1; RUN: llc < %s -mtriple=i686-windows | FileCheck %s -check-prefix=NORMAL 2; RUN: llc < %s -mtriple=i686-windows -no-x86-call-frame-opt | FileCheck %s -check-prefix=NOPUSH 3; RUN: llc < %s -mtriple=x86_64-windows | FileCheck %s -check-prefix=X64 4; RUN: llc < %s -mtriple=i686-windows -stackrealign -stack-alignment=32 | FileCheck %s -check-prefix=ALIGNED 5; RUN: llc < %s -mtriple=i686-pc-linux | FileCheck %s -check-prefix=LINUX 6 7%class.Class = type { i32 } 8%struct.s = type { i64 } 9 10declare void @good(i32 %a, i32 %b, i32 %c, i32 %d) 11declare void @inreg(i32 %a, i32 inreg %b, i32 %c, i32 %d) 12declare x86_thiscallcc void @thiscall(%class.Class* %class, i32 %a, i32 %b, i32 %c, i32 %d) 13declare void @oneparam(i32 %a) 14declare void @eightparams(i32 %a, i32 %b, i32 %c, i32 %d, i32 %e, i32 %f, i32 %g, i32 %h) 15declare void @struct(%struct.s* byval %a, i32 %b, i32 %c, i32 %d) 16declare void @inalloca(<{ %struct.s }>* inalloca) 17 18declare i8* @llvm.stacksave() 19declare void @llvm.stackrestore(i8*) 20 21; We should get pushes for x86, even though there is a reserved call frame. 22; Make sure we don't touch x86-64, and that turning it off works. 23; NORMAL-LABEL: test1: 24; NORMAL-NOT: subl {{.*}} %esp 25; NORMAL: pushl $4 26; NORMAL-NEXT: pushl $3 27; NORMAL-NEXT: pushl $2 28; NORMAL-NEXT: pushl $1 29; NORMAL-NEXT: call 30; NORMAL-NEXT: addl $16, %esp 31; X64-LABEL: test1: 32; X64: movl $1, %ecx 33; X64-NEXT: movl $2, %edx 34; X64-NEXT: movl $3, %r8d 35; X64-NEXT: movl $4, %r9d 36; X64-NEXT: callq good 37; NOPUSH-LABEL: test1: 38; NOPUSH: subl $16, %esp 39; NOPUSH-NEXT: movl $4, 12(%esp) 40; NOPUSH-NEXT: movl $3, 8(%esp) 41; NOPUSH-NEXT: movl $2, 4(%esp) 42; NOPUSH-NEXT: movl $1, (%esp) 43; NOPUSH-NEXT: call 44; NOPUSH-NEXT: addl $16, %esp 45define void @test1() { 46entry: 47 call void @good(i32 1, i32 2, i32 3, i32 4) 48 ret void 49} 50 51; If we have a reserved frame, we should have pushes 52; NORMAL-LABEL: test2: 53; NORMAL-NOT: subl {{.*}} %esp 54; NORMAL: pushl $4 55; NORMAL-NEXT: pushl $3 56; NORMAL-NEXT: pushl $2 57; NORMAL-NEXT: pushl $1 58; NORMAL-NEXT: call 59define void @test2(i32 %k) { 60entry: 61 %a = alloca i32, i32 %k 62 call void @good(i32 1, i32 2, i32 3, i32 4) 63 ret void 64} 65 66; Again, we expect a sequence of 4 immediate pushes 67; Checks that we generate the right pushes for >8bit immediates 68; NORMAL-LABEL: test2b: 69; NORMAL-NOT: subl {{.*}} %esp 70; NORMAL: pushl $4096 71; NORMAL-NEXT: pushl $3072 72; NORMAL-NEXT: pushl $2048 73; NORMAL-NEXT: pushl $1024 74; NORMAL-NEXT: call 75; NORMAL-NEXT: addl $16, %esp 76define void @test2b() optsize { 77entry: 78 call void @good(i32 1024, i32 2048, i32 3072, i32 4096) 79 ret void 80} 81 82; The first push should push a register 83; NORMAL-LABEL: test3: 84; NORMAL-NOT: subl {{.*}} %esp 85; NORMAL: pushl $4 86; NORMAL-NEXT: pushl $3 87; NORMAL-NEXT: pushl $2 88; NORMAL-NEXT: pushl %e{{..