| Joerg Sonnenberger | d8f7fb9 | 2014-01-29 13:46:28 +0000 | [diff] [blame] | 1 | /* ===-- clzdi2.c - Implement __clzdi2 -------------------------------------=== |
| 2 | * |
| 3 | * The LLVM Compiler Infrastructure |
| 4 | * |
| 5 | * This file is dual licensed under the MIT and the University of Illinois Open |
| 6 | * Source Licenses. See LICENSE.TXT for details. |
| 7 | * |
| 8 | * ===----------------------------------------------------------------------=== |
| 9 | * |
| 10 | * This file implements count leading zeros for 64bit arguments. |
| 11 | * |
| 12 | * ===----------------------------------------------------------------------=== |
| 13 | */ |
| 14 | #include "../assembly.h" |
| 15 | |
| Joerg Sonnenberger | d8f7fb9 | 2014-01-29 13:46:28 +0000 | [diff] [blame] | 16 | .syntax unified |
| Joerg Sonnenberger | d8f7fb9 | 2014-01-29 13:46:28 +0000 | [diff] [blame] | 17 | .text |
| Saleem Abdulrasool | 1b8f1a4 | 2014-06-16 16:05:24 +0000 | [diff] [blame] | 18 | #if __ARM_ARCH_ISA_THUMB == 2 |
| 19 | .thumb |
| 20 | #endif |
| 21 | |
| Saleem Abdulrasool | 57aa97f | 2014-06-01 04:07:03 +0000 | [diff] [blame] | 22 | |
| Saleem Abdulrasool | 310874a | 2014-05-12 15:23:37 +0000 | [diff] [blame] | 23 | .p2align 2 |
| Saleem Abdulrasool | 48d4e4d | 2014-10-07 02:39:13 +0000 | [diff] [blame] | 24 | #if __ARM_ARCH_ISA_THUMB == 2 |
| Steven Wu | 84610ba | 2014-10-04 00:18:59 +0000 | [diff] [blame] | 25 | DEFINE_COMPILERRT_THUMB_FUNCTION(__clzdi2) |
| Saleem Abdulrasool | 48d4e4d | 2014-10-07 02:39:13 +0000 | [diff] [blame] | 26 | #else |
| 27 | DEFINE_COMPILERRT_FUNCTION(__clzdi2) |
| 28 | #endif |
| Joerg Sonnenberger | d8f7fb9 | 2014-01-29 13:46:28 +0000 | [diff] [blame] | 29 | #ifdef __ARM_FEATURE_CLZ |
| 30 | #ifdef __ARMEB__ |
| 31 | cmp r0, 0 |
| 32 | itee ne |
| 33 | clzne r0, r0 |
| 34 | clzeq r0, r1 |
| 35 | addeq r0, r0, 32 |
| 36 | #else |
| 37 | cmp r1, 0 |
| 38 | itee ne |
| 39 | clzne r0, r1 |
| 40 | clzeq r0, r0 |
| 41 | addeq r0, r0, 32 |
| 42 | #endif |
| 43 | JMP(lr) |
| 44 | #else |
| 45 | /* Assumption: n != 0 */ |
| 46 | |
| 47 | /* |
| 48 | * r0: n |
| 49 | * r1: upper half of n, overwritten after check |
| 50 | * r1: count of leading zeros in n + 1 |
| 51 | * r2: scratch register for shifted r0 |
| 52 | */ |
| 53 | #ifdef __ARMEB__ |
| 54 | cmp r0, 0 |
| 55 | moveq r0, r1 |
| 56 | #else |
| 57 | cmp r1, 0 |
| 58 | movne r0, r1 |
| 59 | #endif |
| 60 | movne r1, 1 |
| 61 | moveq r1, 33 |
| 62 | |
| 63 | /* |
| 64 | * Basic block: |
| 65 | * if ((r0 >> SHIFT) == 0) |
| 66 | * r1 += SHIFT; |
| 67 | * else |
| 68 | * r0 >>= SHIFT; |
| 69 | * for descending powers of two as SHIFT. |
| 70 | */ |
| 71 | #define BLOCK(shift) \ |
| 72 | lsrs r2, r0, shift; \ |
| 73 | movne r0, r2; \ |
| 74 | addeq r1, shift \ |
| 75 | |
| 76 | BLOCK(16) |
| 77 | BLOCK(8) |
| 78 | BLOCK(4) |
| 79 | BLOCK(2) |
| 80 | |
| 81 | /* |
| 82 | * The basic block invariants at this point are (r0 >> 2) == 0 and |
| 83 | * r0 != 0. This means 1 <= r0 <= 3 and 0 <= (r0 >> 1) <= 1. |
| 84 | * |
| 85 | * r0 | (r0 >> 1) == 0 | (r0 >> 1) == 1 | -(r0 >> 1) | 1 - (r0 >> 1) |
| 86 | * ---+----------------+----------------+------------+-------------- |
| 87 | * 1 | 1 | 0 | 0 | 1 |
| 88 | * 2 | 0 | 1 | -1 | 0 |
| 89 | * 3 | 0 | 1 | -1 | 0 |
| 90 | * |
| 91 | * The r1's initial value of 1 compensates for the 1 here. |
| 92 | */ |
| 93 | sub r0, r1, r0, lsr #1 |
| 94 | |
| 95 | JMP(lr) |
| 96 | #endif // __ARM_FEATURE_CLZ |
| 97 | END_COMPILERRT_FUNCTION(__clzdi2) |