#include <sysdep.h>
-#ifndef STRLEN
-# define STRLEN __strlen
-# define DEFINE_STRLEN_HIDDEN_DEF 1
-#endif
+/* To reuse the code for rawmemchr, we have some extra steps compared to the
+ strlen implementation:
+ - Sum the initial value of r3 with the position at which the char was
+ found, to guarantee we return a pointer and not the length.
+ - In the main loop, subtract each byte by the char we are looking for,
+ so we can keep using vminub to quickly check 64B at once. */
+#ifdef USE_AS_RAWMEMCHR
+# ifndef RAWMEMCHR
+# define FUNCNAME __rawmemchr
+# else
+# define FUNCNAME RAWMEMCHR
+# endif
+# define MCOUNT_NARGS 2
+# define VREG_ZERO v20
+# define OFF_START_LOOP 256
+# define RAWMEMCHR_SUBTRACT_VECTORS \
+ vsububm v4,v4,v18; \
+ vsububm v5,v5,v18; \
+ vsububm v6,v6,v18; \
+ vsububm v7,v7,v18;
+# define TAIL(vreg,increment) \
+ vctzlsbb r4,vreg; \
+ addi r4,r4,increment; \
+ add r3,r5,r4; \
+ blr
+
+#else /* strlen */
+
+# ifndef STRLEN
+# define FUNCNAME __strlen
+# define DEFINE_STRLEN_HIDDEN_DEF 1
+# else
+# define FUNCNAME STRLEN
+# endif
+# define MCOUNT_NARGS 1
+# define VREG_ZERO v18
+# define OFF_START_LOOP 192
+# define TAIL(vreg,increment) \
+ vctzlsbb r4,vreg; \
+ subf r3,r3,r5; \
+ addi r4,r4,increment; \
+ add r3,r3,r4; \
+ blr
+#endif /* USE_AS_RAWMEMCHR */
/* TODO: Replace macros by the actual instructions when minimum binutils becomes
>= 2.35. This is used to keep compatibility with older versions. */
li r6,offset; \
LXVP(v4+32,offset,addr); \
LXVP(v6+32,offset+32,addr); \
+ RAWMEMCHR_SUBTRACT_VECTORS; \
vminub v14,v4,v5; \
vminub v15,v6,v7; \
vminub v16,v14,v15; \
- vcmpequb. v0,v16,v18; \
+ vcmpequb. v0,v16,VREG_ZERO; \
bne cr6,L(label)
-#define TAIL(vreg,increment) \
- vctzlsbb r4,vreg; \
- subf r3,r3,r5; \
- addi r4,r4,increment; \
- add r3,r3,r4; \
- blr
-
/* Implements the function
int [r3] strlen (const void *s [r3])
+ but when USE_AS_RAWMEMCHR is set, implements the function
+
+ void* [r3] rawmemchr (const void *s [r3], int c [r4])
+
The implementation can load bytes past a matching byte, but only
up to the next 64B boundary, so it never crosses a page. */
.machine power9
-ENTRY_TOCLESS (STRLEN, 4)
- CALL_MCOUNT 1
+ENTRY_TOCLESS (FUNCNAME, 4)
+ CALL_MCOUNT MCOUNT_NARGS
- vspltisb v18,0
+#ifdef USE_AS_RAWMEMCHR
+ xori r5,r4,0xff
+
+ mtvsrd v18+32,r4 /* matching char in v18 */
+ mtvsrd v19+32,r5 /* non matching char in v19 */
+
+ vspltb v18,v18,7 /* replicate */
+ vspltb v19,v19,7 /* replicate */
+#else
vspltisb v19,-1
+#endif
+ vspltisb VREG_ZERO,0
/* Next 16B-aligned address. Prepare address for L(aligned). */
addi r5,r3,16
vcmpequb. v6,v0,v18
beq cr6,L(aligned)
+#ifdef USE_AS_RAWMEMCHR
+ vctzlsbb r6,v6
+ add r3,r3,r6
+#else
vctzlsbb r3,v6
+#endif
blr
- /* Test next 176B, 16B at a time. The main loop is optimized for longer
- strings, so checking the first bytes in 16B chunks benefits a lot
