1; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=corei7 -disable-fp-elim | FileCheck %s 2 3; Trivial patchpoint codegen 4; 5define i64 @trivial_patchpoint_codegen(i64 %p1, i64 %p2, i64 %p3, i64 %p4) { 6entry: 7; CHECK-LABEL: trivial_patchpoint_codegen: 8; CHECK: movabsq $-559038736, %r11 9; CHECK-NEXT: callq *%r11 10; CHECK-NEXT: xchgw %ax, %ax 11; CHECK: movq %rax, %[[REG:r.+]] 12; CHECK: callq *%r11 13; CHECK-NEXT: xchgw %ax, %ax 14; CHECK: movq %[[REG]], %rax 15; CHECK: ret 16 %resolveCall2 = inttoptr i64 -559038736 to i8* 17 %result = tail call i64 (i64, i32, i8*, i32, ...)* @llvm.experimental.patchpoint.i64(i64 2, i32 15, i8* %resolveCall2, i32 4, i64 %p1, i64 %p2, i64 %p3, i64 %p4) 18 %resolveCall3 = inttoptr i64 -559038737 to i8* 19 tail call void (i64, i32, i8*, i32, ...)* @llvm.experimental.patchpoint.void(i64 3, i32 15, i8* %resolveCall3, i32 2, i64 %p1, i64 %result) 20 ret i64 %result 21} 22 23; Caller frame metadata with stackmaps. This should not be optimized 24; as a leaf function. 25; 26; CHECK-LABEL: caller_meta_leaf 27; CHECK: subq $32, %rsp 28; CHECK: Ltmp 29; CHECK: addq $32, %rsp 30; CHECK: ret 31define void @caller_meta_leaf() { 32entry: 33 %metadata = alloca i64, i32 3, align 8 34 store i64 11, i64* %metadata 35 store i64 12, i64* %metadata 36 store i64 13, i64* %metadata 37 call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 4, i32 0, i64* %metadata) 38 ret void 39} 40 41; Test the webkit_jscc calling convention. 42; One argument will be passed in register, the other will be pushed on the stack. 43; Return value in $rax. 44define void @jscall_patchpoint_codegen(i64 %p1, i64 %p2, i64 %p3, i64 %p4) { 45entry: 46; CHECK-LABEL: jscall_patchpoint_codegen: 47; CHECK: Ltmp 48; CHECK: movq %r{{.+}}, (%rsp) 49; CHECK: movq %r{{.+}}, %rax 50; CHECK: Ltmp 51; CHECK-NEXT: movabsq $-559038736, %r11 52; CHECK-NEXT: callq *%r11 53; CHECK: movq %rax, (%rsp) 54; CHECK: callq 55 %resolveCall2 = inttoptr i64 -559038736 to i8* 56 %result = tail call webkit_jscc i64 (i64, i32, i8*, i32, ...)* @llvm.experimental.patchpoint.i64(i64 5, i32 15, i8* %resolveCall2, i32 2, i64 %p4, i64 %p2) 57 %resolveCall3 = inttoptr i64 -559038737 to i8* 58 tail call webkit_jscc void (i64, i32, i8*, i32, ...)* @llvm.experimental.patchpoint.void(i64 6, i32 15, i8* %resolveCall3, i32 2, i64 %p4, i64 %result) 59 ret void 60} 61 62; Test if the arguments are properly aligned and that we don't store undef arguments. 63define i64 @jscall_patchpoint_codegen2(i64 %callee) { 64entry: 65; CHECK-LABEL: jscall_patchpoint_codegen2: 66; CHECK: Ltmp 67; CHECK: movq $6, 24(%rsp) 68; CHECK-NEXT: movl $4, 16(%rsp) 69; CHECK-NEXT: movq $2, (%rsp) 70; CHECK: Ltmp 71; CHECK-NEXT: movabsq $-559038736, %r11 72; CHECK-NEXT: callq *%r11 73 %call = inttoptr i64 -559038736 to i8* 74 %result = call webkit_jscc i64 (i64, i32, i8*, i32, ...)