1; RUN: llc < %s -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -verify-machineinstrs | FileCheck %s 2; RUN: llc < %s -asm-verbose=false -disable-wasm-fallthrough-return-opt -verify-machineinstrs | FileCheck %s --check-prefix=NOREGS 3 4; Test the register stackifier pass. 5 6; We have two sets of tests, one with registers and implicit locals, and 7; a stack / explicit locals based version (NOREGS). 8 9target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128" 10target triple = "wasm32-unknown-unknown" 11 12; No because of pointer aliasing. 13 14; CHECK-LABEL: no0: 15; CHECK: return $1{{$}} 16; NOREGS-LABEL: no0: 17; NOREGS: return{{$}} 18define i32 @no0(i32* %p, i32* %q) { 19 %t = load i32, i32* %q 20 store i32 0, i32* %p 21 ret i32 %t 22} 23 24; No because of side effects. 25 26; CHECK-LABEL: no1: 27; CHECK: return $1{{$}} 28; NOREGS-LABEL: no1: 29; NOREGS: return{{$}} 30define i32 @no1(i32* %p, i32* dereferenceable(4) %q) { 31 %t = load volatile i32, i32* %q, !invariant.load !0 32 store volatile i32 0, i32* %p 33 ret i32 %t 34} 35 36; Yes because of invariant load and no side effects. 37 38; CHECK-LABEL: yes0: 39; CHECK: return $pop{{[0-9]+}}{{$}} 40; NOREGS-LABEL: yes0: 41; NOREGS: return{{$}} 42define i32 @yes0(i32* %p, i32* dereferenceable(4) %q) { 43 %t = load i32, i32* %q, !invariant.load !0 44 store i32 0, i32* %p 45 ret i32 %t 46} 47 48; Yes because of no intervening side effects. 49 50; CHECK-LABEL: yes1: 51; CHECK: return $pop0{{$}} 52; NOREGS-LABEL: yes1: 53; NOREGS: return{{$}} 54define i32 @yes1(i32* %q) { 55 %t = load volatile i32, i32* %q 56 ret i32 %t 57} 58 59; Yes because undefined behavior can be sunk past a store. 60 61; CHECK-LABEL: sink_trap: 62; CHECK: return $pop{{[0-9]+}}{{$}} 63; NOREGS-LABEL: sink_trap: 64; NOREGS: return{{$}} 65define i32 @sink_trap(i32 %x, i32 %y, i32* %p) { 66 %t = sdiv i32 %x, %y 67 store volatile i32 0, i32* %p 68 ret i32 %t 69} 70 71; Yes because the call is readnone. 72 73; CHECK-LABEL: sink_readnone_call: 74; CHECK: return $pop1{{$}} 75; NOREGS-LABEL: sink_readnone_call: 76; NOREGS: return{{$}} 77declare i32 @readnone_callee() readnone nounwind 78define i32 @sink_readnone_call(i32 %x, i32 %y, i32* %p) { 79 %t = call i32 @readnone_callee() 80 store volatile i32 0, i32* %p 81 ret i32 %t 82} 83 84; No because the call is readonly and there's an intervening store. 85 86; CHECK-LABEL: no_sink_readonly_call: 87; CHECK: return ${{[0-9]+}}{{$}} 88; NOREGS-LABEL: no_sink_readonly_call: 89; NOREGS: return{{$}} 90declare i32 @readonly_callee() readonly nounwind 91define i32 @no_sink_readonly_call(i32 %x, i32 %y, i32* %p) { 92 %t = call i32 @readonly_callee() 93 store i32 0, i32* %p 94 ret i32 %t 95} 96 97; Don't schedule stack uses into the stack. To reduce register pressure, the 98; scheduler might be tempted to move the definition of $2 down. However, this 99; would risk getting incorrect liveness if the instructions are later 100; rearranged to make the stack contiguous. 