Home | History | Annotate | Line # | Download | only in arm
      1      1.1  joerg /* ===-- clzdi2.c - Implement __clzdi2 -------------------------------------===
      2      1.1  joerg  *
      3      1.1  joerg  *               The LLVM Compiler Infrastructure
      4      1.1  joerg  *
      5      1.1  joerg  * This file is dual licensed under the MIT and the University of Illinois Open
      6      1.1  joerg  * Source Licenses. See LICENSE.TXT for details.
      7      1.1  joerg  *
      8      1.1  joerg  * ===----------------------------------------------------------------------===
      9      1.1  joerg  *
     10      1.1  joerg  * This file implements count leading zeros for 64bit arguments.
     11      1.1  joerg  *
     12      1.1  joerg  * ===----------------------------------------------------------------------===
     13      1.1  joerg  */
     14      1.1  joerg #include "../assembly.h"
     15      1.1  joerg 
     16      1.1  joerg 	.syntax unified
     17      1.1  joerg 	.text
     18  1.1.1.3  joerg #if __ARM_ARCH_ISA_THUMB == 2
     19  1.1.1.3  joerg 	.thumb
     20  1.1.1.3  joerg #endif
     21  1.1.1.3  joerg 
     22  1.1.1.3  joerg 
     23  1.1.1.2  joerg 	.p2align	2
     24  1.1.1.4  joerg #if __ARM_ARCH_ISA_THUMB == 2
     25  1.1.1.4  joerg DEFINE_COMPILERRT_THUMB_FUNCTION(__clzdi2)
     26  1.1.1.4  joerg #else
     27      1.1  joerg DEFINE_COMPILERRT_FUNCTION(__clzdi2)
     28  1.1.1.4  joerg #endif
     29      1.1  joerg #ifdef __ARM_FEATURE_CLZ
     30      1.1  joerg #ifdef __ARMEB__
     31      1.1  joerg 	cmp	r0, 0
     32      1.1  joerg 	itee ne
     33      1.1  joerg 	clzne	r0, r0
     34      1.1  joerg 	clzeq	r0, r1
     35      1.1  joerg 	addeq	r0, r0, 32
     36      1.1  joerg #else
     37      1.1  joerg 	cmp	r1, 0
     38      1.1  joerg 	itee ne
     39      1.1  joerg 	clzne	r0, r1
     40      1.1  joerg 	clzeq	r0, r0
     41      1.1  joerg 	addeq	r0, r0, 32
     42      1.1  joerg #endif
     43      1.1  joerg 	JMP(lr)
     44      1.1  joerg #else
     45      1.1  joerg 	/* Assumption: n != 0 */
     46      1.1  joerg 
     47      1.1  joerg 	/*
     48      1.1  joerg 	 * r0: n
     49      1.1  joerg 	 * r1: upper half of n, overwritten after check
     50      1.1  joerg 	 * r1: count of leading zeros in n + 1
     51      1.1  joerg 	 * r2: scratch register for shifted r0
     52      1.1  joerg 	 */
     53      1.1  joerg #ifdef __ARMEB__
     54      1.1  joerg 	cmp	r0, 0
     55      1.1  joerg 	moveq	r0, r1
     56      1.1  joerg #else
     57      1.1  joerg 	cmp	r1, 0
     58      1.1  joerg 	movne	r0, r1
     59      1.1  joerg #endif
     60      1.1  joerg 	movne	r1, 1
     61      1.1  joerg 	moveq	r1, 33
     62      1.1  joerg 
     63      1.1  joerg 	/*
     64      1.1  joerg 	 * Basic block:
     65      1.1  joerg 	 * if ((r0 >> SHIFT) == 0)
     66      1.1  joerg 	 *   r1 += SHIFT;
     67      1.1  joerg 	 * else
     68      1.1  joerg 	 *   r0 >>= SHIFT;
     69      1.1  joerg 	 * for descending powers of two as SHIFT.
     70      1.1  joerg 	 */
     71      1.1  joerg #define BLOCK(shift) \
     72      1.1  joerg 	lsrs	r2, r0, shift; \
     73      1.1  joerg 	movne	r0, r2; \
     74      1.1  joerg 	addeq	r1, shift \
     75      1.1  joerg 
     76      1.1  joerg 	BLOCK(16)
     77      1.1  joerg 	BLOCK(8)
     78      1.1  joerg 	BLOCK(4)
     79      1.1  joerg 	BLOCK(2)
     80      1.1  joerg 
     81      1.1  joerg 	/*
     82      1.1  joerg 	 * The basic block invariants at this point are (r0 >> 2) == 0 and
     83      1.1  joerg 	 * r0 != 0. This means 1 <= r0 <= 3 and 0 <= (r0 >> 1) <= 1.
     84      1.1  joerg 	 *
     85      1.1  joerg 	 * r0 | (r0 >> 1) == 0 | (r0 >> 1) == 1 | -(r0 >> 1) | 1 - (r0 >> 1)
     86      1.1  joerg 	 * ---+----------------+----------------+------------+--------------
     87      1.1  joerg 	 * 1  | 1              | 0              | 0          | 1
     88      1.1  joerg 	 * 2  | 0              | 1              | -1         | 0
     89      1.1  joerg 	 * 3  | 0              | 1              | -1         | 0
     90      1.1  joerg 	 *
     91      1.1  joerg 	 * The r1's initial value of 1 compensates for the 1 here.
     92      1.1  joerg 	 */
     93      1.1  joerg 	sub	r0, r1, r0, lsr #1
     94      1.1  joerg 
     95      1.1  joerg 	JMP(lr)
     96      1.1  joerg #endif // __ARM_FEATURE_CLZ
     97      1.1  joerg END_COMPILERRT_FUNCTION(__clzdi2)
     98