patch-2.1.30 linux/arch/sparc64/lib/checksum.S
Next file: linux/arch/sparc64/lib/locks.S
Previous file: linux/arch/sparc64/lib/Makefile
Back to the patch index
Back to the overall index
- Lines: 568
- Date:
Thu Mar 20 16:43:32 1997
- Orig file:
v2.1.29/linux/arch/sparc64/lib/checksum.S
- Orig date:
Wed Dec 31 16:00:00 1969
diff -u --recursive --new-file v2.1.29/linux/arch/sparc64/lib/checksum.S linux/arch/sparc64/lib/checksum.S
@@ -0,0 +1,567 @@
+/* checksum.S: Sparc V9 optimized checksum code.
+ *
+ * Copyright(C) 1995 Linus Torvalds
+ * Copyright(C) 1995 Miguel de Icaza
+ * Copyright(C) 1996 David S. Miller
+ * Copyright(C) 1997 Jakub Jelinek
+ *
+ * derived from:
+ * Linux/Alpha checksum c-code
+ * Linux/ix86 inline checksum assembly
+ * RFC1071 Computing the Internet Checksum (esp. Jacobsons m68k code)
+ * David Mosberger-Tang for optimized reference c-code
+ * BSD4.4 portable checksum routine
+ */
+
+#include <asm/errno.h>
+#include <asm/head.h>
+
+#define CSUM_BIGCHUNK(buf, offset, sum, t0, t1, t2, t3, t4, t5) \
+ ldd [buf + offset + 0x00], t0; \
+ ldd [buf + offset + 0x08], t2; \
+ addccc t0, sum, sum; \
+ addccc t1, sum, sum; \
+ ldd [buf + offset + 0x10], t4; \
+ addccc t2, sum, sum; \
+ addccc t3, sum, sum; \
+ ldd [buf + offset + 0x18], t0; \
+ addccc t4, sum, sum; \
+ addccc t5, sum, sum; \
+ addccc t0, sum, sum; \
+ addccc t1, sum, sum;
+
+#define CSUM_LASTCHUNK(buf, offset, sum, t0, t1, t2, t3) \
+ ldd [buf - offset - 0x08], t0; \
+ ldd [buf - offset - 0x00], t2; \
+ addccc t0, sum, sum; \
+ addccc t1, sum, sum; \
+ addccc t2, sum, sum; \
+ addccc t3, sum, sum;
+
+ /* Do end cruft out of band to get better cache patterns. */
+csum_partial_end_cruft:
+ andcc %o1, 8, %g0 ! check how much
+ be,pn %icc, 1f ! caller asks %o1 & 0x8
+ and %o1, 4, %g3 ! nope, check for word remaining
+ ldd [%o0], %g2 ! load two
+ addcc %g2, %o2, %o2 ! add first word to sum
+ addccc %g3, %o2, %o2 ! add second word as well
+ add %o0, 8, %o0 ! advance buf ptr
+ addc %g0, %o2, %o2 ! add in final carry
+1: brz,pn %g3, 1f ! nope, skip this code
+ andcc %o1, 3, %o1 ! check for trailing bytes
+ ld [%o0], %g2 ! load it
+ addcc %g2, %o2, %o2 ! add to sum
+ add %o0, 4, %o0 ! advance buf ptr
+ addc %g0, %o2, %o2 ! add in final carry
+1: brz,pn %o1, 1f ! no trailing bytes, return
+ addcc %o1, -1, %g0 ! only one byte remains?
+ bne,pn %icc, 2f ! at least two bytes more
+ subcc %o1, 2, %o1 ! only two bytes more?
