authorgravatar for jan.hafer@rwth-aachen.deJan Philipp Hafer <jan.hafer@rwth-aachen.de> 2021-08-24 23:26:37+02:00
committergravatar for andrew@ziglang.orgAndrew Kelley <andrew@ziglang.org> 2021-09-01 16:15:21-04:00
log81e2034d4ad935185344030e5ecedbd1388a65e9
tree2417cdc654c74985ad8527df5f14fd15777f073e
parent04b801655cbd9b4f6f00027fc1f638b237890149

compiler_rt: add __clzdi2 and __clzti2

- structure derived from shift.zig - rename clzsi2.zig to count0bits.zig - test cases derived from clzsi2_test.zig See #1290

7 files changed, 894 insertions(+), 130 deletions(-)

CMakeLists.txt+1-1
......@@ -447,8 +447,8 @@ set(ZIG_STAGE2_SOURCES
447447 "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/addXf3.zig"
448448 "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/atomics.zig"
449449 "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/clear_cache.zig"
450 "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/clzsi2.zig"
451450 "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/compareXf2.zig"
451 "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/count0bits.zig"
452452 "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/divdf3.zig"
453453 "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/divsf3.zig"
454454 "${CMAKE_SOURCE_DIR}/lib/std/special/compiler_rt/divtf3.zig"
lib/std/special/compiler_rt.zig+5-1
......@@ -269,8 +269,12 @@ comptime {
269269 const __negdf2 = @import("compiler_rt/negXf2.zig").__negdf2;
270270 @export(__negdf2, .{ .name = "__negdf2", .linkage = linkage });
271271
272 const __clzsi2 = @import("compiler_rt/clzsi2.zig").__clzsi2;
272 const __clzsi2 = @import("compiler_rt/count0bits.zig").__clzsi2;
273273 @export(__clzsi2, .{ .name = "__clzsi2", .linkage = linkage });
274 const __clzdi2 = @import("compiler_rt/count0bits.zig").__clzdi2;
275 @export(__clzdi2, .{ .name = "__clzdi2", .linkage = linkage });
276 const __clzti2 = @import("compiler_rt/count0bits.zig").__clzti2;
277 @export(__clzti2, .{ .name = "__clzti2", .linkage = linkage });
274278
275279 if (builtin.link_libc and os_tag == .openbsd) {
276280 const __emutls_get_address = @import("compiler_rt/emutls.zig").__emutls_get_address;
lib/std/special/compiler_rt/clzdi2_test.zig created+329
......@@ -0,0 +1,329 @@
1const clz = @import("count0bits.zig");
2const testing = @import("std").testing;
3
4fn test__clzdi2(a: u64, expected: i64) !void {
5 // XXX At high optimization levels this test may be horribly miscompiled if
6 // one of the naked implementations is selected.
7 var nakedClzdi2 = clz.__clzdi2;
8 var actualClzdi2 = @ptrCast(fn (a: i64) callconv(.C) i32, nakedClzdi2);
9 var x = @bitCast(i64, a);
10 var result = actualClzdi2(x);
11 try testing.expectEqual(expected, result);
12}
13
14test "clzdi2" {
15 try test__clzdi2(0x00800000_00000000, 8);
16 try test__clzdi2(0x01000000_00000000, 7);
17 try test__clzdi2(0x02000000_00000000, 6);
18 try test__clzdi2(0x03000000_00000000, 6);
19 try test__clzdi2(0x04000000_00000000, 5);
20 try test__clzdi2(0x05000000_00000000, 5);
21 try test__clzdi2(0x06000000_00000000, 5);
22 try test__clzdi2(0x07000000_00000000, 5);
23 try test__clzdi2(0x08000000_00000000, 4);
24 try test__clzdi2(0x09000000_00000000, 4);
25 try test__clzdi2(0x0A000000_00000000, 4);
26 try test__clzdi2(0x0B000000_00000000, 4);
27 try test__clzdi2(0x0C000000_00000000, 4);
28 try test__clzdi2(0x0D000000_00000000, 4);
29 try test__clzdi2(0x0E000000_00000000, 4);
30 try test__clzdi2(0x0F000000_00000000, 4);
31 try test__clzdi2(0x10000000_00000000, 3);
32 try test__clzdi2(0x11000000_00000000, 3);
33 try test__clzdi2(0x12000000_00000000, 3);
34 try test__clzdi2(0x13000000_00000000, 3);
35 try test__clzdi2(0x14000000_00000000, 3);
36 try test__clzdi2(0x15000000_00000000, 3);
37 try test__clzdi2(0x16000000_00000000, 3);
38 try test__clzdi2(0x17000000_00000000, 3);
39 try test__clzdi2(0x18000000_00000000, 3);
40 try test__clzdi2(0x19000000_00000000, 3);
41 try test__clzdi2(0x1A000000_00000000, 3);
42 try test__clzdi2(0x1B000000_00000000, 3);
43 try test__clzdi2(0x1C000000_00000000, 3);
44 try test__clzdi2(0x1D000000_00000000, 3);
45 try test__clzdi2(0x1E000000_00000000, 3);
46 try test__clzdi2(0x1F000000_00000000, 3);
47 try test__clzdi2(0x20000000_00000000, 2);
48 try test__clzdi2(0x21000000_00000000, 2);
49 try test__clzdi2(0x22000000_00000000, 2);
50 try test__clzdi2(0x23000000_00000000, 2);
51 try test__clzdi2(0x24000000_00000000, 2);
52 try test__clzdi2(0x25000000_00000000, 2);
53 try test__clzdi2(0x26000000_00000000, 2);
54 try test__clzdi2(0x27000000_00000000, 2);
55 try test__clzdi2(0x28000000_00000000, 2);
56 try test__clzdi2(0x29000000_00000000, 2);
57 try test__clzdi2(0x2A000000_00000000, 2);
58 try test__clzdi2(0x2B000000_00000000, 2);
59 try test__clzdi2(0x2C000000_00000000, 2);
60 try test__clzdi2(0x2D000000_00000000, 2);
61 try test__clzdi2(0x2E000000_00000000, 2);
62 try test__clzdi2(0x2F000000_00000000, 2);
63 try test__clzdi2(0x30000000_00000000, 2);
64 try test__clzdi2(0x31000000_00000000, 2);
65 try test__clzdi2(0x32000000_00000000, 2);
66 try test__clzdi2(0x33000000_00000000, 2);
67 try test__clzdi2(0x34000000_00000000, 2);
68 try test__clzdi2(0x35000000_00000000, 2);
69 try test__clzdi2(0x36000000_00000000, 2);
70 try test__clzdi2(0x37000000_00000000, 2);
71 try test__clzdi2(0x38000000_00000000, 2);
72 try test__clzdi2(0x39000000_00000000, 2);
73 try test__clzdi2(0x3A000000_00000000, 2);
74 try test__clzdi2(0x3B000000_00000000, 2);
75 try test__clzdi2(0x3C000000_00000000, 2);
76 try test__clzdi2(0x3D000000_00000000, 2);
77 try test__clzdi2(0x3E000000_00000000, 2);
78 try test__clzdi2(0x3F000000_00000000, 2);
79 try test__clzdi2(0x40000000_00000000, 1);
80 try test__clzdi2(0x41000000_00000000, 1);
81 try test__clzdi2(0x42000000_00000000, 1);
82 try test__clzdi2(0x43000000_00000000, 1);
83 try test__clzdi2(0x44000000_00000000, 1);
84 try test__clzdi2(0x45000000_00000000, 1);
85 try test__clzdi2(0x46000000_00000000, 1);
86 try test__clzdi2(0x47000000_00000000, 1);
87 try test__clzdi2(0x48000000_00000000, 1);
88 try test__clzdi2(0x49000000_00000000, 1);
89 try test__clzdi2(0x4A000000_00000000, 1);
90 try test__clzdi2(0x4B000000_00000000, 1);
91 try test__clzdi2(0x4C000000_00000000, 1);
