Loading...
--- Libc/Libc-391.2.3/ppc/string/strcmp.s
+++ Libc/Libc-320.1.3/ppc/string/strcmp.s
@@ -24,17 +24,6 @@
#include <mach/ppc/asm.h>
#undef ASSEMBLER
-#define __APPLE_API_PRIVATE
-#include <machine/cpu_capabilities.h>
-#undef __APPLE_API_PRIVATE
-
-/* We use mode-independent "g" opcodes such as "srgi". These expand
- * into word operations when targeting __ppc__, and into doubleword
- * operations when targeting __ppc64__.
- */
-#include <architecture/ppc/mode_independent_asm.h>
-
-
// ***************
// * S T R C M P *
// ***************
@@ -55,36 +44,31 @@
// The test maps any non-zero byte to zero, and any zero byte to 0x80,
// with one exception: 0x01 bytes preceeding the first zero are also
// mapped to 0x80.
-//
-// In 64-bit mode, the algorithm is doubleword parallel.
.text
.globl EXT(strcmp)
.align 5
-LEXT(strcmp) // int strcmp(const char *s1, const char *s2);
- andi. r0,r3,GPR_BYTES-1 // is LHS aligned?
-#if defined(__ppc__)
- lis r5,hi16(0xFEFEFEFF) // start to generate 32-bit magic constants
+LEXT(strcmp) // int strcmp(const char *s1, const char *s2);
+ andi. r0,r3,3 // is LHS aligned?
+ dcbt 0,r3 // touch in LHS
+ lis r5,hi16(0xFEFEFEFF) // start to load magic constants
lis r6,hi16(0x80808080)
+ dcbt 0,r4 // touch in RHS
ori r5,r5,lo16(0xFEFEFEFF)
ori r6,r6,lo16(0x80808080)
-#else
- ld r5,_COMM_PAGE_MAGIC_FE(0) // get 0xFEFEFEFE FEFEFEFF from commpage
- ld r6,_COMM_PAGE_MAGIC_80(0) // get 0x80808080 80808080 from commpage
-#endif
- subi r3,r3,GPR_BYTES // we use "lgu" in the inner loops
- subi r4,r4,GPR_BYTES
+ subi r3,r3,4 // we use "lwzu" in the inner loops
+ subi r4,r4,4
beq Laligned // LHS is aligned
- subfic r0,r0,GPR_BYTES // r0 <- #bytes to align LHS
+ subfic r0,r0,4 // r0 <- #bytes to word align LHS
mtctr r0
// Loop over bytes.
Lbyteloop:
- lbz r7,GPR_BYTES(r3) // r7 <- next LHS byte
+ lbz r7,4(r3) // r7 <- next LHS byte
addi r3,r3,1
- lbz r8,GPR_BYTES(r4) // r8 <- next RHS byte
+ lbz r8,4(r4) // r8 <- next RHS byte
addi r4,r4,1
cntlzw r9,r7 // is r7 zero?
sub r0,r7,r8 // different?
@@ -94,38 +78,37 @@
bne Ldone // done if different or 0
-// LHS is aligned. If RHS also is, we need not worry about page
-// crossing. Otherwise, we must stop the loop before page is crossed.
+// LHS is word aligned. If RHS also is, we need not worry about page
+// crossing. Otherwise, we must stop the word loop before page is crossed.
Laligned:
- andi. r0,r4,GPR_BYTES-1 // is RHS now aligned too?
- addi r9,r4,GPR_BYTES // restore true address of next RHS byte
+ andi. r0,r4,3 // is RHS now word aligned too?
+ addi r9,r4,4 // restore true address of next RHS byte
rlwinm r9,r9,0,0xFFF // get RHS offset in page
- beq Lalignedloop // RHS also aligned, use simple loop
+ beq Lalignedloop // RHS word aligned, use simple loop
subfic r9,r9,4096 // get #bytes left in RHS page
- srwi. r0,r9,LOG2_GPR_BYTES// get #words or doublewords left in RHS page
+ srwi. r0,r9,2 // get #words left in RHS page
mtctr r0 // set up loop count
bne++ Lunalignedloop // at least one word left in RHS page
- li r0,GPR_BYTES // must check GPR_BYTES, a byte at a time...
- mtctr r0 // ...in order to keep LHS aligned
+ li r0,4 // must check 4 bytes, a byte at a time...
