/*
   *
   * Optimized version of the standard strlen() function
   *
   *
   * Inputs:
   *	in0	address of string
   *
   * Outputs:
   *	ret0	the number of characters in the string (0 if empty string)
   *	does not count the \0
   *
   * Copyright (C) 1999, 2001 Hewlett-Packard Co
   *	Stephane Eranian <eranian@hpl.hp.com>
   *
   * 09/24/99 S.Eranian add speculation recovery code
   */
  
  #include <asm/asmmacro.h>
  
  //
  //
  // This is an enhanced version of the basic strlen. it includes a combination
  // of compute zero index (czx), parallel comparisons, speculative loads and
  // loop unroll using rotating registers.
  //
  // General Ideas about the algorithm:
  //	  The goal is to look at the string in chunks of 8 bytes.
  //	  so we need to do a few extra checks at the beginning because the
  //	  string may not be 8-byte aligned. In this case we load the 8byte
  //	  quantity which includes the start of the string and mask the unused
  //	  bytes with 0xff to avoid confusing czx.
  //	  We use speculative loads and software pipelining to hide memory
  //	  latency and do read ahead safely. This way we defer any exception.
  //
  //	  Because we don't want the kernel to be relying on particular
  //	  settings of the DCR register, we provide recovery code in case
  //	  speculation fails. The recovery code is going to "redo" the work using
  //	  only normal loads. If we still get a fault then we generate a
  //	  kernel panic. Otherwise we return the strlen as usual.
  //
  //	  The fact that speculation may fail can be caused, for instance, by
  //	  the DCR.dm bit being set. In this case TLB misses are deferred, i.e.,
  //	  a NaT bit will be set if the translation is not present. The normal
  //	  load, on the other hand, will cause the translation to be inserted
  //	  if the mapping exists.
  //
  //	  It should be noted that we execute recovery code only when we need
  //	  to use the data that has been speculatively loaded: we don't execute
  //	  recovery code on pure read ahead data.
  //
  // Remarks:
  //	- the cmp r0,r0 is used as a fast way to initialize a predicate
  //	  register to 1. This is required to make sure that we get the parallel
  //	  compare correct.
  //
  //	- we don't use the epilogue counter to exit the loop but we need to set
  //	  it to zero beforehand.
  //
  //	- after the loop we must test for Nat values because neither the
  //	  czx nor cmp instruction raise a NaT consumption fault. We must be
  //	  careful not to look too far for a Nat for which we don't care.
  //	  For instance we don't need to look at a NaT in val2 if the zero byte
  //	  was in val1.
  //
  //	- Clearly performance tuning is required.
  //
  //
  //
  #define saved_pfs	r11
  #define	tmp		r10
  #define base		r16
  #define orig		r17
  #define saved_pr	r18
  #define src		r19
  #define mask		r20
  #define val		r21
  #define val1		r22
  #define val2		r23
  
  GLOBAL_ENTRY(strlen)
  	.prologue
  	.save ar.pfs, saved_pfs
  	alloc saved_pfs=ar.pfs,11,0,0,8 // rotating must be multiple of 8
  
  	.rotr v[2], w[2]	// declares our 4 aliases
  
  	extr.u tmp=in0,0,3	// tmp=least significant 3 bits
  	mov orig=in0		// keep trackof initial byte address
  	dep src=0,in0,0,3	// src=8byte-aligned in0 address
  	.save pr, saved_pr
  	mov saved_pr=pr		// preserve predicates (rotation)
  	;;
  
