diff options
Diffstat (limited to '')
| -rw-r--r-- | private/ntos/tdi/tcpip/tcp/ppc/xsum.s | 257 |
1 files changed, 257 insertions, 0 deletions
diff --git a/private/ntos/tdi/tcpip/tcp/ppc/xsum.s b/private/ntos/tdi/tcpip/tcp/ppc/xsum.s new file mode 100644 index 000000000..831ef6b53 --- /dev/null +++ b/private/ntos/tdi/tcpip/tcp/ppc/xsum.s @@ -0,0 +1,257 @@ +// TITLE("Compute Checksum") +//++ +// +// Copyright (c) 1994 IBM Corporation +// +// Module Name: +// +// xsum.s +// +// Abstract: +// +// This module implement a function to compute the checksum of a buffer. +// +// Author: +// +// David N. Cutler (davec) 27-Jan-1992 +// +// Environment: +// +// User mode. +// +// Revision History: +// +// Michael W. Thomas 02/14/94 Converted from MIPS +// Peter L. Johnston 07/19/94 Updated for Daytona Lvl 734 and +// optimized for PowerPC. +// +//-- + +#include "ksppc.h" + + SBTTL("Compute Checksum") +//++ +// +// ULONG +// tcpxsum ( +// IN ULONG Checksum, +// IN PUCHAR Source, +// IN ULONG Length +// ) +// +// Routine Description: +// +// This function computes the checksum of the specified buffer. +// +// N.B. The checksum is the 16 bit checksum of the 16 bit aligned +// buffer. If the buffer is not 16 bit aligned the first byte is +// moved to high order position to be added to the correct half. +// +// Arguments: +// +// Checksum (r3) - Supplies the initial checksum value. +// +// Source (r4) - Supplies a pointer to the checksum buffer. +// +// Length (r5) - Supplies the length of the buffer in bytes. +// +// Return Value: +// +// The computed checksum is returned as the function value. +// +//-- + + LEAF_ENTRY(tcpxsum) + + cmpwi r.5, 0 // check if bytes to checksum + mtcrf 0x01, r.4 // set up for alignment check + li r.6, 0 // initialize partial checksum + beqlr- // return if no bytes to checksum + + andi. r.7, r.5, 1 // check if length is even + crmove 7, 31 // remember original alignment + bf 31, evenalign // jif 16 bit aligned + +// +// Initialize the checksum to the first byte shifted up a byte. +// + lbz r.6, 0(r.4) // get first byte of buffer + subi r.5, r.5, 1 // reduce count of bytes to checksum + cmpwi cr.6, r.5, 0 // check if done + crnot eq, eq // invert odd/even length check + addi r.4, r.4, 1 // advance buffer address + mtcrf 0x01, r.4 // reset 32 bit alignment check + slwi r.6, r.6, 8 // shift byte up in computed checksum + // max current checksum is 0x0ff00 + beq cr.6, combine // jif no more bytes to checksum + +evenalign: + +// +// Check if the length of the buffer is an even number of bytes. +// +// If the buffer is not an even number of bytes, add the last byte to the +// computed checksum. +// + + beq evenlength + subic. r.5, r.5, 1 // reduce count of bytes to checksum + lbzx r.7, r.4, r.5 // get last byte from buffer + add r.6, r.6, r.7 // add last byte to computed checksum + // max current checksum is 0x0ffff + beq combine // jif no more bytes in buffer + +evenlength: + +// +// Check if we are 4 byte aligned, if not add first 2 byte word into +// checksum so the buffer is then 4 byte aligned. +// + + bf 30, fourbytealigned // jif 4 byte aligned + + lhz r.7, 0(r.4) // get 2 byte word + subic. r.5, r.5, 2 // reduce length + addi r.4, r.4, 2 // bump address + add r.6, r.6, r.7 // add 2 bytes to computed checksum + // max current checksum is 0x1fffe + beq combine // jif no more bytes to checksum + +// +// Attempt to sum the remainder of the buffer in sets of 32 bytes. This +// should achieve 2 bytes per clock on 601 and 603, and 3.2 bytes per clock +// on 604. (A seperate implementation will be required to take advantage +// of 64 bit loads on the 620). +// + +fourbytealigned: + + srwi. r.7, r.5, 5 // get count of 32 byte sets + mtcrf 0x03, r.5 // break length into block for + // various run lengths. + subi r.4, r.4, 4 // adjust buffer address for lwzu + mtctr r.7 + addic r.6, r.6, 0 // clear carry bit + beq try16 // jif no 32 byte sets + +do32: lwz r.8, 4(r.4) // get 1st 4 bytes in set + lwz r.9, 8(r.4) // get 2nd 4 + adde r.6, r.6, r.8 // add 1st 4 to checksum + lwz r.10, 12(r.4) // get 3rd 4 + adde r.6, r.6, r.9 // add 2nd 4 + lwz r.11, 16(r.4) // get 4th 4 + adde r.6, r.6, r.10 // add 3rd 4 + lwz r.8, 20(r.4) // get 5th 4 + adde r.6, r.6, r.11 // add 4th 4 + lwz r.9, 24(r.4) // get 6th 4 + adde r.6, r.6, r.8 // add 5th 4 + lwz r.10, 28(r.4) // get 7th 4 + adde r.6, r.6, r.9 // add 6th 4 + lwzu r.11, 32(r.4) // get 8th 4 and update address + adde r.6, r.6, r.10 // add 7th 4 + adde r.6, r.6, r.11 // add 8th 4 + bdnz do32 + +try16: bf 27, try8 // jif no 16 byte block + + lwz r.8, 4(r.4) // get 1st 4 + lwz r.9, 8(r.4) // get 2nd 4 + adde r.6, r.6, r.8 // add 1st 4 + lwz r.10, 12(r.4) // get 3rd 4 + adde r.6, r.6, r.9 // add 2nd 4 + lwzu r.11, 16(r.4) // get 4th 4 and update address + adde r.6, r.6, r.10 // add 3rd 4 + adde r.6, r.6, r.11 // add 4th 4 + +try8: bf 28, try4 // jif no 8 byte block + lwz r.8, 4(r.4) // get 1st 4 + lwzu r.9, 8(r.4) // get 2nd 4 and update address + adde r.6, r.6, r.8 // add 1st 4 + adde r.6, r.6, r.9 // add 2nd 4 + +try4: bf 29, try2 // jif no 4 byte block + lwzu r.8, 4(r.4) // get 4 bytes and update address + adde r.6, r.6, r.8 + +try2: bf 30, fold // jif no 2 byte block + +// +// At this point, r.4 is pointing at the last 4 byte block processed (or +// not processed if there were no 4 byte blocks). We need to add when we +// pull the last two bytes. +// + lhz r.8, 4(r.4) // get last two bytes + adde r.6, r.6, r.8 // add last two bytes + +// +// Collapse 33 bit (1 carry bit, 32 bits in r.6) into 17 bit checksum. +// + +fold: rlwinm r.7, r.6, 16, 0xffff // get 16 most significant bits (upper) + rlwinm r.6, r.6, 0, 0xffff // get least significant 16 bits (lower) + adde r.6, r.6, r.7 // upper + lower + carry + // max current checksum is 0x1ffff + +// +// Combine input checksum and partial checksum. +// +// If the input buffer was byte aligned, then word swap bytes in computed +// checksum before combination with input chewcksum. +// + +combine: + + bf 7, waseven // jif original alignment was 16 bit + +// +// Swap bytes within upper and lower halves. +// eg: AA BB CC DD becomes BB AA DD CC +// +// As the current maximum partial checksum is 0x1ffff don't worry about AA. +// ie: want BB 00 DD CC +// + + rlwimi r.6, r.6, 16, 0xff000000// r.7 = CC BB CC DD + rlwinm r.6, r.6, 8, 0xff00ffff// r.7 = BB 00 DD CC + +waseven: + + add r.3, r.3, r.6 // combine checksums + // max current checksum is 0x101fffe + rotlwi r.4, r.3, 16 // swap checksum words + add r.3, r.3, r.4 // add words with carry into high word + srwi r.3, r.3, 16 // extract final checksum + + LEAF_EXIT(tcpxsum) + + .debug$S + .ualong 1 + + .uashort 15 + .uashort 0x9 # S_OBJNAME + .ualong 0 + .byte 8, "xsum.obj" + + .uashort 24 + .uashort 0x1 # S_COMPILE + .byte 0x42 # Target processor = PPC 604 + .byte 3 # Language = ASM + .byte 0 + .byte 0 + .byte 17, "PowerPC Assembler" + + .uashort 43 + .uashort 0x205 # S_GPROC32 + .ualong 0 + .ualong 0 + .ualong 0 + .ualong tcpxsum.end-..tcpxsum + .ualong 0 + .ualong tcpxsum.end-..tcpxsum + .ualong [secoff]..tcpxsum + .uashort [secnum]..tcpxsum + .uashort 0x1000 + .byte 0x00 + .byte 7, "tcpxsum" + + .uashort 2, 0x6 # S_END |