92 try test__clzdi2(0x4D000000_00000000, 1);
93 try test__clzdi2(0x4E000000_00000000, 1);
94 try test__clzdi2(0x4F000000_00000000, 1);
95 try test__clzdi2(0x50000000_00000000, 1);
96 try test__clzdi2(0x51000000_00000000, 1);
97 try test__clzdi2(0x52000000_00000000, 1);
98 try test__clzdi2(0x53000000_00000000, 1);
99 try test__clzdi2(0x54000000_00000000, 1);
100 try test__clzdi2(0x55000000_00000000, 1);
101 try test__clzdi2(0x56000000_00000000, 1);
102 try test__clzdi2(0x57000000_00000000, 1);
103 try test__clzdi2(0x58000000_00000000, 1);
104 try test__clzdi2(0x59000000_00000000, 1);
105 try test__clzdi2(0x5A000000_00000000, 1);
106 try test__clzdi2(0x5B000000_00000000, 1);
107 try test__clzdi2(0x5C000000_00000000, 1);
108 try test__clzdi2(0x5D000000_00000000, 1);
109 try test__clzdi2(0x5E000000_00000000, 1);
110 try test__clzdi2(0x5F000000_00000000, 1);
111 try test__clzdi2(0x60000000_00000000, 1);
112 try test__clzdi2(0x61000000_00000000, 1);
113 try test__clzdi2(0x62000000_00000000, 1);
114 try test__clzdi2(0x63000000_00000000, 1);
115 try test__clzdi2(0x64000000_00000000, 1);
116 try test__clzdi2(0x65000000_00000000, 1);
117 try test__clzdi2(0x66000000_00000000, 1);
118 try test__clzdi2(0x67000000_00000000, 1);
119 try test__clzdi2(0x68000000_00000000, 1);
120 try test__clzdi2(0x69000000_00000000, 1);
121 try test__clzdi2(0x6A000000_00000000, 1);
122 try test__clzdi2(0x6B000000_00000000, 1);
123 try test__clzdi2(0x6C000000_00000000, 1);
124 try test__clzdi2(0x6D000000_00000000, 1);
125 try test__clzdi2(0x6E000000_00000000, 1);
126 try test__clzdi2(0x6F000000_00000000, 1);
127 try test__clzdi2(0x70000000_00000000, 1);
128 try test__clzdi2(0x71000000_00000000, 1);
129 try test__clzdi2(0x72000000_00000000, 1);
130 try test__clzdi2(0x73000000_00000000, 1);
131 try test__clzdi2(0x74000000_00000000, 1);
132 try test__clzdi2(0x75000000_00000000, 1);
133 try test__clzdi2(0x76000000_00000000, 1);
134 try test__clzdi2(0x77000000_00000000, 1);
135 try test__clzdi2(0x78000000_00000000, 1);
136 try test__clzdi2(0x79000000_00000000, 1);
137 try test__clzdi2(0x7A000000_00000000, 1);
138 try test__clzdi2(0x7B000000_00000000, 1);
139 try test__clzdi2(0x7C000000_00000000, 1);
140 try test__clzdi2(0x7D000000_00000000, 1);
141 try test__clzdi2(0x7E000000_00000000, 1);
142 try test__clzdi2(0x7F000000_00000000, 1);
143 try test__clzdi2(0x80000000_00000000, 0);
144 try test__clzdi2(0x81000000_00000000, 0);
145 try test__clzdi2(0x82000000_00000000, 0);
146 try test__clzdi2(0x83000000_00000000, 0);
147 try test__clzdi2(0x84000000_00000000, 0);
148 try test__clzdi2(0x85000000_00000000, 0);
149 try test__clzdi2(0x86000000_00000000, 0);
150 try test__clzdi2(0x87000000_00000000, 0);
151 try test__clzdi2(0x88000000_00000000, 0);
152 try test__clzdi2(0x89000000_00000000, 0);
153 try test__clzdi2(0x8A000000_00000000, 0);
154 try test__clzdi2(0x8B000000_00000000, 0);
155 try test__clzdi2(0x8C000000_00000000, 0);
156 try test__clzdi2(0x8D000000_00000000, 0);
157 try test__clzdi2(0x8E000000_00000000, 0);
158 try test__clzdi2(0x8F000000_00000000, 0);
159 try test__clzdi2(0x90000000_00000000, 0);
160 try test__clzdi2(0x91000000_00000000, 0);
161 try test__clzdi2(0x92000000_00000000, 0);
162 try test__clzdi2(0x93000000_00000000, 0);
163 try test__clzdi2(0x94000000_00000000, 0);
164 try test__clzdi2(0x95000000_00000000, 0);
165 try test__clzdi2(0x96000000_00000000, 0);
166 try test__clzdi2(0x97000000_00000000, 0);
167 try test__clzdi2(0x98000000_00000000, 0);
168 try test__clzdi2(0x99000000_00000000, 0);
169 try test__clzdi2(0x9A000000_00000000, 0);
170 try test__clzdi2(0x9B000000_00000000, 0);
171 try test__clzdi2(0x9C000000_00000000, 0);
172 try test__clzdi2(0x9D000000_00000000, 0);
173 try test__clzdi2(0x9E000000_00000000, 0);
174 try test__clzdi2(0x9F000000_00000000, 0);
175 try test__clzdi2(0xA0000000_00000000, 0);
176 try test__clzdi2(0xA1000000_00000000, 0);
177 try test__clzdi2(0xA2000000_00000000, 0);
178 try test__clzdi2(0xA3000000_00000000, 0);
179 try test__clzdi2(0xA4000000_00000000, 0);
180 try test__clzdi2(0xA5000000_00000000, 0);
181 try test__clzdi2(0xA6000000_00000000, 0);
182 try test__clzdi2(0xA7000000_00000000, 0);
183 try test__clzdi2(0xA8000000_00000000, 0);
184 try test__clzdi2(0xA9000000_00000000, 0);
185 try test__clzdi2(0xAA000000_00000000, 0);
186 try test__clzdi2(0xAB000000_00000000, 0);
187 try test__clzdi2(0xAC000000_00000000, 0);
188 try test__clzdi2(0xAD000000_00000000, 0);
189 try test__clzdi2(0xAE000000_00000000, 0);
190 try test__clzdi2(0xAF000000_00000000, 0);
191 try test__clzdi2(0xB0000000_00000000, 0);
192 try test__clzdi2(0xB1000000_00000000, 0);
193 try test__clzdi2(0xB2000000_00000000, 0);
194 try test__clzdi2(0xB3000000_00000000, 0);
195 try test__clzdi2(0xB4000000_00000000, 0);
196 try test__clzdi2(0xB5000000_00000000, 0);
197 try test__clzdi2(0xB6000000_00000000, 0);
198 try test__clzdi2(0xB7000000_00000000, 0);
199 try test__clzdi2(0xB8000000_00000000, 0);
200 try test__clzdi2(0xB9000000_00000000, 0);
201 try test__clzdi2(0xBA000000_00000000, 0);
202 try test__clzdi2(0xBB000000_00000000, 0);
203 try test__clzdi2(0xBC000000_00000000, 0);
204 try test__clzdi2(0xBD000000_00000000, 0);
205 try test__clzdi2(0xBE000000_00000000, 0);
206 try test__clzdi2(0xBF000000_00000000, 0);
207 try test__clzdi2(0xC0000000_00000000, 0);
208 try test__clzdi2(0xC1000000_00000000, 0);
209 try test__clzdi2(0xC2000000_00000000, 0);
210 try test__clzdi2(0xC3000000_00000000, 0);
211 try test__clzdi2(0xC4000000_00000000, 0);
212 try test__clzdi2(0xC5000000_00000000, 0);
213 try test__clzdi2(0xC6000000_00000000, 0);
214 try test__clzdi2(0xC7000000_00000000, 0);
215 try test__clzdi2(0xC8000000_00000000, 0);
216 try test__clzdi2(0xC9000000_00000000, 0);
217 try test__clzdi2(0xCA000000_00000000, 0);
218 try test__clzdi2(0xCB000000_00000000, 0);
219 try test__clzdi2(0xCC000000_00000000, 0);
220 try test__clzdi2(0xCD000000_00000000, 0);
221 try test__clzdi2(0xCE000000_00000000, 0);
222 try test__clzdi2(0xCF000000_00000000, 0);
223 try test__clzdi2(0xD0000000_00000000, 0);
224 try test__clzdi2(0xD1000000_00000000, 0);
225 try test__clzdi2(0xD2000000_00000000, 0);
226 try test__clzdi2(0xD3000000_00000000, 0);
227 try test__clzdi2(0xD4000000_00000000, 0);
