OLD | NEW |
1 ; This tests the optimization of atomic cmpxchg w/ following cmp + branches. | 1 ; This tests the optimization of atomic cmpxchg w/ following cmp + branches. |
2 | 2 |
3 ; RUN: %p2i -i %s --filetype=obj --disassemble --args -O2 \ | 3 ; RUN: %p2i -i %s --filetype=obj --disassemble --args -O2 \ |
4 ; RUN: | FileCheck --check-prefix=O2 %s | 4 ; RUN: -allow-externally-defined-symbols | FileCheck --check-prefix=O2 %s |
5 ; RUN: %p2i -i %s --filetype=obj --disassemble --args -Om1 \ | 5 ; RUN: %p2i -i %s --filetype=obj --disassemble --args -Om1 \ |
6 ; RUN: | FileCheck --check-prefix=OM1 %s | 6 ; RUN: -allow-externally-defined-symbols | FileCheck --check-prefix=OM1 %s |
7 | 7 |
8 declare i32 @llvm.nacl.atomic.cmpxchg.i32(i32*, i32, i32, i32, i32) | 8 declare i32 @llvm.nacl.atomic.cmpxchg.i32(i32*, i32, i32, i32, i32) |
9 | 9 |
10 | 10 |
11 ; Test that a cmpxchg followed by icmp eq and branch can be optimized to | 11 ; Test that a cmpxchg followed by icmp eq and branch can be optimized to |
12 ; reuse the flags set by the cmpxchg instruction itself. | 12 ; reuse the flags set by the cmpxchg instruction itself. |
13 ; This is only expected to work w/ O2, based on lightweight liveness. | 13 ; This is only expected to work w/ O2, based on lightweight liveness. |
14 ; (Or if we had other means to detect the only use). | 14 ; (Or if we had other means to detect the only use). |
15 declare void @use_value(i32) | 15 declare void @use_value(i32) |
16 | 16 |
17 define i32 @test_atomic_cmpxchg_loop(i32 %iptr, i32 %expected, i32 %desired) { | 17 define internal i32 @test_atomic_cmpxchg_loop(i32 %iptr, i32 %expected, |
| 18 i32 %desired) { |
18 entry: | 19 entry: |
19 br label %loop | 20 br label %loop |
20 | 21 |
21 loop: | 22 loop: |
22 %expected_loop = phi i32 [ %expected, %entry ], [ %old, %loop ] | 23 %expected_loop = phi i32 [ %expected, %entry ], [ %old, %loop ] |
23 %succeeded_first_try = phi i32 [ 1, %entry ], [ 2, %loop ] | 24 %succeeded_first_try = phi i32 [ 1, %entry ], [ 2, %loop ] |
24 %ptr = inttoptr i32 %iptr to i32* | 25 %ptr = inttoptr i32 %iptr to i32* |
25 %old = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %expected_loop, | 26 %old = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %expected_loop, |
26 i32 %desired, i32 6, i32 6) | 27 i32 %desired, i32 6, i32 6) |
27 %success = icmp eq i32 %expected_loop, %old | 28 %success = icmp eq i32 %expected_loop, %old |
(...skipping 10 matching lines...) Expand all Loading... |
38 ; O2: call | 39 ; O2: call |
39 ; | 40 ; |
40 ; Check that the unopt version does have a cmp | 41 ; Check that the unopt version does have a cmp |
41 ; OM1-LABEL: test_atomic_cmpxchg_loop | 42 ; OM1-LABEL: test_atomic_cmpxchg_loop |
42 ; OM1: lock cmpxchg DWORD PTR [e{{[^a].}}],e{{[^a]}} | 43 ; OM1: lock cmpxchg DWORD PTR [e{{[^a].}}],e{{[^a]}} |
43 ; OM1: cmp | 44 ; OM1: cmp |
44 ; OM1: sete | 45 ; OM1: sete |
