/*
* arch/alpha/lib/strrchr.S
* Contributed by Richard Henderson (rth@tamu.edu)
*
* Return the address of the last occurrence of a given character
* within a null-terminated string, or null if it is not found.
*/
#include <asm/regdef.h>
.set noreorder
.set noat
.align 3
.ent strrchr
.globl strrchr
strrchr:
.frame sp, 0, ra
.prologue 0
zapnot a1, 1, a1 # e0 : zero extend our test character
mov zero, t6 # .. e1 : t6 is last match aligned addr
sll a1, 8, t5 # e0 : replicate our test character
mov zero, t8 # .. e1 : t8 is last match byte compare mask
or t5, a1, a1 # e0 :
ldq_u t0, 0(a0) # .. e1 : load first quadword
sll a1, 16, t5 # e0 :
andnot a0, 7, v0 # .. e1 : align source addr
or t5, a1, a1 # e0 :
lda t4, -1 # .. e1 : build garbage mask
sll a1, 32, t5 # e0 :
cmpbge zero, t0, t1 # .. e1 : bits set iff byte == zero
mskqh t4, a0, t4 # e0 :
or t5, a1, a1 # .. e1 : character replication complete
xor t0, a1, t2 # e0 : make bytes == c zero
cmpbge zero, t4, t4 # .. e1 : bits set iff byte is garbage
cmpbge zero, t2, t3 # e0 : bits set iff byte == c
andnot t1, t4, t1 # .. e1 : clear garbage from null test
andnot t3, t4, t3 # e0 : clear garbage from char test
bne t1, $eos # .. e1 : did we already hit the terminator?
/* Character search main loop */
$loop:
ldq t0, 8(v0) # e0 : load next quadword
cmovne t3, v0, t6 # .. e1 : save previous comparisons match
cmovne t3, t3, t8 # e0 :
addq v0, 8, v0 # .. e1 :
xor t0, a1, t2 # e0 :
cmpbge zero, t0, t1 # .. e1 : bits set iff byte == zero
cmpbge zero, t2, t3 # e0 : bits set iff byte == c
beq t1, $loop # .. e1 : if we havnt seen a null, loop
/* Mask out character matches after terminator */
$eos:
negq t1, t4 # e0 : isolate first null byte match
and t1, t4, t4 # e1 :
subq t4, 1, t5 # e0 : build a mask of the bytes upto...
or t4, t5, t4 # e1 : ... and including the null
and t3, t4, t3 # e0 : mask out char matches after null
cmovne t3, t3, t8 # .. e1 : save it, if match found
cmovne t3, v0, t6 # e0 :
/* Locate the address of the last matched character */
/* Retain the early exit for the ev4 -- the ev5 mispredict penalty
is 5 cycles -- the same as just falling through. */
beq t8, $retnull # .. e1 :
and t8, 0xf0, t2 # e0 : binary search for the high bit set
cmovne t2, t2, t8 # .. e1 (zdb)
cmovne t2, 4, t2 # e0 :
and t8, 0xcc, t1 # .. e1 :
cmovne t1, t1, t8 # e0 :
cmovne t1, 2, t1 # .. e1 :
and t8, 0xaa, t0 # e0 :
cmovne t0, 1, t0 # .. e1 (zdb)
addq t2, t1, t1 # e0 :
addq t6, t0, v0 # .. e1 : add our aligned base ptr to the mix
addq v0, t1, v0 # e0 :
ret # .. e1 :
$retnull:
mov zero, v0 # e0 :
ret # .. e1 :
.end strrchr
