1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt < %s -instcombine -S | FileCheck %s 3 4; Given pattern: 5; icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0 6; we should move shifts to the same hand of 'and', i.e. e.g. rewrite as 7; icmp eq/ne (and (((x shift Q) shift K), y)), 0 8; We are only interested in opposite logical shifts here. 9; We still can handle the case where there is a truncation between a shift 10; and an 'and', thought the legality check isn't obvious. 11 12;------------------------------------------------------------------------------- 13; Basic scalar tests 14;------------------------------------------------------------------------------- 15 16; This fold can't be performed for fully variable %x and %y 17define i1 @n0(i32 %x, i64 %y, i32 %len) { 18; CHECK-LABEL: @n0( 19; CHECK-NEXT: [[T0:%.*]] = sub i32 32, [[LEN:%.*]] 20; CHECK-NEXT: [[T1:%.*]] = shl i32 [[X:%.*]], [[T0]] 21; CHECK-NEXT: [[T2:%.*]] = add i32 [[LEN]], -16 22; CHECK-NEXT: [[T2_WIDE:%.*]] = zext i32 [[T2]] to i64 23; CHECK-NEXT: [[T3:%.*]] = lshr i64 [[Y:%.*]], [[T2_WIDE]] 24; CHECK-NEXT: [[T3_TRUNC:%.*]] = trunc i64 [[T3]] to i32 25; CHECK-NEXT: [[T4:%.*]] = and i32 [[T1]], [[T3_TRUNC]] 26; CHECK-NEXT: [[T5:%.*]] = icmp ne i32 [[T4]], 0 27; CHECK-NEXT: ret i1 [[T5]] 28; 29 %t0 = sub i32 32, %len 30 %t1 = shl i32 %x, %t0 31 %t2 = add i32 %len, -16 32 %t2_wide = zext i32 %t2 to i64 33 %t3 = lshr i64 %y, %t2_wide 34 %t3_trunc = trunc i64 %t3 to i32 35 %t4 = and i32 %t1, %t3_trunc 36 %t5 = icmp ne i32 %t4, 0 37 ret i1 %t5 38} 39 40; However we can fold if %x/%y are constants that pass extra legality check. 41 42; New shift amount would be 16, %x has 16 leading zeros - can fold. 43define i1 @t1(i64 %y, i32 %len) { 44; CHECK-LABEL: @t1( 45; CHECK-NEXT: [[TMP1:%.*]] = and i64 [[Y:%.*]], 4294901760 46; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i64 [[TMP1]], 0 47; CHECK-NEXT: ret i1 [[TMP2]] 48; 49 %t0 = sub i32 32, %len 50 %t1 = shl i32 65535, %t0 51 %t2 = add i32 %len, -16 52 %t2_wide = zext i32 %t2 to i64 53 %t3 = lshr i64 %y, %t2_wide 54 %t3_trunc = trunc i64 %t3 to i32 55 %t4 = and i32 %t1, %t3_trunc 56 %t5 = icmp ne i32 %t4, 0 57 ret i1 %t5 58} 59; Note that we indeed look at leading zeros! 60define i1 @t1_single_bit(i64 %y, i32 %len) { 61; CHECK-LABEL: @t1_single_bit( 62; CHECK-NEXT: [[TMP1:%.*]] = and i64 [[Y:%.*]], 2147483648 63; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i64 [[TMP1]], 0 64; CHECK-NEXT: ret i1 [[TMP2]] 65; 66 %t0 = sub i32 32, %len 67 %t1 = shl i32 32768, %t0 68 %t2 = add i32 %len, -16 69 %t2_wide = zext i32 %t2 to i64 70 %t3 = lshr i64 %y, %t2_wide 71 %t3_trunc = trunc i64 %t3 to i32 72 %t4 = and i32 %t1, %t3_trunc 73 %t5 = icmp ne i32 %t4, 0 74 ret i1 %t5 75} 76; New shift amount would be 16, %x has 15 leading zeros - can not fold. 77define i1 @n2(i64 %y, i32 %len) { 78; CHECK-LABEL: @n2( 79; CHECK-NEXT: [[T0:%.