+ ba,pt %xcc, 4f ! only one byte remains
+ clr %o4 ! clear fake hword value
+2: lduh [%o0], %o4 ! get hword
+ be,pn %icc, 6f ! jmp if only hword remains
+ add %o0, 2, %o0 ! advance buf ptr either way
+ sll %o4, 16, %o4 ! create upper hword
+4: ldub [%o0], %o5 ! get final byte
+ sll %o5, 8, %o5 ! put into place
+ or %o5, %o4, %o4 ! coalese with hword (if any)
+6: addcc %o4, %o2, %o2 ! add to sum
+1: sllx %g4, 32, %g4 ! give gfp back
+ retl ! get outta here
+ addc %g0, %o2, %o0 ! add final carry into retval
+
+ /* Also do alignment out of band to get better cache patterns. */
+csum_partial_fix_alignment:
+
+ /* The common case is to get called with a nicely aligned
+ * buffer of size 0x20. Follow the code path for that case.
+ */
+ .globl C_LABEL(csum_partial)
+C_LABEL(csum_partial): /* %o0=buf, %o1=len, %o2=sum */
+ andcc %o0, 0x7, %g0 ! alignment problems?
+ be,pt %icc, csum_partial_fix_aligned ! yep, handle it
+ andn %o1, 0x7f, %o3 ! num loop iterations
+ cmp %o1, 6
+ bl,pn %icc, cpte - 0x4
+ andcc %o0, 0x2, %g0
+ be,pn %icc, 1f
+ and %o0, 0x4, %g7
+ lduh [%o0 + 0x00], %g2
+ sub %o1, 2, %o1
+ add %o0, 2, %o0
+ sll %g2, 16, %g2
+ addcc %g2, %o2, %o2
+ srl %o2, 16, %g3
+ addc %g0, %g3, %g2
+ sll %o2, 16, %o2
+ sll %g2, 16, %g3
+ srl %o2, 16, %o2
+ or %g3, %o2, %o2
+1: brz,pn %g7, csum_partial_fix_aligned
+ nop
+ ld [%o0 + 0x00], %g2
+ sub %o1, 4, %o1
+ addcc %g2, %o2, %o2
+ add %o0, 4, %o0
+ addc %g0, %o2, %o2
+csum_partial_fix_aligned:
+ brz,pt %o3, 3f ! none to do
+ andcc %o1, 0x70, %g1 ! clears carry flag too
+5: CSUM_BIGCHUNK(%o0, 0x00, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
+ CSUM_BIGCHUNK(%o0, 0x20, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
+ CSUM_BIGCHUNK(%o0, 0x40, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
+ CSUM_BIGCHUNK(%o0, 0x60, %o2, %o4, %o5, %g2, %g3, %g4, %g5)
+ sub %o3, 128, %o3 ! detract from loop iters
+ addc %g0, %o2, %o2 ! sink in final carry
+ brnz,pt %o3, 5b ! more to do
+ add %o0, 128, %o0 ! advance buf ptr
+3: brz,pn %g1, cpte ! nope
+ andcc %o1, 0xf, %o3 ! anything left at all?
+10: rd %pc, %g7 ! get pc
+ srl %g1, 1, %o4 ! compute offset
+ sub %g7, %g1, %g7 ! adjust jmp ptr
+ sub %g7, %o4, %g7 ! final jmp ptr adjust
+ jmp %g7 + (cpte - 8 - 10b) ! enter the table
+ add %o0, %g1, %o0 ! advance buf ptr
+cptbl: CSUM_LASTCHUNK(%o0, 0x68, %o2, %g2, %g3, %g4, %g5)
+ CSUM_LASTCHUNK(%o0, 0x58, %o2, %g2, %g3, %g4, %g5)
+ CSUM_LASTCHUNK(%o0, 0x48, %o2, %g2, %g3, %g4, %g5)
+ CSUM_LASTCHUNK(%o0, 0x38, %o2, %g2, %g3, %g4, %g5)
+ CSUM_LASTCHUNK(%o0, 0x28, %o2, %g2, %g3, %g4, %g5)
+ CSUM_LASTCHUNK(%o0, 0x18, %o2, %g2, %g3, %g4, %g5)
+ CSUM_LASTCHUNK(%o0, 0x08, %o2, %g2, %g3, %g4, %g5)
+ addc %g0, %o2, %o2 ! fetch final carry
+ andcc %o1, 0xf, %g0 ! anything left at all?
