| 1 | module optimize |
| 2 | |
| 3 | import v3.ssa |
| 4 | |
| 5 | // constant_fold supports constant fold handling for optimize. |
| 6 | fn constant_fold(mut m ssa.Module) bool { |
| 7 | mut changed := false |
| 8 | for fi in 0 .. m.funcs.len { |
| 9 | for blk_id in m.funcs[fi].blocks { |
| 10 | for val_id in m.blocks[blk_id].instrs { |
| 11 | if m.values[val_id].kind != .instruction { |
| 12 | continue |
| 13 | } |
| 14 | instr := m.instrs[m.values[val_id].index] |
| 15 | if instr.operands.len != 2 { |
| 16 | continue |
| 17 | } |
| 18 | lhs := m.values[instr.operands[0]] |
| 19 | rhs := m.values[instr.operands[1]] |
| 20 | |
| 21 | // Skip undef values - can't fold/simplify with undefined. |
| 22 | if lhs.kind == .constant && lhs.name == 'undef' { |
| 23 | continue |
| 24 | } |
| 25 | if rhs.kind == .constant && rhs.name == 'undef' { |
| 26 | continue |
| 27 | } |
| 28 | |
| 29 | // Algebraic simplifications first (work even with non-constant operands): |
| 30 | // x+0, x*1, x*0, x-x, x^x, x&x, x|x, x<<0, x*2 -> x<<1, etc. |
| 31 | repl, needs_zero := try_algebraic_simplify(m, val_id, instr, lhs, rhs) |
| 32 | if repl >= 0 || repl == -2 { |
| 33 | if repl == -2 { |
| 34 | // x * 2 -> x << 1 |
| 35 | other_id := if lhs.kind == .constant { |
| 36 | instr.operands[1] |
| 37 | } else { |
| 38 | instr.operands[0] |
| 39 | } |
| 40 | typ := m.values[val_id].typ |
| 41 | one_val := m.get_or_add_const(typ, '1') |
| 42 | mut shl_instr := m.instrs[m.values[val_id].index] |
| 43 | shl_instr.op = .shl |
| 44 | shl_instr.operands = [other_id, one_val] |
| 45 | m.instrs[m.values[val_id].index] = shl_instr |
| 46 | changed = true |
| 47 | continue |
| 48 | } else if needs_zero { |
| 49 | // x * 0, x & 0, x - x, x ^ x -> 0 |
| 50 | typ := m.values[val_id].typ |
| 51 | zero_val := m.get_or_add_const(typ, '0') |
| 52 | m.replace_uses(val_id, zero_val) |
| 53 | } else { |
| 54 | m.replace_uses(val_id, repl) |
| 55 | } |
| 56 | changed = true |
| 57 | continue |
| 58 | } |
| 59 | |
| 60 | if lhs.kind != .constant || rhs.kind != .constant { |
| 61 | continue |
| 62 | } |
| 63 | l_int := lhs.name.i64() |
| 64 | r_int := rhs.name.i64() |
| 65 | |
| 66 | mut result := i64(0) |
| 67 | mut folded := false |
| 68 | |
| 69 | match instr.op { |
| 70 | .add { |
| 71 | result = l_int + r_int |
| 72 | folded = true |
| 73 | } |
| 74 | .sub { |
| 75 | result = l_int - r_int |
| 76 | folded = true |
| 77 | } |
| 78 | .mul { |
| 79 | result = l_int * r_int |
| 80 | folded = true |
| 81 | } |
| 82 | .sdiv { |
| 83 | if r_int != 0 { |
| 84 | result = l_int / r_int |
| 85 | folded = true |
| 86 | } |
| 87 | } |
| 88 | .srem { |
| 89 | if r_int != 0 { |
| 90 | result = l_int % r_int |
| 91 | folded = true |
| 92 | } |
| 93 | } |
| 94 | .udiv { |
| 95 | if r_int != 0 { |
| 96 | result = i64(u64(l_int) / u64(r_int)) |
| 97 | folded = true |
| 98 | } |
| 99 | } |
| 100 | .urem { |
| 101 | if r_int != 0 { |
| 102 | result = i64(u64(l_int) % u64(r_int)) |
| 103 | folded = true |
| 104 | } |
| 105 | } |
| 106 | .and_ { |
