xref: /aosp_15_r20/external/compiler-rt/lib/builtins/arm/clzsi2.S (revision 7c3d14c8b49c529e04be81a3ce6f5cc23712e4c6)
1*7c3d14c8STreehugger Robot/* ===-- clzsi2.c - Implement __clzsi2 -------------------------------------===
2*7c3d14c8STreehugger Robot *
3*7c3d14c8STreehugger Robot *               The LLVM Compiler Infrastructure
4*7c3d14c8STreehugger Robot *
5*7c3d14c8STreehugger Robot * This file is dual licensed under the MIT and the University of Illinois Open
6*7c3d14c8STreehugger Robot * Source Licenses. See LICENSE.TXT for details.
7*7c3d14c8STreehugger Robot *
8*7c3d14c8STreehugger Robot * ===----------------------------------------------------------------------===
9*7c3d14c8STreehugger Robot *
10*7c3d14c8STreehugger Robot * This file implements count leading zeros for 32bit arguments.
11*7c3d14c8STreehugger Robot *
12*7c3d14c8STreehugger Robot * ===----------------------------------------------------------------------===
13*7c3d14c8STreehugger Robot */
14*7c3d14c8STreehugger Robot#include "../assembly.h"
15*7c3d14c8STreehugger Robot
16*7c3d14c8STreehugger Robot	.syntax unified
17*7c3d14c8STreehugger Robot	.text
18*7c3d14c8STreehugger Robot#if __ARM_ARCH_ISA_THUMB == 2
19*7c3d14c8STreehugger Robot	.thumb
20*7c3d14c8STreehugger Robot#endif
21*7c3d14c8STreehugger Robot
22*7c3d14c8STreehugger Robot	.p2align	2
23*7c3d14c8STreehugger Robot#if __ARM_ARCH_ISA_THUMB == 2
24*7c3d14c8STreehugger RobotDEFINE_COMPILERRT_THUMB_FUNCTION(__clzsi2)
25*7c3d14c8STreehugger Robot#else
26*7c3d14c8STreehugger RobotDEFINE_COMPILERRT_FUNCTION(__clzsi2)
27*7c3d14c8STreehugger Robot#endif
28*7c3d14c8STreehugger Robot#ifdef __ARM_FEATURE_CLZ
29*7c3d14c8STreehugger Robot	clz	r0, r0
30*7c3d14c8STreehugger Robot	JMP(lr)
31*7c3d14c8STreehugger Robot#else
32*7c3d14c8STreehugger Robot	/* Assumption: n != 0 */
33*7c3d14c8STreehugger Robot
34*7c3d14c8STreehugger Robot	/*
35*7c3d14c8STreehugger Robot	 * r0: n
36*7c3d14c8STreehugger Robot	 * r1: count of leading zeros in n + 1
37*7c3d14c8STreehugger Robot	 * r2: scratch register for shifted r0
38*7c3d14c8STreehugger Robot	 */
39*7c3d14c8STreehugger Robot	mov	r1, 1
40*7c3d14c8STreehugger Robot
41*7c3d14c8STreehugger Robot	/*
42*7c3d14c8STreehugger Robot	 * Basic block:
43*7c3d14c8STreehugger Robot	 * if ((r0 >> SHIFT) == 0)
44*7c3d14c8STreehugger Robot	 *   r1 += SHIFT;
45*7c3d14c8STreehugger Robot	 * else
46*7c3d14c8STreehugger Robot	 *   r0 >>= SHIFT;
47*7c3d14c8STreehugger Robot	 * for descending powers of two as SHIFT.
48*7c3d14c8STreehugger Robot	 */
49*7c3d14c8STreehugger Robot
50*7c3d14c8STreehugger Robot#define BLOCK(shift) \
51*7c3d14c8STreehugger Robot	lsrs	r2, r0, shift; \
52*7c3d14c8STreehugger Robot	movne	r0, r2; \
53*7c3d14c8STreehugger Robot	addeq	r1, shift \
54*7c3d14c8STreehugger Robot
55*7c3d14c8STreehugger Robot	BLOCK(16)
56*7c3d14c8STreehugger Robot	BLOCK(8)
57*7c3d14c8STreehugger Robot	BLOCK(4)
58*7c3d14c8STreehugger Robot	BLOCK(2)
59*7c3d14c8STreehugger Robot
60*7c3d14c8STreehugger Robot	/*
61*7c3d14c8STreehugger Robot	 * The basic block invariants at this point are (r0 >> 2) == 0 and
62*7c3d14c8STreehugger Robot	 * r0 != 0. This means 1 <= r0 <= 3 and 0 <= (r0 >> 1) <= 1.
63*7c3d14c8STreehugger Robot	 *
64*7c3d14c8STreehugger Robot	 * r0 | (r0 >> 1) == 0 | (r0 >> 1) == 1 | -(r0 >> 1) | 1 - (r0 >> 1)
65*7c3d14c8STreehugger Robot	 * ---+----------------+----------------+------------+--------------
66*7c3d14c8STreehugger Robot	 * 1  | 1              | 0              | 0          | 1
67*7c3d14c8STreehugger Robot	 * 2  | 0              | 1              | -1         | 0
68*7c3d14c8STreehugger Robot	 * 3  | 0              | 1              | -1         | 0
69*7c3d14c8STreehugger Robot	 *
70*7c3d14c8STreehugger Robot	 * The r1's initial value of 1 compensates for the 1 here.
71*7c3d14c8STreehugger Robot	 */
72*7c3d14c8STreehugger Robot	sub	r0, r1, r0, lsr #1
73*7c3d14c8STreehugger Robot
74*7c3d14c8STreehugger Robot	JMP(lr)
75*7c3d14c8STreehugger Robot#endif // __ARM_FEATURE_CLZ
76*7c3d14c8STreehugger RobotEND_COMPILERRT_FUNCTION(__clzsi2)
77*7c3d14c8STreehugger Robot
78*7c3d14c8STreehugger RobotNO_EXEC_STACK_DIRECTIVE
79*7c3d14c8STreehugger Robot
80