+cpte: brnz,pn %o3, csum_partial_end_cruft ! yep, handle it
+ sethi %uhi(KERNBASE), %g4
+ mov %o2, %o0 ! return computed csum
+ retl ! get outta here
+ sllx %g4, 32, %g4 ! give gfp back
+
+ .globl C_LABEL(__csum_partial_copy_start), C_LABEL(__csum_partial_copy_end)
+C_LABEL(__csum_partial_copy_start):
+
+#define EX(x,y,a,b,z) \
+98: x,y; \
+ .section .fixup,z##alloc,z##execinstr; \
+ .align 4; \
+99: ba,pt %xcc, 30f; \
+ a, b, %o3; \
+ .section __ex_table,z##alloc; \
+ .align 4; \
+ .word 98b, 99b; \
+ .text; \
+ .align 4
+
+#define EX2(x,y,z) \
+98: x,y; \
+ .section __ex_table,z##alloc; \
+ .align 4; \
+ .word 98b, 30f; \
+ .text; \
+ .align 4
+
+#define EX3(x,y,z) \
+98: x,y; \
+ .section __ex_table,z##alloc; \
+ .align 4; \
+ .word 98b, 96f; \
+ .text; \
+ .align 4
+
+#define EXT(start,end,handler,z) \
+ .section __ex_table,z##alloc; \
+ .align 4; \
+ .word start, 0, end, handler; \
+ .text; \
+ .align 4
+
+ /* This aligned version executes typically in 8.5 superscalar cycles, this
+ * is the best I can do. I say 8.5 because the final add will pair with
+ * the next ldd in the main unrolled loop. Thus the pipe is always full.
+ * If you change these macros (including order of instructions),
+ * please check the fixup code below as well.
+ */
+#define CSUMCOPY_BIGCHUNK_ALIGNED(src, dst, sum, off, t0, t1, t2, t3, t4, t5, t6, t7) \
+ ldd [src + off + 0x00], t0; \
+ ldd [src + off + 0x08], t2; \
+ addccc t0, sum, sum; \
+ ldd [src + off + 0x10], t4; \
+ addccc t1, sum, sum; \
+ ldd [src + off + 0x18], t6; \
+ addccc t2, sum, sum; \
+ std t0, [dst + off + 0x00]; \
+ addccc t3, sum, sum; \
+ std t2, [dst + off + 0x08]; \
+ addccc t4, sum, sum; \
+ std t4, [dst + off + 0x10]; \
+ addccc t5, sum, sum; \
+ std t6, [dst + off + 0x18]; \
+ addccc t6, sum, sum; \
+ addccc t7, sum, sum;
+
+ /* 12 superscalar cycles seems to be the limit for this case,
+ * because of this we thus do all the ldd's together to get
+ * Viking MXCC into streaming mode. Ho hum...
