OLD | NEW |
(Empty) | |
| 1 ; This tests the optimization of atomic cmpxchg w/ following cmp + branches. |
| 2 |
| 3 ; RUN: %llvm2ice -O2 --verbose none %s | FileCheck %s --check-prefix=O2 |
| 4 ; RUN: %llvm2ice -Om1 --verbose none %s | FileCheck %s --check-prefix=OM1 |
| 5 ; RUN: %llvm2ice -O2 --verbose none %s \ |
| 6 ; RUN: | llvm-mc -arch=x86 -x86-asm-syntax=intel -filetype=obj |
| 7 ; RUN: %llvm2ice -Om1 --verbose none %s \ |
| 8 ; RUN: | llvm-mc -arch=x86 -x86-asm-syntax=intel -filetype=obj |
| 9 ; RUN: %llvm2ice --verbose none %s | FileCheck --check-prefix=ERRORS %s |
| 10 ; RUN: %llvm2iceinsts %s | %szdiff %s | FileCheck --check-prefix=DUMP %s |
| 11 ; RUN: %llvm2iceinsts --pnacl %s | %szdiff %s \ |
| 12 ; RUN: | FileCheck --check-prefix=DUMP %s |
| 13 |
| 14 declare i32 @llvm.nacl.atomic.cmpxchg.i32(i32*, i32, i32, i32, i32) |
| 15 |
| 16 |
| 17 ; Test that a cmpxchg followed by icmp eq and branch can be optimized to |
| 18 ; reuse the flags set by the cmpxchg instruction itself. |
| 19 ; This is only expected to work w/ O2, based on lightweight liveness. |
| 20 ; (Or if we had other means to detect the only use). |
| 21 declare void @use_value(i32); |
| 22 |
| 23 define i32 @test_atomic_cmpxchg_loop(i32 %iptr, i32 %expected, i32 %desired) { |
| 24 entry: |
| 25 br label %loop |
| 26 |
| 27 loop: |
| 28 %expected_loop = phi i32 [ %expected, %entry ], [ %old, %loop ] |
| 29 %succeeded_first_try = phi i32 [ 1, %entry ], [ 2, %loop ] |
| 30 %ptr = inttoptr i32 %iptr to i32* |
| 31 %old = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %expected_loop, |
| 32 i32 %desired, i32 6, i32 6) |
| 33 %success = icmp eq i32 %expected_loop, %old |
| 34 br i1 %success, label %done, label %loop |
| 35 |
| 36 done: |
| 37 call void @use_value(i32 %old) |
| 38 ret i32 %succeeded_first_try |
| 39 } |
| 40 ; O2-LABEL: .Ltest_atomic_cmpxchg_loop{{.*}}loop |
| 41 ; O2: lock cmpxchg dword ptr [e{{[^a].}}], e{{[^a]}} |
| 42 ; O2-NOT: cmp |
| 43 ; Make sure the phi assignment for succeeded_first_try is still there. |
| 44 ; O2: mov {{.*}}, 2 |
| 45 ; O2-NOT: cmp |
| 46 ; O2: je |
| 47 ; O2-LABEL: .Ltest_atomic_cmpxchg_loop{{.*}}done |
| 48 ; Make sure the call isn't accidentally deleted. |
| 49 ; O2: call |
| 50 ; |
| 51 ; Check that the unopt version does have a cmp |
| 52 ; OM1-LABEL: .Ltest_atomic_cmpxchg_loop{{.*}}loop |
| 53 ; OM1: lock cmpxchg dword ptr [e{{[^a].}}], e{{[^a]}} |
| 54 ; OM1: cmp |
| 55 ; OM1: je |
| 56 ; OM1-LABEL: .Ltest_atomic_cmpxchg_loop{{.*}}done |
| 57 ; OM1: call |
| 58 |
| 59 ; Still works if the compare operands are flipped. |
| 60 define i32 @test_atomic_cmpxchg_loop2(i32 %iptr, i32 %expected, i32 %desired) { |
| 61 entry: |
| 62 br label %loop |
| 63 |
| 64 loop: |
| 65 %expected_loop = phi i32 [ %expected, %entry ], [ %old, %loop ] |
| 66 %ptr = inttoptr i32 %iptr to i32* |
| 67 %old = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %expected_loop, |