101 102; CHECK-LABEL: stack_uses: 103; CHECK: .functype stack_uses (i32, i32, i32, i32) -> (i32){{$}} 104; CHECK-NEXT: block {{$}} 105; CHECK-NEXT: i32.const $push[[L13:[0-9]+]]=, 1{{$}} 106; CHECK-NEXT: i32.lt_s $push[[L0:[0-9]+]]=, $0, $pop[[L13]]{{$}} 107; CHECK-NEXT: i32.const $push[[L1:[0-9]+]]=, 2{{$}} 108; CHECK-NEXT: i32.lt_s $push[[L2:[0-9]+]]=, $1, $pop[[L1]]{{$}} 109; CHECK-NEXT: i32.xor $push[[L5:[0-9]+]]=, $pop[[L0]], $pop[[L2]]{{$}} 110; CHECK-NEXT: i32.const $push[[L12:[0-9]+]]=, 1{{$}} 111; CHECK-NEXT: i32.lt_s $push[[L3:[0-9]+]]=, $2, $pop[[L12]]{{$}} 112; CHECK-NEXT: i32.const $push[[L11:[0-9]+]]=, 2{{$}} 113; CHECK-NEXT: i32.lt_s $push[[L4:[0-9]+]]=, $3, $pop[[L11]]{{$}} 114; CHECK-NEXT: i32.xor $push[[L6:[0-9]+]]=, $pop[[L3]], $pop[[L4]]{{$}} 115; CHECK-NEXT: i32.xor $push[[L7:[0-9]+]]=, $pop[[L5]], $pop[[L6]]{{$}} 116; CHECK-NEXT: i32.const $push10=, 1{{$}} 117; CHECK-NEXT: i32.ne $push8=, $pop7, $pop10{{$}} 118; CHECK-NEXT: br_if 0, $pop8{{$}} 119; CHECK-NEXT: i32.const $push9=, 0{{$}} 120; CHECK-NEXT: return $pop9{{$}} 121; CHECK-NEXT: .LBB{{[0-9]+}}_2: 122; CHECK-NEXT: end_block{{$}} 123; CHECK-NEXT: i32.const $push14=, 1{{$}} 124; CHECK-NEXT: return $pop14{{$}} 125; NOREGS-LABEL: stack_uses: 126; NOREGS: .functype stack_uses (i32, i32, i32, i32) -> (i32){{$}} 127; NOREGS-NEXT: block {{$}} 128; NOREGS-NEXT: local.get 0{{$}} 129; NOREGS-NEXT: i32.const 1{{$}} 130; NOREGS-NEXT: i32.lt_s 131; NOREGS-NEXT: local.get 1{{$}} 132; NOREGS-NEXT: i32.const 2{{$}} 133; NOREGS-NEXT: i32.lt_s 134; NOREGS-NEXT: i32.xor {{$}} 135; NOREGS-NEXT: local.get 2{{$}} 136; NOREGS-NEXT: i32.const 1{{$}} 137; NOREGS-NEXT: i32.lt_s 138; NOREGS-NEXT: local.get 3{{$}} 139; NOREGS-NEXT: i32.const 2{{$}} 140; NOREGS-NEXT: i32.lt_s 141; NOREGS-NEXT: i32.xor {{$}} 142; NOREGS-NEXT: i32.xor {{$}} 143; NOREGS-NEXT: i32.const 1{{$}} 144; NOREGS-NEXT: i32.ne {{$}} 145; NOREGS-NEXT: br_if 0{{$}} 146; NOREGS-NEXT: i32.const 0{{$}} 147; NOREGS-NEXT: return{{$}} 148; NOREGS-NEXT: .LBB{{[0-9]+}}_2: 149; NOREGS-NEXT: end_block{{$}} 150; NOREGS-NEXT: i32.const 1{{$}} 151; NOREGS-NEXT: return{{$}} 152define i32 @stack_uses(i32 %x, i32 %y, i32 %z, i32 %w) { 153entry: 154 %c = icmp sle i32 %x, 0 155 %d = icmp sle i32 %y, 1 156 %e = icmp sle i32 %z, 0 157 %f = icmp sle i32 %w, 1 158 %g = xor i1 %c, %d 159 %h = xor i1 %e, %f 160 %i = xor i1 %g, %h 161 br i1 %i, label %true, label %false 162true: 163 ret i32 0 164false: 165 ret i32 1 166} 167 168; Test an interesting case where the load has multiple uses and cannot 169; be trivially stackified. However, it can be stackified with a local.tee. 170 171; CHECK-LABEL: multiple_uses: 172; CHECK: .functype multiple_uses (i32, i32, i32) -> (){{$}} 173; CHECK-NEXT: block {{$}} 174; CHECK-NEXT: i32.load $push[[NUM0:[0-9]+]]=, 0($2){{$}} 175; CHECK-NEXT: local.tee $push[[NUM1:[0-9]+]]=, $3=, $pop[[NUM0]]{{$}} 176; CHECK-NEXT: i32.ge_u $push[[NUM2:[0-9]+]]=, $pop[[NUM1]], $1{{$}} 177; CHECK-NEXT: br_if 0, $pop[[NUM2]]{{$}} 178; CHECK-NEXT: i32.lt_u $push[[NUM3:[0-9]+]]=, $3, $0{{$}} 179; CHECK-NEXT: br_if 