Blame sysdeps/s390/strcspn-vx.S

Packit Service 93dc7a
/* Vector optimized 32/64 bit S/390 version of strcspn.
Packit Service 93dc7a
   Copyright (C) 2015-2018 Free Software Foundation, Inc.
Packit Service 93dc7a
   This file is part of the GNU C Library.
Packit Service 93dc7a
Packit Service 93dc7a
   The GNU C Library is free software; you can redistribute it and/or
Packit Service 93dc7a
   modify it under the terms of the GNU Lesser General Public
Packit Service 93dc7a
   License as published by the Free Software Foundation; either
Packit Service 93dc7a
   version 2.1 of the License, or (at your option) any later version.
Packit Service 93dc7a
Packit Service 93dc7a
   The GNU C Library is distributed in the hope that it will be useful,
Packit Service 93dc7a
   but WITHOUT ANY WARRANTY; without even the implied warranty of
Packit Service 93dc7a
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
Packit Service 93dc7a
   Lesser General Public License for more details.
Packit Service 93dc7a
Packit Service 93dc7a
   You should have received a copy of the GNU Lesser General Public
Packit Service 93dc7a
   License along with the GNU C Library; if not, see
Packit Service 93dc7a
   <http://www.gnu.org/licenses/>.  */
Packit Service 93dc7a
Packit Service 188d20
#include <ifunc-strcspn.h>
Packit Service 188d20
Packit Service 188d20
#if HAVE_STRCSPN_Z13
Packit Service 93dc7a
Packit Service 93dc7a
# include "sysdep.h"
Packit Service 93dc7a
# include "asm-syntax.h"
Packit Service 93dc7a
Packit Service 93dc7a
	.text
Packit Service 93dc7a
Packit Service 93dc7a
/* size_t strcspn (const char *s, const char * reject)
Packit Service 93dc7a
   The strcspn() function calculates the length of the initial segment
Packit Service 93dc7a
   of s which consists entirely of characters not in reject.
Packit Service 93dc7a
Packit Service 93dc7a
   This method checks the length of reject string. If it fits entirely
Packit Service 93dc7a
   in one vector register, a fast algorithm is used, which does not need
Packit Service 93dc7a
   to check multiple parts of accept-string. Otherwise a slower full
Packit Service 93dc7a
   check of accept-string is used.
Packit Service 93dc7a
Packit Service 93dc7a
   register overview:
Packit Service 93dc7a
   r3:  pointer to start of reject-string
Packit Service 93dc7a
   r2:  pointer to start of search-string
Packit Service 93dc7a
   r0:  loaded byte count of vlbb search-string
Packit Service 93dc7a
   r4:  found byte index
Packit Service 93dc7a
   r1:  current return len
Packit Service 93dc7a
   v16: search-string
Packit Service 93dc7a
   v17: reject-string
Packit Service 93dc7a
   v18: temp-vreg
Packit Service 93dc7a
Packit Service 93dc7a
   ONLY FOR SLOW:
Packit Service 93dc7a
   v19: first reject-string
Packit Service 93dc7a
   v20: zero for preparing acc-vector
Packit Service 93dc7a
   v21: global mask; 1 indicates a match between
Packit Service 93dc7a
	search-string-vreg and any reject-character
Packit Service 93dc7a
   v22: current mask; 1 indicates a match between
Packit Service 93dc7a
	search-string-vreg and any reject-character in current acc-vreg
Packit Service 93dc7a
   v24: one for result-checking of former string-part
