1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=-bmi,-tbm,-bmi2 < %s | FileCheck %s --check-prefixes=CHECK,X86,NOBMI2,X86-NOBMI2,FALLBACK0,X86-FALLBACK0 3; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,-tbm,-bmi2 < %s | FileCheck %s --check-prefixes=CHECK,X86,NOBMI2,X86-NOBMI2,FALLBACK1,X86-FALLBACK1 4; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,+tbm,-bmi2 < %s | FileCheck %s --check-prefixes=CHECK,X86,NOBMI2,X86-NOBMI2,FALLBACK2,X86-FALLBACK2 5; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,+tbm,+bmi2 < %s | FileCheck %s --check-prefixes=CHECK,X86,BMI2,X86-BMI2,FALLBACK3,X86-FALLBACK3 6; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,-tbm,+bmi2 < %s | FileCheck %s --check-prefixes=CHECK,X86,BMI2,X86-BMI2,FALLBACK4,X86-FALLBACK4 7; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=-bmi,-tbm,-bmi2 < %s | FileCheck %s --check-prefixes=CHECK,X64,NOBMI2,X64-NOBMI2,FALLBACK0,X64-FALLBACK0 8; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,-tbm,-bmi2 < %s | FileCheck %s --check-prefixes=CHECK,X64,NOBMI2,X64-NOBMI2,FALLBACK1,X64-FALLBACK1 9; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,+tbm,-bmi2 < %s | FileCheck %s --check-prefixes=CHECK,X64,NOBMI2,X64-NOBMI2,FALLBACK2,X64-FALLBACK2 10; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,+tbm,+bmi2 < %s | FileCheck %s --check-prefixes=CHECK,X64,BMI2,X64-BMI2,FALLBACK3,X64-FALLBACK3 11; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,-tbm,+bmi2 < %s | FileCheck %s --check-prefixes=CHECK,X64,BMI2,X64-BMI2,FALLBACK4,X64-FALLBACK4 12 13; Patterns: 14; c) x & (-1 >> y) 15; d) x << y >> y 16; are equivalent, but we prefer the second variant if we have BMI2. 17 18; We do not test the variant where y = (32 - z), because that is BMI2's BZHI. 19 20; ---------------------------------------------------------------------------- ; 21; 8-bit 22; ---------------------------------------------------------------------------- ; 23 24define i8 @clear_highbits8_c0(i8 %val, i8 %numhighbits) nounwind { 25; X86-LABEL: clear_highbits8_c0: 26; X86: # %bb.0: 27; X86-NEXT: movb {{[0-9]+}}(%esp), %cl 28; X86-NEXT: movb {{[0-9]+}}(%esp), %al 29; X86-NEXT: shlb %cl, %al 30; X86-NEXT: shrb %cl, %al 31; X86-NEXT: retl 32; 33; X64-LABEL: clear_highbits8_c0: 34; X64: # %bb.0: 35; X64-NEXT: movl %esi, %ecx 36; X64-NEXT: shlb %cl, %dil 37; X64-NEXT: # kill: def $cl killed $cl killed $ecx 38; X64-NEXT: shrb %cl, %dil 39; X64-NEXT: movl %edi, %eax 40; X64-NEXT: retq 41 %mask = lshr i8 -1, %numhighbits 42 %masked = and i8 %mask, %val 43 ret i8 %masked 44} 45 46define i8 @clear_highbits8_c2_load(i8* %w, i8 %numhighbits) nounwind { 47; X86-LABEL: clear_highbits8_c2_load: 48; X86: # %bb.0: 49; X86-NEXT: movb {{[0-9]+}}(%esp), %cl 50; X86-NEXT: movl {{[0-9]+}}(%esp), %eax 51; X86-NEXT: movb (%eax), %al 52; X86-NEXT: shlb %cl, %al 53; X86-NEXT: shrb %cl, %al 54; X86-NEXT: retl 55; 56; X64-LABEL: clear_highbits8_c2_load: 57; X64: # %bb.0: 58; X64-NEXT: movl %esi, %ecx 59; X64-NEXT: movb (%rdi), %al 60; X64-NEXT: shlb %cl, %al 61; X64-NEXT: # kill: def $cl killed $cl killed $ecx 62; X64-NEXT: shrb %cl, %al 63; X64-NEXT: retq 64 %val = load i8, i8* %w 65 %mask = lshr i8 -1, %numhighbits 66 %masked = and i8 %mask, %val 67 ret i8 %masked 68} 69 70define i8 @clear_highbits8_c4_commutative(i8 %val, i8 %numhighbits) nounwind { 71; X86-LABEL: clear_highbits8_c4_commutative: 72; X86: # %bb.0: 73; X86-NEXT: movb {{[0-9]+}}(%esp), %cl 74; X86-NEXT: movb {{[0-9]+}}(%esp), %al 75; X86-NEXT: shlb %cl, %al 76; X86-NEXT: shrb %cl, %al 77; X86-NEXT: retl 78; 79; X64-LABEL: clear_highbits8_c4_commutative: 80; X64: # %bb.0: 81; X64-NEXT: movl %esi, %ecx 82; X64-NEXT: shlb %cl, %dil 83; X64-NEXT: # kill: def $cl killed $cl killed $ecx 84; X64-NEXT: shrb %cl, %dil 85; X64-NEXT: movl %edi, %eax 86; X64-NEXT: retq 87 %mask = lshr i8 -1, %numhighbits 88 %masked = and i8 %val, %mask ; swapped order 89 ret i8 %masked 90} 91 92; ---------------------------------------------------------------------------- ; 93; 16-bit 94; ---------------------------------------------------------------------------- ; 95 96define i16 @clear_highbits16_c0(i16 %val, i16 %numhighbits) nounwind { 97; X86-NOBMI2-LABEL: clear_highbits16_c0: 98; X86-NOBMI2: # %bb.0: 99; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 100; X86-NOBMI2-NEXT: movl {{[0-9]+}}(%esp), %eax 101; X86-NOBMI2-NEXT: shll %cl, %eax 102; X86-NOBMI2-NEXT: movzwl %ax, %eax 103; X86-NOBMI2-NEXT: shrl %cl, %eax 104; X86-NOBMI2-NEXT: # kill: def $ax killed $ax killed $eax 105; X86-NOBMI2-NEXT: retl 106; 107; X86-BMI2-LABEL: clear_highbits16_c0: 108; X86-BMI2: # %bb.0: 109; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %al 110; X86-BMI2-NEXT: shlxl %eax, {{[0-9]+}}(%esp), %ecx 111; X86-BMI2-NEXT: movzwl %cx, %ecx 112; X86-BMI2-NEXT: shrxl %eax, %ecx, %eax 113; X86-BMI2-NEXT: # kill: def $ax killed $ax killed $eax 114; X86-BMI2-NEXT: retl 115; 116; X64-NOBMI2-LABEL: clear_highbits16_c0: 117; X64-NOBMI2: # %bb.0: 118; X64-NOBMI2-NEXT: movl %esi, %ecx 119; X64-NOBMI2-NEXT: shll %cl, %edi 120; X64-NOBMI2-NEXT: movzwl %di, %eax 121; X64-NOBMI2-NEXT: # kill: def $cl killed $cl killed $ecx 122; X64-NOBMI2-NEXT: shrl %cl, %eax 123; X64-NOBMI2-NEXT: # kill: def $ax killed $ax killed $eax 124; X64-NOBMI2-NEXT: retq 125; 126; X64-BMI2-LABEL: clear_highbits16_c0: 127; X64-BMI2: # %bb.0: 128; X64-BMI2-NEXT: shlxl %esi, %edi, %eax 129; X64-BMI2-NEXT: movzwl %ax, %eax 130; X64-BMI2-NEXT: shrxl %esi, %eax, %eax 131; X64-BMI2-NEXT: # kill: def $ax killed $ax killed $eax 132; X64-BMI2-NEXT: retq 133 %mask = lshr i16 -1, %numhighbits 134 %masked = and i16 %mask, %val 135 ret i16 %masked 136} 137 138define i16 @clear_highbits16_c1_indexzext(i16 %val, i8 %numhighbits) nounwind { 139; X86-NOBMI2-LABEL: clear_highbits16_c1_indexzext: 140; X86-NOBMI2: # %bb.0: 141; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 142; X86-NOBMI2-NEXT: movl {{[0-9]+}}(%esp), %eax 143; X86-NOBMI2-NEXT: shll %cl, %eax 144; X86-NOBMI2-NEXT: movzwl %ax, %eax 145; X86-NOBMI2-NEXT: shrl %cl, %eax 146; X86-NOBMI2-NEXT: # kill: def $ax killed $ax killed $eax 147; X86-NOBMI2-NEXT: retl 148; 149; X86-BMI2-LABEL: clear_highbits16_c1_indexzext: 150; X86-BMI2: # %bb.0: 151; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %al 152; X86-BMI2-NEXT: shlxl %eax, {{[0-9]+}}(%esp), %ecx 153; X86-BMI2-NEXT: movzwl %cx, %ecx 154; X86-BMI2-NEXT: shrxl %eax, %ecx, %eax 155; X86-BMI2-NEXT: # kill: def $ax killed $ax killed $eax 156; X86-BMI2-NEXT: retl 157; 158; X64-NOBMI2-LABEL: clear_highbits16_c1_indexzext: 159; X64-NOBMI2: # %bb.0: 160; X64-NOBMI2-NEXT: movl %esi, %ecx 161; X64-NOBMI2-NEXT: shll %cl, %edi 162; X64-NOBMI2-NEXT: movzwl %di, %eax 163; X64-NOBMI2-NEXT: # kill: def $cl killed $cl killed $ecx 164; X64-NOBMI2-NEXT: shrl %cl, %eax 165; X64-NOBMI2-NEXT: # kill: def $ax killed $ax killed $eax 166; X64-NOBMI2-NEXT: retq 167; 168; X64-BMI2-LABEL: clear_highbits16_c1_indexzext: 169; X64-BMI2: # %bb.0: 170; X64-BMI2-NEXT: shlxl %esi, %edi, %eax 171; X64-BMI2-NEXT: movzwl %ax, %eax 172; X64-BMI2-NEXT: shrxl %esi, %eax, %eax 173; X64-BMI2-NEXT: # kill: def $ax killed $ax killed $eax 174; X64-BMI2-NEXT: retq 175 %sh_prom = zext i8 %numhighbits to i16 176 %mask = lshr i16 -1, %sh_prom 177 %masked = and i16 %mask, %val 178 ret i16 %masked 179} 180 181define i16 @clear_highbits16_c2_load(i16* %w, i16 %numhighbits) nounwind { 182; X86-NOBMI2-LABEL: clear_highbits16_c2_load: 183; X86-NOBMI2: # %bb.0: 184; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 185; X86-NOBMI2-NEXT: movl {{[0-9]+}}(%esp), %eax 186; X86-NOBMI2-NEXT: movzwl (%eax), %eax 187; X86-NOBMI2-NEXT: shll %cl, %eax 188; X86-NOBMI2-NEXT: movzwl %ax, %eax 189; X86-NOBMI2-NEXT: shrl %cl, %eax 190; X86-NOBMI2-NEXT: # kill: def $ax killed $ax killed $eax 191; X86-NOBMI2-NEXT: retl 192; 193; X86-BMI2-LABEL: clear_highbits16_c2_load: 194; X86-BMI2: # %bb.0: 195; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %al 196; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %ecx 197; X86-BMI2-NEXT: movzwl (%ecx), %ecx 198; X86-BMI2-NEXT: shlxl %eax, %ecx, %ecx 199; X86-BMI2-NEXT: movzwl %cx, %ecx 200; X86-BMI2-NEXT: shrxl %eax, %ecx, %eax 201; X86-BMI2-NEXT: # kill: def $ax killed $ax killed $eax 202; X86-BMI2-NEXT: retl 203; 204; X64-NOBMI2-LABEL: clear_highbits16_c2_load: 205; X64-NOBMI2: # %bb.0: 206; X64-NOBMI2-NEXT: movl %esi, %ecx 207; X64-NOBMI2-NEXT: movzwl (%rdi), %eax 208; X64-NOBMI2-NEXT: shll %cl, %eax 209; X64-NOBMI2-NEXT: movzwl %ax, %eax 210; X64-NOBMI2-NEXT: # kill: def $cl killed $cl killed $ecx 211; X64-NOBMI2-NEXT: shrl %cl, %eax 212; X64-NOBMI2-NEXT: # kill: def $ax killed $ax killed $eax 213; X64-NOBMI2-NEXT: retq 214; 215; X64-BMI2-LABEL: clear_highbits16_c2_load: 216; X64-BMI2: # %bb.0: 217; X64-BMI2-NEXT: movzwl (%rdi), %eax 218; X64-BMI2-NEXT: shlxl %esi, %eax, %eax 219; X64-BMI2-NEXT: movzwl %ax, %eax 220; X64-BMI2-NEXT: shrxl %esi, %eax, %eax 221; X64-BMI2-NEXT: # kill: def $ax killed $ax killed $eax 222; X64-BMI2-NEXT: retq 223 %val = load i16, i16* %w 224 %mask = lshr i16 -1, %numhighbits 225 %masked = and i16 %mask, %val 226 ret i16 %masked 227} 228 229define i16 @clear_highbits16_c3_load_indexzext(i16* %w, i8 %numhighbits) nounwind { 230; X86-NOBMI2-LABEL: clear_highbits16_c3_load_indexzext: 231; X86-NOBMI2: # %bb.0: 232; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 233; X86-NOBMI2-NEXT: movl {{[0-9]+}}(%esp), %eax 234; X86-NOBMI2-NEXT: movzwl (%eax), %eax 235; X86-NOBMI2-NEXT: shll %cl, %eax 236; X86-NOBMI2-NEXT: movzwl %ax, %eax 237; X86-NOBMI2-NEXT: shrl %cl, %eax 238; X86-NOBMI2-NEXT: # kill: def $ax killed $ax killed $eax 239; X86-NOBMI2-NEXT: retl 240; 241; X86-BMI2-LABEL: clear_highbits16_c3_load_indexzext: 242; X86-BMI2: # %bb.0: 243; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %al 244; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %ecx 245; X86-BMI2-NEXT: movzwl (%ecx), %ecx 246; X86-BMI2-NEXT: shlxl %eax, %ecx, %ecx 247; X86-BMI2-NEXT: movzwl %cx, %ecx 248; X86-BMI2-NEXT: shrxl %eax, %ecx, %eax 249; X86-BMI2-NEXT: # kill: def $ax killed $ax killed $eax 250; X86-BMI2-NEXT: retl 251; 252; X64-NOBMI2-LABEL: clear_highbits16_c3_load_indexzext: 253; X64-NOBMI2: # %bb.0: 254; X64-NOBMI2-NEXT: movl %esi, %ecx 255; X64-NOBMI2-NEXT: movzwl (%rdi), %eax 256; X64-NOBMI2-NEXT: shll %cl, %eax 257; X64-NOBMI2-NEXT: movzwl %ax, %eax 258; X64-NOBMI2-NEXT: # kill: def $cl killed $cl killed $ecx 259; X64-NOBMI2-NEXT: shrl %cl, %eax 260; X64-NOBMI2-NEXT: # kill: def $ax killed $ax killed $eax 261; X64-NOBMI2-NEXT: retq 262; 263; X64-BMI2-LABEL: clear_highbits16_c3_load_indexzext: 264; X64-BMI2: # %bb.0: 265; X64-BMI2-NEXT: movzwl (%rdi), %eax 266; X64-BMI2-NEXT: shlxl %esi, %eax, %eax 267; X64-BMI2-NEXT: movzwl %ax, %eax 268; X64-BMI2-NEXT: shrxl %esi, %eax, %eax 269; X64-BMI2-NEXT: # kill: def $ax killed $ax killed $eax 270; X64-BMI2-NEXT: retq 271 %val = load i16, i16* %w 272 %sh_prom = zext i8 %numhighbits to i16 273 %mask = lshr i16 -1, %sh_prom 274 %masked = and i16 %mask, %val 275 ret i16 %masked 276} 277 278define i16 @clear_highbits16_c4_commutative(i16 %val, i16 %numhighbits) nounwind { 279; X86-NOBMI2-LABEL: clear_highbits16_c4_commutative: 280; X86-NOBMI2: # %bb.0: 281; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 282; X86-NOBMI2-NEXT: movl {{[0-9]+}}(%esp), %eax 283; X86-NOBMI2-NEXT: shll %cl, %eax 284; X86-NOBMI2-NEXT: movzwl %ax, %eax 285; X86-NOBMI2-NEXT: shrl %cl, %eax 286; X86-NOBMI2-NEXT: # kill: def $ax killed $ax killed $eax 287; X86-NOBMI2-NEXT: retl 288; 289; X86-BMI2-LABEL: clear_highbits16_c4_commutative: 290; X86-BMI2: # %bb.0: 291; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %al 292; X86-BMI2-NEXT: shlxl %eax, {{[0-9]+}}(%esp), %ecx 293; X86-BMI2-NEXT: movzwl %cx, %ecx 294; X86-BMI2-NEXT: shrxl %eax, %ecx, %eax 295; X86-BMI2-NEXT: # kill: def $ax killed $ax killed $eax 296; X86-BMI2-NEXT: retl 297; 298; X64-NOBMI2-LABEL: clear_highbits16_c4_commutative: 299; X64-NOBMI2: # %bb.0: 300; X64-NOBMI2-NEXT: movl %esi, %ecx 301; X64-NOBMI2-NEXT: shll %cl, %edi 302; X64-NOBMI2-NEXT: movzwl %di, %eax 303; X64-NOBMI2-NEXT: # kill: def $cl killed $cl killed $ecx 304; X64-NOBMI2-NEXT: shrl %cl, %eax 305; X64-NOBMI2-NEXT: # kill: def $ax killed $ax killed $eax 306; X64-NOBMI2-NEXT: retq 307; 308; X64-BMI2-LABEL: clear_highbits16_c4_commutative: 309; X64-BMI2: # %bb.0: 310; X64-BMI2-NEXT: shlxl %esi, %edi, %eax 311; X64-BMI2-NEXT: movzwl %ax, %eax 312; X64-BMI2-NEXT: shrxl %esi, %eax, %eax 313; X64-BMI2-NEXT: # kill: def $ax killed $ax killed $eax 314; X64-BMI2-NEXT: retq 315 %mask = lshr i16 -1, %numhighbits 316 %masked = and i16 %val, %mask ; swapped order 317 ret i16 %masked 318} 319 320; ---------------------------------------------------------------------------- ; 321; 32-bit 322; ---------------------------------------------------------------------------- ; 323 324define i32 @clear_highbits32_c0(i32 %val, i32 %numhighbits) nounwind { 325; X86-NOBMI2-LABEL: clear_highbits32_c0: 326; X86-NOBMI2: # %bb.0: 327; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 328; X86-NOBMI2-NEXT: movl {{[0-9]+}}(%esp), %eax 329; X86-NOBMI2-NEXT: shll %cl, %eax 330; X86-NOBMI2-NEXT: shrl %cl, %eax 331; X86-NOBMI2-NEXT: retl 332; 333; X86-BMI2-LABEL: clear_highbits32_c0: 334; X86-BMI2: # %bb.0: 335; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %al 336; X86-BMI2-NEXT: shlxl %eax, {{[0-9]+}}(%esp), %ecx 337; X86-BMI2-NEXT: shrxl %eax, %ecx, %eax 