  	.body
  
  	ld8 v[1]=[src],8	// must not speculate: can fail here
  	shl tmp=tmp,3		// multiply by 8bits/byte
  	mov mask=-1		// our mask
  	;;
  	ld8.s w[1]=[src],8	// speculatively load next
  	cmp.eq p6,p0=r0,r0	// sets p6 to true for cmp.and
  	sub tmp=64,tmp		// how many bits to shift our mask on the right
  	;;
  	shr.u	mask=mask,tmp	// zero enough bits to hold v[1] valuable part
  	mov ar.ec=r0		// clear epilogue counter (saved in ar.pfs)
  	;;
  	add base=-16,src	// keep track of aligned base
  	or v[1]=v[1],mask	// now we have a safe initial byte pattern
  	;;
  1:
  	ld8.s v[0]=[src],8	// speculatively load next
  	czx1.r val1=v[1]	// search 0 byte from right
  	czx1.r val2=w[1]	// search 0 byte from right following 8bytes
  	;;
  	ld8.s w[0]=[src],8	// speculatively load next to next
  	cmp.eq.and p6,p0=8,val1	// p6 = p6 and val1==8
  	cmp.eq.and p6,p0=8,val2	// p6 = p6 and mask==8
  (p6)	br.wtop.dptk 1b		// loop until p6 == 0
  	;;
  	//
  	// We must return try the recovery code iff
  	// val1_is_nat || (val1==8 && val2_is_nat)
  	//
  	// XXX Fixme
  	//	- there must be a better way of doing the test
  	//
  	cmp.eq  p8,p9=8,val1	// p6 = val1 had zero (disambiguate)
  	tnat.nz p6,p7=val1	// test NaT on val1
  (p6)	br.cond.spnt .recover	// jump to recovery if val1 is NaT
  	;;
  	//
  	// if we come here p7 is true, i.e., initialized for // cmp
  	//
  	cmp.eq.and  p7,p0=8,val1// val1==8?
  	tnat.nz.and p7,p0=val2	// test NaT if val2
  (p7)	br.cond.spnt .recover	// jump to recovery if val2 is NaT
  	;;
  (p8)	mov val1=val2		// the other test got us out of the loop
  (p8)	adds src=-16,src	// correct position when 3 ahead
  (p9)	adds src=-24,src	// correct position when 4 ahead
  	;;
  	sub ret0=src,orig	// distance from base
  	sub tmp=8,val1		// which byte in word
  	mov pr=saved_pr,0xffffffffffff0000
  	;;
  	sub ret0=ret0,tmp	// adjust
  	mov ar.pfs=saved_pfs	// because of ar.ec, restore no matter what
  	br.ret.sptk.many rp	// end of normal execution
  
  	//
  	// Outlined recovery code when speculation failed
  	//
  	// This time we don't use speculation and rely on the normal exception
  	// mechanism. that's why the loop is not as good as the previous one
  	// because read ahead is not possible
  	//
  	// IMPORTANT:
  	// Please note that in the case of strlen() as opposed to strlen_user()
  	// we don't use the exception mechanism, as this function is not
  	// supposed to fail. If that happens it means we have a bug and the
  	// code will cause of kernel fault.
  	//
  	// XXX Fixme
  	//	- today we restart from the beginning of the string instead
  	//	  of trying to continue where we left off.
  	//
  .recover:
  	ld8 val=[base],8	// will fail if unrecoverable fault
  	;;
  	or val=val,mask		// remask first bytes
  	cmp.eq p0,p6=r0,r0	// nullify first ld8 in loop
  	;;
  	//
  	// ar.ec is still zero here
  	//
  2:
  (p6)	ld8 val=[base],8	// will fail if unrecoverable fault
  	;;
  	czx1.r val1=val		// search 0 byte from right
  	;;
  	cmp.eq p6,p0=8,val1	// val1==8 ?
  (p6)	br.wtop.dptk 2b		// loop until p6 == 0
  	;;			// (avoid WAW on p63)
  	sub ret0=base,orig	// distance from base
  	sub tmp=8,val1
  	mov pr=saved_pr,0xffffffffffff0000
  	;;
  	sub ret0=ret0,tmp	// length=now - back -1
  	mov ar.pfs=saved_pfs	// because of ar.ec, restore no matter what
  	br.ret.sptk.many rp	// end of successful recovery code
  END(strlen)