| 107 | result = l_int & r_int |
| 108 | folded = true |
| 109 | } |
| 110 | .or_ { |
| 111 | result = l_int | r_int |
| 112 | folded = true |
| 113 | } |
| 114 | .xor { |
| 115 | result = l_int ^ r_int |
| 116 | folded = true |
| 117 | } |
| 118 | .shl { |
| 119 | if r_int >= 0 && r_int < 64 { |
| 120 | result = i64(u64(l_int) << u64(r_int)) |
| 121 | folded = true |
| 122 | } |
| 123 | } |
| 124 | .ashr { |
| 125 | if r_int >= 0 && r_int < 64 { |
| 126 | result = l_int >> u64(r_int) |
| 127 | folded = true |
| 128 | } |
| 129 | } |
| 130 | .lshr { |
| 131 | if r_int >= 0 && r_int < 64 { |
| 132 | result = i64(u64(l_int) >> u64(r_int)) |
| 133 | folded = true |
| 134 | } |
| 135 | } |
| 136 | .eq { |
| 137 | result = if l_int == r_int { 1 } else { 0 } |
| 138 | folded = true |
| 139 | } |
| 140 | .ne { |
| 141 | result = if l_int != r_int { 1 } else { 0 } |
| 142 | folded = true |
| 143 | } |
| 144 | .lt { |
| 145 | result = if l_int < r_int { 1 } else { 0 } |
| 146 | folded = true |
| 147 | } |
| 148 | .gt { |
| 149 | result = if l_int > r_int { 1 } else { 0 } |
| 150 | folded = true |
| 151 | } |
| 152 | .le { |
| 153 | result = if l_int <= r_int { 1 } else { 0 } |
| 154 | folded = true |
| 155 | } |
| 156 | .ge { |
| 157 | result = if l_int >= r_int { 1 } else { 0 } |
| 158 | folded = true |
| 159 | } |
| 160 | .ult { |
| 161 | result = if u64(l_int) < u64(r_int) { 1 } else { 0 } |
| 162 | folded = true |
| 163 | } |
| 164 | .ugt { |
| 165 | result = if u64(l_int) > u64(r_int) { 1 } else { 0 } |
| 166 | folded = true |
| 167 | } |
| 168 | .ule { |
| 169 | result = if u64(l_int) <= u64(r_int) { 1 } else { 0 } |
| 170 | folded = true |
| 171 | } |
| 172 | .uge { |
| 173 | result = if u64(l_int) >= u64(r_int) { 1 } else { 0 } |
| 174 | folded = true |
| 175 | } |
| 176 | else {} |
| 177 | } |
| 178 | |
| 179 | if folded { |
| 180 | typ := m.values[val_id].typ |
| 181 | const_val := m.get_or_add_const(typ, '${result}') |
| 182 | m.replace_uses(val_id, const_val) |
| 183 | changed = true |
| 184 | } |
| 185 | } |
| 186 | } |
| 187 | } |
| 188 | return changed |
| 189 | } |
| 190 | |
| 191 | // try_algebraic_simplify implements identity/strength-reduction simplifications: |
| 192 | // x+0=x, x*1=x, x*0=0, x-x=0, x^x=0, x&x=x, x|x=x, x<<0=x, x*2=x<<1, x/1=x. |
| 193 | // Returns (replacement_id, needs_zero): |
| 194 | // repl >= 0 -> replace all uses of val_id with `repl` |
| 195 | // repl == -2 -> rewrite as `x << 1` (caller builds the shift) |
| 196 | // needs_zero == true -> replace all uses with a fresh zero constant |
| 197 | // repl == -1 -> no simplification |
| 198 | fn try_algebraic_simplify(_m &ssa.Module, val_id int, instr ssa.Instruction, lhs ssa.Value, rhs ssa.Value) (int, bool) { |
| 199 | lhs_id := instr.operands[0] |
| 200 | rhs_id := instr.operands[1] |
| 201 | |
| 202 | // Same-operand simplifications (x op x). Integer ops only. |
| 203 | if lhs_id == rhs_id { |
| 204 | match instr.op { |
| 205 | .sub, .xor { |
| 206 | return val_id, true // x - x = 0, x ^ x = 0 |
| 207 | } |
