1*4882a593Smuzhiyun/* SPDX-License-Identifier: GPL-2.0 */ 2*4882a593Smuzhiyun/* 3*4882a593Smuzhiyun * arch/alpha/lib/ev67-strchr.S 4*4882a593Smuzhiyun * 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com> 5*4882a593Smuzhiyun * 6*4882a593Smuzhiyun * Return the address of a given character within a null-terminated 7*4882a593Smuzhiyun * string, or null if it is not found. 8*4882a593Smuzhiyun * 9*4882a593Smuzhiyun * Much of the information about 21264 scheduling/coding comes from: 10*4882a593Smuzhiyun * Compiler Writer's Guide for the Alpha 21264 11*4882a593Smuzhiyun * abbreviated as 'CWG' in other comments here 12*4882a593Smuzhiyun * ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html 13*4882a593Smuzhiyun * Scheduling notation: 14*4882a593Smuzhiyun * E - either cluster 15*4882a593Smuzhiyun * U - upper subcluster; U0 - subcluster U0; U1 - subcluster U1 16*4882a593Smuzhiyun * L - lower subcluster; L0 - subcluster L0; L1 - subcluster L1 17*4882a593Smuzhiyun * Try not to change the actual algorithm if possible for consistency. 18*4882a593Smuzhiyun */ 19*4882a593Smuzhiyun#include <asm/export.h> 20*4882a593Smuzhiyun#include <asm/regdef.h> 21*4882a593Smuzhiyun 22*4882a593Smuzhiyun .set noreorder 23*4882a593Smuzhiyun .set noat 24*4882a593Smuzhiyun 25*4882a593Smuzhiyun .align 4 26*4882a593Smuzhiyun .globl strchr 27*4882a593Smuzhiyun .ent strchr 28*4882a593Smuzhiyunstrchr: 29*4882a593Smuzhiyun .frame sp, 0, ra 30*4882a593Smuzhiyun .prologue 0 31*4882a593Smuzhiyun 32*4882a593Smuzhiyun ldq_u t0, 0(a0) # L : load first quadword Latency=3 33*4882a593Smuzhiyun and a1, 0xff, t3 # E : 00000000000000ch 34*4882a593Smuzhiyun insbl a1, 1, t5 # U : 000000000000ch00 35*4882a593Smuzhiyun insbl a1, 7, a2 # U : ch00000000000000 36*4882a593Smuzhiyun 37*4882a593Smuzhiyun insbl t3, 6, a3 # U : 00ch000000000000 38*4882a593Smuzhiyun or t5, t3, a1 # E : 000000000000chch 39*4882a593Smuzhiyun andnot a0, 7, v0 # E : align our loop pointer 40*4882a593Smuzhiyun lda t4, -1 # E : build garbage mask 41*4882a593Smuzhiyun 42*4882a593Smuzhiyun mskqh t4, a0, t4 # U : only want relevant part of first quad 43*4882a593Smuzhiyun or a2, a3, a2 # E : chch000000000000 44*4882a593Smuzhiyun inswl a1, 2, t5 # E : 00000000chch0000 45*4882a593Smuzhiyun inswl a1, 4, a3 # E : 0000chch00000000 46*4882a593Smuzhiyun 47*4882a593Smuzhiyun or a1, a2, a1 # E : chch00000000chch 48*4882a593Smuzhiyun or a3, t5, t5 # E : 0000chchchch0000 49*4882a593Smuzhiyun cmpbge zero, t0, t2 # E : bits set iff byte == zero 50*4882a593Smuzhiyun cmpbge zero, t4, t4 # E : bits set iff byte is garbage 51*4882a593Smuzhiyun 52*4882a593Smuzhiyun /* This quad is _very_ serialized. Lots of stalling happens */ 53*4882a593Smuzhiyun or t5, a1, a1 # E : chchchchchchchch 54*4882a593Smuzhiyun xor t0, a1, t1 # E : make bytes == c zero 55*4882a593Smuzhiyun cmpbge zero, t1, t3 # E : bits set iff byte == c 56*4882a593Smuzhiyun or t2, t3, t0 # E : bits set iff char match or zero match 57*4882a593Smuzhiyun 58*4882a593Smuzhiyun andnot t0, t4, t0 # E : clear garbage bits 59*4882a593Smuzhiyun cttz t0, a2 # U0 : speculative (in case we get a match) 60*4882a593Smuzhiyun nop # E : 61*4882a593Smuzhiyun bne t0, $found # U : 62*4882a593Smuzhiyun 63*4882a593Smuzhiyun /* 64*4882a593Smuzhiyun * Yuk. This loop is going to stall like crazy waiting for the 65*4882a593Smuzhiyun * data to be loaded. Not much can be done about it unless it's 66*4882a593Smuzhiyun * unrolled multiple times - is that safe to do in kernel space? 67*4882a593Smuzhiyun * Or would exception handling recovery code do the trick here? 68*4882a593Smuzhiyun */ 69*4882a593Smuzhiyun$loop: ldq t0, 8(v0) # L : Latency=3 70*4882a593Smuzhiyun addq v0, 8, v0 # E : 71*4882a593Smuzhiyun xor t0, a1, t1 # E : 72*4882a593Smuzhiyun cmpbge zero, t0, t2 # E : bits set iff byte == 0 73*4882a593Smuzhiyun 74*4882a593Smuzhiyun cmpbge zero, t1, t3 # E : bits set iff byte == c 75*4882a593Smuzhiyun or t2, t3, t0 # E : 76*4882a593Smuzhiyun cttz t3, a2 # U0 : speculative (in case we get a match) 77*4882a593Smuzhiyun beq t0, $loop # U : 78*4882a593Smuzhiyun 79*4882a593Smuzhiyun$found: negq t0, t1 # E : clear all but least set bit 80*4882a593Smuzhiyun and t0, t1, t0 # E : 81*4882a593Smuzhiyun and t0, t3, t1 # E : bit set iff byte was the char 82*4882a593Smuzhiyun addq v0, a2, v0 # E : Add in the bit number from above 83*4882a593Smuzhiyun 84*4882a593Smuzhiyun cmoveq t1, $31, v0 # E : Two mapping slots, latency = 2 85*4882a593Smuzhiyun nop 86*4882a593Smuzhiyun nop 87*4882a593Smuzhiyun ret # L0 : 88*4882a593Smuzhiyun 89*4882a593Smuzhiyun .end strchr 90*4882a593Smuzhiyun EXPORT_SYMBOL(strchr) 91