogue counter (saved in ar.pfs)
109	;;
110	add base=-16,src	// keep track of aligned base
111	chk.s v[1], .recover	// if already NaT, then directly skip to recover
112	or v[1]=v[1],mask	// now we have a safe initial byte pattern
113	;;
1141:
115	ld8.s v[0]=[src],8	// speculatively load next
116	czx1.r val1=v[1]	// search 0 byte from right
117	czx1.r val2=w[1]	// search 0 byte from right following 8bytes
118	;;
119	ld8.s w[0]=[src],8	// speculatively load next to next
120	cmp.eq.and p6,p0=8,val1	// p6 = p6 and val1==8
121	cmp.eq.and p6,p0=8,val2	// p6 = p6 and mask==8
122(p6)	br.wtop.dptk.few 1b	// loop until p6 == 0
123	;;
124	//
125	// We must return try the recovery code iff
126	// val1_is_nat || (val1==8 && val2_is_nat)
127	//
128	// XXX Fixme
129	//	- there must be a better way of doing the test
130	//
131	cmp.eq  p8,p9=8,val1	// p6 = val1 had zero (disambiguate)
132	tnat.nz p6,p7=val1	// test NaT on val1
133(p6)	br.cond.spnt .recover	// jump to recovery if val1 is NaT
134	;;
135	//
136	// if we come here p7 is true, i.e., initialized for // cmp
137	//
138	cmp.eq.and  p7,p0=8,val1// val1==8?
139	tnat.nz.and p7,p0=val2	// test NaT if val2
140(p7)	br.cond.spnt .recover	// jump to recovery if val2 is NaT
141	;;
142(p8)	mov val1=val2		// val2 contains the value
143(p8)	adds src=-16,src	// correct position when 3 ahead
144(p9)	adds src=-24,src	// correct position when 4 ahead
145	;;
146	sub ret0=src,orig	// distance from origin
147	sub tmp=7,val1		// 7=8-1 because this strlen returns strlen+1
148	mov pr=saved_pr,0xffffffffffff0000
149	;;
150	sub ret0=ret0,tmp	// length=now - back -1
151	mov ar.pfs=saved_pfs	// because of ar.ec, restore no matter what
152	br.ret.sptk.many rp	// end of normal execution
153
154	//
155	// Outlined recovery code when speculation failed
156	//
157	// This time we don't use speculation and rely on the normal exception
158	// mechanism. that's why the loop is not as good as the previous one
159	// because read ahead is not possible
160	//
161	// XXX Fixme
162	//	- today we restart from the beginning of the string instead
163	//	  of trying to continue where we left off.
164	//
165.recover:
166	EX(.Lexit1, ld8 val=[base],8)	// load the initial bytes
167	;;
168	or val=val,mask			// remask first bytes
169	cmp.eq p0,p6=r0,r0		// nullify first ld8 in loop
170	;;
171	//
172	// ar.ec is still zero here
173	//
1742:
175	EX(.Lexit1, (p6) ld8 val=[base],8)
176	;;
177	czx1.r val1=val		// search 0 byte from right
178	;;
179	cmp.eq p6,p0=8,val1	// val1==8 ?
180(p6)	br.wtop.dptk.few 2b	// loop until p6 == 0
181	;;
182	sub ret0=base,orig	// distance from base
183	sub tmp=7,val1		// 7=8-1 because this strlen returns strlen+1
184	mov pr=saved_pr,0xffffffffffff0000
185	;;
186	sub ret0=ret0,tmp	// length=now - back -1
187	mov ar.pfs=saved_pfs	// because of ar.ec, restore no matter what
188	br.ret.sptk.many rp	// end of successful recovery code
189
190	//
191	// We failed even on the normal load (called from exception handler)
192	//
193.Lexit1:
194	mov ret0=0
195	mov pr=saved_pr,0xffffffffffff0000
196	mov ar.pfs=saved_pfs	// because of ar.ec, restore no matter what
197	br.ret.sptk.many rp
198END(__strlen_user)