338; X86-BMI2-NEXT: retl 339; 340; X64-NOBMI2-LABEL: clear_highbits32_c0: 341; X64-NOBMI2: # %bb.0: 342; X64-NOBMI2-NEXT: movl %esi, %ecx 343; X64-NOBMI2-NEXT: shll %cl, %edi 344; X64-NOBMI2-NEXT: # kill: def $cl killed $cl killed $ecx 345; X64-NOBMI2-NEXT: shrl %cl, %edi 346; X64-NOBMI2-NEXT: movl %edi, %eax 347; X64-NOBMI2-NEXT: retq 348; 349; X64-BMI2-LABEL: clear_highbits32_c0: 350; X64-BMI2: # %bb.0: 351; X64-BMI2-NEXT: shlxl %esi, %edi, %eax 352; X64-BMI2-NEXT: shrxl %esi, %eax, %eax 353; X64-BMI2-NEXT: retq 354 %mask = lshr i32 -1, %numhighbits 355 %masked = and i32 %mask, %val 356 ret i32 %masked 357} 358 359define i32 @clear_highbits32_c1_indexzext(i32 %val, i8 %numhighbits) nounwind { 360; X86-NOBMI2-LABEL: clear_highbits32_c1_indexzext: 361; X86-NOBMI2: # %bb.0: 362; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 363; X86-NOBMI2-NEXT: movl {{[0-9]+}}(%esp), %eax 364; X86-NOBMI2-NEXT: shll %cl, %eax 365; X86-NOBMI2-NEXT: shrl %cl, %eax 366; X86-NOBMI2-NEXT: retl 367; 368; X86-BMI2-LABEL: clear_highbits32_c1_indexzext: 369; X86-BMI2: # %bb.0: 370; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %al 371; X86-BMI2-NEXT: shlxl %eax, {{[0-9]+}}(%esp), %ecx 372; X86-BMI2-NEXT: shrxl %eax, %ecx, %eax 373; X86-BMI2-NEXT: retl 374; 375; X64-NOBMI2-LABEL: clear_highbits32_c1_indexzext: 376; X64-NOBMI2: # %bb.0: 377; X64-NOBMI2-NEXT: movl %esi, %ecx 378; X64-NOBMI2-NEXT: shll %cl, %edi 379; X64-NOBMI2-NEXT: # kill: def $cl killed $cl killed $ecx 380; X64-NOBMI2-NEXT: shrl %cl, %edi 381; X64-NOBMI2-NEXT: movl %edi, %eax 382; X64-NOBMI2-NEXT: retq 383; 384; X64-BMI2-LABEL: clear_highbits32_c1_indexzext: 385; X64-BMI2: # %bb.0: 386; X64-BMI2-NEXT: shlxl %esi, %edi, %eax 387; X64-BMI2-NEXT: shrxl %esi, %eax, %eax 388; X64-BMI2-NEXT: retq 389 %sh_prom = zext i8 %numhighbits to i32 390 %mask = lshr i32 -1, %sh_prom 391 %masked = and i32 %mask, %val 392 ret i32 %masked 393} 394 395define i32 @clear_highbits32_c2_load(i32* %w, i32 %numhighbits) nounwind { 396; X86-NOBMI2-LABEL: clear_highbits32_c2_load: 397; X86-NOBMI2: # %bb.0: 398; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 399; X86-NOBMI2-NEXT: movl {{[0-9]+}}(%esp), %eax 400; X86-NOBMI2-NEXT: movl (%eax), %eax 401; X86-NOBMI2-NEXT: shll %cl, %eax 402; X86-NOBMI2-NEXT: shrl %cl, %eax 403; X86-NOBMI2-NEXT: retl 404; 405; X86-BMI2-LABEL: clear_highbits32_c2_load: 406; X86-BMI2: # %bb.0: 407; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax 408; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 409; X86-BMI2-NEXT: shlxl %ecx, (%eax), %eax 410; X86-BMI2-NEXT: shrxl %ecx, %eax, %eax 411; X86-BMI2-NEXT: retl 412; 413; X64-NOBMI2-LABEL: clear_highbits32_c2_load: 414; X64-NOBMI2: # %bb.0: 415; X64-NOBMI2-NEXT: movl %esi, %ecx 416; X64-NOBMI2-NEXT: movl (%rdi), %eax 417; X64-NOBMI2-NEXT: shll %cl, %eax 418; X64-NOBMI2-NEXT: # kill: def $cl killed $cl killed $ecx 419; X64-NOBMI2-NEXT: shrl %cl, %eax 420; X64-NOBMI2-NEXT: retq 421; 422; X64-BMI2-LABEL: clear_highbits32_c2_load: 423; X64-BMI2: # %bb.0: 424; X64-BMI2-NEXT: shlxl %esi, (%rdi), %eax 425; X64-BMI2-NEXT: shrxl %esi, %eax, %eax 426; X64-BMI2-NEXT: retq 427 %val = load i32, i32* %w 428 %mask = lshr i32 -1, %numhighbits 429 %masked = and i32 %mask, %val 430 ret i32 %masked 431} 432 433define i32 @clear_highbits32_c3_load_indexzext(i32* %w, i8 %numhighbits) nounwind { 434; X86-NOBMI2-LABEL: clear_highbits32_c3_load_indexzext: 435; X86-NOBMI2: # %bb.0: 436; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 437; X86-NOBMI2-NEXT: movl {{[0-9]+}}(%esp), %eax 438; X86-NOBMI2-NEXT: movl (%eax), %eax 439; X86-NOBMI2-NEXT: shll %cl, %eax 440; X86-NOBMI2-NEXT: shrl %cl, %eax 441; X86-NOBMI2-NEXT: retl 442; 443; X86-BMI2-LABEL: clear_highbits32_c3_load_indexzext: 444; X86-BMI2: # %bb.0: 445; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax 446; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 447; X86-BMI2-NEXT: shlxl %ecx, (%eax), %eax 448; X86-BMI2-NEXT: shrxl %ecx, %eax, %eax 449; X86-BMI2-NEXT: retl 450; 451; X64-NOBMI2-LABEL: clear_highbits32_c3_load_indexzext: 452; X64-NOBMI2: # %bb.0: 453; X64-NOBMI2-NEXT: movl %esi, %ecx 454; X64-NOBMI2-NEXT: movl (%rdi), %eax 455; X64-NOBMI2-NEXT: shll %cl, %eax 456; X64-NOBMI2-NEXT: # kill: def $cl killed $cl killed $ecx 457; X64-NOBMI2-NEXT: shrl %cl, %eax 458; X64-NOBMI2-NEXT: retq 459; 460; X64-BMI2-LABEL: clear_highbits32_c3_load_indexzext: 461; X64-BMI2: # %bb.0: 462; X64-BMI2-NEXT: shlxl %esi, (%rdi), %eax 463; X64-BMI2-NEXT: shrxl %esi, %eax, %eax 464; X64-BMI2-NEXT: retq 465 %val = load i32, i32* %w 466 %sh_prom = zext i8 %numhighbits to i32 467 %mask = lshr i32 -1, %sh_prom 468 %masked = and i32 %mask, %val 469 ret i32 %masked 470} 471 472define i32 @clear_highbits32_c4_commutative(i32 %val, i32 %numhighbits) nounwind { 473; X86-NOBMI2-LABEL: clear_highbits32_c4_commutative: 474; X86-NOBMI2: # %bb.0: 475; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 476; X86-NOBMI2-NEXT: movl {{[0-9]+}}(%esp), %eax 477; X86-NOBMI2-NEXT: shll %cl, %eax 478; X86-NOBMI2-NEXT: shrl %cl, %eax 479; X86-NOBMI2-NEXT: retl 480; 481; X86-BMI2-LABEL: clear_highbits32_c4_commutative: 482; X86-BMI2: # %bb.0: 483; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %al 484; X86-BMI2-NEXT: shlxl %eax, {{[0-9]+}}(%esp), %ecx 485; X86-BMI2-NEXT: shrxl %eax, %ecx, %eax 486; X86-BMI2-NEXT: retl 487; 488; X64-NOBMI2-LABEL: clear_highbits32_c4_commutative: 489; X64-NOBMI2: # %bb.0: 490; X64-NOBMI2-NEXT: movl %esi, %ecx 491; X64-NOBMI2-NEXT: shll %cl, %edi 492; X64-NOBMI2-NEXT: # kill: def $cl killed $cl killed $ecx 493; X64-NOBMI2-NEXT: shrl %cl, %edi 494; X64-NOBMI2-NEXT: movl %edi, %eax 495; X64-NOBMI2-NEXT: retq 496; 497; X64-BMI2-LABEL: clear_highbits32_c4_commutative: 498; X64-BMI2: # %bb.0: 499; X64-BMI2-NEXT: shlxl %esi, %edi, %eax 500; X64-BMI2-NEXT: shrxl %esi, %eax, %eax 501; X64-BMI2-NEXT: retq 502 %mask = lshr i32 -1, %numhighbits 503 %masked = and i32 %val, %mask ; swapped order 504 ret i32 %masked 505} 506 507; ---------------------------------------------------------------------------- ; 508; 64-bit 509; ---------------------------------------------------------------------------- ; 510 511define i64 @clear_highbits64_c0(i64 %val, i64 %numhighbits) nounwind { 512; X86-NOBMI2-LABEL: clear_highbits64_c0: 513; X86-NOBMI2: # %bb.0: 514; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 515; X86-NOBMI2-NEXT: movl $-1, %eax 516; X86-NOBMI2-NEXT: movl $-1, %edx 517; X86-NOBMI2-NEXT: shrl %cl, %edx 518; X86-NOBMI2-NEXT: shrdl %cl, %eax, %eax 519; X86-NOBMI2-NEXT: testb $32, %cl 520; X86-NOBMI2-NEXT: je .LBB13_2 521; X86-NOBMI2-NEXT: # %bb.1: 522; X86-NOBMI2-NEXT: movl %edx, %eax 523; X86-NOBMI2-NEXT: xorl %edx, %edx 524; X86-NOBMI2-NEXT: .LBB13_2: 525; X86-NOBMI2-NEXT: andl {{[0-9]+}}(%esp), %eax 526; X86-NOBMI2-NEXT: andl {{[0-9]+}}(%esp), %edx 527; X86-NOBMI2-NEXT: retl 528; 529; X86-BMI2-LABEL: clear_highbits64_c0: 530; X86-BMI2: # %bb.0: 531; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 532; X86-BMI2-NEXT: movl $-1, %eax 533; X86-BMI2-NEXT: shrxl %ecx, %eax, %edx 534; X86-BMI2-NEXT: shrdl %cl, %eax, %eax 535; X86-BMI2-NEXT: testb $32, %cl 536; X86-BMI2-NEXT: je .LBB13_2 537; X86-BMI2-NEXT: # %bb.1: 538; X86-BMI2-NEXT: movl %edx, %eax 539; X86-BMI2-NEXT: xorl %edx, %edx 540; X86-BMI2-NEXT: .LBB13_2: 541; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %eax 542; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %edx 543; X86-BMI2-NEXT: retl 544; 545; X64-NOBMI2-LABEL: clear_highbits64_c0: 546; X64-NOBMI2: # %bb.0: 547; X64-NOBMI2-NEXT: movq %rsi, %rcx 548; X64-NOBMI2-NEXT: shlq %cl, %rdi 549; X64-NOBMI2-NEXT: # kill: def $cl killed $cl killed $rcx 550; X64-NOBMI2-NEXT: shrq %cl, %rdi 551; X64-NOBMI2-NEXT: movq %rdi, %rax 552; X64-NOBMI2-NEXT: retq 553; 554; X64-BMI2-LABEL: clear_highbits64_c0: 555; X64-BMI2: # %bb.0: 556; X64-BMI2-NEXT: shlxq %rsi, %rdi, %rax 557; X64-BMI2-NEXT: shrxq %rsi, %rax, %rax 558; X64-BMI2-NEXT: retq 559 %mask = lshr i64 -1, %numhighbits 560 %masked = and i64 %mask, %val 561 ret i64 %masked 562} 563 564define i64 @clear_highbits64_c1_indexzext(i64 %val, i8 %numhighbits) nounwind { 565; X86-NOBMI2-LABEL: clear_highbits64_c1_indexzext: 566; X86-NOBMI2: # %bb.0: 567; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 568; X86-NOBMI2-NEXT: movl $-1, %eax 569; X86-NOBMI2-NEXT: movl $-1, %edx 570; X86-NOBMI2-NEXT: shrl %cl, %edx 571; X86-NOBMI2-NEXT: shrdl %cl, %eax, %eax 572; X86-NOBMI2-NEXT: testb $32, %cl 573; X86-NOBMI2-NEXT: je .LBB14_2 574; X86-NOBMI2-NEXT: # %bb.1: 575; X86-NOBMI2-NEXT: movl %edx, %eax 576; X86-NOBMI2-NEXT: xorl %edx, %edx 577; X86-NOBMI2-NEXT: .LBB14_2: 578; X86-NOBMI2-NEXT: andl {{[0-9]+}}(%esp), %eax 579; X86-NOBMI2-NEXT: andl {{[0-9]+}}(%esp), %edx 580; X86-NOBMI2-NEXT: retl 581; 582; X86-BMI2-LABEL: clear_highbits64_c1_indexzext: 583; X86-BMI2: # %bb.0: 584; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 585; X86-BMI2-NEXT: movl $-1, %eax 586; X86-BMI2-NEXT: shrxl %ecx, %eax, %edx 587; X86-BMI2-NEXT: shrdl %cl, %eax, %eax 588; X86-BMI2-NEXT: testb $32, %cl 589; X86-BMI2-NEXT: je .LBB14_2 590; X86-BMI2-NEXT: # %bb.1: 591; X86-BMI2-NEXT: movl %edx, %eax 592; X86-BMI2-NEXT: xorl %edx, %edx 593; X86-BMI2-NEXT: .LBB14_2: 594; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %eax 595; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %edx 596; X86-BMI2-NEXT: retl 597; 598; X64-NOBMI2-LABEL: clear_highbits64_c1_indexzext: 599; X64-NOBMI2: # %bb.0: 600; X64-NOBMI2-NEXT: movl %esi, %ecx 601; X64-NOBMI2-NEXT: shlq %cl, %rdi 602; X64-NOBMI2-NEXT: # kill: def $cl killed $cl killed $ecx 603; X64-NOBMI2-NEXT: shrq %cl, %rdi 604; X64-NOBMI2-NEXT: movq %rdi, %rax 605; X64-NOBMI2-NEXT: retq 606; 607; X64-BMI2-LABEL: clear_highbits64_c1_indexzext: 608; X64-BMI2: # %bb.0: 609; X64-BMI2-NEXT: # kill: def $esi killed $esi def $rsi 610; X64-BMI2-NEXT: shlxq %rsi, %rdi, %rax 611; X64-BMI2-NEXT: shrxq %rsi, %rax, %rax 612; X64-BMI2-NEXT: retq 613 %sh_prom = zext i8 %numhighbits to i64 614 %mask = lshr i64 -1, %sh_prom 615 %masked = and i64 %mask, %val 616 ret i64 %masked 617} 618 619define i64 @clear_highbits64_c2_load(i64* %w, i64 %numhighbits) nounwind { 620; X86-NOBMI2-LABEL: clear_highbits64_c2_load: 621; X86-NOBMI2: # %bb.0: 622; X86-NOBMI2-NEXT: pushl %esi 623; X86-NOBMI2-NEXT: movl {{[0-9]+}}(%esp), %esi 624; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 625; X86-NOBMI2-NEXT: movl $-1, %eax 626; X86-NOBMI2-NEXT: movl $-1, %edx 627; X86-NOBMI2-NEXT: shrl %cl, %edx 628; X86-NOBMI2-NEXT: shrdl %cl, %eax, %eax 629; X86-NOBMI2-NEXT: testb $32, %cl 630; X86-NOBMI2-NEXT: je .LBB15_2 631; X86-NOBMI2-NEXT: # %bb.1: 632; X86-NOBMI2-NEXT: movl %edx, %eax 633; X86-NOBMI2-NEXT: xorl %edx, %edx 634; X86-NOBMI2-NEXT: .LBB15_2: 635; X86-NOBMI2-NEXT: andl (%esi), %eax 636; X86-NOBMI2-NEXT: andl 4(%esi), %edx 637; X86-NOBMI2-NEXT: popl %esi 638; X86-NOBMI2-NEXT: retl 639; 640; X86-BMI2-LABEL: clear_highbits64_c2_load: 641; X86-BMI2: # %bb.0: 642; X86-BMI2-NEXT: pushl %esi 643; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %esi 644; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 645; X86-BMI2-NEXT: movl $-1, %eax 646; X86-BMI2-NEXT: shrxl %ecx, %eax, %edx 647; X86-BMI2-NEXT: shrdl %cl, %eax, %eax 648; X86-BMI2-NEXT: testb $32, %cl 649; X86-BMI2-NEXT: je .LBB15_2 650; X86-BMI2-NEXT: # %bb.1: 651; X86-BMI2-NEXT: movl %edx, %eax 652; X86-BMI2-NEXT: xorl %edx, %edx 653; X86-BMI2-NEXT: .LBB15_2: 654; X86-BMI2-NEXT: andl (%esi), %eax 655; X86-BMI2-NEXT: andl 4(%esi), %edx 656; X86-BMI2-NEXT: popl %esi 657; X86-BMI2-NEXT: retl 658; 659; X64-NOBMI2-LABEL: clear_highbits64_c2_load: 660; X64-NOBMI2: # %bb.0: 661; X64-NOBMI2-NEXT: movq %rsi, %rcx 662; X64-NOBMI2-NEXT: movq (%rdi), %rax 663; X64-NOBMI2-NEXT: shlq %cl, %rax 664; X64-NOBMI2-NEXT: # kill: def $cl killed $cl killed $rcx 665; X64-NOBMI2-NEXT: shrq %cl, %rax 666; X64-NOBMI2-NEXT: retq 667; 668; X64-BMI2-LABEL: clear_highbits64_c2_load: 669; X64-BMI2: # %bb.0: 670; X64-BMI2-NEXT: shlxq %rsi, (%rdi), %rax 671; X64-BMI2-NEXT: shrxq %rsi, %rax, %rax 672; X64-BMI2-NEXT: retq 673 %val = load i64, i64* %w 674 %mask = lshr i64 -1, %numhighbits 675 %masked = and i64 %mask, %val 676 ret i64 %masked 677} 678 679define i64 @clear_highbits64_c3_load_indexzext(i64* %w, i8 %numhighbits) nounwind { 680; X86-NOBMI2-LABEL: clear_highbits64_c3_load_indexzext: 681; X86-NOBMI2: # %bb.0: 682; X86-NOBMI2-NEXT: pushl %esi 683; X86-NOBMI2-NEXT: movl {{[0-9]+}}(%esp), %esi 684; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 685; X86-NOBMI2-NEXT: movl $-1, %eax 686; X86-NOBMI2-NEXT: movl $-1, %edx 687; X86-NOBMI2-NEXT: shrl %cl, %edx 688; X86-NOBMI2-NEXT: shrdl %cl, %eax, %eax 689; X86-NOBMI2-NEXT: testb $32, %cl 690; X86-NOBMI2-NEXT: je .LBB16_2 691; X86-NOBMI2-NEXT: # %bb.1: 692; X86-NOBMI2-NEXT: movl %edx, %eax 693; X86-NOBMI2-NEXT: xorl %edx, %edx 694; X86-NOBMI2-NEXT: .LBB16_2: 695; X86-NOBMI2-NEXT: andl (%esi), %eax 696; X86-NOBMI2-NEXT: andl 4(%esi), %edx 697; X86-NOBMI2-NEXT: popl %esi 698; X86-NOBMI2-NEXT: retl 699; 700; X86-BMI2-LABEL: clear_highbits64_c3_load_indexzext: 701; X86-BMI2: # %bb.0: 702; X86-BMI2-NEXT: pushl %esi 703; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %esi 704; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 705; X86-BMI2-NEXT: movl $-1, %eax 706; X86-BMI2-NEXT: shrxl %ecx, %eax, %edx 707; X86-BMI2-NEXT: shrdl %cl, %eax, %eax 708; X86-BMI2-NEXT: testb $32, %cl 709; X86-BMI2-NEXT: je .LBB16_2 710; X86-BMI2-NEXT: # %bb.1: 711; X86-BMI2-NEXT: movl %edx, %eax 712; X86-BMI2-NEXT: xorl %edx, %edx 713; X86-BMI2-NEXT: .LBB16_2: 714; X86-BMI2-NEXT: andl (%esi), %eax 715; X86-BMI2-NEXT: andl 4(%esi), %edx 716; X86-BMI2-NEXT: popl %esi 717; X86-BMI2-NEXT: retl 718; 719; X64-NOBMI2-LABEL: clear_highbits64_c3_load_indexzext: 720; X64-NOBMI2: # %bb.0: 721; X64-NOBMI2-NEXT: movl %esi, %ecx 722; X64-NOBMI2-NEXT: movq (%rdi), %rax 723; X64-NOBMI2-NEXT: shlq %cl, %rax 724; X64-NOBMI2-NEXT: # kill: def $cl killed $cl killed $ecx 725; X64-NOBMI2-NEXT: shrq %cl, %rax 726; X64-NOBMI2-NEXT: retq 727; 728; X64-BMI2-LABEL: clear_highbits64_c3_load_indexzext: 729; X64-BMI2: # %bb.0: 730; X64-BMI2-NEXT: # kill: def $esi killed $esi def $rsi 731; X64-BMI2-NEXT: shlxq %rsi, (%rdi), %rax 732; X64-BMI2-NEXT: shrxq %rsi, %rax, %rax 733; X64-BMI2-NEXT: retq 734 %val = load i64, i64* %w 735 %sh_prom = zext i8 %numhighbits to i64 736 %mask = lshr i64 -1, %sh_prom 737 %masked = and i64 %mask, %val 738 ret i64 %masked 739} 740 741define i64 @clear_highbits64_c4_commutative(i64 %val, i64 %numhighbits) nounwind { 742; X86-NOBMI2-LABEL: clear_highbits64_c4_commutative: 743; X86-NOBMI2: # %bb.0: 744; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 745; X86-NOBMI2-NEXT: movl $-1, %eax 746; X86-NOBMI2-NEXT: movl $-1, %edx 747; X86-NOBMI2-NEXT: shrl %cl, %edx 748; X86-NOBMI2-NEXT: shrdl %cl, %eax, %eax 749; X86-NOBMI2-NEXT: testb $32, %cl 750; X86-NOBMI2-NEXT: je .LBB17_2 751; X86-NOBMI2-NEXT: # %bb.1: 752; X86-NOBMI2-NEXT: movl %edx, %eax 753; X86-NOBMI2-NEXT: xorl %edx, %edx 754; X86-NOBMI2-NEXT: .LBB17_2: 755; X86-NOBMI2-NEXT: andl {{[0-9]+}}(%esp), %eax 756; X86-NOBMI2-NEXT: andl {{[0-9]+}}(%esp), %edx 757; X86-NOBMI2-NEXT: retl 758; 759; X86-BMI2-LABEL: clear_highbits64_c4_commutative: 760; X86-BMI2: # %bb.0: 761; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 762; X86-BMI2-NEXT: movl $-1, %eax 763; X86-BMI2-NEXT: shrxl %ecx, %eax, %edx 764; X86-BMI2-NEXT: shrdl %cl, %eax, %eax 765; X86-BMI2-NEXT: testb $32, %cl 766; X86-BMI2-NEXT: je .LBB17_2 767; X86-BMI2-NEXT: # %bb.1: 768; X86-BMI2-NEXT: movl %edx, %eax 769; X86-BMI2-NEXT: xorl %edx, %edx 770; X86-BMI2-NEXT: .LBB17_2: 771; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %eax 772; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %edx 773; X86-BMI2-NEXT: retl 774; 775; X64-NOBMI2-LABEL: clear_highbits64_c4_commutative: 776; X64-NOBMI2: # %bb.0: 777; X64-NOBMI2-NEXT: movq %rsi, %rcx 778; X64-NOBMI2-NEXT: shlq %cl, %rdi 779; X64-NOBMI2-NEXT: # kill: def $cl killed $cl killed $rcx 780; X64-NOBMI2-NEXT: shrq %cl, %rdi 781; X64-NOBMI2-NEXT: movq %rdi, %rax 782; X64-NOBMI2-NEXT: retq 783; 784; X64-BMI2-LABEL: clear_highbits64_c4_commutative: 785; X64-BMI2: # %bb.0: 786; X64-BMI2-NEXT: shlxq %rsi, %rdi, %rax 787; X64-BMI2-NEXT: shrxq %rsi, %rax, %rax 788; X64-BMI2-NEXT: retq 789 %mask = lshr i64 -1, %numhighbits 790 %masked = and i64 %val, %mask ; swapped order 791 ret i64 %masked 792} 793 794; ---------------------------------------------------------------------------- ; 795; Multi-use tests 796; ---------------------------------------------------------------------------- ; 797 798declare void @use32(i32) 799declare void @use64(i64) 800 801define i32 @oneuse32(i32 %val, i32 %numhighbits) nounwind { 802; X86-NOBMI2-LABEL: oneuse32: 803; X86-NOBMI2: # %bb.0: 804; X86-NOBMI2-NEXT: pushl %esi 805; X86-NOBMI2-NEXT: subl $8, %esp 806; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 807; X86-NOBMI2-NEXT: movl $-1, %esi 808; X86-NOBMI2-NEXT: shrl %cl, %esi 809; X86-NOBMI2-NEXT: movl %esi, (%esp) 810; X86-NOBMI2-NEXT: calll use32 811; X86-NOBMI2-NEXT: andl {{[0-9]+}}(%esp), %esi 812; X86-NOBMI2-NEXT: movl %esi, %eax 813; X86-NOBMI2-NEXT: addl $8, %esp 814; X86-NOBMI2-NEXT: popl %esi 815; X86-NOBMI2-NEXT: retl 816; 817; X86-BMI2-LABEL: oneuse32: 818; X86-BMI2: # %bb.0: 819; X86-BMI2-NEXT: pushl %esi 820; X86-BMI2-NEXT: subl $8, %esp 821; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %al 822; X86-BMI2-NEXT: movl $-1, %ecx 823; X86-BMI2-NEXT: shrxl %eax, %ecx, %esi 824; X86-BMI2-NEXT: movl %esi, (%esp) 825; X86-BMI2-NEXT: calll use32 826; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %esi 827; X86-BMI2-NEXT: movl %esi, %eax 828; X86-BMI2-NEXT: addl $8, %esp 829; X86-BMI2-NEXT: popl %esi 830; X86-BMI2-NEXT: retl 831; 832; X64-NOBMI2-LABEL: oneuse32: 833; X64-NOBMI2: # %bb.0: 834; X64-NOBMI2-NEXT: pushq %rbp 835; X64-NOBMI2-NEXT: pushq %rbx 