|
| 1 | +// Ported from: |
| 2 | +// |
| 3 | +// https://github.com/llvm-mirror/compiler-rt/blob/f0745e8476f069296a7c71accedd061dce4cdf79/lib/builtins/clzsi2.c |
| 4 | +// https://github.com/llvm-mirror/compiler-rt/blob/f0745e8476f069296a7c71accedd061dce4cdf79/lib/builtins/arm/clzsi2.S |
| 5 | +const builtin = @import("builtin"); |
| 6 | + |
| 7 | +// Precondition: a != 0 |
| 8 | +fn __clzsi2_generic(a: i32) callconv(.C) i32 { |
| 9 | + @setRuntimeSafety(builtin.is_test); |
| 10 | + |
| 11 | + var x = @bitCast(u32, a); |
| 12 | + var n: i32 = 32; |
| 13 | + |
| 14 | + // Count first bit set using binary search, from Hacker's Delight |
| 15 | + var y: u32 = 0; |
| 16 | + inline for ([_]i32{ 16, 8, 4, 2, 1 }) |shift| { |
| 17 | + y = x >> shift; |
| 18 | + if (y != 0) { |
| 19 | + n = n - shift; |
| 20 | + x = y; |
| 21 | + } |
| 22 | + } |
| 23 | + |
| 24 | + return n - @bitCast(i32, x); |
| 25 | +} |
| 26 | + |
| 27 | +fn __clzsi2_arm_clz(a: i32) callconv(.Naked) noreturn { |
| 28 | + asm volatile ( |
| 29 | + \\ clz r0,r0 |
| 30 | + \\ bx lr |
| 31 | + ); |
| 32 | + unreachable; |
| 33 | +} |
| 34 | + |
| 35 | +fn __clzsi2_arm32(a: i32) callconv(.Naked) noreturn { |
| 36 | + asm volatile ( |
| 37 | + \\ // Assumption: n != 0 |
| 38 | + \\ // r0: n |
| 39 | + \\ // r1: count of leading zeros in n + 1 |
| 40 | + \\ // r2: scratch register for shifted r0 |
| 41 | + \\ mov r1, #1 |
| 42 | + \\ |
| 43 | + \\ // Basic block: |
| 44 | + \\ // if ((r0 >> SHIFT) == 0) |
| 45 | + \\ // r1 += SHIFT; |
| 46 | + \\ // else |
| 47 | + \\ // r0 >>= SHIFT; |
| 48 | + \\ // for descending powers of two as SHIFT. |
| 49 | + \\ lsrs r2, r0, #16 |
| 50 | + \\ movne r0, r2 |
| 51 | + \\ addeq r1, #16 |
| 52 | + \\ |
| 53 | + \\ lsrs r2, r0, #8 |
| 54 | + \\ movne r0, r2 |
| 55 | + \\ addeq r1, #8 |
| 56 | + \\ |
| 57 | + \\ lsrs r2, r0, #4 |
| 58 | + \\ movne r0, r2 |
| 59 | + \\ addeq r1, #4 |
| 60 | + \\ |
| 61 | + \\ lsrs r2, r0, #2 |
| 62 | + \\ movne r0, r2 |
| 63 | + \\ addeq r1, #2 |
| 64 | + \\ |
| 65 | + \\ // The basic block invariants at this point are (r0 >> 2) == 0 and |
| 66 | + \\ // r0 != 0. This means 1 <= r0 <= 3 and 0 <= (r0 >> 1) <= 1. |
| 67 | + \\ // |
| 68 | + \\ // r0 | (r0 >> 1) == 0 | (r0 >> 1) == 1 | -(r0 >> 1) | 1 - (r0 >> 1)f |
| 69 | + \\ // ---+----------------+----------------+------------+-------------- |
| 70 | + \\ // 1 | 1 | 0 | 0 | 1 |
| 71 | + \\ // 2 | 0 | 1 | -1 | 0 |
| 72 | + \\ // 3 | 0 | 1 | -1 | 0 |
| 73 | + \\ // |
| 74 | + \\ // The r1's initial value of 1 compensates for the 1 here. |
| 75 | + \\ sub r0, r1, r0, lsr #1 |
| 76 | + \\ bx lr |
| 77 | + ); |
| 78 | + unreachable; |
| 79 | +} |
| 80 | + |
| 81 | +const can_use_arm_clz = switch (builtin.arch) { |
| 82 | + .arm, .armeb => |sub_arch| switch (sub_arch) { |
| 83 | + .v4t => false, |
| 84 | + .v6m => false, |
| 85 | + else => true, |
| 86 | + }, |
| 87 | + .thumb, .thumbeb => |sub_arch| switch (sub_arch) { |
| 88 | + .v6, |
| 89 | + .v6k, |
| 90 | + .v5, |
| 91 | + .v5te, |
| 92 | + .v4t, |
| 93 | + => false, |
| 94 | + else => true, |
| 95 | + }, |
| 96 | + else => false, |
| 97 | +}; |
| 98 | + |
| 99 | +const is_arm32_no_thumb = switch (builtin.arch) { |
| 100 | + builtin.Arch.arm, |
| 101 | + builtin.Arch.armeb, |
| 102 | + => true, |
| 103 | + else => false, |
| 104 | +}; |
| 105 | + |
| 106 | +pub const __clzsi2 = blk: { |
| 107 | + if (comptime can_use_arm_clz) { |
| 108 | + break :blk __clzsi2_arm_clz; |
| 109 | + } else if (comptime is_arm32_no_thumb) { |
| 110 | + break :blk __clzsi2_arm32; |
| 111 | + } else { |
| 112 | + break :blk __clzsi2_generic; |
| 113 | + } |
| 114 | +}; |
| 115 | + |
| 116 | +test "test clzsi2" { |
| 117 | + _ = @import("clzsi2_test.zig"); |
| 118 | +} |
0 commit comments