228 try test__clzdi2(0xD5000000_00000000, 0);
229 try test__clzdi2(0xD6000000_00000000, 0);
230 try test__clzdi2(0xD7000000_00000000, 0);
231 try test__clzdi2(0xD8000000_00000000, 0);
232 try test__clzdi2(0xD9000000_00000000, 0);
233 try test__clzdi2(0xDA000000_00000000, 0);
234 try test__clzdi2(0xDB000000_00000000, 0);
235 try test__clzdi2(0xDC000000_00000000, 0);
236 try test__clzdi2(0xDD000000_00000000, 0);
237 try test__clzdi2(0xDE000000_00000000, 0);
238 try test__clzdi2(0xDF000000_00000000, 0);
239 try test__clzdi2(0xE0000000_00000000, 0);
240 try test__clzdi2(0xE1000000_00000000, 0);
241 try test__clzdi2(0xE2000000_00000000, 0);
242 try test__clzdi2(0xE3000000_00000000, 0);
243 try test__clzdi2(0xE4000000_00000000, 0);
244 try test__clzdi2(0xE5000000_00000000, 0);
245 try test__clzdi2(0xE6000000_00000000, 0);
246 try test__clzdi2(0xE7000000_00000000, 0);
247 try test__clzdi2(0xE8000000_00000000, 0);
248 try test__clzdi2(0xE9000000_00000000, 0);
249 try test__clzdi2(0xEA000000_00000000, 0);
250 try test__clzdi2(0xEB000000_00000000, 0);
251 try test__clzdi2(0xEC000000_00000000, 0);
252 try test__clzdi2(0xED000000_00000000, 0);
253 try test__clzdi2(0xEE000000_00000000, 0);
254 try test__clzdi2(0xEF000000_00000000, 0);
255 try test__clzdi2(0xF0000000_00000000, 0);
256 try test__clzdi2(0xF1000000_00000000, 0);
257 try test__clzdi2(0xF2000000_00000000, 0);
258 try test__clzdi2(0xF3000000_00000000, 0);
259 try test__clzdi2(0xF4000000_00000000, 0);
260 try test__clzdi2(0xF5000000_00000000, 0);
261 try test__clzdi2(0xF6000000_00000000, 0);
262 try test__clzdi2(0xF7000000_00000000, 0);
263 try test__clzdi2(0xF8000000_00000000, 0);
264 try test__clzdi2(0xF9000000_00000000, 0);
265 try test__clzdi2(0xFA000000_00000000, 0);
266 try test__clzdi2(0xFB000000_00000000, 0);
267 try test__clzdi2(0xFC000000_00000000, 0);
268 try test__clzdi2(0xFD000000_00000000, 0);
269 try test__clzdi2(0xFE000000_00000000, 0);
270 try test__clzdi2(0xFF000000_00000000, 0);
271
272 try test__clzdi2(0x00000000_00000000, 64);
273 try test__clzdi2(0x00000000_00000001, 63);
274 try test__clzdi2(0x00000000_00000002, 62);
275 try test__clzdi2(0x00000000_00000004, 61);
276 try test__clzdi2(0x00000000_00000008, 60);
277 try test__clzdi2(0x00000000_00000010, 59);
278 try test__clzdi2(0x00000000_00000020, 58);
279 try test__clzdi2(0x00000000_00000040, 57);
280 try test__clzdi2(0x00000000_00000080, 56);
281 try test__clzdi2(0x00000000_00000100, 55);
282 try test__clzdi2(0x00000000_00000200, 54);
283 try test__clzdi2(0x00000000_00000400, 53);
284 try test__clzdi2(0x00000000_00000800, 52);
285 try test__clzdi2(0x00000000_00001000, 51);
286 try test__clzdi2(0x00000000_00002000, 50);
287 try test__clzdi2(0x00000000_00004000, 49);
288 try test__clzdi2(0x00000000_00008000, 48);
289 try test__clzdi2(0x00000000_00010000, 47);
290 try test__clzdi2(0x00000000_00020000, 46);
291 try test__clzdi2(0x00000000_00040000, 45);
292 try test__clzdi2(0x00000000_00080000, 44);
293 try test__clzdi2(0x00000000_00100000, 43);
294 try test__clzdi2(0x00000000_00200000, 42);
295 try test__clzdi2(0x00000000_00400000, 41);
296 try test__clzdi2(0x00000000_00800000, 40);
297 try test__clzdi2(0x00000000_01000000, 39);
298 try test__clzdi2(0x00000000_02000000, 38);
299 try test__clzdi2(0x00000000_04000000, 37);
300 try test__clzdi2(0x00000000_08000000, 36);
301 try test__clzdi2(0x00000000_10000000, 35);
302 try test__clzdi2(0x00000000_20000000, 34);
303 try test__clzdi2(0x00000000_40000000, 33);
304 try test__clzdi2(0x00000000_80000000, 32);
305
306 try test__clzdi2(0x00000001_00000000, 31);
307 try test__clzdi2(0x00000002_00000000, 30);
308 try test__clzdi2(0x00000004_00000000, 29);
309 try test__clzdi2(0x00000008_00000000, 28);
310 try test__clzdi2(0x00000010_00000000, 27);
311 try test__clzdi2(0x00000020_00000000, 26);
312 try test__clzdi2(0x00000040_00000000, 25);
313 try test__clzdi2(0x00000080_00000000, 24);
314 try test__clzdi2(0x00000100_00000000, 23);
315 try test__clzdi2(0x00000200_00000000, 22);
316 try test__clzdi2(0x00000400_00000000, 21);
317 try test__clzdi2(0x00000800_00000000, 20);
318 try test__clzdi2(0x00001000_00000000, 19);
319 try test__clzdi2(0x00002000_00000000, 18);
320 try test__clzdi2(0x00004000_00000000, 17);
321 try test__clzdi2(0x00008000_00000000, 16);
322 try test__clzdi2(0x00010000_00000000, 15);
323 try test__clzdi2(0x00020000_00000000, 14);
324 try test__clzdi2(0x00040000_00000000, 13);
325 try test__clzdi2(0x00080000_00000000, 12);
326 try test__clzdi2(0x00100000_00000000, 11);
327 try test__clzdi2(0x00200000_00000000, 10);
328 try test__clzdi2(0x00400000_00000000, 9);
329}
lib/std/special/compiler_rt/clzsi2.zig deleted-126
......@@ -1,126 +0,0 @@
1const std = @import("std");
2const builtin = std.builtin;
3
4fn __clzsi2_generic(a: i32) callconv(.C) i32 {
5 @setRuntimeSafety(builtin.is_test);
6
7 var x = @bitCast(u32, a);
8 var n: i32 = 32;
9
10 // Count first bit set using binary search, from Hacker's Delight
11 var y: u32 = 0;
12 inline for ([_]i32{ 16, 8, 4, 2, 1 }) |shift| {
13 y = x >> shift;
14 if (y != 0) {
15 n = n - shift;
16 x = y;
17 }
18 }
19
20 return n - @bitCast(i32, x);
21}
22
23fn __clzsi2_thumb1() callconv(.Naked) void {
24 @setRuntimeSafety(false);
25
26 // Similar to the generic version with the last two rounds replaced by a LUT
27 asm volatile (
28 \\ movs r1, #32
29 \\ lsrs r2, r0, #16
30 \\ beq 1f
31 \\ subs r1, #16
32 \\ movs r0, r2
33 \\ 1:
34 \\ lsrs r2, r0, #8
35 \\ beq 1f
36 \\ subs r1, #8
37 \\ movs r0, r2
38 \\ 1:
39 \\ lsrs r2, r0, #4
40 \\ beq 1f
41 \\ subs r1, #4
42 \\ movs r0, r2
43 \\ 1:
44 \\ ldr r3, =LUT
45 \\ ldrb r0, [r3, r0]
46 \\ subs r0, r1, r0
47 \\ bx lr
48 \\ .p2align 2
49 \\ // Number of bits set in the 0-15 range
50 \\ LUT:
51 \\ .byte 0,1,2,2,3,3,3,3,4,4,4,4,4,4,4,4
52 );
53
54 unreachable;
55}
56
57fn __clzsi2_arm32() callconv(.Naked) void {
58 @setRuntimeSafety(false);
59
60 asm volatile (
61 \\ // Assumption: n != 0
62 \\ // r0: n
63 \\ // r1: count of leading zeros in n + 1
64 \\ // r2: scratch register for shifted r0
65 \\ mov r1, #1
66 \\
67 \\ // Basic block:
68 \\ // if ((r0 >> SHIFT) == 0)
69 \\ // r1 += SHIFT;
70 \\ // else
71 \\ // r0 >>= SHIFT;
72 \\ // for descending powers of two as SHIFT.
