| David Brazdil | 0f672f6 | 2019-12-10 10:32:29 +0000 | [diff] [blame^] | 1 | ;; SPDX-License-Identifier: GPL-2.0-or-later |
| Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 2 | ;; Copyright 2010 Free Software Foundation, Inc. |
| 3 | ;; Contributed by Bernd Schmidt <bernds@codesourcery.com>. |
| 4 | ;; |
| Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 5 | |
| 6 | #include <linux/linkage.h> |
| 7 | |
| 8 | ;; ABI considerations for the divide functions |
| 9 | ;; The following registers are call-used: |
| 10 | ;; __c6xabi_divi A0,A1,A2,A4,A6,B0,B1,B2,B4,B5 |
| 11 | ;; __c6xabi_divu A0,A1,A2,A4,A6,B0,B1,B2,B4 |
| 12 | ;; __c6xabi_remi A1,A2,A4,A5,A6,B0,B1,B2,B4 |
| 13 | ;; __c6xabi_remu A1,A4,A5,A7,B0,B1,B2,B4 |
| 14 | ;; |
| 15 | ;; In our implementation, divu and remu are leaf functions, |
| 16 | ;; while both divi and remi call into divu. |
| 17 | ;; A0 is not clobbered by any of the functions. |
| 18 | ;; divu does not clobber B2 either, which is taken advantage of |
| 19 | ;; in remi. |
| 20 | ;; divi uses B5 to hold the original return address during |
| 21 | ;; the call to divu. |
| 22 | ;; remi uses B2 and A5 to hold the input values during the |
| 23 | ;; call to divu. It stores B3 in on the stack. |
| 24 | |
| 25 | .text |
| 26 | ENTRY(__c6xabi_divu) |
| 27 | ;; We use a series of up to 31 subc instructions. First, we find |
| 28 | ;; out how many leading zero bits there are in the divisor. This |
| 29 | ;; gives us both a shift count for aligning (shifting) the divisor |
| 30 | ;; to the, and the number of times we have to execute subc. |
| 31 | |
| 32 | ;; At the end, we have both the remainder and most of the quotient |
| 33 | ;; in A4. The top bit of the quotient is computed first and is |
| 34 | ;; placed in A2. |
| 35 | |
| 36 | ;; Return immediately if the dividend is zero. |
| 37 | mv .s2x A4, B1 |
| 38 | [B1] lmbd .l2 1, B4, B1 |
| 39 | || [!B1] b .s2 B3 ; RETURN A |
| 40 | || [!B1] mvk .d2 1, B4 |
| 41 | mv .l1x B1, A6 |
| 42 | || shl .s2 B4, B1, B4 |
| 43 | |
| 44 | ;; The loop performs a maximum of 28 steps, so we do the |
| 45 | ;; first 3 here. |
| 46 | cmpltu .l1x A4, B4, A2 |
| 47 | [!A2] sub .l1x A4, B4, A4 |
| 48 | || shru .s2 B4, 1, B4 |
| 49 | || xor .s1 1, A2, A2 |
| 50 | |
| 51 | shl .s1 A2, 31, A2 |
| 52 | || [B1] subc .l1x A4,B4,A4 |
| 53 | || [B1] add .s2 -1, B1, B1 |
| 54 | [B1] subc .l1x A4,B4,A4 |
| 55 | || [B1] add .s2 -1, B1, B1 |
| 56 | |
| 57 | ;; RETURN A may happen here (note: must happen before the next branch) |
| 58 | _divu_loop: |
| 59 | cmpgt .l2 B1, 7, B0 |
| 60 | || [B1] subc .l1x A4,B4,A4 |
| 61 | || [B1] add .s2 -1, B1, B1 |
| 62 | [B1] subc .l1x A4,B4,A4 |
| 63 | || [B1] add .s2 -1, B1, B1 |
| 64 | || [B0] b .s1 _divu_loop |
| 65 | [B1] subc .l1x A4,B4,A4 |
| 66 | || [B1] add .s2 -1, B1, B1 |
| 67 | [B1] subc .l1x A4,B4,A4 |
| 68 | || [B1] add .s2 -1, B1, B1 |
| 69 | [B1] subc .l1x A4,B4,A4 |
| 70 | || [B1] add .s2 -1, B1, B1 |
| 71 | [B1] subc .l1x A4,B4,A4 |
| 72 | || [B1] add .s2 -1, B1, B1 |
| 73 | [B1] subc .l1x A4,B4,A4 |
| 74 | || [B1] add .s2 -1, B1, B1 |
| 75 | ;; loop backwards branch happens here |
| 76 | |
| 77 | ret .s2 B3 |
| 78 | || mvk .s1 32, A1 |
| 79 | sub .l1 A1, A6, A6 |
| 80 | shl .s1 A4, A6, A4 |
| 81 | shru .s1 A4, 1, A4 |
| 82 | || sub .l1 A6, 1, A6 |
| 83 | or .l1 A2, A4, A4 |
| 84 | shru .s1 A4, A6, A4 |
| 85 | nop |
| 86 | ENDPROC(__c6xabi_divu) |