Blame sysdeps/s390/multiarch/wcscspn-vx.S

Packit Service 8bbac7
/* Vector optimized 32/64 bit S/390 version of wcscspn.
Packit Service 8bbac7
   Copyright (C) 2015-2018 Free Software Foundation, Inc.
Packit Service 8bbac7
   This file is part of the GNU C Library.
Packit Service 8bbac7
Packit Service 8bbac7
   The GNU C Library is free software; you can redistribute it and/or
Packit Service 8bbac7
   modify it under the terms of the GNU Lesser General Public
Packit Service 8bbac7
   License as published by the Free Software Foundation; either
Packit Service 8bbac7
   version 2.1 of the License, or (at your option) any later version.
Packit Service 8bbac7
Packit Service 8bbac7
   The GNU C Library is distributed in the hope that it will be useful,
Packit Service 8bbac7
   but WITHOUT ANY WARRANTY; without even the implied warranty of
Packit Service 8bbac7
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
Packit Service 8bbac7
   Lesser General Public License for more details.
Packit Service 8bbac7
Packit Service 8bbac7
   You should have received a copy of the GNU Lesser General Public
Packit Service 8bbac7
   License along with the GNU C Library; if not, see
Packit Service 8bbac7
   <http://www.gnu.org/licenses/>.  */
Packit Service 8bbac7
Packit Service 8bbac7
#if defined HAVE_S390_VX_ASM_SUPPORT && IS_IN (libc)
Packit Service 8bbac7
Packit Service 8bbac7
# include "sysdep.h"
Packit Service 8bbac7
# include "asm-syntax.h"
Packit Service 8bbac7
Packit Service 8bbac7
	.text
Packit Service 8bbac7
Packit Service 8bbac7
/* size_t wcscspn (const wchar_t *s, const wchar_t * reject)
Packit Service 8bbac7
   The wcscspn() function calculates the length of the initial segment
Packit Service 8bbac7
   of s which consists entirely of characters not in reject.
Packit Service 8bbac7
Packit Service 8bbac7
   This method checks the length of reject string. If it fits entirely
Packit Service 8bbac7
   in one vector register, a fast algorithm is used, which does not need
Packit Service 8bbac7
   to check multiple parts of accept-string. Otherwise a slower full
Packit Service 8bbac7
   check of accept-string is used.
Packit Service 8bbac7
Packit Service 8bbac7
   register overview:
Packit Service 8bbac7
   r3:  pointer to start of reject-string
Packit Service 8bbac7
   r2:  pointer to start of search-string
Packit Service 8bbac7
   r0:  loaded byte count of vlbb search-string
Packit Service 8bbac7
   r4:  found byte index
Packit Service 8bbac7
   r1:  current return len
Packit Service 8bbac7
   v16: search-string
Packit Service 8bbac7
   v17: reject-string
Packit Service 8bbac7
   v18: temp-vreg
Packit Service 8bbac7
Packit Service 8bbac7
   ONLY FOR SLOW:
Packit Service 8bbac7
   v19: first reject-string
Packit Service 8bbac7
   v20: zero for preparing acc-vector
Packit Service 8bbac7
   v21: global mask; 1 indicates a match between
Packit Service 8bbac7
	search-string-vreg and any reject-character
Packit Service 8bbac7
   v22: current mask; 1 indicates a match between
Packit Service 8bbac7
	search-string-vreg and any reject-character in current acc-vreg
Packit Service 8bbac7
   v30, v31: for re-/storing registers r6, r8, r9
Packit Service 8bbac7
   r5:  current len of reject-string
Packit Service 8bbac7
   r6:  zero-index in search-string or 16 if no zero