0, $pop[[NUM3]]{{$}} 180; CHECK-NEXT: i32.store 0($2), $3{{$}} 181; CHECK-NEXT: .LBB{{[0-9]+}}_3: 182; CHECK-NEXT: end_block{{$}} 183; CHECK-NEXT: return{{$}} 184; NOREGS-LABEL: multiple_uses: 185; NOREGS: .functype multiple_uses (i32, i32, i32) -> (){{$}} 186; NOREGS: .local i32{{$}} 187; NOREGS-NEXT: block {{$}} 188; NOREGS-NEXT: local.get 2{{$}} 189; NOREGS-NEXT: i32.load 0{{$}} 190; NOREGS-NEXT: local.tee 3{{$}} 191; NOREGS-NEXT: local.get 1{{$}} 192; NOREGS-NEXT: i32.ge_u 193; NOREGS-NEXT: br_if 0{{$}} 194; NOREGS-NEXT: local.get 3{{$}} 195; NOREGS-NEXT: local.get 0{{$}} 196; NOREGS-NEXT: i32.lt_u 197; NOREGS-NEXT: br_if 0{{$}} 198; NOREGS-NEXT: local.get 2{{$}} 199; NOREGS-NEXT: local.get 3{{$}} 200; NOREGS-NEXT: i32.store 0{{$}} 201; NOREGS-NEXT: .LBB{{[0-9]+}}_3: 202; NOREGS-NEXT: end_block{{$}} 203; NOREGS-NEXT: return{{$}} 204define void @multiple_uses(i32* %arg0, i32* %arg1, i32* %arg2) nounwind { 205bb: 206 br label %loop 207 208loop: 209 %tmp7 = load i32, i32* %arg2 210 %tmp8 = inttoptr i32 %tmp7 to i32* 211 %tmp9 = icmp uge i32* %tmp8, %arg1 212 %tmp10 = icmp ult i32* %tmp8, %arg0 213 %tmp11 = or i1 %tmp9, %tmp10 214 br i1 %tmp11, label %back, label %then 215 216then: 217 store i32 %tmp7, i32* %arg2 218 br label %back 219 220back: 221 br i1 undef, label %return, label %loop 222 223return: 224 ret void 225} 226 227; Don't stackify stores effects across other instructions with side effects. 228 229; CHECK: side_effects: 230; CHECK: store 231; CHECK: call 232; CHECK: store 233; CHECK-NEXT: call 234; NOREGS: side_effects: 235; NOREGS: store 236; NOREGS: call 237; NOREGS: store 238; NOREGS-NEXT: call 239declare void @evoke_side_effects() 240define hidden void @stackify_store_across_side_effects(double* nocapture %d) { 241entry: 242 store double 2.0, double* %d 243 call void @evoke_side_effects() 244 store double 2.0, double* %d 245 call void @evoke_side_effects() 246 ret void 247} 248 249; Div instructions have side effects and can't be reordered, but this entire 250; function should still be able to be stackified because it's already in 251; tree order. 252 253; CHECK-LABEL: div_tree: 254; CHECK: .functype div_tree (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> (i32){{$}} 255; CHECK-NEXT: i32.div_s $push[[L0:[0-9]+]]=, $0, $1{{$}} 256; CHECK-NEXT: i32.div_s $push[[L1:[0-9]+]]=, $2, $3{{$}} 257; CHECK-NEXT: i32.div_s $push[[L2:[0-9]+]]=, $pop[[L0]], $pop[[L1]]{{$}} 258; CHECK-NEXT: i32.div_s $push[[L3:[0-9]+]]=, $4, $5{{$}} 259; CHECK-NEXT: i32.div_s $push[[L4:[0-9]+]]=, $6, $7{{$}} 260; CHECK-NEXT: i32.div_s $push[[L5:[0-9]+]]=, $pop[[L3]], $pop[[L4]]{{$}} 261; CHECK-NEXT: i32.div_s $push[[L6:[0-9]+]]=, $pop[[L2]], $pop[[L5]]{{$}} 262; CHECK-NEXT: i32.div_s $push[[L7:[0-9]+]]=, $8, $9{{$}} 263; CHECK-NEXT: i32.div_s $push[[L8:[0-9]+]]=, $10, $11{{$}} 264; CHECK-NEXT: i32.div_s $push[[L9:[0-9]+]]=, $pop[[L7]], $pop[[L8]]{{$}} 265; CHECK-NEXT: i32.div_s $push[[L10:[0-9]+]]=, $12, $13{{$}} 266; CHECK-NEXT: i32.div_s $push[[L11:[0-9]+]]=, $14, $15{{$}} 267; CHECK-NEXT: i32.div_s $push[[L12:[0-9]+]]=, $pop[[L10]], $pop[[L11]]{{$}} 268; CHECK-NEXT: i32.div_s $push[[L13:[0-9]+]]=, $pop[[L9]], $pop[[L12]]{{$}} 269; CHECK-NEXT: i32.div_s $push[[L14:[0-9]+]]=, $pop[[L6]], $pop[[L13]]{{$}} 270; CHECK-NEXT: return $pop[[L14]]{{$}} 271; NOREGS-LABEL: div_tree: 272; NOREGS: .functype div_tree (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> (i32){{$}} 273; NOREGS-NEXT: local.get 0{{$}} 274; NOREGS-NEXT: local.get 1{{$}} 275; NOREGS-NEXT: i32.div_s{{$}} 276; NOREGS-NEXT: local.get 2{{$}} 277; NOREGS-NEXT: local.get 3{{$}} 278; NOREGS-NEXT: i32.div_s{{$}} 279; NOREGS-NEXT: i32.div_s{{$}} 280; NOREGS-NEXT: local.get 4{{$}} 281; NOREGS-NEXT: local.get 5{{$}} 282; NOREGS-NEXT: i32.div_s{{$}} 283; NOREGS-NEXT: local.get 6{{$}} 284; NOREGS-NEXT: local.get 7{{$}} 285; NOREGS-NEXT: i32.div_s{{$}} 286; NOREGS-NEXT: i32.div_s{{$}} 287; NOREGS-NEXT: i32.div_s{{$}} 288; NOREGS-NEXT: local.get 8{{$}} 289; NOREGS-NEXT: local.get 9{{$}} 290; NOREGS-NEXT: i32.div_s{{$}} 291; NOREGS-NEXT: local.get 10{{$}} 292; NOREGS-NEXT: local.get 11{{$}} 293; NOREGS-NEXT: i32.div_s{{$}} 294; NOREGS-NEXT: i32.div_s{{$}} 295; NOREGS-NEXT: local.get 12{{$}} 296; NOREGS-NEXT: local.get 13{{$}} 297; NOREGS-NEXT: i32.div_s{{$}} 298; NOREGS-NEXT: local.get 14{{$}} 299; NOREGS-NEXT: local.get 15{{$}} 300; NOREGS-NEXT: i32.div_s{{$}} 301; NOREGS-NEXT: i32.div_s{{$}} 302; NOREGS-NEXT: i32.div_s{{$}} 303; NOREGS-NEXT: i32.div_s{{$}} 304; NOREGS-NEXT: return{{$}} 305define i32 @div_tree(i32 %a, i32 %b, i32 %c, i32 %d, i32 %e, i32 %f, i32 %g, i32 %h, i32 %i, i32 %j, i32 %k, i32 %l, i32 %m, i32 %n, i32 %o, i32 %p) { 306entry: 307 %div = sdiv i32 %a, %b 308 %div1 = sdiv i32 %c, %d 309 %div2 = sdiv i32 %div, %div1 310 %div3 = sdiv i32 %e, %f 311 %div4 = sdiv i32 %g, %h 312 %div5 = sdiv i32 %div3, %div4 313 %div6 = sdiv i32 %div2, %div5 314 %div7 = sdiv i32 %i, %j 315 %div8 = sdiv i32 %k, %l 316 %div9 = sdiv i32 %div7, %div8 317 %div10 = sdiv i32 %m, %n 318 %div11 = sdiv i32 %o, %p 319 %div12 = sdiv i32 %div10, %div11 320 %div13 = sdiv i32 %div9, %div12 321 %div14 = sdiv i32 %div6, %div13 322 ret i32 %div14 323} 324 325; A simple multiple-use case. 326 327; CHECK-LABEL: simple_multiple_use: 328; CHECK: .functype simple_multiple_use (i32, i32) -> (){{$}} 329; CHECK-NEXT: i32.mul $push[[NUM0:[0-9]+]]=, $1, $0{{$}} 330; CHECK-NEXT: local.tee $push[[NUM1:[0-9]+]]=, $[[NUM2:[0-9]+]]=, $pop[[NUM0]]{{$}} 331; CHECK-NEXT: call use_a, $pop[[NUM1]]{{$}} 332; CHECK-NEXT: call use_b, $[[NUM2]]{{$}} 333; CHECK-NEXT: return{{$}} 334; NOREGS-LABEL: simple_multiple_use: 335; NOREGS: .functype simple_multiple_use (i32, i32) -> (){{$}} 336; NOREGS-NEXT: local.get 1{{$}} 337; NOREGS-NEXT: local.get 0{{$}} 338; NOREGS-NEXT: i32.mul 339; NOREGS-NEXT: local.tee 1{{$}} 340; NOREGS-NEXT: call use_a{{$}} 341; NOREGS-NEXT: local.get 1{{$}} 342; NOREGS-NEXT: call use_b{{$}} 343; NOREGS-NEXT: return{{$}} 344declare void @use_a(i32) 345declare void @use_b(i32) 346define void @simple_multiple_use(i32 %x, i32 %y) { 347 %mul = mul i32 %y, %x 348 call void @use_a(i32 %mul) 349 call void @use_b(i32 %mul) 350 ret void 351} 352 353; Multiple uses of the same value in one instruction. 