xnu-344.tar.gz

[apple/xnu.git] / osfmk / ppc / bzero.s

diff --git a/osfmk/ppc/bzero.s b/osfmk/ppc/bzero.s
index ac0afddf99dce9aa181c40584ef7a38f0748592b..2faf894a06e6e29cc4277620f7c1f54b8bc1c475 100644 (file)
--- a/osfmk/ppc/bzero.s
+++ b/osfmk/ppc/bzero.s
@@ -169,3 +169,119 @@ ENTRY(memset, TAG_NO_FRAME_USED)
        /* Return original address in ARG0 */
        
        blr
+
+/*
+ *      void   bzero_nc(char *addr, unsigned int length)
+ *
+ * bzero implementation for PowerPC
+ *   - assumes non-pic code
+ *
+ * returns start address in r3, as per memset (called by memset)
+ */    
+       
+ENTRY(bzero_nc, TAG_NO_FRAME_USED)
+
+       cmpwi   cr0,    r4,     0 /* no bytes to zero? */
+       mr      r7,     r3
+       mr      r8,     r3      /* use r8 as counter to where we are */
+       beqlr-
+       cmpwi   cr0,    r4,     CACHE_LINE_SIZE /* clear less than a block? */
+       li      r0,     0        /* use r0 as source of zeros */
+       blt     .L_bzeroNCEndWord
+
+/* first, clear bytes up to the next word boundary */
+       addis   r6,     0,      HIGH_CADDR(.L_bzeroNCBeginWord)
+       addi    r6,     r6,     LOW_ADDR(.L_bzeroNCBeginWord)
+                /* extract byte offset as word offset */
+       rlwinm. r5,     r8,     2,      28,     29
+       addi    r8,     r8,     -1 /* adjust for update */
+       beq     .L_bzeroNCBeginWord /* no bytes to zero */
+       subfic  r5,     r5,     16 /* compute the number of instructions */
+       sub     r6,     r6,     r5 /* back from word clear to execute */
+       mtctr   r6
+       bctr
+
+       stbu    r0,     1(r8)
+       stbu    r0,     1(r8)
+       stbu    r0,     1(r8)
+
+/* clear words up to the next block boundary */
+.L_bzeroNCBeginWord:
+       addis   r6,     0,      HIGH_CADDR(.L_bzeroNCBlock)
+       addi    r6,     r6,     LOW_ADDR(.L_bzeroNCBlock)
+       addi    r8,     r8,     1
+       rlwinm. r5,     r8,     0,      27,     29 /* extract word offset */
+       addi    r8,     r8,     -4              /* adjust for update */
+       beq     .L_bzeroNCBlock                 /* no words to zero */
+               /* compute the number of instructions */
+       subfic  r5,     r5,     CACHE_LINE_SIZE
+       sub     r6,     r6,     r5 /* back from word clear to execute */
+       mtctr   r6
+       bctr
+
+       stwu    r0,     4(r8)
+       stwu    r0,     4(r8)
+       stwu    r0,     4(r8)
+       stwu    r0,     4(r8)
+       stwu    r0,     4(r8)
+       stwu    r0,     4(r8)
+       stwu    r0,     4(r8)
+
+ /* clear cache blocks */
+.L_bzeroNCBlock:
+       addi    r8,     r8,     4 /* remove update adjust */
+       sub     r5,     r8,     r7 /* bytes zeroed */
+       sub     r4,     r4,     r5
+       srwi.   r5,     r4,     CACHE_LINE_POW2 /* blocks to zero */
+       beq     .L_bzeroNCEndWord
+       mtctr   r5
+
+.L_bzeroNCBlock1:
+       stw     r0,     0(r8)
+       stw     r0,     4(r8)
+       stw     r0,     8(r8)
+       stw     r0,     12(r8)
+       stw     r0,     16(r8)
+       stw     r0,     20(r8)
+       stw     r0,     24(r8)
+       stw     r0,     28(r8)
+       addi    r8,     r8,     CACHE_LINE_SIZE
+       bdnz    .L_bzeroNCBlock1
+
+ /* clear remaining words */
+.L_bzeroNCEndWord:
+       addis   r6,     0,      HIGH_CADDR(.L_bzeroNCEndByte)
+       addi    r6,     r6,     LOW_ADDR(.L_bzeroNCEndByte)
+       rlwinm. r5,     r4,     0,      27,     29 /* extract word offset */
+       addi    r8,     r8,     -4                 /* adjust for update */
+       beq     .L_bzeroNCEndByte                          /* no words to zero */
+       sub     r6,     r6,     r5 /* back from word clear to execute */
+       mtctr   r6
+       bctr
+
+       stwu    r0,     4(r8)
+       stwu    r0,     4(r8)
+       stwu    r0,     4(r8)
+       stwu    r0,     4(r8)
+       stwu    r0,     4(r8)
+       stwu    r0,     4(r8)
+       stwu    r0,     4(r8)
+
+ /* clear remaining bytes */
+.L_bzeroNCEndByte:
+       addis   r6,     0,      HIGH_CADDR(.L_bzeroNCEnd)
+       addi    r6,     r6,     LOW_ADDR(.L_bzeroNCEnd)
+               /* extract byte offset as word offset */
+       rlwinm. r5,     r4,     2,      28,     29
+       addi    r8,     r8,     3 /* adjust for update */
+       beqlr
+       sub     r6,     r6,     r5 /* back from word clear to execute */
+       mtctr   r6
+       bctr
+
+       stbu    r0,     1(r8)
+       stbu    r0,     1(r8)
+       stbu    r0,     1(r8)
+
+.L_bzeroNCEnd:
+       blr