836; X64-NOBMI2-NEXT: pushq %rax 837; X64-NOBMI2-NEXT: movl %edi, %ebx 838; X64-NOBMI2-NEXT: movl $-1, %ebp 839; X64-NOBMI2-NEXT: movl %esi, %ecx 840; X64-NOBMI2-NEXT: shrl %cl, %ebp 841; X64-NOBMI2-NEXT: movl %ebp, %edi 842; X64-NOBMI2-NEXT: callq use32 843; X64-NOBMI2-NEXT: andl %ebx, %ebp 844; X64-NOBMI2-NEXT: movl %ebp, %eax 845; X64-NOBMI2-NEXT: addq $8, %rsp 846; X64-NOBMI2-NEXT: popq %rbx 847; X64-NOBMI2-NEXT: popq %rbp 848; X64-NOBMI2-NEXT: retq 849; 850; X64-BMI2-LABEL: oneuse32: 851; X64-BMI2: # %bb.0: 852; X64-BMI2-NEXT: pushq %rbp 853; X64-BMI2-NEXT: pushq %rbx 854; X64-BMI2-NEXT: pushq %rax 855; X64-BMI2-NEXT: movl %edi, %ebx 856; X64-BMI2-NEXT: movl $-1, %eax 857; X64-BMI2-NEXT: shrxl %esi, %eax, %ebp 858; X64-BMI2-NEXT: movl %ebp, %edi 859; X64-BMI2-NEXT: callq use32 860; X64-BMI2-NEXT: andl %ebx, %ebp 861; X64-BMI2-NEXT: movl %ebp, %eax 862; X64-BMI2-NEXT: addq $8, %rsp 863; X64-BMI2-NEXT: popq %rbx 864; X64-BMI2-NEXT: popq %rbp 865; X64-BMI2-NEXT: retq 866 %mask = lshr i32 -1, %numhighbits 867 call void @use32(i32 %mask) 868 %masked = and i32 %mask, %val 869 ret i32 %masked 870} 871 872define i64 @oneuse64(i64 %val, i64 %numhighbits) nounwind { 873; X86-NOBMI2-LABEL: oneuse64: 874; X86-NOBMI2: # %bb.0: 875; X86-NOBMI2-NEXT: pushl %edi 876; X86-NOBMI2-NEXT: pushl %esi 877; X86-NOBMI2-NEXT: pushl %eax 878; X86-NOBMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 879; X86-NOBMI2-NEXT: movl $-1, %esi 880; X86-NOBMI2-NEXT: movl $-1, %edi 881; X86-NOBMI2-NEXT: shrl %cl, %edi 882; X86-NOBMI2-NEXT: shrdl %cl, %esi, %esi 883; X86-NOBMI2-NEXT: testb $32, %cl 884; X86-NOBMI2-NEXT: je .LBB19_2 885; X86-NOBMI2-NEXT: # %bb.1: 886; X86-NOBMI2-NEXT: movl %edi, %esi 887; X86-NOBMI2-NEXT: xorl %edi, %edi 888; X86-NOBMI2-NEXT: .LBB19_2: 889; X86-NOBMI2-NEXT: subl $8, %esp 890; X86-NOBMI2-NEXT: pushl %edi 891; X86-NOBMI2-NEXT: pushl %esi 892; X86-NOBMI2-NEXT: calll use64 893; X86-NOBMI2-NEXT: addl $16, %esp 894; X86-NOBMI2-NEXT: andl {{[0-9]+}}(%esp), %esi 895; X86-NOBMI2-NEXT: andl {{[0-9]+}}(%esp), %edi 896; X86-NOBMI2-NEXT: movl %esi, %eax 897; X86-NOBMI2-NEXT: movl %edi, %edx 898; X86-NOBMI2-NEXT: addl $4, %esp 899; X86-NOBMI2-NEXT: popl %esi 900; X86-NOBMI2-NEXT: popl %edi 901; X86-NOBMI2-NEXT: retl 902; 903; X86-BMI2-LABEL: oneuse64: 904; X86-BMI2: # %bb.0: 905; X86-BMI2-NEXT: pushl %edi 906; X86-BMI2-NEXT: pushl %esi 907; X86-BMI2-NEXT: pushl %eax 908; X86-BMI2-NEXT: movb {{[0-9]+}}(%esp), %cl 909; X86-BMI2-NEXT: movl $-1, %esi 910; X86-BMI2-NEXT: shrxl %ecx, %esi, %edi 911; X86-BMI2-NEXT: shrdl %cl, %esi, %esi 912; X86-BMI2-NEXT: testb $32, %cl 913; X86-BMI2-NEXT: je .LBB19_2 914; X86-BMI2-NEXT: # %bb.1: 915; X86-BMI2-NEXT: movl %edi, %esi 916; X86-BMI2-NEXT: xorl %edi, %edi 917; X86-BMI2-NEXT: .LBB19_2: 918; X86-BMI2-NEXT: subl $8, %esp 919; X86-BMI2-NEXT: pushl %edi 920; X86-BMI2-NEXT: pushl %esi 921; X86-BMI2-NEXT: calll use64 922; X86-BMI2-NEXT: addl $16, %esp 923; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %esi 924; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %edi 925; X86-BMI2-NEXT: movl %esi, %eax 926; X86-BMI2-NEXT: movl %edi, %edx 927; X86-BMI2-NEXT: addl $4, %esp 928; X86-BMI2-NEXT: popl %esi 929; X86-BMI2-NEXT: popl %edi 930; X86-BMI2-NEXT: retl 931; 932; X64-NOBMI2-LABEL: oneuse64: 933; X64-NOBMI2: # %bb.0: 934; X64-NOBMI2-NEXT: pushq %r14 935; X64-NOBMI2-NEXT: pushq %rbx 936; X64-NOBMI2-NEXT: pushq %rax 937; X64-NOBMI2-NEXT: movq %rdi, %r14 938; X64-NOBMI2-NEXT: movq $-1, %rbx 939; X64-NOBMI2-NEXT: movl %esi, %ecx 940; X64-NOBMI2-NEXT: shrq %cl, %rbx 941; X64-NOBMI2-NEXT: movq %rbx, %rdi 942; X64-NOBMI2-NEXT: callq use64 943; X64-NOBMI2-NEXT: andq %r14, %rbx 944; X64-NOBMI2-NEXT: movq %rbx, %rax 945; X64-NOBMI2-NEXT: addq $8, %rsp 946; X64-NOBMI2-NEXT: popq %rbx 947; X64-NOBMI2-NEXT: popq %r14 948; X64-NOBMI2-NEXT: retq 949; 950; X64-BMI2-LABEL: oneuse64: 951; X64-BMI2: # %bb.0: 952; X64-BMI2-NEXT: pushq %r14 953; X64-BMI2-NEXT: pushq %rbx 954; X64-BMI2-NEXT: pushq %rax 955; X64-BMI2-NEXT: movq %rdi, %r14 956; X64-BMI2-NEXT: movq $-1, %rax 957; X64-BMI2-NEXT: shrxq %rsi, %rax, %rbx 958; X64-BMI2-NEXT: movq %rbx, %rdi 959; X64-BMI2-NEXT: callq use64 960; X64-BMI2-NEXT: andq %r14, %rbx 961; X64-BMI2-NEXT: movq %rbx, %rax 962; X64-BMI2-NEXT: addq $8, %rsp 963; X64-BMI2-NEXT: popq %rbx 964; X64-BMI2-NEXT: popq %r14 965; X64-BMI2-NEXT: retq 966 %mask = lshr i64 -1, %numhighbits 967 call void @use64(i64 %mask) 968 %masked = and i64 %mask, %val 969 ret i64 %masked 970} 971