Packit Service 93dc7a
   v30, v31: for re-/storing registers r6, r8, r9
Packit Service 93dc7a
   r5:  current len of reject-string
Packit Service 93dc7a
   r6:  zero-index in search-string or 16 if no zero
Packit Service 93dc7a
	or min(zero-index, loaded byte count)
Packit Service 93dc7a
   r8:  >0, if former reject-string-part contains a zero,
Packit Service 93dc7a
			otherwise =0;
Packit Service 93dc7a
   r9:  loaded byte count of vlbb reject-string
Packit Service 93dc7a
*/
Packit Service 188d20
ENTRY(STRCSPN_Z13)
Packit Service 93dc7a
	.machine "z13"
Packit Service 93dc7a
	.machinemode "zarch_nohighgprs"
Packit Service 93dc7a
Packit Service 93dc7a
	/*
Packit Service 93dc7a
	  Check if reject-string fits in one vreg:
Packit Service 93dc7a
	  ----------------------------------------
Packit Service 93dc7a
	*/
Packit Service 93dc7a
	vlbb	%v17,0(%r3),6	/* Load reject.  */
Packit Service 93dc7a
	lghi	%r1,0		/* Zero out current len.  */
Packit Service 93dc7a
	lcbb	%r0,0(%r3),6
Packit Service 93dc7a
	jo	.Lcheck_onbb	/* Special case if reject
Packit Service 93dc7a
				   lays on block-boundary.  */
Packit Service 93dc7a
.Lcheck_notonbb:
Packit Service 93dc7a
	vistrbs	%v17,%v17	/* Fill with zeros after first zero.  */
Packit Service 93dc7a
	je	.Lfast		/* Zero found -> reject fits in one vreg.  */
Packit Service 93dc7a
	j	.Lslow		/* No zero -> reject exceeds one vreg.  */
Packit Service 93dc7a
Packit Service 93dc7a
Packit Service 93dc7a
.Lcheck_onbb:
Packit Service 93dc7a
	/* Reject lays on block-boundary.  */
Packit Service 93dc7a
	vfenezb	%v18,%v17,%v17	/* Search zero in loaded reject bytes.  */
Packit Service 93dc7a
	vlgvb	%r4,%v18,7	/* Get index of zero or 16 if not found.  */
Packit Service 93dc7a
	clrjl	%r4,%r0,.Lcheck_notonbb /* Zero index < loaded bytes count ->
Packit Service 93dc7a
					    Reject fits in one vreg;
Packit Service 93dc7a
					    Fill with zeros and proceed
Packit Service 93dc7a
					    with FAST.  */
Packit Service 93dc7a
	vl	%v17,0(%r3)	/* Load reject, which exceeds loaded bytes.  */
Packit Service 93dc7a
	j	.Lcheck_notonbb /* Check if reject fits in one vreg.  */
Packit Service 93dc7a
Packit Service 93dc7a
Packit Service 93dc7a
	/*
Packit Service 93dc7a
	  Search s for reject in one vreg
Packit Service 93dc7a
	  -------------------------------
Packit Service 93dc7a
	*/
Packit Service 93dc7a
.Lfast:
Packit Service 93dc7a
	/* Complete reject-string in v17 and remaining bytes are zero.  */
Packit Service 93dc7a
Packit Service 93dc7a
	vlbb	%v16,0(%r2),6	/* Load s until next 4k-byte boundary.  */
Packit Service 93dc7a
	lcbb	%r0,0(%r2),6	/* Get bytes to 4k-byte boundary or 16.  */
Packit Service 93dc7a
Packit Service 93dc7a
	vfaezbs	%v18,%v16,%v17,0 /* Find first element in v16
Packit Service 93dc7a
				    unequal to any in v17
Packit Service 93dc7a
				    or first zero element.  */
Packit Service 93dc7a
Packit Service 93dc7a
	vlgvb	%r4,%v18,7	/* Load byte index of found element.  */
Packit Service 93dc7a
	clrjl	%r4,%r0,.Lfast_loop_found2 /* If found index is within loaded
Packit Service 93dc7a
					       bytes, return with found element
Packit Service 93dc7a
					       index (=equal count).  */