45 ; OM1: call | 46 ; OM1: call |
46 | 47 |
47 ; Still works if the compare operands are flipped. | 48 ; Still works if the compare operands are flipped. |
48 define i32 @test_atomic_cmpxchg_loop2(i32 %iptr, i32 %expected, i32 %desired) { | 49 define internal i32 @test_atomic_cmpxchg_loop2(i32 %iptr, i32 %expected, |
| 50 i32 %desired) { |
49 entry: | 51 entry: |
50 br label %loop | 52 br label %loop |
51 | 53 |
52 loop: | 54 loop: |
53 %expected_loop = phi i32 [ %expected, %entry ], [ %old, %loop ] | 55 %expected_loop = phi i32 [ %expected, %entry ], [ %old, %loop ] |
54 %ptr = inttoptr i32 %iptr to i32* | 56 %ptr = inttoptr i32 %iptr to i32* |
55 %old = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %expected_loop, | 57 %old = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %expected_loop, |
56 i32 %desired, i32 6, i32 6) | 58 i32 %desired, i32 6, i32 6) |
57 %success = icmp eq i32 %old, %expected_loop | 59 %success = icmp eq i32 %old, %expected_loop |
58 br i1 %success, label %done, label %loop | 60 br i1 %success, label %done, label %loop |
59 | 61 |
60 done: | 62 done: |
61 ret i32 %old | 63 ret i32 %old |
62 } | 64 } |
63 ; O2-LABEL: test_atomic_cmpxchg_loop2 | 65 ; O2-LABEL: test_atomic_cmpxchg_loop2 |
64 ; O2: lock cmpxchg DWORD PTR [e{{[^a].}}],e{{[^a]}} | 66 ; O2: lock cmpxchg DWORD PTR [e{{[^a].}}],e{{[^a]}} |
65 ; O2-NOT: cmp | 67 ; O2-NOT: cmp |
66 ; O2: jne | 68 ; O2: jne |
67 | 69 |
68 | 70 |
69 ; Still works if the compare operands are constants. | 71 ; Still works if the compare operands are constants. |
70 define i32 @test_atomic_cmpxchg_loop_const(i32 %iptr, i32 %desired) { | 72 define internal i32 @test_atomic_cmpxchg_loop_const(i32 %iptr, i32 %desired) { |
71 entry: | 73 entry: |
72 br label %loop | 74 br label %loop |
73 | 75 |
74 loop: | 76 loop: |
75 %succeeded_first_try = phi i32 [ 1, %entry ], [ 0, %loop ] | 77 %succeeded_first_try = phi i32 [ 1, %entry ], [ 0, %loop ] |
76 %ptr = inttoptr i32 %iptr to i32* | 78 %ptr = inttoptr i32 %iptr to i32* |
77 %old = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 0, | 79 %old = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 0, |
78 i32 %desired, i32 6, i32 6) | 80 i32 %desired, i32 6, i32 6) |
79 %success = icmp eq i32 %old, 0 | 81 %success = icmp eq i32 %old, 0 |
80 br i1 %success, label %done, label %loop | 82 br i1 %success, label %done, label %loop |
81 | 83 |
82 done: | 84 done: |
83 ret i32 %succeeded_first_try | 85 ret i32 %succeeded_first_try |
84 } | 86 } |
85 ; O2-LABEL: test_atomic_cmpxchg_loop_const | 87 ; O2-LABEL: test_atomic_cmpxchg_loop_const |
86 ; O2: lock cmpxchg DWORD PTR [e{{[^a].}}],e{{[^a]}} | 88 ; O2: lock cmpxchg DWORD PTR [e{{[^a].}}],e{{[^a]}} |
87 ; O2-NEXT: j{{e|ne}} | 89 ; O2-NEXT: j{{e|ne}} |
88 | 90 |
89 ; This is a case where the flags cannot be reused (compare is for some | 91 ; This is a case where the flags cannot be reused (compare is for some |