*]] = sub i32 32, [[LEN:%.*]] 80; CHECK-NEXT: [[T1:%.*]] = shl i32 131071, [[T0]] 81; CHECK-NEXT: [[T2:%.*]] = add i32 [[LEN]], -16 82; CHECK-NEXT: [[T2_WIDE:%.*]] = zext i32 [[T2]] to i64 83; CHECK-NEXT: [[T3:%.*]] = lshr i64 [[Y:%.*]], [[T2_WIDE]] 84; CHECK-NEXT: [[T3_TRUNC:%.*]] = trunc i64 [[T3]] to i32 85; CHECK-NEXT: [[T4:%.*]] = and i32 [[T1]], [[T3_TRUNC]] 86; CHECK-NEXT: [[T5:%.*]] = icmp ne i32 [[T4]], 0 87; CHECK-NEXT: ret i1 [[T5]] 88; 89 %t0 = sub i32 32, %len 90 %t1 = shl i32 131071, %t0 91 %t2 = add i32 %len, -16 92 %t2_wide = zext i32 %t2 to i64 93 %t3 = lshr i64 %y, %t2_wide 94 %t3_trunc = trunc i64 %t3 to i32 95 %t4 = and i32 %t1, %t3_trunc 96 %t5 = icmp ne i32 %t4, 0 97 ret i1 %t5 98} 99 100; New shift amount would be 16, %y has 47 leading zeros - can fold. 101define i1 @t3(i32 %x, i32 %len) { 102; CHECK-LABEL: @t3( 103; CHECK-NEXT: [[TMP1:%.*]] = and i32 [[X:%.*]], 1 104; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0 105; CHECK-NEXT: ret i1 [[TMP2]] 106; 107 %t0 = sub i32 32, %len 108 %t1 = shl i32 %x, %t0 109 %t2 = add i32 %len, -16 110 %t2_wide = zext i32 %t2 to i64 111 %t3 = lshr i64 131071, %t2_wide 112 %t3_trunc = trunc i64 %t3 to i32 113 %t4 = and i32 %t1, %t3_trunc 114 %t5 = icmp ne i32 %t4, 0 115 ret i1 %t5 116} 117; Note that we indeed look at leading zeros! 118define i1 @t3_singlebit(i32 %x, i32 %len) { 119; CHECK-LABEL: @t3_singlebit( 120; CHECK-NEXT: [[TMP1:%.*]] = and i32 [[X:%.*]], 1 121; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i32 [[TMP1]], 0 122; CHECK-NEXT: ret i1 [[TMP2]] 123; 124 %t0 = sub i32 32, %len 125 %t1 = shl i32 %x, %t0 126 %t2 = add i32 %len, -16 127 %t2_wide = zext i32 %t2 to i64 128 %t3 = lshr i64 65536, %t2_wide 129 %t3_trunc = trunc i64 %t3 to i32 130 %t4 = and i32 %t1, %t3_trunc 131 %t5 = icmp ne i32 %t4, 0 132 ret i1 %t5 133} 134; New shift amount would be 16, %y has 48 leading zeros - can not fold. 135define i1 @n4(i32 %x, i32 %len) { 136; CHECK-LABEL: @n4( 137; CHECK-NEXT: [[T0:%.*]] = sub i32 32, [[LEN:%.*]] 138; CHECK-NEXT: [[T1:%.*]] = shl i32 [[X:%.*]], [[T0]] 139; CHECK-NEXT: [[T2:%.*]] = add i32 [[LEN]], -16 140; CHECK-NEXT: [[T2_WIDE:%.*]] = zext i32 [[T2]] to i64 141; CHECK-NEXT: [[T3:%.*]] = lshr i64 262143, [[T2_WIDE]] 142; CHECK-NEXT: [[T3_TRUNC:%.*]] = trunc i64 [[T3]] to i32 143; CHECK-NEXT: [[T4:%.*]] = and i32 [[T1]], [[T3_TRUNC]] 144; CHECK-NEXT: [[T5:%.