73 \\ lsrs r2, r0, #16
74 \\ movne r0, r2
75 \\ addeq r1, #16
76 \\
77 \\ lsrs r2, r0, #8
78 \\ movne r0, r2
79 \\ addeq r1, #8
80 \\
81 \\ lsrs r2, r0, #4
82 \\ movne r0, r2
83 \\ addeq r1, #4
84 \\
85 \\ lsrs r2, r0, #2
86 \\ movne r0, r2
87 \\ addeq r1, #2
88 \\
89 \\ // The basic block invariants at this point are (r0 >> 2) == 0 and
90 \\ // r0 != 0. This means 1 <= r0 <= 3 and 0 <= (r0 >> 1) <= 1.
91 \\ //
92 \\ // r0 | (r0 >> 1) == 0 | (r0 >> 1) == 1 | -(r0 >> 1) | 1 - (r0 >> 1)f
93 \\ // ---+----------------+----------------+------------+--------------
94 \\ // 1 | 1 | 0 | 0 | 1
95 \\ // 2 | 0 | 1 | -1 | 0
96 \\ // 3 | 0 | 1 | -1 | 0
97 \\ //
98 \\ // The r1's initial value of 1 compensates for the 1 here.
99 \\ sub r0, r1, r0, lsr #1
100 \\ bx lr
101 );
102
103 unreachable;
104}
105
106pub const __clzsi2 = impl: {
107 switch (std.Target.current.cpu.arch) {
108 .arm, .armeb, .thumb, .thumbeb => {
109 const use_thumb1 =
110 (std.Target.current.cpu.arch.isThumb() or
111 std.Target.arm.featureSetHas(std.Target.current.cpu.features, .noarm)) and
112 !std.Target.arm.featureSetHas(std.Target.current.cpu.features, .thumb2);
113
114 if (use_thumb1) break :impl __clzsi2_thumb1
115 // From here on we're either targeting Thumb2 or ARM.
116 else if (!std.Target.current.cpu.arch.isThumb()) break :impl __clzsi2_arm32
117 // Use the generic implementation otherwise.
118 else break :impl __clzsi2_generic;
119 },
120 else => break :impl __clzsi2_generic,
121 }
122};
123
124test "test clzsi2" {
125 _ = @import("clzsi2_test.zig");
126}
lib/std/special/compiler_rt/clzsi2_test.zig+4-2
......@@ -1,10 +1,10 @@
1const clzsi2 = @import("clzsi2.zig");
1const clz = @import("count0bits.zig");
22const testing = @import("std").testing;
33
44fn test__clzsi2(a: u32, expected: i32) !void {
55 // XXX At high optimization levels this test may be horribly miscompiled if
66 // one of the naked implementations is selected.
7 var nakedClzsi2 = clzsi2.__clzsi2;
7 var nakedClzsi2 = clz.__clzsi2;
88 var actualClzsi2 = @ptrCast(fn (a: i32) callconv(.C) i32, nakedClzsi2);
99 var x = @bitCast(i32, a);
1010 var result = actualClzsi2(x);
......@@ -268,6 +268,8 @@ test "clzsi2" {
268268 try test__clzsi2(0xFD000000, 0);
269269 try test__clzsi2(0xFE000000, 0);
270270 try test__clzsi2(0xFF000000, 0);
271 // arm and thumb1 assume input a != 0
272 //try test__clzsi2(0x00000000, 32);
271273 try test__clzsi2(0x00000001, 31);
272274 try test__clzsi2(0x00000002, 30);
273275 try test__clzsi2(0x00000004, 29);
lib/std/special/compiler_rt/clzti2_test.zig created+395
......@@ -0,0 +1,395 @@
1const clz = @import("count0bits.zig");
2const testing = @import("std").testing;
3
4fn test__clzti2(a: u128, expected: i64) !void {
5 // XXX At high optimization levels this test may be horribly miscompiled if
6 // one of the naked implementations is selected.
7 var nakedClzti2 = clz.__clzti2;
8 var actualClzti2 = @ptrCast(fn (a: i128) callconv(.C) i32, nakedClzti2);
9 var x = @bitCast(i128, a);
10 var result = actualClzti2(x);
11 try testing.expectEqual(expected, result);
12}
13
14test "clzti2" {
15 try test__clzti2(0x00800000_00000000_00000000_00000000, 8);
16 try test__clzti2(0x01000000_00000000_00000000_00000000, 7);
17 try test__clzti2(0x02000000_00000000_00000000_00000000, 6);
18 try test__clzti2(0x03000000_00000000_00000000_00000000, 6);
19 try test__clzti2(0x04000000_00000000_00000000_00000000, 5);
20 try test__clzti2(0x05000000_00000000_00000000_00000000, 5);
21 try test__clzti2(0x06000000_00000000_00000000_00000000, 5);
22 try test__clzti2(0x07000000_00000000_00000000_00000000, 5);
23 try test__clzti2(0x08000000_00000000_00000000_00000000, 4);
24 try test__clzti2(0x09000000_00000000_00000000_00000000, 4);
25 try test__clzti2(0x0A000000_00000000_00000000_00000000, 4);
26 try test__clzti2(0x0B000000_00000000_00000000_00000000, 4);
27 try test__clzti2(0x0C000000_00000000_00000000_00000000, 4);
28 try test__clzti2(0x0D000000_00000000_00000000_00000000, 4);
29 try test__clzti2(0x0E000000_00000000_00000000_00000000, 4);
30 try test__clzti2(0x0F000000_00000000_00000000_00000000, 4);
31 try test__clzti2(0x10000000_00000000_00000000_00000000, 3);
32 try test__clzti2(0x11000000_00000000_00000000_00000000, 3);
33 try test__clzti2(0x12000000_00000000_00000000_00000000, 3);
34 try test__clzti2(0x13000000_00000000_00000000_00000000, 3);
35 try test__clzti2(0x14000000_00000000_00000000_00000000, 3);
36 try test__clzti2(0x15000000_00000000_00000000_00000000, 3);
37 try test__clzti2(0x16000000_00000000_00000000_00000000, 3);
38 try test__clzti2(0x17000000_00000000_00000000_00000000, 3);
39 try test__clzti2(0x18000000_00000000_00000000_00000000, 3);
40 try test__clzti2(0x19000000_00000000_00000000_00000000, 3);
41 try test__clzti2(0x1A000000_00000000_00000000_00000000, 3);
42 try test__clzti2(0x1B000000_00000000_00000000_00000000, 3);
43 try test__clzti2(0x1C000000_00000000_00000000_00000000, 3);
44 try test__clzti2(0x1D000000_00000000_00000000_00000000, 3);
45 try test__clzti2(0x1E000000_00000000_00000000_00000000, 3);
46 try test__clzti2(0x1F000000_00000000_00000000_00000000, 3);
47 try test__clzti2(0x20000000_00000000_00000000_00000000, 2);
48 try test__clzti2(0x21000000_00000000_00000000_00000000, 2);
49 try test__clzti2(0x22000000_00000000_00000000_00000000, 2);
50 try test__clzti2(0x23000000_00000000_00000000_00000000, 2);
51 try test__clzti2(0x24000000_00000000_00000000_00000000, 2);
52 try test__clzti2(0x25000000_00000000_00000000_00000000, 2);
53 try test__clzti2(0x26000000_00000000_00000000_00000000, 2);
54 try test__clzti2(0x27000000_00000000_00000000_00000000, 2);
55 try test__clzti2(0x28000000_00000000_00000000_00000000, 2);
56 try test__clzti2(0x29000000_00000000_00000000_00000000, 2);
57 try test__clzti2(0x2A000000_00000000_00000000_00000000, 2);
58 try test__clzti2(0x2B000000_00000000_00000000_00000000, 2);
59 try test__clzti2(0x2C000000_00000000_00000000_00000000, 2);
60 try test__clzti2(0x2D000000_00000000_00000000_00000000, 2);
61 try test__clzti2(0x2E000000_00000000_00000000_00000000, 2);
62 try test__clzti2(0x2F000000_00000000_00000000_00000000, 2);
63 try test__clzti2(0x30000000_00000000_00000000_00000000, 2);
64 try test__clzti2(0x31000000_00000000_00000000_00000000, 2);