- small strings. */
+ /* Test up to OFF_START_LOOP-16 bytes in 16B chunks. The main loop is
+ optimized for longer strings, so checking the first bytes in 16B
+ chunks benefits a lot small strings. */
.p2align 5
L(aligned):
+#ifdef USE_AS_RAWMEMCHR
+ cmpdi cr5,r4,0 /* Check if c == 0. This will be useful to
+ choose how we will perform the main loop. */
+#endif
/* Prepare address for the loop. */
- addi r4,r3,192
+ addi r4,r3,OFF_START_LOOP
clrrdi r4,r4,6
CHECK16(v0,0,r5,tail1)
CHECK16(v8,128,r5,tail9)
CHECK16(v9,144,r5,tail10)
CHECK16(v10,160,r5,tail11)
+#ifdef USE_AS_RAWMEMCHR
+ CHECK16(v0,176,r5,tail12)
+ CHECK16(v1,192,r5,tail13)
+ CHECK16(v2,208,r5,tail14)
+ CHECK16(v3,224,r5,tail15)
+#endif
addi r5,r4,128
+#ifdef USE_AS_RAWMEMCHR
+ /* If c == 0, use the same loop as strlen, without the vsububm. */
+ beq cr5,L(loop)
+
+ /* This is very similar to the block after L(loop), the difference is
+ that here RAWMEMCHR_SUBTRACT_VECTORS is not empty, and we subtract
+ each byte loaded by the char we are looking for, this way we can keep
+ using vminub to merge the results and checking for nulls. */
+ .p2align 5
+L(rawmemchr_loop):
+ CHECK64(0,r4,pre_tail_64b)
+ CHECK64(64,r4,pre_tail_64b)
+ addi r4,r4,256
+
+ CHECK64(0,r5,tail_64b)
+ CHECK64(64,r5,tail_64b)
+ addi r5,r5,256
+
+ b L(rawmemchr_loop)
+#endif
/* Switch to a more aggressive approach checking 64B each time. Use 2
pointers 128B apart and unroll the loop once to make the pointer
updates and usages separated enough to avoid stalls waiting for
address calculation. */
.p2align 5
L(loop):
+#undef RAWMEMCHR_SUBTRACT_VECTORS
+#define RAWMEMCHR_SUBTRACT_VECTORS /* nothing */
CHECK64(0,r4,pre_tail_64b)
CHECK64(64,r4,pre_tail_64b)
addi r4,r4,256
block and mark it in its corresponding VR. lxvp vx,0(ry) puts the
low 16B bytes into vx+1, and the high into vx, so the order here is
v5, v4, v7, v6. */
- vcmpequb v1,v5,v18
- vcmpequb v2,v4,v18
- vcmpequb v3,v7,v18
- vcmpequb v4,v6,v18
+ vcmpequb v1,v5,VREG_ZERO
+ vcmpequb v2,v4,VREG_ZERO
+ vcmpequb v3,v7,VREG_ZERO
+ vcmpequb v4,v6,VREG_ZERO
/* Take into account the other 64B blocks we had already checked. */
add r5,r5,r6
or r10,r8,r7
cnttzd r0,r10 /* Count trailing zeros before the match. */
+#ifndef USE_AS_RAWMEMCHR
subf r5,r3,r5
+#endif
add r3,r5,r0 /* Compute final length. */
blr
L(tail11):
TAIL(v10,160)
-END (STRLEN)
+#ifdef USE_AS_RAWMEMCHR
+ .p2align 5
+L(tail12):
+ TAIL(v0,176)
+
+ .p2align 5
+L(tail13):
+ TAIL(v1,192)
+
+ .p2align 5
+L(tail14):
+ TAIL(v2,208)
+
+ .p2align 5
+L(tail15):
+ TAIL(v3,224)
+#endif
+
+END (FUNCNAME)
-#ifdef DEFINE_STRLEN_HIDDEN_DEF
+#ifdef USE_AS_RAWMEMCHR
+weak_alias (__rawmemchr,rawmemchr)
+libc_hidden_builtin_def (__rawmemchr)
+#else
+# ifdef DEFINE_STRLEN_HIDDEN_DEF
weak_alias (__strlen, strlen)
libc_hidden_builtin_def (strlen)
+# endif
#endif