*]] = icmp ne i32 [[T4]], 0 145; CHECK-NEXT: ret i1 [[T5]] 146; 147 %t0 = sub i32 32, %len 148 %t1 = shl i32 %x, %t0 149 %t2 = add i32 %len, -16 150 %t2_wide = zext i32 %t2 to i64 151 %t3 = lshr i64 262143, %t2_wide 152 %t3_trunc = trunc i64 %t3 to i32 153 %t4 = and i32 %t1, %t3_trunc 154 %t5 = icmp ne i32 %t4, 0 155 ret i1 %t5 156} 157 158; While we could still deal with arbitrary values if KnownBits can answer 159; the question, it isn't obvious it's worth it, so let's not for now. 160 161;------------------------------------------------------------------------------- 162; Vector tests 163;------------------------------------------------------------------------------- 164 165; New shift amount would be 16, minimal count of leading zeros in %x is 16. Ok. 166define <2 x i1> @t5_vec(<2 x i64> %y, <2 x i32> %len) { 167; CHECK-LABEL: @t5_vec( 168; CHECK-NEXT: [[TMP1:%.*]] = lshr <2 x i64> [[Y:%.*]], <i64 16, i64 16> 169; CHECK-NEXT: [[TMP2:%.*]] = and <2 x i64> [[TMP1]], <i64 65535, i64 32767> 170; CHECK-NEXT: [[TMP3:%.*]] = icmp ne <2 x i64> [[TMP2]], zeroinitializer 171; CHECK-NEXT: ret <2 x i1> [[TMP3]] 172; 173 %t0 = sub <2 x i32> <i32 32, i32 32>, %len 174 %t1 = shl <2 x i32> <i32 65535, i32 32767>, %t0 175 %t2 = add <2 x i32> %len, <i32 -16, i32 -16> 176 %t2_wide = zext <2 x i32> %t2 to <2 x i64> 177 %t3 = lshr <2 x i64> %y, %t2_wide 178 %t3_trunc = trunc <2 x i64> %t3 to <2 x i32> 179 %t4 = and <2 x i32> %t1, %t3_trunc 180 %t5 = icmp ne <2 x i32> %t4, <i32 0, i32 0> 181 ret <2 x i1> %t5 182} 183; New shift amount would be 16, minimal count of leading zeros in %x is 15, not ok to fold. 184define <2 x i1> @n6_vec(<2 x i64> %y, <2 x i32> %len) { 185; CHECK-LABEL: @n6_vec( 186; CHECK-NEXT: [[T0:%.*]] = sub <2 x i32> <i32 32, i32 32>, [[LEN:%.*]] 187; CHECK-NEXT: [[T1:%.*]] = shl <2 x i32> <i32 65535, i32 131071>, [[T0]] 188; CHECK-NEXT: [[T2:%.*]] = add <2 x i32> [[LEN]], <i32 -16, i32 -16> 189; CHECK-NEXT: [[T2_WIDE:%.*]] = zext <2 x i32> [[T2]] to <2 x i64> 190; CHECK-NEXT: [[T3:%.*]] = lshr <2 x i64> [[Y:%.*]], [[T2_WIDE]] 191; CHECK-NEXT: [[T3_TRUNC:%.*]] = trunc <2 x i64> [[T3]] to <2 x i32> 192; CHECK-NEXT: [[T4:%.*]] = and <2 x i32> [[T1]], [[T3_TRUNC]] 193; CHECK-NEXT: [[T5:%.*]] = icmp ne <2 x i32> [[T4]], zeroinitializer 194; CHECK-NEXT: ret <2 x i1> [[T5]] 195; 196 %t0 = sub <2 x i32> <i32 32, i32 32>, %len 197 %t1 = shl <2 x i32> <i32 65535, i32 131071>, %t0 198 %t2 = add <2 x i32> %len, <i32 -16, i32 -16> 199 %t2_wide = zext <2 x i32> %t2 to <2 x i64> 200 %t3 = lshr <2 x i64> %y, %t2_wide 201 %t3_trunc = trunc <2 x i64> %t3 to <2 x i32> 202 %t4 = and <2 x i32> %t1, %t3_trunc 203 %t5 = icmp ne <2 x i32> %t4, <i32 0, i32 0> 204 ret <2 x i1> %t5 205} 206 207; New shift amount would be 16, minimal count of leading zeros in %x is 47. Ok. 208define <2 x i1> @t7_vec(<2 x i32> %x, <2 x i32> %len) { 209; CHECK-LABEL: @t7_vec( 210; CHECK-NEXT: [[TMP1:%.