xref: /llvm-project/llvm/test/Transforms/NewGVN/fence-xfail.ll (revision fa8448c1b18bfe7ffa89f91ab0cfc16f2f83760d)
1; XFAIL: *
2; RUN: opt -S -passes=newgvn < %s | FileCheck %s
3
4@a = external constant i32
5; We can value forward across the fence since we can (semantically)
6; reorder the following load before the fence.
7define i32 @test(ptr %addr.i) {
8; CHECK-LABEL: @test
9; CHECK: store
10; CHECK: fence
11; CHECK-NOT: load
12; CHECK: ret
13  store i32 5, ptr %addr.i, align 4
14  fence release
15  %a = load i32, ptr %addr.i, align 4
16  ret i32 %a
17}
18
19; Same as above
20define i32 @test2(ptr %addr.i) {
21; CHECK-LABEL: @test2
22; CHECK-NEXT: fence
23; CHECK-NOT: load
24; CHECK: ret
25  %a = load i32, ptr %addr.i, align 4
26  fence release
27  %a2 = load i32, ptr %addr.i, align 4
28  %res = sub i32 %a, %a2
29  ret i32 %res
30}
31
32; We can not value forward across an acquire barrier since we might
33; be syncronizing with another thread storing to the same variable
34; followed by a release fence.  This is not so much enforcing an
35; ordering property (though it is that too), but a liveness
36; property.  We expect to eventually see the value of store by
37; another thread when spinning on that location.
38define i32 @test3(ptr noalias %addr.i, ptr noalias %otheraddr) {
39; CHECK-LABEL: @test3
40; CHECK: load
41; CHECK: fence
42; CHECK: load
43; CHECK: ret i32 %res
44  ; the following code is intented to model the unrolling of
45  ; two iterations in a spin loop of the form:
46  ;   do { fence acquire: tmp = *%addr.i; ) while (!tmp);
47  ; It's hopefully clear that allowing PRE to turn this into:
48  ;   if (!*%addr.i) while(true) {} would be unfortunate
49  fence acquire
50  %a = load i32, ptr %addr.i, align 4
51  fence acquire
52  %a2 = load i32, ptr %addr.i, align 4
53  %res = sub i32 %a, %a2
54  ret i32 %res
55}
56
57; We can forward the value forward the load
58; across both the fences, because the load is from
59; a constant memory location.
60define i32 @test4(ptr %addr) {
61; CHECK-LABEL: @test4
62; CHECK-NOT: load
63; CHECK: fence release
64; CHECK: store
65; CHECK: fence seq_cst
66; CHECK: ret i32 0
67  %var = load i32, ptr @a
68  fence release
69  store i32 42, ptr %addr, align 8
70  fence seq_cst
71  %var2 = load i32, ptr @a
72  %var3 = sub i32 %var, %var2
73  ret i32 %var3
74}
75
76; Another example of why forwarding across an acquire fence is problematic
77; can be seen in a normal locking operation.  Say we had:
78; *p = 5; unlock(l); lock(l); use(p);
79; forwarding the store to p would be invalid.  A reasonable implementation
80; of unlock and lock might be:
81; unlock() { atomicrmw sub %l, 1 unordered; fence release }
82; lock() {
83;   do {
84;     %res = cmpxchg %p, 0, 1, monotonic monotonic
85;   } while(!%res.success)
86;   fence acquire;
87; }
88; Given we chose to forward across the release fence, we clearly can't forward
89; across the acquire fence as well.
90
91