Packit Service 8bbac7
	or min(zero-index, loaded byte count)
Packit Service 8bbac7
   r8:  >0, if former reject-string-part contains a zero,
Packit Service 8bbac7
			otherwise =0;
Packit Service 8bbac7
   r9:  loaded byte count of vlbb reject-string
Packit Service 8bbac7
*/
Packit Service 8bbac7
ENTRY(__wcscspn_vx)
Packit Service 8bbac7
	.machine "z13"
Packit Service 8bbac7
	.machinemode "zarch_nohighgprs"
Packit Service 8bbac7
Packit Service 8bbac7
	tmll	%r2,3		/* Test if s is 4-byte aligned?  */
Packit Service 8bbac7
	jne	.Lfallback	/* And use common-code variant if not.  */
Packit Service 8bbac7
Packit Service 8bbac7
	/*
Packit Service 8bbac7
	  Check if reject-string fits in one vreg:
Packit Service 8bbac7
	  ----------------------------------------
Packit Service 8bbac7
	*/
Packit Service 8bbac7
	vlbb	%v17,0(%r3),0	/* Load reject.  */
Packit Service 8bbac7
	lcbb	%r0,0(%r3),0
Packit Service 8bbac7
	jo	.Lcheck_onbb	/* Special case if reject
Packit Service 8bbac7
				   lays on block-boundary.  */
Packit Service 8bbac7
Packit Service 8bbac7
.Lcheck_notonbb:
Packit Service 8bbac7
	lghi	%r1,0		/* Zero out current len.  */
Packit Service 8bbac7
	vistrfs	%v17,%v17	/* Fill with zeros after first zero.  */
Packit Service 8bbac7
	je	.Lfast		/* Zero found -> reject fits in one vreg.  */
Packit Service 8bbac7
	j	.Lslow		/* No zero -> reject exceeds one vreg.  */
Packit Service 8bbac7
Packit Service 8bbac7
Packit Service 8bbac7
.Lcheck_onbb:
Packit Service 8bbac7
	/* Reject lays on block-boundary.  */
Packit Service 8bbac7
	nill	%r0,65532	/* Recognize only fully loaded characters.  */
Packit Service 8bbac7
	je	.Lcheck_onbb2	/* Reload vr, if we loaded no full wchar_t.  */
Packit Service 8bbac7
	vfenezf	%v18,%v17,%v17	/* Search zero in loaded reject bytes.  */
Packit Service 8bbac7
	vlgvb	%r4,%v18,7	/* Get index of zero or 16 if not found.  */
Packit Service 8bbac7
	clrjl	%r4,%r0,.Lcheck_notonbb /* Zero index < loaded bytes count ->
Packit Service 8bbac7
					    Reject fits in one vreg;
Packit Service 8bbac7
					    Fill with zeros and proceed
Packit Service 8bbac7
					    with FAST.  */
Packit Service 8bbac7
.Lcheck_onbb2:
Packit Service 8bbac7
	vl	%v17,0(%r3)	/* Load reject, which exceeds loaded bytes.  */
Packit Service 8bbac7
	j	.Lcheck_notonbb /* Check if reject fits in one vreg.  */
Packit Service 8bbac7
Packit Service 8bbac7
Packit Service 8bbac7
	/*
Packit Service 8bbac7
	  Search s for reject in one vreg
Packit Service 8bbac7
	  -------------------------------
Packit Service 8bbac7
	*/
Packit Service 8bbac7
.Lfast:
Packit Service 8bbac7
	/* Complete reject-string in v17 and remaining bytes are zero.  */
Packit Service 8bbac7
Packit Service 8bbac7
	vlbb	%v16,0(%r2),6	/* Load s until next 4k-byte boundary.  */
Packit Service 8bbac7
	lcbb	%r0,0(%r2),6	/* Get bytes to 4k-byte boundary or 16.  */
Packit Service 8bbac7
Packit Service 8bbac7
	vfaezfs	%v18,%v16,%v17,0 /* Find first element in v16
Packit Service 8bbac7
				    unequal to any in v17
Packit Service 8bbac7
				    or first zero element.  */
Packit Service 8bbac7
	vlgvb	%r4,%v18,7	/* Load byte index of found element.  */
Packit Service 8bbac7
	clrjl	%r4,%r0,.Lfast_loop_found2 /* If found index is within loaded
Packit Service 8bbac7
