diff options
Diffstat (limited to 'private/crt32/misc/alpha/insv.s')
| -rw-r--r-- | private/crt32/misc/alpha/insv.s | 158 |
1 files changed, 158 insertions, 0 deletions
diff --git a/private/crt32/misc/alpha/insv.s b/private/crt32/misc/alpha/insv.s new file mode 100644 index 000000000..7d0956103 --- /dev/null +++ b/private/crt32/misc/alpha/insv.s @@ -0,0 +1,158 @@ + #+ + # Copyright 1991, 1994 Digital Equipment Corporation + # + # int ots_insv(char *addr, int position, unsigned size, int value) + # + # Arbitrary bitfield insertion, longword granularity + # + # Special conventions: No stack space, r0-r1, r16-r19 and r26-r28 ONLY, + # no linkage pointer required. + # (Warning: The auto-loader potentially takes some regs across + # the call if this is being used in a shared lib. environment.) + # + # See also: ots_ext[z]v + # + # 001 5 Sep 1991 KDG Initial version + # + # 002 19 May 1992 KDG Changes for common VMS/OSF sources + # + # 003 22 Sep 1992 KDG Add case-sensitive name + # + # 004 26 Jan 1993 KDG Add underscore + # + # 005 19 Apr 1994 kdg Longword granularity version based on quadword + # granularity version 004 + +#include "ots_defs.hs" + + # Totally general field insertion - arbitrary run-time field of 0-64 bits + # at an unknown alignment target. Longword granularity. + # + # Conceptually, this operation takes a 67 bit bit-address, which is the sum + # of a byte-aligned memory address and the bit offset (which is signed). + # + # +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ + # | | | | | | | | | | | | | | | | | | | | | | | | | |.|.|.|Q|L|W|B| + # +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ + # | | | | | | | | | | | | | | | | | | | | | | | | | | |.|.|.|b|b|b| + # +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ + # + # Inputs: + # r16 - input address + # r17 - input bit offset + # r18 - input size + # r19 - input value + # + # This is based on the original insert routine modified for longword + # granularity. This routine could probably be improved. (It does + # a "reasonable" job, but hasn't had as much attention as the quadword + # granularity version. Fields contained in a single longword are + # roughly the same cost as the quadword granularity version. Fields + # contained in two longwords in the same quadword are somewhat slower, + # while the two longword spanning a quadword case is roughly comparable. + # The 3 longword case is relatively slow [2 mispredicted branches, + # on unnecessary safe speculative load, could potentially use better + # scheduling too.]) + # + .globl _OtsFieldInsert + .ent _OtsFieldInsert +_OtsFieldInsert: + .set noat + .set noreorder + ble r18, noacc # check for zero size - no memory access + sra r17, 3, r27 # get byte part of bit offset (signed!) + addq r16, r27, r16 # add to initial base addr. + and r17, 7, r17 # get bit-in-byte from bit offset + and r16, 3, r27 # get byte-in-longword (must be clean for compares) + bic r16, 3, r16 # get a longword aligned address + s8addq r27, r17, r17 # form the true bit offset in the longword + ldl r28, (r16) # load first or only longword + addq r17, r18, r27 # get bit offset of bit following field + subq r27, 32, r0 # if <=32, field is contained in 1 longword + bgt r0, multiple # handle multi-longword case if not + # Common case of field in single LW - fall through + negq r27, r27 # <5:0> = bits for right shift + negq r18, r0 # bits for left shift (wordlength-is) + not r31, r1 # all ones + sll r1, r0, r1 # shift mask to high bits + sll r19, r0, r19 # shift source to high bits (hand interleaving for better sched) + srl r1, r27, r1 # and into position + srl r19, r27, r19 # and into position + bic r28, r1, r28 # clear the bits... + bis r28, r19, r28 # insert them + stl r28, (r16) # put the value back... +noacc: ret r31, (r26) + + # At this point: + # Field is known to be contained in at least 2 longwords + # r0 is bit position past end of field, -32 + # r1 junk + # r16 is longword aligned + # r17 is bit offset in longword + # r18 is field size + # r19 is value to store + # r27 is bit position past end of field + # r28 first lw from memory + # +multiple: + subq r0, 32, r27 # if <=64, the field is contained in 2 longwords + ldl r1, 4(r16) # load the 2nd longword (safe) + bgt r27, three # handle 3 longword case (rare...) + not r31, r27 # all ones + sll r27, r17, r27 # get mask in correct place + sll r19, r17, r17 # get insert value to top of register + bic r28, r27, r28 # clear bits in target + bis r28, r17, r28 # merge the field in + srl r1, r0, r1 # clear bits in target + negq r18, r27 # + sll r19, r27, r19 # shift to high bits + negq r0, r27 # + srl r19, r27, r19 # and into position + sll r1, r0, r1 # + stl r28, (r16) # store the first longword + bis r1, r19, r1 # merge + stl r1, 4(r16) # store back the second longword + ret r31, (r26) + + # At this point: + # Field is known to be contained in exactly 3 longwords + # r0 is bit position past end of field, -32 + # r1 value loaded for 2nd longword (which will be totally overwritten - i.e. junk) + # r16 is longword aligned + # r17 is bit offset in longword + # r18 is field size + # r19 is value to store + # r27 is bit position past end of field -64 + # r28 first lw from memory + # + # Three word case is roughly similar to two word case, except + # the middle store isn't a merge, just a real store, and the offsets + # for the 3rd word need to be adjusted. (This case hasn't + # received much attention and could probably be improved by + # at least a few instructions...) + # +three: + not r31, r0 # all ones + sll r0, r17, r0 # get mask in correct place + sll r19, r17, r1 # get insert value to top of register + bic r28, r0, r28 # clear bits in target + bis r28, r1, r28 # merge the field in + ldl r1, 8(r16) # load the 3rd longword + lda r0, 32(r31) # load 32 + stl r28, (r16) # store the first longword + subq r0, r17, r0 # shift amount + srl r19, r0, r28 # discard bits already stored + negq r18, r0 # + srl r1, r27, r1 # clear bits in target + sll r19, r0, r19 # shift to high bits + negq r27, r0 # + srl r19, r0, r19 # and into position + sll r1, r27, r1 # + stl r28, 4(r16) # store second complete longword + bis r1, r19, r1 # merge + stl r1, 8(r16) # store back the third longword + ret r31, (r26) + + .set at + .set reorder + .end _OtsFieldInsert |