354 355; CHECK-LABEL: multiple_uses_in_same_insn: 356; CHECK: .functype multiple_uses_in_same_insn (i32, i32) -> (){{$}} 357; CHECK-NEXT: i32.mul $push[[NUM0:[0-9]+]]=, $1, $0{{$}} 358; CHECK-NEXT: local.tee $push[[NUM1:[0-9]+]]=, $[[NUM2:[0-9]+]]=, $pop[[NUM0]]{{$}} 359; CHECK-NEXT: call use_2, $pop[[NUM1]], $[[NUM2]]{{$}} 360; CHECK-NEXT: return{{$}} 361; NOREGS-LABEL: multiple_uses_in_same_insn: 362; NOREGS: .functype multiple_uses_in_same_insn (i32, i32) -> (){{$}} 363; NOREGS-NEXT: local.get 1{{$}} 364; NOREGS-NEXT: local.get 0{{$}} 365; NOREGS-NEXT: i32.mul 366; NOREGS-NEXT: local.tee 1{{$}} 367; NOREGS-NEXT: local.get 1{{$}} 368; NOREGS-NEXT: call use_2{{$}} 369; NOREGS-NEXT: return{{$}} 370declare void @use_2(i32, i32) 371define void @multiple_uses_in_same_insn(i32 %x, i32 %y) { 372 %mul = mul i32 %y, %x 373 call void @use_2(i32 %mul, i32 %mul) 374 ret void 375} 376 377; Commute operands to achieve better stackifying. 378 379; CHECK-LABEL: commute: 380; CHECK: .functype commute () -> (i32){{$}} 381; CHECK-NEXT: call $push0=, red{{$}} 382; CHECK-NEXT: call $push1=, green{{$}} 383; CHECK-NEXT: i32.add $push2=, $pop0, $pop1{{$}} 384; CHECK-NEXT: call $push3=, blue{{$}} 385; CHECK-NEXT: i32.add $push4=, $pop2, $pop3{{$}} 386; CHECK-NEXT: return $pop4{{$}} 387; NOREGS-LABEL: commute: 388; NOREGS: .functype commute () -> (i32){{$}} 389; NOREGS-NEXT: call red{{$}} 390; NOREGS-NEXT: call green{{$}} 391; NOREGS-NEXT: i32.add {{$}} 392; NOREGS-NEXT: call blue{{$}} 393; NOREGS-NEXT: i32.add {{$}} 394; NOREGS-NEXT: return{{$}} 395declare i32 @red() 396declare i32 @green() 397declare i32 @blue() 398define i32 @commute() { 399 %call = call i32 @red() 400 %call1 = call i32 @green() 401 %add = add i32 %call1, %call 402 %call2 = call i32 @blue() 403 %add3 = add i32 %add, %call2 404 ret i32 %add3 405} 406 407; Don't stackify a register when it would move a the def of the register past 408; an implicit local.get for the register. 409 410; CHECK-LABEL: no_stackify_past_use: 411; CHECK: call $1=, callee, $0 412; CHECK-NEXT: i32.const $push0=, 1 413; CHECK-NEXT: i32.add $push1=, $0, $pop0 414; CHECK-NEXT: call $push2=, callee, $pop1 415; CHECK-NEXT: i32.sub $push3=, $pop2, $1 416; CHECK-NEXT: i32.div_s $push4=, $pop3, $1 417; CHECK-NEXT: return $pop4 418; NOREGS-LABEL: no_stackify_past_use: 419; NOREGS: local.get 0{{$}} 420; NOREGS-NEXT: call callee 421; NOREGS-NEXT: local.set 1{{$}} 422; NOREGS-NEXT: local.get 0{{$}} 423; NOREGS-NEXT: i32.const 1 424; NOREGS-NEXT: i32.add 425; NOREGS-NEXT: call callee 426; NOREGS-NEXT: local.get 1{{$}} 427; NOREGS-NEXT: i32.sub 428; NOREGS-NEXT: local.get 1{{$}} 429; NOREGS-NEXT: i32.div_s 430; NOREGS-NEXT: return 431declare i32 @callee(i32) 432define i32 @no_stackify_past_use(i32 %arg) { 433 %tmp1 = call i32 @callee(i32 %arg) 434 %tmp2 = add i32 %arg, 1 435 %tmp3 = call i32 @callee(i32 %tmp2) 436 %tmp5 = sub i32 %tmp3, %tmp1 437 %tmp6 = sdiv i32 %tmp5, %tmp1 438 ret i32 %tmp6 439} 440 441; This is the same as no_stackify_past_use, except using a commutative operator, 442; so we can reorder the operands and stackify. 