*]] = and <2 x i32> [[X:%.*]], <i32 1, i32 0> 211; CHECK-NEXT: [[TMP2:%.*]] = icmp ne <2 x i32> [[TMP1]], zeroinitializer 212; CHECK-NEXT: ret <2 x i1> [[TMP2]] 213; 214 %t0 = sub <2 x i32> <i32 32, i32 32>, %len 215 %t1 = shl <2 x i32> %x, %t0 216 %t2 = add <2 x i32> %len, <i32 -16, i32 -16> 217 %t2_wide = zext <2 x i32> %t2 to <2 x i64> 218 %t3 = lshr <2 x i64> <i64 131071, i64 65535>, %t2_wide 219 %t3_trunc = trunc <2 x i64> %t3 to <2 x i32> 220 %t4 = and <2 x i32> %t1, %t3_trunc 221 %t5 = icmp ne <2 x i32> %t4, <i32 0, i32 0> 222 ret <2 x i1> %t5 223} 224; New shift amount would be 16, minimal count of leading zeros in %x is 48, not ok to fold. 225define <2 x i1> @n8_vec(<2 x i32> %x, <2 x i32> %len) { 226; CHECK-LABEL: @n8_vec( 227; CHECK-NEXT: [[T0:%.*]] = sub <2 x i32> <i32 32, i32 32>, [[LEN:%.*]] 228; CHECK-NEXT: [[T1:%.*]] = shl <2 x i32> [[X:%.*]], [[T0]] 229; CHECK-NEXT: [[T2:%.*]] = add <2 x i32> [[LEN]], <i32 -16, i32 -16> 230; CHECK-NEXT: [[T2_WIDE:%.*]] = zext <2 x i32> [[T2]] to <2 x i64> 231; CHECK-NEXT: [[T3:%.*]] = lshr <2 x i64> <i64 131071, i64 262143>, [[T2_WIDE]] 232; CHECK-NEXT: [[T3_TRUNC:%.*]] = trunc <2 x i64> [[T3]] to <2 x i32> 233; CHECK-NEXT: [[T4:%.*]] = and <2 x i32> [[T1]], [[T3_TRUNC]] 234; CHECK-NEXT: [[T5:%.*]] = icmp ne <2 x i32> [[T4]], zeroinitializer 235; CHECK-NEXT: ret <2 x i1> [[T5]] 236; 237 %t0 = sub <2 x i32> <i32 32, i32 32>, %len 238 %t1 = shl <2 x i32> %x, %t0 239 %t2 = add <2 x i32> %len, <i32 -16, i32 -16> 240 %t2_wide = zext <2 x i32> %t2 to <2 x i64> 241 %t3 = lshr <2 x i64> <i64 131071, i64 262143>, %t2_wide 242 %t3_trunc = trunc <2 x i64> %t3 to <2 x i32> 243 %t4 = and <2 x i32> %t1, %t3_trunc 244 %t5 = icmp ne <2 x i32> %t4, <i32 0, i32 0> 245 ret <2 x i1> %t5 246} 247 248;------------------------------------------------------------------------------- 249 250; Ok if the final shift amount is exactly one less than widest bit width. 251define i1 @t9_highest_bit(i32 %x, i64 %y, i32 %len) { 252; CHECK-LABEL: @t9_highest_bit( 253; CHECK-NEXT: [[TMP1:%.*]] = zext i32 [[X:%.*]] to i64 254; CHECK-NEXT: [[TMP2:%.*]] = lshr i64 [[Y:%.*]], 63 255; CHECK-NEXT: [[TMP3:%.*]] = and i64 [[TMP2]], [[TMP1]] 256; CHECK-NEXT: [[TMP4:%.