+ */
+#define CSUMCOPY_BIGCHUNK(src, dst, sum, off, t0, t1, t2, t3, t4, t5, t6, t7) \
+ ldd [src + off + 0x00], t0; \
+ ldd [src + off + 0x08], t2; \
+ ldd [src + off + 0x10], t4; \
+ ldd [src + off + 0x18], t6; \
+ st t0, [dst + off + 0x00]; \
+ addccc t0, sum, sum; \
+ st t1, [dst + off + 0x04]; \
+ addccc t1, sum, sum; \
+ st t2, [dst + off + 0x08]; \
+ addccc t2, sum, sum; \
+ st t3, [dst + off + 0x0c]; \
+ addccc t3, sum, sum; \
+ st t4, [dst + off + 0x10]; \
+ addccc t4, sum, sum; \
+ st t5, [dst + off + 0x14]; \
+ addccc t5, sum, sum; \
+ st t6, [dst + off + 0x18]; \
+ addccc t6, sum, sum; \
+ st t7, [dst + off + 0x1c]; \
+ addccc t7, sum, sum;
+
+ /* Yuck, 6 superscalar cycles... */
+#define CSUMCOPY_LASTCHUNK(src, dst, sum, off, t0, t1, t2, t3) \
+ ldd [src - off - 0x08], t0; \
+ ldd [src - off - 0x00], t2; \
+ addccc t0, sum, sum; \
+ st t0, [dst - off - 0x08]; \
+ addccc t1, sum, sum; \
+ st t1, [dst - off - 0x04]; \
+ addccc t2, sum, sum; \
+ st t2, [dst - off - 0x00]; \
+ addccc t3, sum, sum; \
+ st t3, [dst - off + 0x04];
+
+ /* Handle the end cruft code out of band for better cache patterns. */
+cc_end_cruft:
+ andcc %o3, 8, %g0 ! begin checks for that code
+ be,pn %icc, 1f
+ and %o3, 4, %g5
+ EX(ldd [%o0 + 0x00], %g2, and %o3, 0xf,#)
+ add %o1, 8, %o1
+ addcc %g2, %g7, %g7
+ add %o0, 8, %o0
+ addccc %g3, %g7, %g7
+ EX2(st %g2, [%o1 - 0x08],#)
+ addc %g0, %g7, %g7
+ EX2(st %g3, [%o1 - 0x04],#)
+1: brz,pt %g5, 1f
+ andcc %o3, 3, %o3
+ EX(ld [%o0 + 0x00], %g2, add %o3, 4,#)
+ add %o1, 4, %o1
+ addcc %g2, %g7, %g7
+ EX2(st %g2, [%o1 - 0x04],#)
+ addc %g0, %g7, %g7
+ add %o0, 4, %o0
+1: brz,pn %o3, 1f
+ addcc %o3, -1, %g0
+ bne,pn %icc, 2f
+ subcc %o3, 2, %o3
+ ba,pt %xcc, 4f
+ clr %o4
+2: EX(lduh [%o0 + 0x00], %o4, add %o3, 2,#)
+ add %o0, 2, %o0
+ EX2(sth %o4, [%o1 + 0x00],#)
+ be,pn %icc, 6f
+ add %o1, 2, %o1
+ sll %o4, 16, %o4
+4: EX(ldub [%o0 + 0x00], %o5, add %g0, 1,#)
+ EX2(stb %o5, [%o1 + 0x00],#)
+ sll %o5, 8, %o5
+ or %o5, %o4, %o4
+6: addcc %o4, %g7, %g7
+1: sllx %g4, 32, %g4
+ retl
+ addc %g0, %g7, %o0
+
+ /* Sun, you just can't beat me, you just can't. Stop trying,
+ * give up. I'm serious, I am going to kick the living shit
+ * out of you, game over, lights out.
+ */
+ .align 8
+ .globl C_LABEL(__csum_partial_copy_sparc_generic)
+C_LABEL(__csum_partial_copy_sparc_generic):
+ /* %o0=src, %o1=dest, %g1=len, %g7=sum */
+ xor %o0, %o1, %o4 ! get changing bits
+ andcc %o4, 3, %g0 ! check for mismatched alignment
+ bne,pn %icc, ccslow ! better this than unaligned/fixups
+ andcc %o0, 7, %g0 ! need to align things?
+ be,pt %icc, cc_dword_aligned ! yes, we check for short lengths there
+ andn %g1, 0x7f, %g2 ! can we use unrolled loop?