| 68 i32 %desired, i32 6, i32 6) |
| 69 %success = icmp eq i32 %old, %expected_loop |
| 70 br i1 %success, label %done, label %loop |
| 71 |
| 72 done: |
| 73 ret i32 %old |
| 74 } |
| 75 ; O2-LABEL: .Ltest_atomic_cmpxchg_loop2{{.*}}loop |
| 76 ; O2: lock cmpxchg dword ptr [e{{[^a].}}], e{{[^a]}} |
| 77 ; O2-NOT: cmp |
| 78 ; O2: je |
| 79 |
| 80 |
| 81 ; Still works if the compare operands are constants. |
| 82 define i32 @test_atomic_cmpxchg_loop_const(i32 %iptr, i32 %desired) { |
| 83 entry: |
| 84 br label %loop |
| 85 |
| 86 loop: |
| 87 %succeeded_first_try = phi i32 [ 1, %entry ], [ 0, %loop ] |
| 88 %ptr = inttoptr i32 %iptr to i32* |
| 89 %old = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 0, |
| 90 i32 %desired, i32 6, i32 6) |
| 91 %success = icmp eq i32 %old, 0 |
| 92 br i1 %success, label %done, label %loop |
| 93 |
| 94 done: |
| 95 ret i32 %succeeded_first_try |
| 96 } |
| 97 ; O2-LABEL: .Ltest_atomic_cmpxchg_loop_const{{.*}}loop |
| 98 ; O2: lock cmpxchg dword ptr [e{{[^a].}}], e{{[^a]}} |
| 99 ; O2-NOT: cmp |
| 100 ; O2: je |
| 101 |
| 102 ; This is a case where the flags cannot be reused (compare is for some |
| 103 ; other condition). |
| 104 define i32 @test_atomic_cmpxchg_no_opt(i32 %iptr, i32 %expected, i32 %desired) { |
| 105 entry: |
| 106 br label %loop |
| 107 |
| 108 loop: |
| 109 %expected_loop = phi i32 [ %expected, %entry ], [ %old, %loop ] |
| 110 %ptr = inttoptr i32 %iptr to i32* |
| 111 %old = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %expected_loop, |
| 112 i32 %desired, i32 6, i32 6) |
| 113 %success = icmp sgt i32 %old, %expected |
| 114 br i1 %success, label %done, label %loop |
| 115 |
| 116 done: |
| 117 ret i32 %old |
| 118 } |
| 119 ; O2-LABEL: .Ltest_atomic_cmpxchg_no_opt{{.*}}loop |
| 120 ; O2: lock cmpxchg dword ptr [e{{[^a].}}], e{{[^a]}} |
| 121 ; O2: mov {{.*}} |
| 122 ; O2: cmp |
| 123 ; O2: jg |
| 124 |
| 125 ; Another case where the flags cannot be reused (the comparison result |
| 126 ; is used somewhere else). |
| 127 define i32 @test_atomic_cmpxchg_no_opt2(i32 %iptr, i32 %expected, i32 %desired)
{ |
| 128 entry: |
| 129 br label %loop |
| 130 |
| 131 loop: |
| 132 %expected_loop = phi i32 [ %expected, %entry ], [ %old, %loop ] |
| 133 %ptr = inttoptr i32 %iptr to i32* |
| 134 %old = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %expected_loop, |
| 135 i32 %desired, i32 6, i32 6) |
| 136 %success = icmp eq i32 %old, %expected |
| 137 br i1 %success, label %done, label %loop |
| 138 |
| 139 done: |
| 140 %r = zext i1 %success to i32 |
| 141 ret i32 %r |
| 142 } |
| 143 ; O2-LABEL: .Ltest_atomic_cmpxchg_no_opt2{{.*}}loop |
| 144 ; O2: lock cmpxchg dword ptr [e{{[^a].}}], e{{[^a]}} |
| 145 ; O2: mov {{.*}} |
| 146 ; O2: cmp |
| 147 ; O2: je |
| 148 |
| 149 ; ERRORS-NOT: ICE translation error |
| 150 ; DUMP-NOT: SZ |
OLD | NEW |