Packit Service 93dc7a
Packit Service 93dc7a
	/* Align s to 16 byte.  */
Packit Service 93dc7a
	risbgn	%r4,%r2,60,128+63,0 /* %r3 = bits 60-63 of %r2 'and' 15.  */
Packit Service 93dc7a
	lghi	%r1,16		/* current_len = 16.  */
Packit Service 93dc7a
	slr	%r1,%r4		/* Compute bytes to 16bytes boundary.  */
Packit Service 93dc7a
Packit Service 93dc7a
	/* Process s in 16byte aligned loop.  */
Packit Service 93dc7a
.Lfast_loop:
Packit Service 93dc7a
	vl	%v16,0(%r1,%r2)	/* Load search-string.  */
Packit Service 93dc7a
	vfaezbs	%v18,%v16,%v17,0 /* Find first element in v16 equal to any
Packit Service 93dc7a
				    in v17 or first zero element.  */
Packit Service 93dc7a
	jno	.Lfast_loop_found
Packit Service 93dc7a
Packit Service 93dc7a
	vl	%v16,16(%r1,%r2)
Packit Service 93dc7a
	vfaezbs	%v18,%v16,%v17,0
Packit Service 93dc7a
	jno	.Lfast_loop_found16
Packit Service 93dc7a
Packit Service 93dc7a
	vl	%v16,32(%r1,%r2)
Packit Service 93dc7a
	vfaezbs	%v18,%v16,%v17,0
Packit Service 93dc7a
	jno	.Lfast_loop_found32
Packit Service 93dc7a
Packit Service 93dc7a
	vl	%v16,48(%r1,%r2)
Packit Service 93dc7a
	vfaezbs	%v18,%v16,%v17,0
Packit Service 93dc7a
	jno	.Lfast_loop_found48
Packit Service 93dc7a
Packit Service 93dc7a
	aghi	%r1,64
Packit Service 93dc7a
	j	.Lfast_loop	/* Loop if no element was unequal to reject
Packit Service 93dc7a
				   and not zero.  */
Packit Service 93dc7a
Packit Service 93dc7a
	/* Found equal or zero element.  */
Packit Service 93dc7a
.Lfast_loop_found48:
Packit Service 93dc7a
	aghi	%r1,16
Packit Service 93dc7a
.Lfast_loop_found32:
Packit Service 93dc7a
	aghi	%r1,16
Packit Service 93dc7a
.Lfast_loop_found16:
Packit Service 93dc7a
	aghi	%r1,16
Packit Service 93dc7a
.Lfast_loop_found:
Packit Service 93dc7a
	vlgvb	%r4,%v18,7	/* Load byte index of found element or zero.  */
Packit Service 93dc7a
.Lfast_loop_found2:
Packit Service 93dc7a
	algrk	%r2,%r1,%r4	/* Add found index to current len.  */
Packit Service 93dc7a
	br	%r14
Packit Service 93dc7a
Packit Service 93dc7a
Packit Service 93dc7a
Packit Service 93dc7a
	/*
Packit Service 93dc7a
	  Search s for reject in multiple vregs
Packit Service 93dc7a
	  -------------------------------------
Packit Service 93dc7a
	*/
Packit Service 93dc7a
.Lslow:
Packit Service 93dc7a
	/* Save registers.  */
Packit Service 93dc7a
	vlvgg	%v30,%r6,0
Packit Service 93dc7a
	vlvgp	%v31,%r8,%r9
Packit Service 93dc7a
Packit Service 93dc7a
	/* Reject in v17 without zero.  */
Packit Service 93dc7a
	vlr	%v19,%v17	/* Save first acc-part for a fast reload.  */
Packit Service 93dc7a
	vzero	%v20		/* Zero for preparing acc-vector.  */
Packit Service 93dc7a
	vone	%v24		/* One for checking result of former
Packit Service 93dc7a
				   string-part.  */
Packit Service 93dc7a
Packit Service 93dc7a
	/* Align s to 16 byte.  */
Packit Service 93dc7a
	risbg	%r4,%r2,60,128+63,0 /* Test if s is aligned and
Packit Service 93dc7a
				       %r4 = bits 60-63 'and' 15.   */
Packit Service 93dc7a
	je	.Lslow_loop_str /* If s is aligned, loop aligned.  */
Packit Service 93dc7a
	lghi	%r0,15
Packit Service 93dc7a
	slr	%r0,%r4		/* Compute highest index to load (15-x).  */
Packit Service 93dc7a
	vll	%v16,%r0,0(%r2) /* Load up to 16 byte boundary (vll needs