					       bytes, return with found element
Packit Service 8bbac7
					       index (=equal count).  */
Packit Service 8bbac7
Packit Service 8bbac7
	/* Align s to 16 byte.  */
Packit Service 8bbac7
	risbgn	%r4,%r2,60,128+63,0 /* %r3 = bits 60-63 of %r2 'and' 15.  */
Packit Service 8bbac7
	lghi	%r1,16		/* current_len = 16.  */
Packit Service 8bbac7
	slr	%r1,%r4		/* Compute bytes to 16bytes boundary.  */
Packit Service 8bbac7
Packit Service 8bbac7
	/* Process s in 16byte aligned loop.  */
Packit Service 8bbac7
.Lfast_loop:
Packit Service 8bbac7
	vl	%v16,0(%r1,%r2)	/* Load search-string.  */
Packit Service 8bbac7
	vfaezfs	%v18,%v16,%v17,0 /* Find first element in v16 equal to any
Packit Service 8bbac7
				    in v17 or first zero element.  */
Packit Service 8bbac7
	jno	.Lfast_loop_found
Packit Service 8bbac7
Packit Service 8bbac7
	vl	%v16,16(%r1,%r2)
Packit Service 8bbac7
	vfaezfs	%v18,%v16,%v17,0
Packit Service 8bbac7
	jno	.Lfast_loop_found16
Packit Service 8bbac7
Packit Service 8bbac7
	vl	%v16,32(%r1,%r2)
Packit Service 8bbac7
	vfaezfs	%v18,%v16,%v17,0
Packit Service 8bbac7
	jno	.Lfast_loop_found32
Packit Service 8bbac7
Packit Service 8bbac7
	vl	%v16,48(%r1,%r2)
Packit Service 8bbac7
	vfaezfs	%v18,%v16,%v17,0
Packit Service 8bbac7
	jno	.Lfast_loop_found48
Packit Service 8bbac7
Packit Service 8bbac7
	aghi	%r1,64
Packit Service 8bbac7
	j	.Lfast_loop	/* Loop if no element was unequal to reject
Packit Service 8bbac7
				   and not zero.  */
Packit Service 8bbac7
Packit Service 8bbac7
	/* Found equal or zero element.  */
Packit Service 8bbac7
.Lfast_loop_found48:
Packit Service 8bbac7
	aghi	%r1,16
Packit Service 8bbac7
.Lfast_loop_found32:
Packit Service 8bbac7
	aghi	%r1,16
Packit Service 8bbac7
.Lfast_loop_found16:
Packit Service 8bbac7
	aghi	%r1,16
Packit Service 8bbac7
.Lfast_loop_found:
Packit Service 8bbac7
	vlgvb	%r4,%v18,7	/* Load byte index of found element or zero.  */
Packit Service 8bbac7
.Lfast_loop_found2:
Packit Service 8bbac7
	algrk	%r2,%r1,%r4	/* Add found index to current len.  */
Packit Service 8bbac7
	srlg	%r2,%r2,2	/* Convert byte-count to character-count.  */
Packit Service 8bbac7
	br	%r14
Packit Service 8bbac7
Packit Service 8bbac7
Packit Service 8bbac7
Packit Service 8bbac7
	/*
Packit Service 8bbac7
	  Search s for reject in multiple vregs
Packit Service 8bbac7
	  -------------------------------------
Packit Service 8bbac7
	*/
Packit Service 8bbac7
.Lslow:
Packit Service 8bbac7
	/* Save registers.  */
Packit Service 8bbac7
	vlvgg	%v30,%r6,0
Packit Service 8bbac7
	vlvgp	%v31,%r8,%r9
Packit Service 8bbac7
Packit Service 8bbac7
	/* Reject in v17 without zero.  */
Packit Service 8bbac7
	vlr	%v19,%v17	/* Save first acc-part for a fast reload.  */
Packit Service 8bbac7
	vzero	%v20		/* Zero for preparing acc-vector.  */
Packit Service 8bbac7
	vone	%v24		/* One for checking result of former
Packit Service 8bbac7
				   string-part.  */
Packit Service 8bbac7
Packit Service 8bbac7
	/* Align s to 16 byte.  */
Packit Service 8bbac7
	risbg	%r4,%r2,60,128+63,0 /* Test if s is aligned and
Packit Service 8bbac7
				       %r4 = bits 60-63 'and' 15.   */
Packit Service 8bbac7
	je	.Lslow_loop_str /* If s is aligned, loop aligned.  */
Packit Service 8bbac7
	lghi	%r0,15
Packit Service 8bbac7
	slr	%r0,%r4		/* Compute highest index to load (15-x).  */