+ cmp %g1, 6
+ bl,a,pn %icc, ccte
+ andcc %g1, 0xf, %o3
+ andcc %o0, 0x1, %g0
+ bne,pn %icc, ccslow
+ andcc %o0, 0x2, %g0
+ be,pn %icc, 1f
+ andcc %o0, 0x4, %g0
+ EX(lduh [%o0 + 0x00], %g4, add %g1, 0,#)
+ sub %g1, 2, %g1
+ EX2(sth %g4, [%o1 + 0x00],#)
+ add %o0, 2, %o0
+ sll %g4, 16, %g4
+ addcc %g4, %g7, %g7
+ add %o1, 2, %o1
+ srl %g7, 16, %g3
+ addc %g0, %g3, %g4
+ sll %g7, 16, %g7
+ sll %g4, 16, %g3
+ srl %g7, 16, %g7
+ andcc %o0, 0x4, %g0
+ or %g3, %g7, %g7
+1: be,pt %icc, 3f
+ andn %g1, 0x7f, %g0
+ EX(ld [%o0 + 0x00], %g4, add %g1, 0,#)
+ sub %g1, 4, %g1
+ EX2(st %g4, [%o1 + 0x00],#)
+ add %o0, 4, %o0
+ addcc %g4, %g7, %g7
+ add %o1, 4, %o1
+ addc %g0, %g7, %g7
+cc_dword_aligned:
+3: brz,pn %g2, 3f ! nope, less than one loop remains
+ andcc %o1, 4, %g0 ! dest aligned on 4 or 8 byte boundry?
+ be,pn %icc, ccdbl + 4 ! 8 byte aligned, kick ass
+5: CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x00,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
+ CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x20,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
+ CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x40,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
+ CSUMCOPY_BIGCHUNK(%o0,%o1,%g7,0x60,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
+10: EXT(5b, 10b, 20f,#) ! note for exception handling
+ sub %g1, 128, %g1 ! detract from length
+ addc %g0, %g7, %g7 ! add in last carry bit
+ andncc %g1, 0x7f, %g0 ! more to csum?
+ add %o0, 128, %o0 ! advance src ptr
+ bne,pt %icc, 5b ! we did not go negative, continue looping
+ add %o1, 128, %o1 ! advance dest ptr
+3: andcc %g1, 0x70, %o2 ! can use table?
+ccmerge:be,pn %icc, ccte ! nope, go and check for end cruft
+ andcc %g1, 0xf, %o3 ! get low bits of length (clears carry btw)
+ srl %o2, 1, %o4 ! begin negative offset computation
+13: rd %pc, %o5 ! set up table ptr end
+ add %o0, %o2, %o0 ! advance src ptr
+ sub %o5, %o4, %o5 ! continue table calculation
+ sll %o2, 1, %g2 ! constant multiplies are fun...
+ sub %o5, %g2, %o5 ! some more adjustments
+ jmpl %o5 + (12f-13b), %g0 ! jump into it, duff style, wheee...
+ add %o1, %o2, %o1 ! advance dest ptr (carry is clear btw)
+cctbl: CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x68,%g2,%g3,%g4,%g5)
+ CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x58,%g2,%g3,%g4,%g5)
+ CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x48,%g2,%g3,%g4,%g5)
+ CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x38,%g2,%g3,%g4,%g5)
+ CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x28,%g2,%g3,%g4,%g5)
+ CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x18,%g2,%g3,%g4,%g5)
+ CSUMCOPY_LASTCHUNK(%o0,%o1,%g7,0x08,%g2,%g3,%g4,%g5)
+12: EXT(cctbl, 12b, 22f,#) ! note for exception table handling
+ addc %g0, %g7, %g7
+ andcc %o3, 0xf, %g0 ! check for low bits set
+ccte: bne,pn %icc, cc_end_cruft ! something left, handle it out of band
+ sethi %uhi(KERNBASE), %g4 ! restore gfp
+ mov %g7, %o0 ! give em the computed checksum
+ retl ! return
+ sllx %g4, 32, %g4 ! finish gfp restoration
+ccdbl: CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x00,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
+ CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x20,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
+ CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x40,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
+ CSUMCOPY_BIGCHUNK_ALIGNED(%o0,%o1,%g7,0x60,%o4,%o5,%g2,%g3,%g4,%g5,%o2,%o3)
+11: EXT(ccdbl, 11b, 21f,#) ! note for exception table handling
+ sub %g1, 128, %g1 ! detract from length
+ addc %g0, %g7, %g7 ! add in last carry bit
+ andncc %g1, 0x7f, %g0 ! more to csum?