*]] = icmp ne i64 [[TMP3]], 0 257; CHECK-NEXT: ret i1 [[TMP4]] 258; 259 %t0 = sub i32 64, %len 260 %t1 = shl i32 %x, %t0 261 %t2 = add i32 %len, -1 262 %t2_wide = zext i32 %t2 to i64 263 %t3 = lshr i64 %y, %t2_wide 264 %t3_trunc = trunc i64 %t3 to i32 265 %t4 = and i32 %t1, %t3_trunc 266 %t5 = icmp ne i32 %t4, 0 267 ret i1 %t5 268} 269; Not highest bit. 270define i1 @t10_almost_highest_bit(i32 %x, i64 %y, i32 %len) { 271; CHECK-LABEL: @t10_almost_highest_bit( 272; CHECK-NEXT: [[T0:%.*]] = sub i32 64, [[LEN:%.*]] 273; CHECK-NEXT: [[T1:%.*]] = shl i32 [[X:%.*]], [[T0]] 274; CHECK-NEXT: [[T2:%.*]] = add i32 [[LEN]], -2 275; CHECK-NEXT: [[T2_WIDE:%.*]] = zext i32 [[T2]] to i64 276; CHECK-NEXT: [[T3:%.*]] = lshr i64 [[Y:%.*]], [[T2_WIDE]] 277; CHECK-NEXT: [[T3_TRUNC:%.*]] = trunc i64 [[T3]] to i32 278; CHECK-NEXT: [[T4:%.*]] = and i32 [[T1]], [[T3_TRUNC]] 279; CHECK-NEXT: [[T5:%.*]] = icmp ne i32 [[T4]], 0 280; CHECK-NEXT: ret i1 [[T5]] 281; 282 %t0 = sub i32 64, %len 283 %t1 = shl i32 %x, %t0 284 %t2 = add i32 %len, -2 285 %t2_wide = zext i32 %t2 to i64 286 %t3 = lshr i64 %y, %t2_wide 287 %t3_trunc = trunc i64 %t3 to i32 288 %t4 = and i32 %t1, %t3_trunc 289 %t5 = icmp ne i32 %t4, 0 290 ret i1 %t5 291} 292 293; Ok if the final shift amount is zero. 294define i1 @t11_no_shift(i32 %x, i64 %y, i32 %len) { 295; CHECK-LABEL: @t11_no_shift( 296; CHECK-NEXT: [[TMP1:%.*]] = zext i32 [[X:%.*]] to i64 297; CHECK-NEXT: [[TMP2:%.*]] = and i64 [[TMP1]], [[Y:%.*]] 298; CHECK-NEXT: [[TMP3:%.*]] = icmp ne i64 [[TMP2]], 0 299; CHECK-NEXT: ret i1 [[TMP3]] 300; 301 %t0 = sub i32 64, %len 302 %t1 = shl i32 %x, %t0 303 %t2 = add i32 %len, -64 304 %t2_wide = zext i32 %t2 to i64 305 %t3 = lshr i64 %y, %t2_wide 306 %t3_trunc = trunc i64 %t3 to i32 307 %t4 = and i32 %t1, %t3_trunc 308 %t5 = icmp ne i32 %t4, 0 309 ret i1 %t5 310} 311; Not zero-shift. 312define i1 @t10_shift_by_one(i32 %x, i64 %y, i32 %len) { 313; CHECK-LABEL: @t10_shift_by_one( 314; CHECK-NEXT: [[T0:%.*]] = sub i32 64, [[LEN:%.*]] 315; CHECK-NEXT: [[T1:%.*]] = shl i32 [[X:%.*]], [[T0]] 316; CHECK-NEXT: [[T2:%.*]] = add i32 [[LEN]], -63 317; CHECK-NEXT: [[T2_WIDE:%.*]] = zext i32 [[T2]] to i64 318; CHECK-NEXT: [[T3:%.*]] = lshr i64 [[Y:%.*]], [[T2_WIDE]] 319; CHECK-NEXT: [[T3_TRUNC:%.*]] = trunc i64 [[T3]] to i32 320; CHECK-NEXT: [[T4:%.*]] = and i32 [[T1]], [[T3_TRUNC]] 321; CHECK-NEXT: [[T5:%.*]] = icmp ne i32 [[T4]], 0 322; CHECK-NEXT: ret i1 [[T5]] 323; 324 %t0 = sub i32 64, %len 325 %t1 = shl i32 %x, %t0 326 %t2 = add i32 %len, -63 327 %t2_wide = zext i32 %t2 to i64 328 %t3 = lshr i64 %y, %t2_wide 329 %t3_trunc = trunc i64 %t3 to i32 330 %t4 = and i32 %t1, %t3_trunc 331 %t5 = icmp ne i32 %t4, 0 332 ret i1 %t5 333} 334 335; A mix of those conditions is ok. 336define <2 x i1> @t11_zero_and_almost_bitwidth(<2 x i32> %x, <2 x i64> %y, <2 x i32> %len) { 337; CHECK-LABEL: @t11_zero_and_almost_bitwidth( 338; CHECK-NEXT: [[T0:%.