+ mtctr r0 // ...in order to keep LHS word aligned
b Lbyteloop // go cross RHS page
-// Unaligned inner loop: compare a word or doubleword at a time, until one of
-// three conditions:
+// Unaligned inner loop: compare a word at a time, until one of three conditions:
// - a difference is found
// - a zero byte is found
// - end of RHS page (we dare not touch next page until we must)
// At this point, registers are as follows:
-// r3 = LHS ptr - GPR_BYTES (aligned)
-// r4 = RHS ptr - GPR_BYTES (not aligned)
+// r3 = LHS ptr - 4 (word aligned)
+// r4 = RHS ptr - 4 (not aligned)
// r5 = 0xFEFEFEFF
// r6 = 0x80808080
-// ctr = words or doublewords left in RHS page
+// ctr = whole words left in RHS page
.align 5 // align inner loop, which is 8 words long
Lunalignedloop:
- lgu r7,GPR_BYTES(r3) // r7 <- next LHS bytes
- lgu r8,GPR_BYTES(r4) // r8 <- next RHS bytes
+ lwzu r7,4(r3) // r7 <- next 4 LHS bytes
+ lwzu r8,4(r4) // r8 <- next 4 RHS bytes
add r10,r7,r5 // r10 <- LHS + 0xFEFEFEFF
andc r12,r6,r7 // r12 <- ~LHS & 0x80808080
xor r11,r7,r8 // r11 <- compare the words
@@ -134,7 +117,7 @@
bdnzt eq,Lunalignedloop // loop if ctr!=0 and cr0_eq
bne++ Ldifferent // done if we found a 0 or difference
- li r0,GPR_BYTES // must check GPR_BYTES, a byte at a time...
+ li r0,4 // must check 4 bytes, a byte at a time...
mtctr r0 // ...in order to keep LHS word aligned
b Lbyteloop // cross RHS page, then resume word loop
@@ -149,8 +132,8 @@
.align 5 // align inner loop, which is 8 words ling
Lalignedloop:
- lgu r7,GPR_BYTES(r3) // r7 <- next LHS bytes
- lgu r8,GPR_BYTES(r4) // r8 <- next RHS bytes
+ lwzu r7,4(r3) // r7 <- next 4 LHS bytes
+ lwzu r8,4(r4) // r8 <- next 4 RHS bytes
add r10,r7,r5 // r10 <- LHS + 0xFEFEFEFF
andc r12,r6,r7 // r12 <- ~LHS & 0x80808080
xor r11,r7,r8 // r11 <- compare the words
@@ -164,21 +147,18 @@
// the 0-byte.
Ldifferent: // r0 == 0-test vector (with 0x01 false hits)
- slgi r9,r7,7 // move 0x01 bits in LHS into position 0x80
+ slwi r9,r7,7 // move 0x01 bits in LHS into position 0x80
andc r0,r0,r9 // mask out the false 0-hits from 0x01 bytes
or r11,r11,r0 // recompute difference vector
- cntlzg r9,r11 // find 1st difference (r9 = 0..31 or 63)
- rlwinm r9,r9,0,0x38 // byte align bit offset (now, r9 = 0,8,16, or 24 etc)
- addi r0,r9,8 // now, r0 = 8, 16, 24, or 32
-#if defined(__ppc__)
- rlwnm r7,r7,r0,24,31 // right justify differing bytes and mask off rest
- rlwnm r8,r8,r0,24,31
-#else
- rldcl r7,r7,r0,56 // right justify differing bytes and mask off rest
- rldcl r8,r8,r0,56
-#endif
-
-Ldone: // differing bytes are in r7 and r8
- sub r3,r7,r8 // compute difference (0, +, or -)
+ cntlzw r9,r11 // find 1st difference (r9 = 0..31)
+ rlwinm r9,r9,0,0x18 // byte align bit offset (now, r9 = 0,8,16, or 24)
+ addi r9,r9,8 // now, r9 = 8, 16, 24, or 32
+ rlwnm r5,r7,r9,24,31 // right justify differing bytes and mask off rest
+ rlwnm r6,r8,r9,24,31
+ sub r3,r5,r6 // compute difference (0, +, or -)
blr
+Ldone: // r0 = return value
+ mr r3,r0 // return in r3
+ blr
+