65 try test__clzti2(0x32000000_00000000_00000000_00000000, 2);
66 try test__clzti2(0x33000000_00000000_00000000_00000000, 2);
67 try test__clzti2(0x34000000_00000000_00000000_00000000, 2);
68 try test__clzti2(0x35000000_00000000_00000000_00000000, 2);
69 try test__clzti2(0x36000000_00000000_00000000_00000000, 2);
70 try test__clzti2(0x37000000_00000000_00000000_00000000, 2);
71 try test__clzti2(0x38000000_00000000_00000000_00000000, 2);
72 try test__clzti2(0x39000000_00000000_00000000_00000000, 2);
73 try test__clzti2(0x3A000000_00000000_00000000_00000000, 2);
74 try test__clzti2(0x3B000000_00000000_00000000_00000000, 2);
75 try test__clzti2(0x3C000000_00000000_00000000_00000000, 2);
76 try test__clzti2(0x3D000000_00000000_00000000_00000000, 2);
77 try test__clzti2(0x3E000000_00000000_00000000_00000000, 2);
78 try test__clzti2(0x3F000000_00000000_00000000_00000000, 2);
79 try test__clzti2(0x40000000_00000000_00000000_00000000, 1);
80 try test__clzti2(0x41000000_00000000_00000000_00000000, 1);
81 try test__clzti2(0x42000000_00000000_00000000_00000000, 1);
82 try test__clzti2(0x43000000_00000000_00000000_00000000, 1);
83 try test__clzti2(0x44000000_00000000_00000000_00000000, 1);
84 try test__clzti2(0x45000000_00000000_00000000_00000000, 1);
85 try test__clzti2(0x46000000_00000000_00000000_00000000, 1);
86 try test__clzti2(0x47000000_00000000_00000000_00000000, 1);
87 try test__clzti2(0x48000000_00000000_00000000_00000000, 1);
88 try test__clzti2(0x49000000_00000000_00000000_00000000, 1);
89 try test__clzti2(0x4A000000_00000000_00000000_00000000, 1);
90 try test__clzti2(0x4B000000_00000000_00000000_00000000, 1);
91 try test__clzti2(0x4C000000_00000000_00000000_00000000, 1);
92 try test__clzti2(0x4D000000_00000000_00000000_00000000, 1);
93 try test__clzti2(0x4E000000_00000000_00000000_00000000, 1);
94 try test__clzti2(0x4F000000_00000000_00000000_00000000, 1);
95 try test__clzti2(0x50000000_00000000_00000000_00000000, 1);
96 try test__clzti2(0x51000000_00000000_00000000_00000000, 1);
97 try test__clzti2(0x52000000_00000000_00000000_00000000, 1);
98 try test__clzti2(0x53000000_00000000_00000000_00000000, 1);
99 try test__clzti2(0x54000000_00000000_00000000_00000000, 1);
100 try test__clzti2(0x55000000_00000000_00000000_00000000, 1);
101 try test__clzti2(0x56000000_00000000_00000000_00000000, 1);
102 try test__clzti2(0x57000000_00000000_00000000_00000000, 1);
103 try test__clzti2(0x58000000_00000000_00000000_00000000, 1);
104 try test__clzti2(0x59000000_00000000_00000000_00000000, 1);
105 try test__clzti2(0x5A000000_00000000_00000000_00000000, 1);
106 try test__clzti2(0x5B000000_00000000_00000000_00000000, 1);
107 try test__clzti2(0x5C000000_00000000_00000000_00000000, 1);
108 try test__clzti2(0x5D000000_00000000_00000000_00000000, 1);
109 try test__clzti2(0x5E000000_00000000_00000000_00000000, 1);
110 try test__clzti2(0x5F000000_00000000_00000000_00000000, 1);
111 try test__clzti2(0x60000000_00000000_00000000_00000000, 1);
112 try test__clzti2(0x61000000_00000000_00000000_00000000, 1);
113 try test__clzti2(0x62000000_00000000_00000000_00000000, 1);
114 try test__clzti2(0x63000000_00000000_00000000_00000000, 1);
115 try test__clzti2(0x64000000_00000000_00000000_00000000, 1);
116 try test__clzti2(0x65000000_00000000_00000000_00000000, 1);
117 try test__clzti2(0x66000000_00000000_00000000_00000000, 1);
118 try test__clzti2(0x67000000_00000000_00000000_00000000, 1);
119 try test__clzti2(0x68000000_00000000_00000000_00000000, 1);
120 try test__clzti2(0x69000000_00000000_00000000_00000000, 1);
121 try test__clzti2(0x6A000000_00000000_00000000_00000000, 1);
122 try test__clzti2(0x6B000000_00000000_00000000_00000000, 1);
123 try test__clzti2(0x6C000000_00000000_00000000_00000000, 1);
124 try test__clzti2(0x6D000000_00000000_00000000_00000000, 1);
125 try test__clzti2(0x6E000000_00000000_00000000_00000000, 1);
126 try test__clzti2(0x6F000000_00000000_00000000_00000000, 1);
127 try test__clzti2(0x70000000_00000000_00000000_00000000, 1);
128 try test__clzti2(0x71000000_00000000_00000000_00000000, 1);
129 try test__clzti2(0x72000000_00000000_00000000_00000000, 1);
130 try test__clzti2(0x73000000_00000000_00000000_00000000, 1);
131 try test__clzti2(0x74000000_00000000_00000000_00000000, 1);
132 try test__clzti2(0x75000000_00000000_00000000_00000000, 1);
133 try test__clzti2(0x76000000_00000000_00000000_00000000, 1);
134 try test__clzti2(0x77000000_00000000_00000000_00000000, 1);
135 try test__clzti2(0x78000000_00000000_00000000_00000000, 1);
136 try test__clzti2(0x79000000_00000000_00000000_00000000, 1);
137 try test__clzti2(0x7A000000_00000000_00000000_00000000, 1);
138 try test__clzti2(0x7B000000_00000000_00000000_00000000, 1);
139 try test__clzti2(0x7C000000_00000000_00000000_00000000, 1);
140 try test__clzti2(0x7D000000_00000000_00000000_00000000, 1);
141 try test__clzti2(0x7E000000_00000000_00000000_00000000, 1);
142 try test__clzti2(0x7F000000_00000000_00000000_00000000, 1);
143 try test__clzti2(0x80000000_00000000_00000000_00000000, 0);
144 try test__clzti2(0x81000000_00000000_00000000_00000000, 0);
145 try test__clzti2(0x82000000_00000000_00000000_00000000, 0);
146 try test__clzti2(0x83000000_00000000_00000000_00000000, 0);
147 try test__clzti2(0x84000000_00000000_00000000_00000000, 0);
148 try test__clzti2(0x85000000_00000000_00000000_00000000, 0);
149 try test__clzti2(0x86000000_00000000_00000000_00000000, 0);
150 try test__clzti2(0x87000000_00000000_00000000_00000000, 0);
151 try test__clzti2(0x88000000_00000000_00000000_00000000, 0);
152 try test__clzti2(0x89000000_00000000_00000000_00000000, 0);
153 try test__clzti2(0x8A000000_00000000_00000000_00000000, 0);
154 try test__clzti2(0x8B000000_00000000_00000000_00000000, 0);
155 try test__clzti2(0x8C000000_00000000_00000000_00000000, 0);
156 try test__clzti2(0x8D000000_00000000_00000000_00000000, 0);
157 try test__clzti2(0x8E000000_00000000_00000000_00000000, 0);
158 try test__clzti2(0x8F000000_00000000_00000000_00000000, 0);
159 try test__clzti2(0x90000000_00000000_00000000_00000000, 0);
160 try test__clzti2(0x91000000_00000000_00000000_00000000, 0);
161 try test__clzti2(0x92000000_00000000_00000000_00000000, 0);
162 try test__clzti2(0x93000000_00000000_00000000_00000000, 0);
163 try test__clzti2(0x94000000_00000000_00000000_00000000, 0);