}} 89; NORMAL-NEXT: call 90; NORMAL-NEXT: addl $16, %esp 91define void @test3(i32 %k) optsize { 92entry: 93 %f = add i32 %k, 1 94 call void @good(i32 %f, i32 2, i32 3, i32 4) 95 ret void 96} 97 98; We support weird calling conventions 99; NORMAL-LABEL: test4: 100; NORMAL: movl $2, %eax 101; NORMAL-NEXT: pushl $4 102; NORMAL-NEXT: pushl $3 103; NORMAL-NEXT: pushl $1 104; NORMAL-NEXT: call 105; NORMAL-NEXT: addl $12, %esp 106define void @test4() optsize { 107entry: 108 call void @inreg(i32 1, i32 2, i32 3, i32 4) 109 ret void 110} 111 112; NORMAL-LABEL: test4b: 113; NORMAL: movl 4(%esp), %ecx 114; NORMAL-NEXT: pushl $4 115; NORMAL-NEXT: pushl $3 116; NORMAL-NEXT: pushl $2 117; NORMAL-NEXT: pushl $1 118; NORMAL-NEXT: call 119; NORMAL-NEXT: ret 120define void @test4b(%class.Class* %f) optsize { 121entry: 122 call x86_thiscallcc void @thiscall(%class.Class* %f, i32 1, i32 2, i32 3, i32 4) 123 ret void 124} 125 126; When there is no reserved call frame, check that additional alignment 127; is added when the pushes don't add up to the required alignment. 128; ALIGNED-LABEL: test5: 129; ALIGNED: subl $16, %esp 130; ALIGNED-NEXT: pushl $4 131; ALIGNED-NEXT: pushl $3 132; ALIGNED-NEXT: pushl $2 133; ALIGNED-NEXT: pushl $1 134; ALIGNED-NEXT: call 135define void @test5(i32 %k) { 136entry: 137 %a = alloca i32, i32 %k 138 call void @good(i32 1, i32 2, i32 3, i32 4) 139 ret void 140} 141 142; When the alignment adds up, do the transformation 143; ALIGNED-LABEL: test5b: 144; ALIGNED: pushl $8 145; ALIGNED-NEXT: pushl $7 146; ALIGNED-NEXT: pushl $6 147; ALIGNED-NEXT: pushl $5 148; ALIGNED-NEXT: pushl $4 149; ALIGNED-NEXT: pushl $3 150; ALIGNED-NEXT: pushl $2 151; ALIGNED-NEXT: pushl $1 152; ALIGNED-NEXT: call 153define void @test5b() optsize { 154entry: 155 call void @eightparams(i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8) 156 ret void 157} 158 159; When having to compensate for the alignment isn't worth it, 160; don't use pushes. 161; ALIGNED-LABEL: test5c: 162; ALIGNED: movl $1, (%esp) 163; ALIGNED-NEXT: call 164define void @test5c() optsize { 165entry: 166 call void @oneparam(i32 1) 167 ret void 168} 169 170; Check that pushing the addresses of globals (Or generally, things that 171; aren't exactly immediates) isn't broken. 172; Fixes PR21878. 173; NORMAL-LABEL: test6: 174; NORMAL: pushl $_ext 175; NORMAL-NEXT: call 176declare void @f(i8*) 177@ext = external constant i8 178 179define void @test6() { 180 call void @f(i8* @ext) 181 br label %bb 182bb: 183 alloca i32 184 ret void 185} 186 187; Check that we fold simple cases into the push 188; NORMAL-LABEL: test7: 189; NORMAL-NOT: subl {{.*}} %esp 190; NORMAL: movl 4(%esp), [[EAX:%e..]] 191; NORMAL-NEXT: pushl $4 192; NORMAL-NEXT: pushl ([[EAX]]) 193; NORMAL-NEXT: pushl $2 194; NORMAL-NEXT: pushl $1 195; NORMAL-NEXT: call 196; NORMAL-NEXT: addl $16, %esp 197define void @test7(i32* %ptr) optsize { 198entry: 199 %val = load i32, i32* %ptr 200 call void @good(i32 1, i32 2, i32 %val, i32 4) 201 ret void 202} 203 204; Fold stack-relative loads into the push, with correct offset 205; In particular, at the second push, %b was at 12(%esp) and 206; %a wast at 8(%esp), but the second push bumped %esp, so %a 207; is now it at 12(%esp) 208; NORMAL-LABEL: test8: 209; NORMAL: pushl $4 210; NORMAL-NEXT: pushl 12(%esp) 211; NORMAL-NEXT: pushl 12(%esp) 212; NORMAL-NEXT: pushl $1 213; NORMAL-NEXT: call 214; NORMAL-NEXT: addl $16, %esp 215define void @test8(i32 %a, i32 %b) optsize { 216entry: 217 call void @good(i32 1, i32 %a, i32 %b, i32 4) 218 ret void 219} 220 221; If one function is using push instructions, and the other isn't 222; (because it has frame-index references), then we must resolve 223; these references correctly. 