*]] = sub <2 x i32> <i32 64, i32 64>, [[LEN:%.*]] 339; CHECK-NEXT: [[T1:%.*]] = shl <2 x i32> [[X:%.*]], [[T0]] 340; CHECK-NEXT: [[T2:%.*]] = add <2 x i32> [[LEN]], <i32 -1, i32 -64> 341; CHECK-NEXT: [[T2_WIDE:%.*]] = zext <2 x i32> [[T2]] to <2 x i64> 342; CHECK-NEXT: [[T3:%.*]] = lshr <2 x i64> [[Y:%.*]], [[T2_WIDE]] 343; CHECK-NEXT: [[T3_TRUNC:%.*]] = trunc <2 x i64> [[T3]] to <2 x i32> 344; CHECK-NEXT: [[T4:%.*]] = and <2 x i32> [[T1]], [[T3_TRUNC]] 345; CHECK-NEXT: [[T5:%.*]] = icmp ne <2 x i32> [[T4]], zeroinitializer 346; CHECK-NEXT: ret <2 x i1> [[T5]] 347; 348 %t0 = sub <2 x i32> <i32 64, i32 64>, %len 349 %t1 = shl <2 x i32> %x, %t0 350 %t2 = add <2 x i32> %len, <i32 -1, i32 -64> 351 %t2_wide = zext <2 x i32> %t2 to <2 x i64> 352 %t3 = lshr <2 x i64> %y, %t2_wide 353 %t3_trunc = trunc <2 x i64> %t3 to <2 x i32> 354 %t4 = and <2 x i32> %t1, %t3_trunc 355 %t5 = icmp ne <2 x i32> %t4, <i32 0, i32 0> 356 ret <2 x i1> %t5 357} 358define <2 x i1> @n12_bad(<2 x i32> %x, <2 x i64> %y, <2 x i32> %len) { 359; CHECK-LABEL: @n12_bad( 360; CHECK-NEXT: [[T0:%.*]] = sub <2 x i32> <i32 64, i32 64>, [[LEN:%.*]] 361; CHECK-NEXT: [[T1:%.*]] = shl <2 x i32> [[X:%.*]], [[T0]] 362; CHECK-NEXT: [[T2:%.*]] = add <2 x i32> [[LEN]], <i32 -2, i32 -64> 363; CHECK-NEXT: [[T2_WIDE:%.*]] = zext <2 x i32> [[T2]] to <2 x i64> 364; CHECK-NEXT: [[T3:%.*]] = lshr <2 x i64> [[Y:%.*]], [[T2_WIDE]] 365; CHECK-NEXT: [[T3_TRUNC:%.*]] = trunc <2 x i64> [[T3]] to <2 x i32> 366; CHECK-NEXT: [[T4:%.*]] = and <2 x i32> [[T1]], [[T3_TRUNC]] 367; CHECK-NEXT: [[T5:%.*]] = icmp ne <2 x i32> [[T4]], zeroinitializer 368; CHECK-NEXT: ret <2 x i1> [[T5]] 369; 370 %t0 = sub <2 x i32> <i32 64, i32 64>, %len 371 %t1 = shl <2 x i32> %x, %t0 372 %t2 = add <2 x i32> %len, <i32 -2, i32 -64> 373 %t2_wide = zext <2 x i32> %t2 to <2 x i64> 374 %t3 = lshr <2 x i64> %y, %t2_wide 375 %t3_trunc = trunc <2 x i64> %t3 to <2 x i32> 376 %t4 = and <2 x i32> %t1, %t3_trunc 377 %t5 = icmp ne <2 x i32> %t4, <i32 0, i32 0> 378 ret <2 x i1> %t5 379} 380 381;------------------------------------------------------------------------------; 382 383; Ok if one of the values being shifted is 1 384define i1 @t13_x_is_one(i64 %y, i32 %len) { 385; CHECK-LABEL: @t13_x_is_one( 386; CHECK-NEXT: [[TMP1:%.*]] = and i64 [[Y:%.