164 try test__clzti2(0x95000000_00000000_00000000_00000000, 0);
165 try test__clzti2(0x96000000_00000000_00000000_00000000, 0);
166 try test__clzti2(0x97000000_00000000_00000000_00000000, 0);
167 try test__clzti2(0x98000000_00000000_00000000_00000000, 0);
168 try test__clzti2(0x99000000_00000000_00000000_00000000, 0);
169 try test__clzti2(0x9A000000_00000000_00000000_00000000, 0);
170 try test__clzti2(0x9B000000_00000000_00000000_00000000, 0);
171 try test__clzti2(0x9C000000_00000000_00000000_00000000, 0);
172 try test__clzti2(0x9D000000_00000000_00000000_00000000, 0);
173 try test__clzti2(0x9E000000_00000000_00000000_00000000, 0);
174 try test__clzti2(0x9F000000_00000000_00000000_00000000, 0);
175 try test__clzti2(0xA0000000_00000000_00000000_00000000, 0);
176 try test__clzti2(0xA1000000_00000000_00000000_00000000, 0);
177 try test__clzti2(0xA2000000_00000000_00000000_00000000, 0);
178 try test__clzti2(0xA3000000_00000000_00000000_00000000, 0);
179 try test__clzti2(0xA4000000_00000000_00000000_00000000, 0);
180 try test__clzti2(0xA5000000_00000000_00000000_00000000, 0);
181 try test__clzti2(0xA6000000_00000000_00000000_00000000, 0);
182 try test__clzti2(0xA7000000_00000000_00000000_00000000, 0);
183 try test__clzti2(0xA8000000_00000000_00000000_00000000, 0);
184 try test__clzti2(0xA9000000_00000000_00000000_00000000, 0);
185 try test__clzti2(0xAA000000_00000000_00000000_00000000, 0);
186 try test__clzti2(0xAB000000_00000000_00000000_00000000, 0);
187 try test__clzti2(0xAC000000_00000000_00000000_00000000, 0);
188 try test__clzti2(0xAD000000_00000000_00000000_00000000, 0);
189 try test__clzti2(0xAE000000_00000000_00000000_00000000, 0);
190 try test__clzti2(0xAF000000_00000000_00000000_00000000, 0);
191 try test__clzti2(0xB0000000_00000000_00000000_00000000, 0);
192 try test__clzti2(0xB1000000_00000000_00000000_00000000, 0);
193 try test__clzti2(0xB2000000_00000000_00000000_00000000, 0);
194 try test__clzti2(0xB3000000_00000000_00000000_00000000, 0);
195 try test__clzti2(0xB4000000_00000000_00000000_00000000, 0);
196 try test__clzti2(0xB5000000_00000000_00000000_00000000, 0);
197 try test__clzti2(0xB6000000_00000000_00000000_00000000, 0);
198 try test__clzti2(0xB7000000_00000000_00000000_00000000, 0);
199 try test__clzti2(0xB8000000_00000000_00000000_00000000, 0);
200 try test__clzti2(0xB9000000_00000000_00000000_00000000, 0);
201 try test__clzti2(0xBA000000_00000000_00000000_00000000, 0);
202 try test__clzti2(0xBB000000_00000000_00000000_00000000, 0);
203 try test__clzti2(0xBC000000_00000000_00000000_00000000, 0);
204 try test__clzti2(0xBD000000_00000000_00000000_00000000, 0);
205 try test__clzti2(0xBE000000_00000000_00000000_00000000, 0);
206 try test__clzti2(0xBF000000_00000000_00000000_00000000, 0);
207 try test__clzti2(0xC0000000_00000000_00000000_00000000, 0);
208 try test__clzti2(0xC1000000_00000000_00000000_00000000, 0);
209 try test__clzti2(0xC2000000_00000000_00000000_00000000, 0);
210 try test__clzti2(0xC3000000_00000000_00000000_00000000, 0);
211 try test__clzti2(0xC4000000_00000000_00000000_00000000, 0);
212 try test__clzti2(0xC5000000_00000000_00000000_00000000, 0);
213 try test__clzti2(0xC6000000_00000000_00000000_00000000, 0);
214 try test__clzti2(0xC7000000_00000000_00000000_00000000, 0);
215 try test__clzti2(0xC8000000_00000000_00000000_00000000, 0);
216 try test__clzti2(0xC9000000_00000000_00000000_00000000, 0);
217 try test__clzti2(0xCA000000_00000000_00000000_00000000, 0);
218 try test__clzti2(0xCB000000_00000000_00000000_00000000, 0);
219 try test__clzti2(0xCC000000_00000000_00000000_00000000, 0);
220 try test__clzti2(0xCD000000_00000000_00000000_00000000, 0);
221 try test__clzti2(0xCE000000_00000000_00000000_00000000, 0);
222 try test__clzti2(0xCF000000_00000000_00000000_00000000, 0);
223 try test__clzti2(0xD0000000_00000000_00000000_00000000, 0);
224 try test__clzti2(0xD1000000_00000000_00000000_00000000, 0);
225 try test__clzti2(0xD2000000_00000000_00000000_00000000, 0);
226 try test__clzti2(0xD3000000_00000000_00000000_00000000, 0);
227 try test__clzti2(0xD4000000_00000000_00000000_00000000, 0);
228 try test__clzti2(0xD5000000_00000000_00000000_00000000, 0);
229 try test__clzti2(0xD6000000_00000000_00000000_00000000, 0);
230 try test__clzti2(0xD7000000_00000000_00000000_00000000, 0);
231 try test__clzti2(0xD8000000_00000000_00000000_00000000, 0);
232 try test__clzti2(0xD9000000_00000000_00000000_00000000, 0);
233 try test__clzti2(0xDA000000_00000000_00000000_00000000, 0);
234 try test__clzti2(0xDB000000_00000000_00000000_00000000, 0);
235 try test__clzti2(0xDC000000_00000000_00000000_00000000, 0);
236 try test__clzti2(0xDD000000_00000000_00000000_00000000, 0);
237 try test__clzti2(0xDE000000_00000000_00000000_00000000, 0);
238 try test__clzti2(0xDF000000_00000000_00000000_00000000, 0);
239 try test__clzti2(0xE0000000_00000000_00000000_00000000, 0);
240 try test__clzti2(0xE1000000_00000000_00000000_00000000, 0);
241 try test__clzti2(0xE2000000_00000000_00000000_00000000, 0);
242 try test__clzti2(0xE3000000_00000000_00000000_00000000, 0);
243 try test__clzti2(0xE4000000_00000000_00000000_00000000, 0);
244 try test__clzti2(0xE5000000_00000000_00000000_00000000, 0);
245 try test__clzti2(0xE6000000_00000000_00000000_00000000, 0);
246 try test__clzti2(0xE7000000_00000000_00000000_00000000, 0);
247 try test__clzti2(0xE8000000_00000000_00000000_00000000, 0);
248 try test__clzti2(0xE9000000_00000000_00000000_00000000, 0);
249 try test__clzti2(0xEA000000_00000000_00000000_00000000, 0);
250 try test__clzti2(0xEB000000_00000000_00000000_00000000, 0);
251 try test__clzti2(0xEC000000_00000000_00000000_00000000, 0);
252 try test__clzti2(0xED000000_00000000_00000000_00000000, 0);
253 try test__clzti2(0xEE000000_00000000_00000000_00000000, 0);
254 try test__clzti2(0xEF000000_00000000_00000000_00000000, 0);
255 try test__clzti2(0xF0000000_00000000_00000000_00000000, 0);
256 try test__clzti2(0xF1000000_00000000_00000000_00000000, 0);
257 try test__clzti2(0xF2000000_00000000_00000000_00000000, 0);
258 try test__clzti2(0xF3000000_00000000_00000000_00000000, 0);
259 try test__clzti2(0xF4000000_00000000_00000000_00000000, 0);
260 try test__clzti2(0xF5000000_00000000_00000000_00000000, 0);