224; NORMAL-LABEL: test9: 225; NORMAL-NOT: leal (%esp), 226; NORMAL: pushl $4 227; NORMAL-NEXT: pushl $3 228; NORMAL-NEXT: pushl $2 229; NORMAL-NEXT: pushl $1 230; NORMAL-NEXT: call 231; NORMAL-NEXT: subl $4, %esp 232; NORMAL-NEXT: movl 20(%esp), [[E1:%e..]] 233; NORMAL-NEXT: movl 24(%esp), [[E2:%e..]] 234; NORMAL-NEXT: movl [[E2]], 4(%esp) 235; NORMAL-NEXT: movl [[E1]], (%esp) 236; NORMAL-NEXT: leal 32(%esp), [[E3:%e..]] 237; NORMAL-NEXT: movl [[E3]], 16(%esp) 238; NORMAL-NEXT: leal 28(%esp), [[E4:%e..]] 239; NORMAL-NEXT: movl [[E4]], 12(%esp) 240; NORMAL-NEXT: movl $6, 8(%esp) 241; NORMAL-NEXT: call 242; NORMAL-NEXT: addl $20, %esp 243define void @test9() optsize { 244entry: 245 %p = alloca i32, align 4 246 %q = alloca i32, align 4 247 %s = alloca %struct.s, align 4 248 call void @good(i32 1, i32 2, i32 3, i32 4) 249 %pv = ptrtoint i32* %p to i32 250 %qv = ptrtoint i32* %q to i32 251 call void @struct(%struct.s* byval %s, i32 6, i32 %qv, i32 %pv) 252 ret void 253} 254 255; We can end up with an indirect call which gets reloaded on the spot. 256; Make sure we reference the correct stack slot - we spill into (%esp) 257; and reload from 16(%esp) due to the pushes. 258; NORMAL-LABEL: test10: 259; NORMAL: movl $_good, [[ALLOC:.*]] 260; NORMAL-NEXT: movl [[ALLOC]], [[EAX:%e..]] 261; NORMAL-NEXT: movl [[EAX]], (%esp) # 4-byte Spill 262; NORMAL: nop 263; NORMAL: pushl $4 264; NORMAL-NEXT: pushl $3 265; NORMAL-NEXT: pushl $2 266; NORMAL-NEXT: pushl $1 267; NORMAL-NEXT: calll *16(%esp) 268; NORMAL-NEXT: addl $24, %esp 269define void @test10() optsize { 270 %stack_fptr = alloca void (i32, i32, i32, i32)* 271 store void (i32, i32, i32, i32)* @good, void (i32, i32, i32, i32)** %stack_fptr 272 %good_ptr = load volatile void (i32, i32, i32, i32)*, void (i32, i32, i32, i32)** %stack_fptr 273 call void asm sideeffect "nop", "~{ax},~{bx},~{cx},~{dx},~{bp},~{si},~{di}"() 274 call void (i32, i32, i32, i32) %good_ptr(i32 1, i32 2, i32 3, i32 4) 275 ret void 276} 277 278; We can't fold the load from the global into the push because of 279; interference from the store 280; NORMAL-LABEL: test11: 281; NORMAL: movl _the_global, [[EAX:%e..]] 282; NORMAL-NEXT: movl $42, _the_global 283; NORMAL-NEXT: pushl $4 284; NORMAL-NEXT: pushl $3 285; NORMAL-NEXT: pushl $2 286; NORMAL-NEXT: pushl [[EAX]] 287; NORMAL-NEXT: call 288; NORMAL-NEXT: addl $16, %esp 289@the_global = external global i32 290define void @test11() optsize { 291 %myload = load i32, i32* @the_global 292 store i32 42, i32* @the_global 293 call void @good(i32 %myload, i32 2, i32 3, i32 4) 294 ret void 295} 296 297; Converting one mov into a push isn't worth it when 298; doing so forces too much overhead for other calls. 299; NORMAL-LABEL: test12: 300; NORMAL: movl $8, 12(%esp) 301; NORMAL-NEXT: movl $7, 8(%esp) 302; NORMAL-NEXT: movl $6, 4(%esp) 303; NORMAL-NEXT: movl $5, (%esp) 304; NORMAL-NEXT: calll _good 305define void @test12() optsize { 306entry: 307 %s = alloca %struct.s, align 4 308 call void @struct(%struct.s* %s, i32 2, i32 3, i32 4) 309 call void @good(i32 5, i32 6, i32 7, i32 8) 310 call void @struct(%struct.s* %s, i32 10, i32 11, i32 12) 311 ret void 312} 313 314; But if the gains outweigh the overhead, we should do it 315; NORMAL-LABEL: test12b: 316; NORMAL: pushl $4 317; NORMAL-NEXT: pushl $3 318; NORMAL-NEXT: pushl $2 319; NORMAL-NEXT: pushl $1 320; NORMAL-NEXT: calll _good 321; NORMAL-NEXT: subl $4, %esp 322; NORMAL: movl $8, 16(%esp) 323; NORMAL-NEXT: movl $7, 12(%esp) 324; NORMAL-NEXT: movl $6, 8(%esp) 325; NORMAL-NEXT: calll _struct 326; NORMAL-NEXT: addl $20, %esp 327; NORMAL-NEXT: pushl $12 328; NORMAL-NEXT: pushl $11 329; NORMAL-NEXT: pushl $10 330; NORMAL-NEXT: pushl $9 331; NORMAL-NEXT: calll _good 332; NORMAL-NEXT: addl $16, %esp 333define void @test12b() optsize { 334entry: 335 %s = alloca %struct.s, align 4 336 call void @good(i32 1, i32 2, i32 3, i32 4) 337 call void @struct(%struct.s* %s, i32 6, i32 7, i32 8) 338 call void @good(i32 9, i32 10, i32 11, i32 12) 339 ret void 340} 341 342; Make sure the add does not prevent folding loads into pushes. 343; val1 and val2 will not be folded into pushes since they have 344; an additional use, but val3 should be. 345; NORMAL-LABEL: test13: 346; NORMAL: movl ([[P1:%e..]]), [[V1:%e..]] 347; NORMAL-NEXT: movl ([[P2:%e..]]), [[V2:%e..]] 348; NORMAL-NEXT: , [[ADD:%e..]] 349; NORMAL-NEXT: pushl [[ADD]] 350; NORMAL-NEXT: pushl ([[P3:%e..]]) 351; NORMAL-NEXT: pushl [[V2]] 352; NORMAL-NEXT: pushl [[V1]] 353; NORMAL-NEXT: calll _good 354; NORMAL: movl [[P3]], %eax 355define i32* @test13(i32* inreg %ptr1, i32* inreg %ptr2, i32* inreg %ptr3) optsize { 356entry: 357 %val1 = load i32, i32* %ptr1 358 %val2 = load i32, i32* %ptr2 359 %val3 = load i32, i32* %ptr3 360 %add = add i32 %val1, %val2 361 call void @good(i32 %val1, i32 %val2, i32 %val3, i32 %add) 362 ret i32* %ptr3 363} 364 365; Make sure to fold adjacent stack adjustments. 366; LINUX-LABEL: pr27140: 367; LINUX: subl $12, %esp 368; LINUX: .cfi_def_cfa_offset 16 369; LINUX-NOT: sub 370; LINUX: pushl $4 371; LINUX: .cfi_adjust_cfa_offset 4 372; LINUX: pushl $3 373; LINUX: .cfi_adjust_cfa_offset 4 374; LINUX: pushl $2 375; LINUX: .cfi_adjust_cfa_offset 4 376; LINUX: pushl $1 377; LINUX: .cfi_adjust_cfa_offset 4 378; LINUX: calll good 379; LINUX: addl $28, %esp 380; LINUX: .cfi_adjust_cfa_offset -16 381; LINUX-NOT: add 382; LINUX: retl 383define void @pr27140() optsize { 384entry: 385 tail call void @good(i32 1, i32 2, i32 3, i32 4) 386 ret void 387} 388 389; Check that a stack restore (leal -4(%ebp), %esp) doesn't get merged with a 390; stack adjustment (addl $12, %esp). Just because it's a lea doesn't mean it's 391; simply decreasing the stack pointer. 392; NORMAL-LABEL: test14: 393; NORMAL: calll _B_func 394; NORMAL: leal -4(%ebp), %esp 395; NORMAL-NOT: %esp 396; NORMAL: retl 397%struct.A = type { i32, i32 } 398%struct.B = type { i8 } 399declare x86_thiscallcc %struct.B* @B_ctor(%struct.B* returned, %struct.A* byval) 400declare void @B_func(%struct.B* sret, %struct.B*, i32) 401define void @test14(%struct.A* %a) { 402entry: 403 %ref.tmp = alloca %struct.B, align 1 404 %agg.tmp = alloca i64, align 4 405 %tmpcast = bitcast i64* %agg.tmp to %struct.A* 406 %tmp = alloca %struct.B, align 1 407 %0 = bitcast %struct.A* %a to i64* 408 %1 = load i64, i64* %0, align 4 409 store i64 %1, i64* %agg.tmp, align 4 410 %call = call x86_thiscallcc %struct.B* @B_ctor(%struct.B* %ref.tmp, %struct.A* byval %tmpcast) 411 %2 = getelementptr inbounds %struct.B, %struct.B* %tmp, i32 0, i32 0 412 call void @B_func(%struct.B* sret %tmp, %struct.B* %ref.tmp, i32 1) 413 ret void 414} 415