*]], 65536 387; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i64 [[TMP1]], 0 388; CHECK-NEXT: ret i1 [[TMP2]] 389; 390 %t0 = sub i32 32, %len 391 %t1 = shl i32 1, %t0 392 %t2 = add i32 %len, -16 393 %t2_wide = zext i32 %t2 to i64 394 %t3 = lshr i64 %y, %t2_wide 395 %t3_trunc = trunc i64 %t3 to i32 396 %t4 = and i32 %t1, %t3_trunc 397 %t5 = icmp ne i32 %t4, 0 398 ret i1 %t5 399} 400define i1 @t14_x_is_one(i32 %x, i32 %len) { 401; CHECK-LABEL: @t14_x_is_one( 402; CHECK-NEXT: ret i1 false 403; 404 %t0 = sub i32 32, %len 405 %t1 = shl i32 %x, %t0 406 %t2 = add i32 %len, -16 407 %t2_wide = zext i32 %t2 to i64 408 %t3 = lshr i64 1, %t2_wide 409 %t3_trunc = trunc i64 %t3 to i32 410 %t4 = and i32 %t1, %t3_trunc 411 %t5 = icmp ne i32 %t4, 0 412 ret i1 %t5 413} 414 415define <2 x i1> @t15_vec_x_is_one_or_zero(<2 x i64> %y, <2 x i32> %len) { 416; CHECK-LABEL: @t15_vec_x_is_one_or_zero( 417; CHECK-NEXT: [[TMP1:%.*]] = lshr <2 x i64> [[Y:%.*]], <i64 48, i64 48> 418; CHECK-NEXT: [[TMP2:%.*]] = and <2 x i64> [[TMP1]], <i64 1, i64 0> 419; CHECK-NEXT: [[TMP3:%.*]] = icmp ne <2 x i64> [[TMP2]], zeroinitializer 420; CHECK-NEXT: ret <2 x i1> [[TMP3]] 421; 422 %t0 = sub <2 x i32> <i32 64, i32 64>, %len 423 %t1 = shl <2 x i32> <i32 1, i32 0>, %t0 424 %t2 = add <2 x i32> %len, <i32 -16, i32 -16> 425 %t2_wide = zext <2 x i32> %t2 to <2 x i64> 426 %t3 = lshr <2 x i64> %y, %t2_wide 427 %t3_trunc = trunc <2 x i64> %t3 to <2 x i32> 428 %t4 = and <2 x i32> %t1, %t3_trunc 429 %t5 = icmp ne <2 x i32> %t4, <i32 0, i32 0> 430 ret <2 x i1> %t5 431} 432define <2 x i1> @t16_vec_y_is_one_or_zero(<2 x i32> %x, <2 x i32> %len) { 433; CHECK-LABEL: @t16_vec_y_is_one_or_zero( 434; CHECK-NEXT: ret <2 x i1> zeroinitializer 435; 436 %t0 = sub <2 x i32> <i32 64, i32 64>, %len 437 %t1 = shl <2 x i32> %x, %t0 438 %t2 = add <2 x i32> %len, <i32 -16, i32 -16> 439 %t2_wide = zext <2 x i32> %t2 to <2 x i64> 440 %t3 = lshr <2 x i64> <i64 1, i64 0>, %t2_wide 441 %t3_trunc = trunc <2 x i64> %t3 to <2 x i32> 442 %t4 = and <2 x i32> %t1, %t3_trunc 443 %t5 = icmp ne <2 x i32> %t4, <i32 0, i32 0> 444 ret <2 x i1> %t5 445} 446 447;------------------------------------------------------------------------------; 448 449; All other tests - extra uses, etc are already covered in 450; shift-amount-reassociation-in-bittest-with-truncation-shl.ll and 451; shift-amount-reassociation-in-bittest.ll 452 453; And that's the main motivational pattern: 454define i1 @rawspeed_signbit(i64 %storage, i32 %nbits) { 455; CHECK-LABEL: @rawspeed_signbit( 456; CHECK-NEXT: [[TMP1:%.*]] = icmp sgt i64 [[STORAGE:%.*]], -1 457; CHECK-NEXT: ret i1 [[TMP1]] 458; 459 %skipnbits = sub nsw i32 64, %nbits 460 %skipnbitswide = zext i32 %skipnbits to i64 461 %datawide = lshr i64 %storage, %skipnbitswide 462 %data = trunc i64 %datawide to i32 463 %nbitsminusone = add nsw i32 %nbits, -1 464 %bitmask = shl i32 1, %nbitsminusone 465 %bitmasked = and i32 %bitmask, %data 466 %isbitunset = icmp eq i32 %bitmasked, 0 467 ret i1 %isbitunset 468} 469