261 try test__clzti2(0xF6000000_00000000_00000000_00000000, 0);
262 try test__clzti2(0xF7000000_00000000_00000000_00000000, 0);
263 try test__clzti2(0xF8000000_00000000_00000000_00000000, 0);
264 try test__clzti2(0xF9000000_00000000_00000000_00000000, 0);
265 try test__clzti2(0xFA000000_00000000_00000000_00000000, 0);
266 try test__clzti2(0xFB000000_00000000_00000000_00000000, 0);
267 try test__clzti2(0xFC000000_00000000_00000000_00000000, 0);
268 try test__clzti2(0xFD000000_00000000_00000000_00000000, 0);
269 try test__clzti2(0xFE000000_00000000_00000000_00000000, 0);
270 try test__clzti2(0xFF000000_00000000_00000000_00000000, 0);
271
272 try test__clzti2(0x00000000_00000000_00000000_00000000, 128);
273 try test__clzti2(0x00000000_00000000_00000000_00000001, 127);
274 try test__clzti2(0x00000000_00000000_00000000_00000002, 126);
275 try test__clzti2(0x00000000_00000000_00000000_00000004, 125);
276 try test__clzti2(0x00000000_00000000_00000000_00000008, 124);
277 try test__clzti2(0x00000000_00000000_00000000_00000010, 123);
278 try test__clzti2(0x00000000_00000000_00000000_00000020, 122);
279 try test__clzti2(0x00000000_00000000_00000000_00000040, 121);
280 try test__clzti2(0x00000000_00000000_00000000_00000080, 120);
281 try test__clzti2(0x00000000_00000000_00000000_00000100, 119);
282 try test__clzti2(0x00000000_00000000_00000000_00000200, 118);
283 try test__clzti2(0x00000000_00000000_00000000_00000400, 117);
284 try test__clzti2(0x00000000_00000000_00000000_00000800, 116);
285 try test__clzti2(0x00000000_00000000_00000000_00001000, 115);
286 try test__clzti2(0x00000000_00000000_00000000_00002000, 114);
287 try test__clzti2(0x00000000_00000000_00000000_00004000, 113);
288 try test__clzti2(0x00000000_00000000_00000000_00008000, 112);
289 try test__clzti2(0x00000000_00000000_00000000_00010000, 111);
290 try test__clzti2(0x00000000_00000000_00000000_00020000, 110);
291 try test__clzti2(0x00000000_00000000_00000000_00040000, 109);
292 try test__clzti2(0x00000000_00000000_00000000_00080000, 108);
293 try test__clzti2(0x00000000_00000000_00000000_00100000, 107);
294 try test__clzti2(0x00000000_00000000_00000000_00200000, 106);
295 try test__clzti2(0x00000000_00000000_00000000_00400000, 105);
296 try test__clzti2(0x00000000_00000000_00000000_00800000, 104);
297 try test__clzti2(0x00000000_00000000_00000000_01000000, 103);
298 try test__clzti2(0x00000000_00000000_00000000_02000000, 102);
299 try test__clzti2(0x00000000_00000000_00000000_04000000, 101);
300 try test__clzti2(0x00000000_00000000_00000000_08000000, 100);
301 try test__clzti2(0x00000000_00000000_00000000_10000000, 99);
302 try test__clzti2(0x00000000_00000000_00000000_20000000, 98);
303 try test__clzti2(0x00000000_00000000_00000000_40000000, 97);
304 try test__clzti2(0x00000000_00000000_00000000_80000000, 96);
305
306 try test__clzti2(0x00000000_00000000_00000001_00000000, 95);
307 try test__clzti2(0x00000000_00000000_00000002_00000000, 94);
308 try test__clzti2(0x00000000_00000000_00000004_00000000, 93);
309 try test__clzti2(0x00000000_00000000_00000008_00000000, 92);
310 try test__clzti2(0x00000000_00000000_00000010_00000000, 91);
311 try test__clzti2(0x00000000_00000000_00000020_00000000, 90);
312 try test__clzti2(0x00000000_00000000_00000040_00000000, 89);
313 try test__clzti2(0x00000000_00000000_00000080_00000000, 88);
314 try test__clzti2(0x00000000_00000000_00000100_00000000, 87);
315 try test__clzti2(0x00000000_00000000_00000200_00000000, 86);
316 try test__clzti2(0x00000000_00000000_00000400_00000000, 85);
317 try test__clzti2(0x00000000_00000000_00000800_00000000, 84);
318 try test__clzti2(0x00000000_00000000_00001000_00000000, 83);
319 try test__clzti2(0x00000000_00000000_00002000_00000000, 82);
320 try test__clzti2(0x00000000_00000000_00004000_00000000, 81);
321 try test__clzti2(0x00000000_00000000_00008000_00000000, 80);
322 try test__clzti2(0x00000000_00000000_00010000_00000000, 79);
323 try test__clzti2(0x00000000_00000000_00020000_00000000, 78);
324 try test__clzti2(0x00000000_00000000_00040000_00000000, 77);
325 try test__clzti2(0x00000000_00000000_00080000_00000000, 76);
326 try test__clzti2(0x00000000_00000000_00100000_00000000, 75);
327 try test__clzti2(0x00000000_00000000_00200000_00000000, 74);
328 try test__clzti2(0x00000000_00000000_00400000_00000000, 73);
329 try test__clzti2(0x00000000_00000000_00800000_00000000, 72);
330 try test__clzti2(0x00000000_00000000_01000000_00000000, 71);
331 try test__clzti2(0x00000000_00000000_02000000_00000000, 70);
332 try test__clzti2(0x00000000_00000000_04000000_00000000, 69);
333 try test__clzti2(0x00000000_00000000_08000000_00000000, 68);
334 try test__clzti2(0x00000000_00000000_10000000_00000000, 67);
335 try test__clzti2(0x00000000_00000000_20000000_00000000, 66);
336 try test__clzti2(0x00000000_00000000_40000000_00000000, 65);
337 try test__clzti2(0x00000000_00000000_80000000_00000000, 64);
338
339 try test__clzti2(0x00000000_00000001_00000000_00000000, 63);
340 try test__clzti2(0x00000000_00000002_00000000_00000000, 62);
341 try test__clzti2(0x00000000_00000004_00000000_00000000, 61);
342 try test__clzti2(0x00000000_00000008_00000000_00000000, 60);
343 try test__clzti2(0x00000000_00000010_00000000_00000000, 59);
344 try test__clzti2(0x00000000_00000020_00000000_00000000, 58);
345 try test__clzti2(0x00000000_00000040_00000000_00000000, 57);
346 try test__clzti2(0x00000000_00000080_00000000_00000000, 56);
347 try test__clzti2(0x00000000_00000100_00000000_00000000, 55);
348 try test__clzti2(0x00000000_00000200_00000000_00000000, 54);
349 try test__clzti2(0x00000000_00000400_00000000_00000000, 53);
350 try test__clzti2(0x00000000_00000800_00000000_00000000, 52);
351 try test__clzti2(0x00000000_00001000_00000000_00000000, 51);
352 try test__clzti2(0x00000000_00002000_00000000_00000000, 50);
353 try test__clzti2(0x00000000_00004000_00000000_00000000, 49);
354 try test__clzti2(0x00000000_00008000_00000000_00000000, 48);
355 try test__clzti2(0x00000000_00010000_00000000_00000000, 47);
356 try test__clzti2(0x00000000_00020000_00000000_00000000, 46);
357 try test__clzti2(0x00000000_00040000_00000000_00000000, 45);
358 try test__clzti2(0x00000000_00080000_00000000_00000000, 44);
359 try test__clzti2(0x00000000_00100000_00000000_00000000, 43);