Packit Service 93dc7a
				   highest index, remaining bytes are 0).  */
Packit Service 93dc7a
	ahi	%r0,1		/* Work with loaded byte count.  */
Packit Service 93dc7a
	vzero	%v21		/* Zero out global mask.  */
Packit Service 93dc7a
	lghi	%r5,0		/* Set current len of reject-string to zero.  */
Packit Service 93dc7a
	vfenezb	%v18,%v16,%v16	/* Find zero in current string-part.  */
Packit Service 93dc7a
	lghi	%r8,0		/* There is no zero in first reject-part.  */
Packit Service 93dc7a
	vlgvb	%r6,%v18,7	/* Load byte index of zero or 16 if no zero.  */
Packit Service 93dc7a
	clije	%r6,0,.Lslow_end /* If first element is zero -> return 0.  */
Packit Service 93dc7a
	clr	%r0,%r6		/* cc==1 if loaded byte count < zero-index.  */
Packit Service 93dc7a
	locrl	%r6,%r0		/* Load on cc==1; zero-index = lbc.  */
Packit Service 93dc7a
	j	.Lslow_loop_acc
Packit Service 93dc7a
Packit Service 93dc7a
Packit Service 93dc7a
	/* Process s in 16byte aligned loop.  */
Packit Service 93dc7a
.Lslow_next_str:
Packit Service 93dc7a
	/* Check results of former processed str-part.  */
Packit Service 93dc7a
	vfeeb	%v18,%v21,%v24	/* Find first equal match in global mask
Packit Service 93dc7a
				   (ones in element).  */
Packit Service 93dc7a
	vlgvb	%r4,%v18,7	/* Get index of first one (=equal) or 16.  */
Packit Service 93dc7a
	/* Equal-index < min(zero-index, loaded byte count)
Packit Service 93dc7a
	   -> Return pointer to equal element.  */
Packit Service 93dc7a
	clrjl	%r4,%r6,.Lslow_index_found
Packit Service 93dc7a
	/* Zero-index < loaded byte count
Packit Service 93dc7a
	   -> Former str-part was last str-part
Packit Service 93dc7a
	   -> Return null */
Packit Service 93dc7a
	clrjl	%r6,%r0,.Lslow_end_not_found
Packit Service 93dc7a
Packit Service 93dc7a
	/* All elements are zero (=no match) -> Proceed with next str-part.  */
Packit Service 93dc7a
	vlr	%v17,%v19	/* Load first part of reject (no zero).  */
Packit Service 93dc7a
	algfr	%r1,%r0		/* Add loaded byte count to current len.  */
Packit Service 93dc7a
Packit Service 93dc7a
.Lslow_loop_str:
Packit Service 93dc7a
	vl	%v16,0(%r1,%r2)	/* Load search-string.  */
Packit Service 93dc7a
	lghi	%r0,16		/* Loaded byte count is 16.  */
Packit Service 93dc7a
	vzero	%v21		/* Zero out global mask.  */
Packit Service 93dc7a
	lghi	%r5,0		/* Set current len of reject to zero.  */
Packit Service 93dc7a
	vfenezb	%v18,%v16,%v16	/* Find zero in current string-part.  */
Packit Service 93dc7a
	lghi	%r8,0		/* There is no zero in first reject-part.  */
Packit Service 93dc7a
	vlgvb	%r6,%v18,7	/* Load byte index of zero or 16 if no zero.  */
Packit Service 93dc7a
	clije	%r6,0,.Lslow_end /* If first element is zero (end of string)
Packit Service 93dc7a
				     -> Return current length.  */
Packit Service 93dc7a
Packit Service 93dc7a
.Lslow_loop_acc:
Packit Service 93dc7a
	vfaeb	%v22,%v16,%v17,4 /* Create matching-mask (1 in mask ->
Packit Service 93dc7a
				    Character matches any rejected character in
Packit Service 93dc7a
				    this reject-string-part) IN=0, RT=1.  */
Packit Service 93dc7a
	vlgvb	%r4,%v22,0	/* Get result of first element.  */
Packit Service 93dc7a
	/* First element is equal to any rejected characters?
