| 1 | /* |
|---|---|
| 2 | * Compute 16-bit sum in ones' complement arithmetic (with end-around carry). |
| 3 | * This sum is often used as a simple checksum in networking. |
| 4 | * |
| 5 | * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 6 | * See https://llvm.org/LICENSE.txt for license information. |
| 7 | * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 8 | */ |
| 9 | |
| 10 | #include "networking.h" |
| 11 | #include "chksum_common.h" |
| 12 | |
| 13 | always_inline |
| 14 | static inline uint32_t |
| 15 | slurp_head32(const void **pptr, uint32_t *nbytes) |
| 16 | { |
| 17 | uint32_t sum = 0; |
| 18 | Assert(*nbytes >= 4); |
| 19 | uint32_t off = (uintptr_t) *pptr % 4; |
| 20 | if (likely(off != 0)) |
| 21 | { |
| 22 | /* Get rid of bytes 0..off-1 */ |
| 23 | const unsigned char *ptr32 = align_ptr(ptr: *pptr, bytes: 4); |
| 24 | uint32_t mask = ~0U << (CHAR_BIT * off); |
| 25 | sum = load32(ptr: ptr32) & mask; |
| 26 | *pptr = ptr32 + 4; |
| 27 | *nbytes -= 4 - off; |
| 28 | } |
| 29 | return sum; |
| 30 | } |
| 31 | |
| 32 | /* Additional loop unrolling would help when not auto-vectorizing */ |
| 33 | unsigned short |
| 34 | __chksum(const void *ptr, unsigned int nbytes) |
| 35 | { |
| 36 | bool swap = false; |
| 37 | uint64_t sum = 0; |
| 38 | |
| 39 | if (nbytes > 300) |
| 40 | { |
| 41 | /* 4-byte align pointer */ |
| 42 | swap = (uintptr_t) ptr & 1; |
| 43 | sum = slurp_head32(pptr: &ptr, nbytes: &nbytes); |
| 44 | } |
| 45 | /* Else benefit of aligning not worth the overhead */ |
| 46 | |
| 47 | /* Sum all 16-byte chunks */ |
| 48 | const char *cptr = ptr; |
| 49 | for (uint32_t nquads = nbytes / 16; nquads != 0; nquads--) |
| 50 | { |
| 51 | uint64_t h0 = load32(ptr: cptr + 0); |
| 52 | uint64_t h1 = load32(ptr: cptr + 4); |
| 53 | uint64_t h2 = load32(ptr: cptr + 8); |
| 54 | uint64_t h3 = load32(ptr: cptr + 12); |
| 55 | sum += h0 + h1 + h2 + h3; |
| 56 | cptr += 16; |
| 57 | } |
| 58 | nbytes %= 16; |
| 59 | Assert(nbytes < 16); |
| 60 | |
| 61 | /* Handle any trailing 4-byte chunks */ |
| 62 | while (nbytes >= 4) |
| 63 | { |
| 64 | sum += load32(ptr: cptr); |
| 65 | cptr += 4; |
| 66 | nbytes -= 4; |
| 67 | } |
| 68 | Assert(nbytes < 4); |
| 69 | |
| 70 | if (nbytes & 2) |
| 71 | { |
| 72 | sum += load16(ptr: cptr); |
| 73 | cptr += 2; |
| 74 | } |
| 75 | |
| 76 | if (nbytes & 1) |
| 77 | { |
| 78 | sum += *(uint8_t *)cptr; |
| 79 | } |
| 80 | |
| 81 | return fold_and_swap(sum, swap); |
| 82 | } |
| 83 |