443 444; CHECK-LABEL: commute_to_fix_ordering: 445; CHECK: call $push[[L0:.+]]=, callee, $0 446; CHECK: local.tee $push[[L1:.+]]=, $1=, $pop[[L0]] 447; CHECK: i32.const $push0=, 1 448; CHECK: i32.add $push1=, $0, $pop0 449; CHECK: call $push2=, callee, $pop1 450; CHECK: i32.add $push3=, $1, $pop2 451; CHECK: i32.mul $push4=, $pop[[L1]], $pop3 452; CHECK: return $pop4 453; NOREGS-LABEL: commute_to_fix_ordering: 454; NOREGS: local.get 0{{$}} 455; NOREGS: call callee 456; NOREGS: local.tee 1 457; NOREGS: local.get 1{{$}} 458; NOREGS: local.get 0{{$}} 459; NOREGS: i32.const 1 460; NOREGS: i32.add 461; NOREGS: call callee 462; NOREGS: i32.add 463; NOREGS: i32.mul 464; NOREGS: return 465define i32 @commute_to_fix_ordering(i32 %arg) { 466 %tmp1 = call i32 @callee(i32 %arg) 467 %tmp2 = add i32 %arg, 1 468 %tmp3 = call i32 @callee(i32 %tmp2) 469 %tmp5 = add i32 %tmp3, %tmp1 470 %tmp6 = mul i32 %tmp5, %tmp1 471 ret i32 %tmp6 472} 473 474; Stackify individual defs of virtual registers with multiple defs. 475 476; CHECK-LABEL: multiple_defs: 477; CHECK: f64.add $push[[NUM0:[0-9]+]]=, ${{[0-9]+}}, $pop{{[0-9]+}}{{$}} 478; CHECK-NEXT: local.tee $push[[NUM1:[0-9]+]]=, $[[NUM2:[0-9]+]]=, $pop[[NUM0]]{{$}} 479; CHECK-NEXT: f64.select $push{{[0-9]+}}=, $pop{{[0-9]+}}, $pop[[NUM1]], ${{[0-9]+}}{{$}} 480; CHECK: $[[NUM2]]=, 481; NOREGS-LABEL: multiple_defs: 482; NOREGS: f64.add 483; NOREGS: local.tee 484; NOREGS: f64.select 485define void @multiple_defs(i32 %arg, i32 %arg1, i1 %arg2, i1 %arg3, i1 %arg4) { 486bb: 487 br label %bb5 488 489bb5: ; preds = %bb21, %bb 490 %tmp = phi double [ 0.000000e+00, %bb ], [ %tmp22, %bb21 ] 491 %tmp6 = phi double [ 0.000000e+00, %bb ], [ %tmp23, %bb21 ] 492 %tmp7 = fcmp olt double %tmp6, 2.323450e+01 493 br i1 %tmp7, label %bb8, label %bb21 494 495bb8: ; preds = %bb17, %bb5 496 %tmp9 = phi double [ %tmp19, %bb17 ], [ %tmp, %bb5 ] 497 %tmp10 = fadd double %tmp6, -1.000000e+00 498 %tmp11 = select i1 %arg2, double -1.135357e+04, double %tmp10 499 %tmp12 = fadd double %tmp11, %tmp9 500 br i1 %arg3, label %bb17, label %bb13 501 502bb13: ; preds = %bb8 503 %tmp14 = or i32 %arg1, 2 504 %tmp15 = icmp eq i32 %tmp14, 14 505 %tmp16 = select i1 %tmp15, double -1.135357e+04, double 0xBFCE147AE147B000 506 br label %bb17 507 508bb17: ; preds = %bb13, %bb8 509 %tmp18 = phi double [ %tmp16, %bb13 ], [ %tmp10, %bb8 ] 510 %tmp19 = fadd double %tmp18, %tmp12 511 %tmp20 = fcmp olt double %tmp6, 2.323450e+01 512 br i1 %tmp20, label %bb8, label %bb21 513 514bb21: ; preds = %bb17, %bb5 515 %tmp22 = phi double [ %tmp, %bb5 ], [ %tmp9, %bb17 ] 516 %tmp23 = fadd double %tmp6, 1.000000e+00 517 br i1 %arg4, label %exit, label %bb5 518exit: 519 ret void 520} 521 522; Don't move calls past loads 523; CHECK-LABEL: no_stackify_call_past_load: 524; CHECK: call $0=, red 