+ add %o0, 128, %o0 ! advance src ptr
+ bne,pt %icc, ccdbl ! we did not go negative, continue looping
+ add %o1, 128, %o1 ! advance dest ptr
+ ba,pt %xcc, ccmerge ! finish it off, above
+ andcc %g1, 0x70, %o2 ! can use table? (clears carry btw)
+
+ccslow: mov 0, %g5
+ brlez,pn %g1, 4f
+ andcc %o0, 1, %o5
+ be,a,pt %icc, 1f
+ srl %g1, 1, %o3
+ sub %g1, 1, %g1
+ EX(ldub [%o0], %g5, add %g1, 1,#)
+ add %o0, 1, %o0
+ EX2(stb %g5, [%o1],#)
+ srl %g1, 1, %o3
+ add %o1, 1, %o1
+1: brz,a,pn %o3, 3f
+ andcc %g1, 1, %g0
+ andcc %o0, 2, %g0
+ be,a,pt %icc, 1f
+ srl %o3, 1, %o3
+ EX(lduh [%o0], %o4, add %g1, 0,#)
+ sub %g1, 2, %g1
+ srl %o4, 8, %g2
+ sub %o3, 1, %o3
+ EX2(stb %g2, [%o1],#)
+ add %o4, %g5, %g5
+ EX2(stb %o4, [%o1 + 1],#)
+ add %o0, 2, %o0
+ srl %o3, 1, %o3
+ add %o1, 2, %o1
+1: brz,a,pn %o3, 2f
+ andcc %g1, 2, %g0
+ EX3(ld [%o0], %o4,#)
+5: srl %o4, 24, %g2
+ srl %o4, 16, %g3
+ EX2(stb %g2, [%o1],#)
+ srl %o4, 8, %g2
+ EX2(stb %g3, [%o1 + 1],#)
+ add %o0, 4, %o0
+ EX2(stb %g2, [%o1 + 2],#)
+ addcc %o4, %g5, %g5
+ EX2(stb %o4, [%o1 + 3],#)
+ addc %g5, %g0, %g5 ! I am now to lazy to optimize this (question is if it
+ add %o1, 4, %o1 ! is worthy). Maybe some day - with the sll/srl
+ subcc %o3, 1, %o3 ! tricks
+ bne,a,pt %icc, 5b
+ EX3(ld [%o0], %o4,#)
+ sll %g5, 16, %g2
+ srl %g5, 16, %g5
+ srl %g2, 16, %g2
+ andcc %g1, 2, %g0
+ add %g2, %g5, %g5
+2: be,a,pt %icc, 3f
+ andcc %g1, 1, %g0
+ EX(lduh [%o0], %o4, and %g1, 3,#)
+ andcc %g1, 1, %g0
+ srl %o4, 8, %g2
+ add %o0, 2, %o0
+ EX2(stb %g2, [%o1],#)
+ add %g5, %o4, %g5
+ EX2(stb %o4, [%o1 + 1],#)
+ add %o1, 2, %o1
+3: be,a,pt %icc, 1f
+ sll %g5, 16, %o4
+ EX(ldub [%o0], %g2, add %g0, 1,#)
+ sll %g2, 8, %o4
+ EX2(stb %g2, [%o1],#)
+ add %g5, %o4, %g5
+ sll %g5, 16, %o4
+1: addcc %o4, %g5, %g5
+ srl %g5, 16, %o4
+ addc %g0, %o4, %g5
+ brz,pt %o5, 4f
+ srl %g5, 8, %o4
+ and %g5, 0xff, %g2
+ and %o4, 0xff, %o4
+ sll %g2, 8, %g2
+ or %g2, %o4, %g5
+4: addcc %g7, %g5, %g7
+ retl
+ addc %g0, %g7, %o0
+C_LABEL(__csum_partial_copy_end):
+
+ .section .fixup,#alloc,#execinstr
+ .align 4
+/* We do these strange calculations for the csum_*_from_user case only, ie.
