github.com/sandwichdev/go-internals@v0.0.0-20210605002614-12311ac6b2c5/bytealg/indexbyte_s390x.s (about)

     1  // Copyright 2018 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  #include "go_asm.h"
     6  #include "textflag.h"
     7  
     8  TEXT ·IndexByte(SB),NOSPLIT|NOFRAME,$0-40
     9  	MOVD	b_base+0(FP), R3// b_base => R3
    10  	MOVD	b_len+8(FP), R4 // b_len => R4
    11  	MOVBZ	c+24(FP), R5    // c => R5
    12  	MOVD	$ret+32(FP), R2 // &ret => R9
    13  	BR	indexbytebody<>(SB)
    14  
    15  TEXT ·IndexByteString(SB),NOSPLIT|NOFRAME,$0-32
    16  	MOVD	s_base+0(FP), R3// s_base => R3
    17  	MOVD	s_len+8(FP), R4 // s_len => R4
    18  	MOVBZ	c+16(FP), R5    // c => R5
    19  	MOVD	$ret+24(FP), R2 // &ret => R9
    20  	BR	indexbytebody<>(SB)
    21  
    22  // input:
    23  // R3: s
    24  // R4: s_len
    25  // R5: c -- byte sought
    26  // R2: &ret -- address to put index into
    27  TEXT indexbytebody<>(SB),NOSPLIT|NOFRAME,$0
    28  	CMPBEQ	R4, $0, notfound
    29  	MOVD	R3, R6          // store base for later
    30  	ADD	R3, R4, R8      // the address after the end of the string
    31  	//if the length is small, use loop; otherwise, use vector or srst search
    32  	CMPBGE	R4, $16, large
    33  
    34  residual:
    35  	CMPBEQ	R3, R8, notfound
    36  	MOVBZ	0(R3), R7
    37  	LA	1(R3), R3
    38  	CMPBNE	R7, R5, residual
    39  
    40  found:
    41  	SUB	R6, R3
    42  	SUB	$1, R3
    43  	MOVD	R3, 0(R2)
    44  	RET
    45  
    46  notfound:
    47  	MOVD	$-1, 0(R2)
    48  	RET
    49  
    50  large:
    51  	MOVBZ	internal∕cpu·S390X+const_offsetS390xHasVX(SB), R1
    52  	CMPBNE	R1, $0, vectorimpl
    53  
    54  srstimpl:                       // no vector facility
    55  	MOVBZ	R5, R0          // c needs to be in R0, leave until last minute as currently R0 is expected to be 0
    56  srstloop:
    57  	WORD	$0xB25E0083     // srst %r8, %r3 (search the range [R3, R8))
    58  	BVS	srstloop        // interrupted - continue
    59  	BGT	notfoundr0
    60  foundr0:
    61  	XOR	R0, R0          // reset R0
    62  	SUB	R6, R8          // remove base
    63  	MOVD	R8, 0(R2)
    64  	RET
    65  notfoundr0:
    66  	XOR	R0, R0          // reset R0
    67  	MOVD	$-1, 0(R2)
    68  	RET
    69  
    70  vectorimpl:
    71  	//if the address is not 16byte aligned, use loop for the header
    72  	MOVD	R3, R8
    73  	AND	$15, R8
    74  	CMPBGT	R8, $0, notaligned
    75  
    76  aligned:
    77  	ADD	R6, R4, R8
    78  	MOVD	R8, R7
    79  	AND	$-16, R7
    80  	// replicate c across V17
    81  	VLVGB	$0, R5, V19
    82  	VREPB	$0, V19, V17
    83  
    84  vectorloop:
    85  	CMPBGE	R3, R7, residual
    86  	VL	0(R3), V16    // load string to be searched into V16
    87  	ADD	$16, R3
    88  	VFEEBS	V16, V17, V18 // search V17 in V16 and set conditional code accordingly
    89  	BVS	vectorloop
    90  
    91  	// when vector search found c in the string
    92  	VLGVB	$7, V18, R7   // load 7th element of V18 containing index into R7
    93  	SUB	$16, R3
    94  	SUB	R6, R3
    95  	ADD	R3, R7
    96  	MOVD	R7, 0(R2)
    97  	RET
    98  
    99  notaligned:
   100  	MOVD	R3, R8
   101  	AND	$-16, R8
   102  	ADD     $16, R8
   103  notalignedloop:
   104  	CMPBEQ	R3, R8, aligned
   105  	MOVBZ	0(R3), R7
   106  	LA	1(R3), R3
   107  	CMPBNE	R7, R5, notalignedloop
   108  	BR	found