| 1 | /* SPDX-License-Identifier: GPL-2.0 */ |
|---|
| 2 | /* Copyright (C) 2023 Rivos Inc. */ |
|---|
| 3 | |
|---|
| 4 | #include <linux/linkage.h> |
|---|
| 5 | #include <asm/asm.h> |
|---|
| 6 | |
|---|
| 7 | .text |
|---|
| 8 | |
|---|
| 9 | /* void __riscv_copy_words_unaligned(void *, const void *, size_t) */ |
|---|
| 10 | /* Performs a memcpy without aligning buffers, using word loads and stores. */ |
|---|
| 11 | /* Note: The size is truncated to a multiple of 8 * SZREG */ |
|---|
| 12 | SYM_FUNC_START(__riscv_copy_words_unaligned) |
|---|
| 13 | andi a4, a2, ~((8*SZREG)-1) |
|---|
| 14 | beqz a4, 2f |
|---|
| 15 | add a3, a1, a4 |
|---|
| 16 | 1: |
|---|
| 17 | REG_L a4, 0(a1) |
|---|
| 18 | REG_L a5, SZREG(a1) |
|---|
| 19 | REG_L a6, 2*SZREG(a1) |
|---|
| 20 | REG_L a7, 3*SZREG(a1) |
|---|
| 21 | REG_L t0, 4*SZREG(a1) |
|---|
| 22 | REG_L t1, 5*SZREG(a1) |
|---|
| 23 | REG_L t2, 6*SZREG(a1) |
|---|
| 24 | REG_L t3, 7*SZREG(a1) |
|---|
| 25 | REG_S a4, 0(a0) |
|---|
| 26 | REG_S a5, SZREG(a0) |
|---|
| 27 | REG_S a6, 2*SZREG(a0) |
|---|
| 28 | REG_S a7, 3*SZREG(a0) |
|---|
| 29 | REG_S t0, 4*SZREG(a0) |
|---|
| 30 | REG_S t1, 5*SZREG(a0) |
|---|
| 31 | REG_S t2, 6*SZREG(a0) |
|---|
| 32 | REG_S t3, 7*SZREG(a0) |
|---|
| 33 | addi a0, a0, 8*SZREG |
|---|
| 34 | addi a1, a1, 8*SZREG |
|---|
| 35 | bltu a1, a3, 1b |
|---|
| 36 | |
|---|
| 37 | 2: |
|---|
| 38 | ret |
|---|
| 39 | SYM_FUNC_END(__riscv_copy_words_unaligned) |
|---|
| 40 | |
|---|
| 41 | /* void __riscv_copy_bytes_unaligned(void *, const void *, size_t) */ |
|---|
| 42 | /* Performs a memcpy without aligning buffers, using only byte accesses. */ |
|---|
| 43 | /* Note: The size is truncated to a multiple of 8 */ |
|---|
| 44 | SYM_FUNC_START(__riscv_copy_bytes_unaligned) |
|---|
| 45 | andi a4, a2, ~(8-1) |
|---|
| 46 | beqz a4, 2f |
|---|
| 47 | add a3, a1, a4 |
|---|
| 48 | 1: |
|---|
| 49 | lb a4, 0(a1) |
|---|
| 50 | lb a5, 1(a1) |
|---|
| 51 | lb a6, 2(a1) |
|---|
| 52 | lb a7, 3(a1) |
|---|
| 53 | lb t0, 4(a1) |
|---|
| 54 | lb t1, 5(a1) |
|---|
| 55 | lb t2, 6(a1) |
|---|
| 56 | lb t3, 7(a1) |
|---|
| 57 | sb a4, 0(a0) |
|---|
| 58 | sb a5, 1(a0) |
|---|
| 59 | sb a6, 2(a0) |
|---|
| 60 | sb a7, 3(a0) |
|---|
| 61 | sb t0, 4(a0) |
|---|
| 62 | sb t1, 5(a0) |
|---|
| 63 | sb t2, 6(a0) |
|---|
| 64 | sb t3, 7(a0) |
|---|
| 65 | addi a0, a0, 8 |
|---|
| 66 | addi a1, a1, 8 |
|---|
| 67 | bltu a1, a3, 1b |
|---|
| 68 | |
|---|
| 69 | 2: |
|---|
| 70 | ret |
|---|
| 71 | SYM_FUNC_END(__riscv_copy_bytes_unaligned) |
|---|
| 72 | |
|---|