Packit Service 8bbac7
	vll	%v16,%r0,0(%r2) /* Load up to 16byte boundary (vll needs
Packit Service 8bbac7
				   highest index, remaining bytes are 0).  */
Packit Service 8bbac7
	ahi	%r0,1		/* Work with loaded byte count.  */
Packit Service 8bbac7
	vzero	%v21		/* Zero out global mask.  */
Packit Service 8bbac7
	lghi	%r5,0		/* Set current len of reject-string to zero.  */
Packit Service 8bbac7
	vfenezf	%v18,%v16,%v16	/* Find zero in current string-part.  */
Packit Service 8bbac7
	lghi	%r8,0		/* There is no zero in first reject-part.  */
Packit Service 8bbac7
	vlgvb	%r6,%v18,7	/* Load byte index of zero or 16 if no zero.  */
Packit Service 8bbac7
	clije	%r6,0,.Lslow_end /* If first element is zero -> return 0.  */
Packit Service 8bbac7
	clr	%r0,%r6		/* cc==1 if loaded byte count < zero-index.  */
Packit Service 8bbac7
	locrl	%r6,%r0		/* Load on cc==1; zero-index = lbc.  */
Packit Service 8bbac7
	j	.Lslow_loop_acc
Packit Service 8bbac7
Packit Service 8bbac7
Packit Service 8bbac7
	/* Process s in 16byte aligned loop.  */
Packit Service 8bbac7
.Lslow_next_str:
Packit Service 8bbac7
	/* Check results of former processed str-part.  */
Packit Service 8bbac7
	vfeef	%v18,%v21,%v24	/* Find first equal match in global mask
Packit Service 8bbac7
				   (ones in element).  */
Packit Service 8bbac7
	vlgvb	%r4,%v18,7	/* Get index of first one (=equal) or 16.  */
Packit Service 8bbac7
	/* Equal-index < min(zero-index, loaded byte count)
Packit Service 8bbac7
	   -> Return pointer to equal element.  */
Packit Service 8bbac7
	clrjl	%r4,%r6,.Lslow_index_found
Packit Service 8bbac7
	/* Zero-index < loaded byte count
Packit Service 8bbac7
	   -> Former str-part was last str-part
Packit Service 8bbac7
	   -> Return null  */
Packit Service 8bbac7
	clrjl	%r6,%r0,.Lslow_end_not_found
Packit Service 8bbac7
Packit Service 8bbac7
	/* All elements are zero (=no match) -> proceed with next str-part.  */
Packit Service 8bbac7
	vlr	%v17,%v19	/* Load first part of reject (no zero).  */
Packit Service 8bbac7
	algfr	%r1,%r0		/* Add loaded byte count to current len.  */
Packit Service 8bbac7
Packit Service 8bbac7
.Lslow_loop_str:
Packit Service 8bbac7
	vl	%v16,0(%r1,%r2)	/* Load search-string.  */
Packit Service 8bbac7
	lghi	%r0,16		/* Loaded byte count is 16.  */
Packit Service 8bbac7
	vzero	%v21		/* Zero out global mask.  */
Packit Service 8bbac7
	lghi	%r5,0		/* Set current len of reject to zero.  */
Packit Service 8bbac7
	vfenezf	%v18,%v16,%v16	/* Find zero in current string-part.  */
Packit Service 8bbac7
	lghi	%r8,0		/* There is no zero in first reject-part.  */
Packit Service 8bbac7
	vlgvb	%r6,%v18,7	/* Load byte index of zero or 16 if no zero.  */
Packit Service 8bbac7
	clije	%r6,0,.Lslow_end /* If first element is zero (end of string)
Packit Service 8bbac7
				     -> Return current length.  */
Packit Service 8bbac7
Packit Service 8bbac7
.Lslow_loop_acc:
Packit Service 8bbac7
	vfaef	%v22,%v16,%v17,4 /* Create matching-mask (1 in mask ->
Packit Service 8bbac7
				    Character matches any rejected character in
Packit Service 8bbac7
				    this reject-string-part) IN=0, RT=1.  */
Packit Service 8bbac7
	vlgvf	%r4,%v22,0	/* Get result of first element.  */
Packit Service 8bbac7
	/* First element is equal to any rejected characters?