+ * we only bother with faults on loads... */
+
+/* o2 = ((g2%20)&3)*8
+ * o3 = g1 - (g2/20)*32 - o2 */
+20:
+ cmp %g2, 20
+ blu,a,pn %icc, 1f
+ and %g2, 3, %o2
+ sub %g1, 32, %g1
+ ba,pt %xcc, 20b
+ sub %g2, 20, %g2
+1:
+ sll %o2, 3, %o2
+ ba,pt %xcc, 31f
+ sub %g1, %o2, %o3
+
+/* o2 = (!(g2 & 15) ? 0 : (((g2 & 15) + 1) & ~1)*8)
+ * o3 = g1 - (g2/16)*32 - o2 */
+21:
+ andcc %g2, 15, %o3
+ srl %g2, 4, %g2
+ be,a,pn %icc, 1f
+ clr %o2
+ add %o3, 1, %o3
+ and %o3, 14, %o3
+ sll %o3, 3, %o2
+1:
+ sll %g2, 5, %g2
+ sub %g1, %g2, %o3
+ ba,pt %xcc, 31f
+ sub %o3, %o2, %o3
+
+/* o0 += (g2/10)*16 - 0x70
+ * 01 += (g2/10)*16 - 0x70
+ * o2 = (g2 % 10) ? 8 : 0
+ * o3 += 0x70 - (g2/10)*16 - o2 */
+22:
+ cmp %g2, 10
+ blu,a,pt %xcc, 1f
+ sub %o0, 0x70, %o0
+ add %o0, 16, %o0
+ add %o1, 16, %o1
+ sub %o3, 16, %o3
+ ba,pt %xcc, 22b
+ sub %g2, 10, %g2
+1:
+ sub %o1, 0x70, %o1
+ add %o3, 0x70, %o3
+ clr %o2
+ movne %g2, 8, %o2
+ ba,pt %xcc, 31f
+ sub %o3, %o2, %o3
+96:
+ and %g1, 3, %g1
+ sll %o3, 2, %o3
+ add %g1, %o3, %o3
+30:
+/* %o1 is dst
+ * %o3 is # bytes to zero out
+ * %o4 is faulting address
+ * %o5 is %pc where fault occured */
+ clr %o2
+31:
+/* %o0 is src
+ * %o1 is dst
+ * %o2 is # of bytes to copy from src to dst
+ * %o3 is # bytes to zero out
+ * %o4 is faulting address
+ * %o5 is %pc where fault occured */
+ save %sp, -136, %sp
+ mov %i5, %o0
+ mov %i7, %o1
+ mov %i4, %o2
+ call C_LABEL(lookup_fault)
+ mov %g7, %i4
+ cmp %o0, 2
+ bne,pn %icc, 1f
+ add %g0, -EFAULT, %i5
+ brz,pn %i2, 2f
+ mov %i0, %o1
+ mov %i1, %o0
+5:
+ call C_LABEL(__memcpy)
+ mov %i2, %o2
+ brnz,a,pn %o0, 2f
+ add %i3, %i2, %i3
+ add %i1, %i2, %i1
+2:
+ mov %i1, %o0
+ call C_LABEL(__bzero)
+ mov %i3, %o1
+1:
+ ldx [%sp + 264], %o2 ! struct_ptr of parent
+ st %i5, [%o2]
+ ret
+ restore
+
+ .section __ex_table,#alloc
+ .align 4
+ .word 5b,2
FUNET's LINUX-ADM group, linux-adm@nic.funet.fi
TCL-scripts by Sam Shen, slshen@lbl.gov