360 try test__clzti2(0x00000000_00200000_00000000_00000000, 42);
361 try test__clzti2(0x00000000_00400000_00000000_00000000, 41);
362 try test__clzti2(0x00000000_00800000_00000000_00000000, 40);
363 try test__clzti2(0x00000000_01000000_00000000_00000000, 39);
364 try test__clzti2(0x00000000_02000000_00000000_00000000, 38);
365 try test__clzti2(0x00000000_04000000_00000000_00000000, 37);
366 try test__clzti2(0x00000000_08000000_00000000_00000000, 36);
367 try test__clzti2(0x00000000_10000000_00000000_00000000, 35);
368 try test__clzti2(0x00000000_20000000_00000000_00000000, 34);
369 try test__clzti2(0x00000000_40000000_00000000_00000000, 33);
370 try test__clzti2(0x00000000_80000000_00000000_00000000, 32);
371
372 try test__clzti2(0x00000001_00000000_00000000_00000000, 31);
373 try test__clzti2(0x00000002_00000000_00000000_00000000, 30);
374 try test__clzti2(0x00000004_00000000_00000000_00000000, 29);
375 try test__clzti2(0x00000008_00000000_00000000_00000000, 28);
376 try test__clzti2(0x00000010_00000000_00000000_00000000, 27);
377 try test__clzti2(0x00000020_00000000_00000000_00000000, 26);
378 try test__clzti2(0x00000040_00000000_00000000_00000000, 25);
379 try test__clzti2(0x00000080_00000000_00000000_00000000, 24);
380 try test__clzti2(0x00000100_00000000_00000000_00000000, 23);
381 try test__clzti2(0x00000200_00000000_00000000_00000000, 22);
382 try test__clzti2(0x00000400_00000000_00000000_00000000, 21);
383 try test__clzti2(0x00000800_00000000_00000000_00000000, 20);
384 try test__clzti2(0x00001000_00000000_00000000_00000000, 19);
385 try test__clzti2(0x00002000_00000000_00000000_00000000, 18);
386 try test__clzti2(0x00004000_00000000_00000000_00000000, 17);
387 try test__clzti2(0x00008000_00000000_00000000_00000000, 16);
388 try test__clzti2(0x00010000_00000000_00000000_00000000, 15);
389 try test__clzti2(0x00020000_00000000_00000000_00000000, 14);
390 try test__clzti2(0x00040000_00000000_00000000_00000000, 13);
391 try test__clzti2(0x00080000_00000000_00000000_00000000, 12);
392 try test__clzti2(0x00100000_00000000_00000000_00000000, 11);
393 try test__clzti2(0x00200000_00000000_00000000_00000000, 10);
394 try test__clzti2(0x00400000_00000000_00000000_00000000, 9);
395}
lib/std/special/compiler_rt/count0bits.zig created+160
......@@ -0,0 +1,160 @@
1const std = @import("std");
2const builtin = std.builtin;
3
4// clz - count leading zeroes
5// - clzXi2_generic for little endian
6// - __clzsi2_thumb1: assume a != 0
7// - __clzsi2_arm32: assume a != 0
8
9fn clzXi2_generic(comptime T: type) fn (a: T) callconv(.C) i32 {
10 return struct {
11 fn f(a: T) callconv(.C) i32 {
12 @setRuntimeSafety(builtin.is_test);
13
14 var x = switch (@bitSizeOf(T)) {
15 32 => @bitCast(u32, a),
16 64 => @bitCast(u64, a),
17 128 => @bitCast(u128, a),
18 else => unreachable,
19 };
20 var n: T = @bitSizeOf(T);
21 // Count first bit set using binary search, from Hacker's Delight
22 var y: @TypeOf(x) = 0;
23 comptime var shift: u8 = @bitSizeOf(T);
24 inline while (shift > 0) {
25 shift = shift >> 1;
26 y = x >> shift;
27 if (y != 0) {
28 n = n - shift;
29 x = y;
30 }
31 }
32 return @intCast(i32, n - @bitCast(T, x));
33 }
34 }.f;
35}
36
37fn __clzsi2_thumb1() callconv(.Naked) void {
38 @setRuntimeSafety(false);
39
40 // Similar to the generic version with the last two rounds replaced by a LUT
41 asm volatile (
42 \\ movs r1, #32
43 \\ lsrs r2, r0, #16
44 \\ beq 1f
45 \\ subs r1, #16
46 \\ movs r0, r2
47 \\ 1:
48 \\ lsrs r2, r0, #8
49 \\ beq 1f
50 \\ subs r1, #8
51 \\ movs r0, r2
52 \\ 1:
53 \\ lsrs r2, r0, #4
54 \\ beq 1f
55 \\ subs r1, #4
56 \\ movs r0, r2
57 \\ 1:
58 \\ ldr r3, =LUT
59 \\ ldrb r0, [r3, r0]
60 \\ subs r0, r1, r0
61 \\ bx lr
62 \\ .p2align 2
63 \\ // Number of bits set in the 0-15 range
64 \\ LUT:
65 \\ .byte 0,1,2,2,3,3,3,3,4,4,4,4,4,4,4,4
66 );
67
68 unreachable;
69}
70
71fn __clzsi2_arm32() callconv(.Naked) void {
72 @setRuntimeSafety(false);
73
74 asm volatile (
75 \\ // Assumption: n != 0
76 \\ // r0: n
77 \\ // r1: count of leading zeros in n + 1
78 \\ // r2: scratch register for shifted r0
79 \\ mov r1, #1
80 \\
81 \\ // Basic block:
82 \\ // if ((r0 >> SHIFT) == 0)
83 \\ // r1 += SHIFT;
84 \\ // else
85 \\ // r0 >>= SHIFT;
86 \\ // for descending powers of two as SHIFT.
87 \\ lsrs r2, r0, #16
88 \\ movne r0, r2
89 \\ addeq r1, #16
90 \\
91 \\ lsrs r2, r0, #8
92 \\ movne r0, r2
93 \\ addeq r1, #8
94 \\
95 \\ lsrs r2, r0, #4
96 \\ movne r0, r2
97 \\ addeq r1, #4
98 \\
99 \\ lsrs r2, r0, #2
100 \\ movne r0, r2
101 \\ addeq r1, #2
102 \\
103 \\ // The basic block invariants at this point are (r0 >> 2) == 0 and
104 \\ // r0 != 0. This means 1 <= r0 <= 3 and 0 <= (r0 >> 1) <= 1.
105 \\ //
106 \\ // r0 | (r0 >> 1) == 0 | (r0 >> 1) == 1 | -(r0 >> 1) | 1 - (r0 >> 1)f
107 \\ // ---+----------------+----------------+------------+--------------
108 \\ // 1 | 1 | 0 | 0 | 1
109 \\ // 2 | 0 | 1 | -1 | 0
110 \\ // 3 | 0 | 1 | -1 | 0
111 \\ //
112 \\ // The r1's initial value of 1 compensates for the 1 here.
113 \\ sub r0, r1, r0, lsr #1
114 \\ bx lr
115 );
116
117 unreachable;
118}
119
120pub const __clzsi2 = impl: {
121 switch (std.Target.current.cpu.arch) {
122 .arm, .armeb, .thumb, .thumbeb => {
123 const use_thumb1 =
124 (std.Target.current.cpu.arch.isThumb() or
125 std.Target.arm.featureSetHas(std.Target.current.cpu.features, .noarm)) and
126 !std.Target.arm.featureSetHas(std.Target.current.cpu.features, .thumb2);
127
128 if (use_thumb1) {
129 break :impl __clzsi2_thumb1;
130 }
131 // From here on we're either targeting Thumb2 or ARM.
132 else if (!std.Target.current.cpu.arch.isThumb()) {
133 break :impl __clzsi2_arm32;
134 }
135 // Use the generic implementation otherwise.
136 else break :impl clzXi2_generic(i32);
137 },
138 else => break :impl clzXi2_generic(i32),
139 }
140};
141
142pub const __clzdi2 = impl: {
143 switch (std.Target.current.cpu.arch) {
144 // TODO architecture optimised versions
145 else => break :impl clzXi2_generic(i64),
146 }
147};
148
149pub const __clzti2 = impl: {
150 switch (std.Target.current.cpu.arch) {
151 // TODO architecture optimised versions
152 else => break :impl clzXi2_generic(i128),
153 }
154};
155
156test {
157 _ = @import("clzsi2_test.zig");
158 _ = @import("clzdi2_test.zig");
159 _ = @import("clzti2_test.zig");
160}