* @llvm.experimental.patchpoint.i64(i64 7, i32 15, i8* %call, i32 6, i64 %callee, i64 2, i64 undef, i32 4, i32 undef, i64 6) 75 ret i64 %result 76} 77 78; Test if the arguments are properly aligned and that we don't store undef arguments. 79define i64 @jscall_patchpoint_codegen3(i64 %callee) { 80entry: 81; CHECK-LABEL: jscall_patchpoint_codegen3: 82; CHECK: Ltmp 83; CHECK: movq $10, 48(%rsp) 84; CHECK-NEXT: movl $8, 36(%rsp) 85; CHECK-NEXT: movq $6, 24(%rsp) 86; CHECK-NEXT: movl $4, 16(%rsp) 87; CHECK-NEXT: movq $2, (%rsp) 88; CHECK: Ltmp 89; CHECK-NEXT: movabsq $-559038736, %r11 90; CHECK-NEXT: callq *%r11 91 %call = inttoptr i64 -559038736 to i8* 92 %result = call webkit_jscc i64 (i64, i32, i8*, i32, ...)* @llvm.experimental.patchpoint.i64(i64 7, i32 15, i8* %call, i32 10, i64 %callee, i64 2, i64 undef, i32 4, i32 undef, i64 6, i32 undef, i32 8, i32 undef, i64 10) 93 ret i64 %result 94} 95 96; Test patchpoints reusing the same TargetConstant. 97; <rdar:15390785> Assertion failed: (CI.getNumArgOperands() >= NumArgs + 4) 98; There is no way to verify this, since it depends on memory allocation. 99; But I think it's useful to include as a working example. 100define i64 @testLowerConstant(i64 %arg, i64 %tmp2, i64 %tmp10, i64* %tmp33, i64 %tmp79) { 101entry: 102 %tmp80 = add i64 %tmp79, -16 103 %tmp81 = inttoptr i64 %tmp80 to i64* 104 %tmp82 = load i64* %tmp81, align 8 105 tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 14, i32 5, i64 %arg, i64 %tmp2, i64 %tmp10, i64 %tmp82) 106 tail call void (i64, i32, i8*, i32, ...)* @llvm.experimental.patchpoint.void(i64 15, i32 30, i8* null, i32 3, i64 %arg, i64 %tmp10, i64 %tmp82) 107 %tmp83 = load i64* %tmp33, align 8 108 %tmp84 = add i64 %tmp83, -24 109 %tmp85 = inttoptr i64 %tmp84 to i64* 110 %tmp86 = load i64* %tmp85, align 8 111 tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 17, i32 5, i64 %arg, i64 %tmp10, i64 %tmp86) 112 tail call void (i64, i32, i8*, i32, ...)* @llvm.experimental.patchpoint.void(i64 18, i32 30, i8* null, i32 3, i64 %arg, i64 %tmp10, i64 %tmp86) 113 ret i64 10 114} 115 116; Test small patchpoints that don't emit calls. 117define void @small_patchpoint_codegen(i64 %p1, i64 %p2, i64 %p3, i64 %p4) { 118entry: 119; CHECK-LABEL: small_patchpoint_codegen: 120; CHECK: Ltmp 121; CHECK: nopl 8(%rax,%rax) 122; CHECK-NEXT: popq 123; CHECK-NEXT: ret 124 %result = tail call i64 (i64, i32, i8*, i32, ...)* @llvm.experimental.patchpoint.i64(i64 5, i32 5, i8* null, i32 2, i64 %p1, i64 %p2) 125 ret void 126} 127 128declare void @llvm.experimental.stackmap(i64, i32, ...) 129declare void @llvm.experimental.patchpoint.void(i64, i32, i8*, i32, ...) 130declare i64 @llvm.experimental.patchpoint.i64(i64, i32, i8*, i32, ...) 131