| 208 | .and_, .or_ { |
| 209 | return lhs_id, false // x & x = x, x | x = x |
| 210 | } |
| 211 | else {} |
| 212 | } |
| 213 | } |
| 214 | |
| 215 | mut const_val := i64(0) |
| 216 | mut const_is_rhs := false |
| 217 | mut other_id := 0 |
| 218 | if lhs.kind == .constant && lhs.name != 'undef' { |
| 219 | const_val = lhs.name.i64() |
| 220 | const_is_rhs = false |
| 221 | other_id = instr.operands[1] |
| 222 | } else if rhs.kind == .constant && rhs.name != 'undef' { |
| 223 | const_val = rhs.name.i64() |
| 224 | const_is_rhs = true |
| 225 | other_id = instr.operands[0] |
| 226 | } else { |
| 227 | return -1, false |
| 228 | } |
| 229 | |
| 230 | match instr.op { |
| 231 | .add { |
| 232 | if const_val == 0 { |
| 233 | return other_id, false // x + 0 = x |
| 234 | } |
| 235 | } |
| 236 | .sub { |
| 237 | if const_is_rhs && const_val == 0 { |
| 238 | return other_id, false // x - 0 = x |
| 239 | } |
| 240 | } |
| 241 | .mul { |
| 242 | if const_val == 0 { |
| 243 | return val_id, true // x * 0 = 0 |
| 244 | } |
| 245 | if const_val == 1 { |
| 246 | return other_id, false // x * 1 = x |
| 247 | } |
| 248 | if const_val == 2 { |
| 249 | return -2, false // x * 2 = x << 1 |
| 250 | } |
| 251 | } |
| 252 | .sdiv, .udiv { |
| 253 | if const_is_rhs && const_val == 1 { |
| 254 | return other_id, false // x / 1 = x |
| 255 | } |
| 256 | } |
| 257 | .and_ { |
| 258 | if const_val == 0 { |
| 259 | return val_id, true // x & 0 = 0 |
| 260 | } |
| 261 | } |
| 262 | .or_ { |
| 263 | if const_val == 0 { |
| 264 | return other_id, false // x | 0 = x |
| 265 | } |
| 266 | } |
| 267 | .xor { |
| 268 | if const_val == 0 { |
| 269 | return other_id, false // x ^ 0 = x |
| 270 | } |
| 271 | } |
| 272 | .shl, .ashr, .lshr { |
| 273 | if const_is_rhs && const_val == 0 { |
| 274 | return other_id, false // x << 0 = x, x >> 0 = x |
| 275 | } |
| 276 | } |
| 277 | else {} |
| 278 | } |
| 279 | |
| 280 | return -1, false |
| 281 | } |
| 282 | |
| 283 | // branch_fold supports branch fold handling for optimize. |
| 284 | fn branch_fold(mut m ssa.Module) bool { |
| 285 | mut changed := false |
| 286 | for fi in 0 .. m.funcs.len { |
| 287 | for blk_id in m.funcs[fi].blocks { |
| 288 | if m.blocks[blk_id].instrs.len == 0 { |
| 289 | continue |
| 290 | } |
| 291 | term_val_id := m.blocks[blk_id].instrs.last() |
| 292 | term := m.instrs[m.values[term_val_id].index] |
| 293 | if term.op == .br && term.operands.len >= 3 { |
| 294 | cond_val := m.values[term.operands[0]] |
| 295 | if cond_val.kind == .constant { |
| 296 | cond_int := cond_val.name.i64() |
| 297 | target := if cond_int != 0 { term.operands[1] } else { term.operands[2] } |
| 298 | mut jmp_instr := m.instrs[m.values[term_val_id].index] |
| 299 | jmp_instr.op = .jmp |
| 300 | mut new_ops := []ssa.ValueID{} |
| 301 | new_ops << target |
| 302 | jmp_instr.operands = new_ops |
| 303 | m.instrs[m.values[term_val_id].index] = jmp_instr |
| 304 | changed = true |
| 305 | } |
| 306 | } |
| 307 | } |
| 308 | } |
| 309 | return changed |
| 310 | } |
| 311 | |