Packit Service 8bbac7
	   (All other parts of reject cannot lead to a match before this one)
Packit Service 8bbac7
	   -> Return current len, which is pointing to this element.  */
Packit Service 8bbac7
	clijh	%r4,0,.Lslow_end
Packit Service 8bbac7
	vo	%v21,%v21,%v22	/* Global-mask = global-|matching-mask.  */
Packit Service 8bbac7
	/* Proceed with next acc until end of acc is reached.  */
Packit Service 8bbac7
Packit Service 8bbac7
Packit Service 8bbac7
.Lslow_next_acc:
Packit Service 8bbac7
	clijh	%r8,0,.Lslow_next_str /* There was a zero in last reject-part
Packit Service 8bbac7
					  -> Add found index to current len
Packit Service 8bbac7
					     and end.  */
Packit Service 8bbac7
	vlbb	%v17,16(%r5,%r3),6 /* Load next reject part.  */
Packit Service 8bbac7
	aghi	%r5,16		/* Increment current len of reject-string.  */
Packit Service 8bbac7
	lcbb	%r9,0(%r5,%r3),6 /* Get loaded byte count of reject-string.  */
Packit Service 8bbac7
	jo	.Lslow_next_acc_onbb /* Jump away if reject-string is
Packit Service 8bbac7
					 on block-boundary.  */
Packit Service 8bbac7
.Lslow_next_acc_notonbb:
Packit Service 8bbac7
	vistrfs	%v17,%v17	/* Fill with zeros after first zero.  */
Packit Service 8bbac7
	jo	.Lslow_loop_acc /* No zero found -> no preparation needed.  */
Packit Service 8bbac7
Packit Service 8bbac7
.Lslow_next_acc_prepare_zero:
Packit Service 8bbac7
	/* Zero in reject-part: fill zeros with first-reject-character.  */
Packit Service 8bbac7
	vlgvf	%r8,%v17,0	/* Load first element of reject-part.  */
Packit Service 8bbac7
	clije	%r8,0,.Lslow_next_str /* Process next str-part if first
Packit Service 8bbac7
					  character in this part of reject
Packit Service 8bbac7
					  is a zero.  */
Packit Service 8bbac7
	/* r8>0 -> zero found in this acc-part.  */
Packit Service 8bbac7
	vrepf	%v18,%v17,0	/* Replicate first char accross all chars.  */
Packit Service 8bbac7
	vceqf	%v22,%v20,%v17	/* Create a mask (v22) of null chars
Packit Service 8bbac7
				   by comparing with 0 (v20).  */
Packit Service 8bbac7
	vsel	%v17,%v18,%v17,%v22 /* Replace null chars with first char.  */
Packit Service 8bbac7
	j	.Lslow_loop_acc /* Reject-string part is prepared.  */
Packit Service 8bbac7
Packit Service 8bbac7
.Lslow_next_acc_onbb:
Packit Service 8bbac7
	nill	%r9,65532	/* Recognize only fully loaded characters.  */
Packit Service 8bbac7
	je	.Lslow_next_acc_onbb2 /* Reload vr, if no full wchar_t
Packit Service 8bbac7
					  loaded.  */
Packit Service 8bbac7
	vfenezf	%v18,%v17,%v17	/* Find zero in loaded bytes of reject part.  */
Packit Service 8bbac7
	vlgvb	%r8,%v18,7	/* Load byte index of zero.  */
Packit Service 8bbac7
	clrjl	%r8,%r9,.Lslow_next_acc_notonbb /* Found a zero in loaded bytes
Packit Service 8bbac7
						    -> Prepare vreg.  */
Packit Service 8bbac7
.Lslow_next_acc_onbb2:
Packit Service 8bbac7
	vl	%v17,0(%r5,%r3)	/* Load over boundary ...  */
Packit Service 8bbac7
	lghi	%r8,0		/* r8=0 -> no zero in this part of acc,
Packit Service 8bbac7
				   check for zero is in jump-target.  */
Packit Service 8bbac7
	j	.Lslow_next_acc_notonbb /* ... and search for zero in
Packit Service 8bbac7
					    fully loaded vreg again.  */
Packit Service 8bbac7
Packit Service 8bbac7
.Lslow_end_not_found:
Packit Service 8bbac7
	algfr	%r1,%r6		/* Add zero-index to current len.  */
Packit Service 8bbac7
	j	.Lslow_end
Packit Service 8bbac7
.Lslow_index_found:
Packit Service 8bbac7
	algfr	%r1,%r4		/* Add found index of char to current len.  */
Packit Service 8bbac7
.Lslow_end:
Packit Service 8bbac7
	srlg	%r2,%r1,2	/* Convert byte-count to character-count.  */
Packit Service 8bbac7
	/* Restore registers.  */
Packit Service 8bbac7
	vlgvg	%r6,%v30,0
Packit Service 8bbac7
	vlgvg	%r8,%v31,0
Packit Service 8bbac7
	vlgvg	%r9,%v31,1
Packit Service 8bbac7
	br	%r14
Packit Service 8bbac7
.Lfallback:
Packit Service 8bbac7
	jg	__wcscspn_c
Packit Service 8bbac7
END(__wcscspn_vx)
Packit Service 8bbac7
#endif /* HAVE_S390_VX_ASM_SUPPORT && IS_IN (libc) */