90 ; other condition). | 92 ; other condition). |
91 define i32 @test_atomic_cmpxchg_no_opt(i32 %iptr, i32 %expected, i32 %desired) { | 93 define internal i32 @test_atomic_cmpxchg_no_opt(i32 %iptr, i32 %expected, |
| 94 i32 %desired) { |
92 entry: | 95 entry: |
93 br label %loop | 96 br label %loop |
94 | 97 |
95 loop: | 98 loop: |
96 %expected_loop = phi i32 [ %expected, %entry ], [ %old, %loop ] | 99 %expected_loop = phi i32 [ %expected, %entry ], [ %old, %loop ] |
97 %ptr = inttoptr i32 %iptr to i32* | 100 %ptr = inttoptr i32 %iptr to i32* |
98 %old = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %expected_loop, | 101 %old = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %expected_loop, |
99 i32 %desired, i32 6, i32 6) | 102 i32 %desired, i32 6, i32 6) |
100 %success = icmp sgt i32 %old, %expected | 103 %success = icmp sgt i32 %old, %expected |
101 br i1 %success, label %done, label %loop | 104 br i1 %success, label %done, label %loop |
102 | 105 |
103 done: | 106 done: |
104 ret i32 %old | 107 ret i32 %old |
105 } | 108 } |
106 ; O2-LABEL: test_atomic_cmpxchg_no_opt | 109 ; O2-LABEL: test_atomic_cmpxchg_no_opt |
107 ; O2: lock cmpxchg DWORD PTR [e{{[^a].}}],e{{[^a]}} | 110 ; O2: lock cmpxchg DWORD PTR [e{{[^a].}}],e{{[^a]}} |
108 ; O2: cmp | 111 ; O2: cmp |
109 ; O2: jle | 112 ; O2: jle |
110 | 113 |
111 ; Another case where the flags cannot be reused (the comparison result | 114 ; Another case where the flags cannot be reused (the comparison result |
112 ; is used somewhere else). | 115 ; is used somewhere else). |
113 define i32 @test_atomic_cmpxchg_no_opt2(i32 %iptr, i32 %expected, i32 %desired)
{ | 116 define internal i32 @test_atomic_cmpxchg_no_opt2(i32 %iptr, i32 %expected, |
| 117 i32 %desired) { |
114 entry: | 118 entry: |
115 br label %loop | 119 br label %loop |
116 | 120 |
117 loop: | 121 loop: |
118 %expected_loop = phi i32 [ %expected, %entry ], [ %old, %loop ] | 122 %expected_loop = phi i32 [ %expected, %entry ], [ %old, %loop ] |
119 %ptr = inttoptr i32 %iptr to i32* | 123 %ptr = inttoptr i32 %iptr to i32* |
120 %old = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %expected_loop, | 124 %old = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %expected_loop, |
121 i32 %desired, i32 6, i32 6) | 125 i32 %desired, i32 6, i32 6) |
122 %success = icmp eq i32 %old, %expected | 126 %success = icmp eq i32 %old, %expected |
123 br i1 %success, label %done, label %loop | 127 br i1 %success, label %done, label %loop |
124 | 128 |
125 done: | 129 done: |
126 %r = zext i1 %success to i32 | 130 %r = zext i1 %success to i32 |
127 ret i32 %r | 131 ret i32 %r |
128 } | 132 } |
129 ; O2-LABEL: test_atomic_cmpxchg_no_opt2 | 133 ; O2-LABEL: test_atomic_cmpxchg_no_opt2 |
130 ; O2: lock cmpxchg DWORD PTR [e{{[^a].}}],e{{[^a]}} | 134 ; O2: lock cmpxchg DWORD PTR [e{{[^a].}}],e{{[^a]}} |
131 ; O2: cmp | 135 ; O2: cmp |
132 ; O2: sete | 136 ; O2: sete |
OLD | NEW |