525; CHECK: i32.const $push0=, 0 526; CHECK: i32.load $1=, count($pop0) 527; NOREGS-LABEL: no_stackify_call_past_load: 528; NOREGS: call red 529; NOREGS: i32.const 0 530; NOREGS: i32.load count 531@count = hidden global i32 0, align 4 532define i32 @no_stackify_call_past_load() { 533 %a = call i32 @red() 534 %b = load i32, i32* @count, align 4 535 call i32 @callee(i32 %a) 536 ret i32 %b 537 ; use of a 538} 539 540; Don't move stores past loads if there may be aliasing 541; CHECK-LABEL: no_stackify_store_past_load 542; CHECK: i32.store 0($1), $0 543; CHECK: i32.load {{.*}}, 0($2) 544; CHECK: call {{.*}}, callee, $0{{$}} 545; NOREGS-LABEL: no_stackify_store_past_load 546; NOREGS: i32.store 0 547; NOREGS: i32.load 0 548; NOREGS: call callee{{$}} 549define i32 @no_stackify_store_past_load(i32 %a, i32* %p1, i32* %p2) { 550 store i32 %a, i32* %p1 551 %b = load i32, i32* %p2, align 4 552 call i32 @callee(i32 %a) 553 ret i32 %b 554} 555 556; Can still stackify past invariant loads. 557; CHECK-LABEL: store_past_invar_load 558; CHECK: i32.store 0($1), $0 559; CHECK: call {{.*}}, callee, $0 560; CHECK: i32.load $push{{.*}}, 0($2) 561; CHECK: return $pop 562; NOREGS-LABEL: store_past_invar_load 563; NOREGS: i32.store 0 564; NOREGS: call callee 565; NOREGS: i32.load 0 566; NOREGS: return 567define i32 @store_past_invar_load(i32 %a, i32* %p1, i32* dereferenceable(4) %p2) { 568 store i32 %a, i32* %p1 569 %b = load i32, i32* %p2, !invariant.load !0 570 call i32 @callee(i32 %a) 571 ret i32 %b 572} 573 574; CHECK-LABEL: ignore_dbg_value: 575; CHECK: .Lfunc_begin 576; CHECK: unreachable 577; NOREGS-LABEL: ignore_dbg_value: 578; NOREGS: .Lfunc_begin 579; NOREGS: unreachable 580declare void @llvm.dbg.value(metadata, i64, metadata, metadata) 581define void @ignore_dbg_value() { 582 call void @llvm.dbg.value(metadata i32 0, i64 0, metadata !7, metadata !9), !dbg !10 583 unreachable 584} 585 586; Don't stackify an expression that might use the stack into a return, since we 587; might insert a prologue before the return. 588 589; CHECK-LABEL: no_stackify_past_epilogue: 590; CHECK: return ${{[0-9]+}}{{$}} 591; NOREGS-LABEL: no_stackify_past_epilogue: 592; NOREGS: return{{$}} 593declare i32 @use_memory(i32*) 594define i32 @no_stackify_past_epilogue() { 595 %x = alloca i32 596 %call = call i32 @use_memory(i32* %x) 597 ret i32 %call 598} 599 600; Stackify a loop induction variable into a loop comparison. 601 602; CHECK-LABEL: stackify_indvar: 603; CHECK: i32.const $push[[L5:.+]]=, 1{{$}} 604; CHECK-NEXT: i32.add $push[[L4:.+]]=, $[[R0:.+]], $pop[[L5]]{{$}} 605; CHECK-NEXT: local.tee $push[[L3:.+]]=, $[[R0]]=, $pop[[L4]]{{$}} 606; CHECK-NEXT: i32.ne $push[[L2:.+]]=, $0, $pop[[L3]]{{$}} 607; NOREGS-LABEL: stackify_indvar: 608; NOREGS: i32.const 1{{$}} 609; NOREGS-NEXT: i32.add 610; NOREGS-NEXT: local.tee 2{{$}} 611; NOREGS-NEXT: i32.ne 612define void @stackify_indvar(i32 %tmp, i32* %v) #0 { 613bb: 614 br label %bb3 615 616bb3: ; preds = %bb3, %bb2 617 %tmp4 = phi i32 [ %tmp7, %bb3 ], [ 0, %bb ] 618 %tmp5 = load volatile i32, i32* %v, align 4 619 %tmp6 = add nsw i32 %tmp5, %tmp4 620 store volatile i32 %tmp6, i32* %v, align 4 621 %tmp7 = add nuw nsw i32 %tmp4, 1 622 %tmp8 = icmp eq i32 %tmp7, %tmp 623 br i1 %tmp8, label %bb10, label %bb3 624 625bb10: ; preds = %bb9, %bb 626 ret void 627} 628 629; Don't stackify a call past a __stack_pointer store. 