268 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -aarch64-load-store-renaming=true -aarch64-enable-sink-fold=true < %s -mtriple=arm64-apple-ios7.0.0 -mcpu=cyclone -enable-misched=false | FileCheck %s3 4; rdar://136255055; Here we have 9 fixed integer arguments the 9th argument in on stack, the6; varargs start right after at 8-byte alignment.7define void @fn9(ptr %a1, i32 %a2, i32 %a3, i32 %a4, i32 %a5, i32 %a6, i32 %a7, i32 %a8, i32 %a9, ...) nounwind noinline ssp {8; CHECK-LABEL: fn9:9; CHECK: ; %bb.0:10; CHECK-NEXT: sub sp, sp, #6411; CHECK-NEXT: ldr w8, [sp, #64]12; CHECK-NEXT: stp w2, w1, [sp, #52]13; CHECK-NEXT: stp w4, w3, [sp, #44]14; CHECK-NEXT: stp w6, w5, [sp, #36]15; CHECK-NEXT: str w7, [sp, #32]16; CHECK-NEXT: str w8, [x0]17; CHECK-NEXT: add x8, sp, #7218; CHECK-NEXT: add x8, x8, #819; CHECK-NEXT: ldr w9, [sp, #72]20; CHECK-NEXT: str w9, [sp, #20]21; CHECK-NEXT: ldr w9, [x8], #822; CHECK-NEXT: str w9, [sp, #16]23; CHECK-NEXT: ldr w9, [x8], #824; CHECK-NEXT: str x8, [sp, #24]25; CHECK-NEXT: str w9, [sp, #12]26; CHECK-NEXT: add sp, sp, #6427; CHECK-NEXT: ret28 %1 = alloca i32, align 429 %2 = alloca i32, align 430 %3 = alloca i32, align 431 %4 = alloca i32, align 432 %5 = alloca i32, align 433 %6 = alloca i32, align 434 %7 = alloca i32, align 435 %8 = alloca i32, align 436 %9 = alloca i32, align 437 %args = alloca ptr, align 838 %a10 = alloca i32, align 439 %a11 = alloca i32, align 440 %a12 = alloca i32, align 441 store i32 %a2, ptr %2, align 442 store i32 %a3, ptr %3, align 443 store i32 %a4, ptr %4, align 444 store i32 %a5, ptr %5, align 445 store i32 %a6, ptr %6, align 446 store i32 %a7, ptr %7, align 447 store i32 %a8, ptr %8, align 448 store i32 %a9, ptr %9, align 449 store i32 %a9, ptr %a150 call void @llvm.va_start(ptr %args)51 %10 = va_arg ptr %args, i3252 store i32 %10, ptr %a10, align 453 %11 = va_arg ptr %args, i3254 store i32 %11, ptr %a11, align 455 %12 = va_arg ptr %args, i3256 store i32 %12, ptr %a12, align 457 ret void58}59 60declare void @llvm.va_start(ptr) nounwind61 62define i32 @main() nounwind ssp {63; CHECK-LABEL: main:64; CHECK: ; %bb.0:65; CHECK-NEXT: sub sp, sp, #9666; CHECK-NEXT: stp x29, x30, [sp, #80] ; 16-byte Folded Spill67; CHECK-NEXT: mov w9, #1 ; =0x168; CHECK-NEXT: mov w8, #2 ; =0x269; CHECK-NEXT: stp w8, w9, [sp, #72]70; CHECK-NEXT: mov w9, #3 ; =0x371; CHECK-NEXT: mov w8, #4 ; =0x472; CHECK-NEXT: stp w8, w9, [sp, #64]73; CHECK-NEXT: mov w9, #5 ; =0x574; CHECK-NEXT: mov w8, #6 ; =0x675; CHECK-NEXT: stp w8, w9, [sp, #56]76; CHECK-NEXT: mov w9, #7 ; =0x777; CHECK-NEXT: mov w8, #8 ; =0x878; CHECK-NEXT: stp w8, w9, [sp, #48]79; CHECK-NEXT: mov w8, #9 ; =0x980; CHECK-NEXT: mov w9, #10 ; =0xa81; CHECK-NEXT: stp w9, w8, [sp, #40]82; CHECK-NEXT: mov w10, #11 ; =0xb83; CHECK-NEXT: mov w11, #12 ; =0xc84; CHECK-NEXT: stp w11, w10, [sp, #32]85; CHECK-NEXT: stp x10, x11, [sp, #16]86; CHECK-NEXT: str x9, [sp, #8]87; CHECK-NEXT: str w8, [sp]88; CHECK-NEXT: add x0, sp, #7689; CHECK-NEXT: mov w1, #2 ; =0x290; CHECK-NEXT: mov w2, #3 ; =0x391; CHECK-NEXT: mov w3, #4 ; =0x492; CHECK-NEXT: mov w4, #5 ; =0x593; CHECK-NEXT: mov w5, #6 ; =0x694; CHECK-NEXT: mov w6, #7 ; =0x795; CHECK-NEXT: mov w7, #8 ; =0x896; CHECK-NEXT: bl _fn997; CHECK-NEXT: mov w0, #0 ; =0x098; CHECK-NEXT: ldp x29, x30, [sp, #80] ; 16-byte Folded Reload99; CHECK-NEXT: add sp, sp, #96100; CHECK-NEXT: ret101 %a1 = alloca i32, align 4102 %a2 = alloca i32, align 4103 %a3 = alloca i32, align 4104 %a4 = alloca i32, align 4105 %a5 = alloca i32, align 4106 %a6 = alloca i32, align 4107 %a7 = alloca i32, align 4108 %a8 = alloca i32, align 4109 %a9 = alloca i32, align 4110 %a10 = alloca i32, align 4111 %a11 = alloca i32, align 4112 %a12 = alloca i32, align 4113 store i32 1, ptr %a1, align 4114 store i32 2, ptr %a2, align 4115 store i32 3, ptr %a3, align 4116 store i32 4, ptr %a4, align 4117 store i32 5, ptr %a5, align 4118 store i32 6, ptr %a6, align 4119 store i32 7, ptr %a7, align 4120 store i32 8, ptr %a8, align 4121 store i32 9, ptr %a9, align 4122 store i32 10, ptr %a10, align 4123 store i32 11, ptr %a11, align 4124 store i32 12, ptr %a12, align 4125 %1 = load i32, ptr %a1, align 4126 %2 = load i32, ptr %a2, align 4127 %3 = load i32, ptr %a3, align 4128 %4 = load i32, ptr %a4, align 4129 %5 = load i32, ptr %a5, align 4130 %6 = load i32, ptr %a6, align 4131 %7 = load i32, ptr %a7, align 4132 %8 = load i32, ptr %a8, align 4133 %9 = load i32, ptr %a9, align 4134 %10 = load i32, ptr %a10, align 4135 %11 = load i32, ptr %a11, align 4136 %12 = load i32, ptr %a12, align 4137 call void (ptr, i32, i32, i32, i32, i32, i32, i32, i32, ...) @fn9(ptr %a1, i32 %2, i32 %3, i32 %4, i32 %5, i32 %6, i32 %7, i32 %8, i32 %9, i32 %10, i32 %11, i32 %12)138 ret i32 0139}140 141;rdar://13668483142@.str = private unnamed_addr constant [4 x i8] c"fmt\00", align 1143define void @foo(ptr %fmt, ...) nounwind {144; CHECK-LABEL: foo:145; CHECK: ; %bb.0: ; %entry146; CHECK-NEXT: sub sp, sp, #48147; CHECK-NEXT: ldr w8, [sp, #48]148; CHECK-NEXT: str w8, [sp, #28]149; CHECK-NEXT: add x8, sp, #48150; CHECK-NEXT: add x8, x8, #23151; CHECK-NEXT: and x8, x8, #0xfffffffffffffff0152; CHECK-NEXT: add x9, x8, #16153; CHECK-NEXT: stp x9, x0, [sp, #32]154; CHECK-NEXT: ldr q0, [x8]155; CHECK-NEXT: str q0, [sp], #48156; CHECK-NEXT: ret157entry:158 %fmt.addr = alloca ptr, align 8159 %args = alloca ptr, align 8160 %vc = alloca i32, align 4161 %vv = alloca <4 x i32>, align 16162 store ptr %fmt, ptr %fmt.addr, align 8163 call void @llvm.va_start(ptr %args)164 %0 = va_arg ptr %args, i32165 store i32 %0, ptr %vc, align 4166 %1 = va_arg ptr %args, <4 x i32>167 store <4 x i32> %1, ptr %vv, align 16168 ret void169}170 171define void @bar(i32 %x, <4 x i32> %y) nounwind {172; CHECK-LABEL: bar:173; CHECK: ; %bb.0: ; %entry174; CHECK-NEXT: sub sp, sp, #80175; CHECK-NEXT: stp x29, x30, [sp, #64] ; 16-byte Folded Spill176; CHECK-NEXT: ; kill: def $w0 killed $w0 def $x0177; CHECK-NEXT: str w0, [sp, #60]178; CHECK-NEXT: stp q0, q0, [sp, #16]179; CHECK-NEXT: str x0, [sp]180; CHECK-NEXT: Lloh0:181; CHECK-NEXT: adrp x0, l_.str@PAGE182; CHECK-NEXT: Lloh1:183; CHECK-NEXT: add x0, x0, l_.str@PAGEOFF184; CHECK-NEXT: bl _foo185; CHECK-NEXT: ldp x29, x30, [sp, #64] ; 16-byte Folded Reload186; CHECK-NEXT: add sp, sp, #80187; CHECK-NEXT: ret188; CHECK-NEXT: .loh AdrpAdd Lloh0, Lloh1189entry:190 %x.addr = alloca i32, align 4191 %y.addr = alloca <4 x i32>, align 16192 store i32 %x, ptr %x.addr, align 4193 store <4 x i32> %y, ptr %y.addr, align 16194 %0 = load i32, ptr %x.addr, align 4195 %1 = load <4 x i32>, ptr %y.addr, align 16196 call void (ptr, ...) @foo(ptr @.str, i32 %0, <4 x i32> %1)197 ret void198}199 200; rdar://13668927201; When passing 16-byte aligned small structs as vararg, make sure the caller202; side is 16-byte aligned on stack.203%struct.s41 = type { i32, i16, i32, i16 }204define void @foo2(ptr %fmt, ...) nounwind {205; CHECK-LABEL: foo2:206; CHECK: ; %bb.0: ; %entry207; CHECK-NEXT: sub sp, sp, #48208; CHECK-NEXT: ldr w8, [sp, #48]209; CHECK-NEXT: str w8, [sp, #28]210; CHECK-NEXT: add x8, sp, #48211; CHECK-NEXT: add x8, x8, #23212; CHECK-NEXT: and x8, x8, #0xfffffffffffffff0213; CHECK-NEXT: add x9, x8, #16214; CHECK-NEXT: stp x9, x0, [sp, #32]215; CHECK-NEXT: ldr q0, [x8]216; CHECK-NEXT: str q0, [sp], #48217; CHECK-NEXT: ret218entry:219 %fmt.addr = alloca ptr, align 8220 %args = alloca ptr, align 8221 %vc = alloca i32, align 4222 %vs = alloca %struct.s41, align 16223 store ptr %fmt, ptr %fmt.addr, align 8224 call void @llvm.va_start(ptr %args)225 %0 = va_arg ptr %args, i32226 store i32 %0, ptr %vc, align 4227 %ap.cur = load ptr, ptr %args228 %1 = getelementptr i8, ptr %ap.cur, i32 15229 %2 = ptrtoint ptr %1 to i64230 %3 = and i64 %2, -16231 %ap.align = inttoptr i64 %3 to ptr232 %ap.next = getelementptr i8, ptr %ap.align, i32 16233 store ptr %ap.next, ptr %args234 call void @llvm.memcpy.p0.p0.i64(ptr align 16 %vs, ptr align 16 %ap.align, i64 16, i1 false)235 ret void236}237declare void @llvm.memcpy.p0.p0.i64(ptr nocapture, ptr nocapture, i64, i1) nounwind238 239define void @bar2(i32 %x, i128 %s41.coerce) nounwind {240; CHECK-LABEL: bar2:241; CHECK: ; %bb.0: ; %entry242; CHECK-NEXT: sub sp, sp, #80243; CHECK-NEXT: stp x29, x30, [sp, #64] ; 16-byte Folded Spill244; CHECK-NEXT: ; kill: def $w0 killed $w0 def $x0245; CHECK-NEXT: str w0, [sp, #60]246; CHECK-NEXT: stp x1, x2, [sp, #32]247; CHECK-NEXT: stp x1, x2, [sp, #16]248; CHECK-NEXT: str x0, [sp]249; CHECK-NEXT: Lloh2:250; CHECK-NEXT: adrp x0, l_.str@PAGE251; CHECK-NEXT: Lloh3:252; CHECK-NEXT: add x0, x0, l_.str@PAGEOFF253; CHECK-NEXT: bl _foo2254; CHECK-NEXT: ldp x29, x30, [sp, #64] ; 16-byte Folded Reload255; CHECK-NEXT: add sp, sp, #80256; CHECK-NEXT: ret257; CHECK-NEXT: .loh AdrpAdd Lloh2, Lloh3258entry:259 %x.addr = alloca i32, align 4260 %s41 = alloca %struct.s41, align 16261 store i32 %x, ptr %x.addr, align 4262 store i128 %s41.coerce, ptr %s41, align 1263 %0 = load i32, ptr %x.addr, align 4264 %1 = load i128, ptr %s41, align 1265 call void (ptr, ...) @foo2(ptr @.str, i32 %0, i128 %1)266 ret void267}268