Packit Service 93dc7a
	   (all other parts of reject cannot lead to a match before this one)
Packit Service 93dc7a
	   -> Return current len, which is pointing to this element.  */
Packit Service 93dc7a
	clijh	%r4,0,.Lslow_end
Packit Service 93dc7a
	vo	%v21,%v21,%v22	/* Global-mask = global-|matching-mask.  */
Packit Service 93dc7a
	/* Proceed with next acc until end of acc is reached.  */
Packit Service 93dc7a
Packit Service 93dc7a
Packit Service 93dc7a
.Lslow_next_acc:
Packit Service 93dc7a
	clijh	%r8,0,.Lslow_next_str /* There was a zero in last reject-part
Packit Service 93dc7a
					  -> Add found index to current len
Packit Service 93dc7a
					     and end.  */
Packit Service 93dc7a
	vlbb	%v17,16(%r5,%r3),6 /* Load next reject part.  */
Packit Service 93dc7a
	aghi	%r5,16		/* Increment current len of reject-string.  */
Packit Service 93dc7a
	lcbb	%r9,0(%r5,%r3),6 /* Get loaded byte count of reject-string.  */
Packit Service 93dc7a
	jo	.Lslow_next_acc_onbb /* Jump away if reject-string is
Packit Service 93dc7a
					 on block-boundary.  */
Packit Service 93dc7a
.Lslow_next_acc_notonbb:
Packit Service 93dc7a
	vistrbs	%v17,%v17	/* Fill with zeros after first zero.  */
Packit Service 93dc7a
	jo	.Lslow_loop_acc /* No zero found -> no preparation needed.  */
Packit Service 93dc7a
Packit Service 93dc7a
.Lslow_next_acc_prepare_zero:
Packit Service 93dc7a
	/* Zero in reject-part: fill zeros with first-reject-character.  */
Packit Service 93dc7a
	vlgvb	%r8,%v17,0	/* Load first element of reject-part.  */
Packit Service 93dc7a
	clije	%r8,0,.Lslow_next_str /* Process next str-part if first
Packit Service 93dc7a
					  character in this part of reject
Packit Service 93dc7a
					  is a zero.  */
Packit Service 93dc7a
	/* r8>0 -> zero found in this acc-part.  */
Packit Service 93dc7a
	vrepb	%v18,%v17,0	/* Replicate first char accross all chars.  */
Packit Service 93dc7a
	vceqb	%v22,%v20,%v17	/* Create a mask (v22) of null chars
Packit Service 93dc7a
				   by comparing with 0 (v20).  */
Packit Service 93dc7a
	vsel	%v17,%v18,%v17,%v22 /* Replace null chars with first char.  */
Packit Service 93dc7a
	j	.Lslow_loop_acc /* Reject-string part is prepared.  */
Packit Service 93dc7a
Packit Service 93dc7a
.Lslow_next_acc_onbb:
Packit Service 93dc7a
	vfenezb	%v18,%v17,%v17	/* Find zero in loaded bytes of reject part.  */
Packit Service 93dc7a
	vlgvb	%r8,%v18,7	/* Load byte index of zero.  */
Packit Service 93dc7a
	clrjl	%r8,%r9,.Lslow_next_acc_notonbb /* Found a zero in loaded bytes
Packit Service 93dc7a
						    -> Prepare vreg.  */
Packit Service 93dc7a
	vl	%v17,0(%r5,%r3)	/* Load over boundary ...  */
Packit Service 93dc7a
	lghi	%r8,0		/* r8=0 -> no zero in this part of acc,
Packit Service 93dc7a
				   check for zero is in jump-target.  */
Packit Service 93dc7a
	j	.Lslow_next_acc_notonbb /* ... and search for zero in
Packit Service 93dc7a
					    fully loaded vreg again.  */
Packit Service 93dc7a
Packit Service 93dc7a
.Lslow_end_not_found:
Packit Service 93dc7a
	algfr	%r1,%r6		/* Add zero-index to current len.  */
Packit Service 93dc7a
	j	.Lslow_end
Packit Service 93dc7a
.Lslow_index_found:
Packit Service 93dc7a
	algfr	%r1,%r4		/* Add found index of char to current len.  */
Packit Service 93dc7a
.Lslow_end:
Packit Service 93dc7a
	lgr	%r2,%r1
Packit Service 93dc7a
	/* Restore registers.  */
Packit Service 93dc7a
	vlgvg	%r6,%v30,0
Packit Service 93dc7a
	vlgvg	%r8,%v31,0
Packit Service 93dc7a
	vlgvg	%r9,%v31,1
Packit Service 93dc7a
	br	%r14
Packit Service 188d20
END(STRCSPN_Z13)
Packit Service 188d20
Packit Service 188d20
# if ! HAVE_STRCSPN_IFUNC
Packit Service 188d20
strong_alias (STRCSPN_Z13, strcspn)
Packit Service 188d20
# endif
Packit Service 188d20
Packit Service 188d20
# if ! HAVE_STRCSPN_C && defined SHARED && IS_IN (libc)
Packit Service 188d20
strong_alias (STRCSPN_Z13, __GI_strcspn)
Packit Service 188d20
# endif
Packit Service 188d20
Packit Service 188d20
#endif /* HAVE_STRCSPN_Z13  */