630 631; CHECK-LABEL: stackpointer_dependency: 632; CHECK: call {{.+}}, stackpointer_callee, 633; CHECK-NEXT: global.set __stack_pointer, 634; NOREGS-LABEL: stackpointer_dependency: 635; NOREGS: call stackpointer_callee 636; NOREGS: global.set __stack_pointer 637declare i32 @stackpointer_callee(i8* readnone, i8* readnone) nounwind readnone 638declare i8* @llvm.frameaddress(i32) 639define i32 @stackpointer_dependency(i8* readnone) { 640 %2 = tail call i8* @llvm.frameaddress(i32 0) 641 %3 = tail call i32 @stackpointer_callee(i8* %0, i8* %2) 642 ret i32 %3 643} 644 645; Stackify a call_indirect with respect to its ordering 646 647; CHECK-LABEL: call_indirect_stackify: 648; CHECK: i32.load $push[[L4:.+]]=, 0($0) 649; CHECK-NEXT: local.tee $push[[L3:.+]]=, $0=, $pop[[L4]] 650; CHECK-NEXT: i32.load $push[[L0:.+]]=, 0($0) 651; CHECK-NEXT: i32.load $push[[L1:.+]]=, 0($pop[[L0]]) 652; CHECK-NEXT: call_indirect $push{{.+}}=, $pop[[L3]], $1, $pop[[L1]] 653; NOREGS-LABEL: call_indirect_stackify: 654; NOREGS: i32.load 0 655; NOREGS-NEXT: local.tee 0 656; NOREGS: i32.load 0 657; NOREGS-NEXT: i32.load 0 658; NOREGS-NEXT: call_indirect (i32, i32) -> (i32) 659%class.call_indirect = type { i32 (...)** } 660define i32 @call_indirect_stackify(%class.call_indirect** %objptr, i32 %arg) { 661 %obj = load %class.call_indirect*, %class.call_indirect** %objptr 662 %addr = bitcast %class.call_indirect* %obj to i32(%class.call_indirect*, i32)*** 663 %vtable = load i32(%class.call_indirect*, i32)**, i32(%class.call_indirect*, i32)*** %addr 664 %vfn = getelementptr inbounds i32(%class.call_indirect*, i32)*, i32(%class.call_indirect*, i32)** %vtable, i32 0 665 %f = load i32(%class.call_indirect*, i32)*, i32(%class.call_indirect*, i32)** %vfn 666 %ret = call i32 %f(%class.call_indirect* %obj, i32 %arg) 667 ret i32 %ret 668} 669 670!llvm.module.flags = !{!0} 671!llvm.dbg.cu = !{!1} 672 673!0 = !{i32 2, !"Debug Info Version", i32 3} 674!1 = distinct !DICompileUnit(language: DW_LANG_C99, file: !2, producer: "clang version 3.9.0 (trunk 266005) (llvm/trunk 266105)", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, enums: !3) 675!2 = !DIFile(filename: "test.c", directory: "/") 676!3 = !{} 677!5 = distinct !DISubprogram(name: "test", scope: !2, file: !2, line: 10, type: !6, isLocal: false, isDefinition: true, scopeLine: 11, flags: DIFlagPrototyped, isOptimized: true, unit: !1, retainedNodes: !3) 678!6 = !DISubroutineType(types: !3) 679!7 = !DILocalVariable(name: "nzcnt", scope: !5, file: !2, line: 15, type: !8) 680!8 = !DIBasicType(name: "int", size: 32, align: 32, encoding: DW_ATE_signed) 681!9 = !DIExpression() 682!10 = !DILocation(line: 15, column: 6, scope: !5) 683