Endilll wrote:
Following the suggestion from @nikic, I prepared two variants of the same test,
one hard-written, and one using `update_cc_test_checks.py` (with heavy manual
editing afterwards). Hand-written:
```
// CHECK-LABEL: define dso_local void @dr1807::f()
// CHECK: invoke void @dr1807::S::S(){{.+}}
// CHECK-NEXT: {{.+}} unwind label %lpad
// CHECK-LABEL: lpad:
// CHECK: br {{.+}}, label {{.+}}, label %arraydestroy.body
// CHECK-LABEL: arraydestroy.body:
// CHECK: [[ARRAYDESTROY_ELEMENT:%.*]] = getelementptr {{.+}}, i64
-1
// CXX98-NEXT: invoke void
@dr1807::S::~S(){{.*}}[[ARRAYDESTROY_ELEMENT]]
// SINCE-CXX11-NEXT: call void @dr1807::S::~S(){{.*}}[[ARRAYDESTROY_ELEMENT]]
```
Script:
```
// CHECK-LABEL: define dso_local void @dr1807::f()(
// CHECK: invoke void @dr1807::S::S()(ptr noundef nonnull align 1
dereferenceable(1) [[ARRAYCTOR_CUR:%.*]])
// CHECK-NEXT: to label [[INVOKE_CONT:%.*]] unwind label
[[LPAD:%.*]]
// CHECK-LABEL: lpad:
// CHECK: br i1 [[ARRAYDESTROY_ISEMPTY:%.*]], label
[[ARRAYDESTROY_DONE2:%.*]], label [[ARRAYDESTROY_BODY:%.*]]
// CHECK-LABEL: arraydestroy.body:
// CHECK: [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds
%"struct.dr1807::S", ptr [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
// CXX98-NEXT: invoke void @dr1807::S::~S()(ptr noundef nonnull align 1
dereferenceable(1) [[ARRAYDESTROY_ELEMENT]])
// CXX98-NEXT: to label [[INVOKE_CONT1]] unwind label
[[TERMINATE_LPAD:%.*]]
// SINCE-CXX11-NEXT: call void @dr1807::S::~S()(ptr noundef nonnull align 1
dereferenceable(1) [[ARRAYDESTROY_ELEMENT5:.*]]) #[[ATTR2:.*]]
```
In my opinion, script captures way too much, diluting the point of the test.
For example, `i64 -1` is a crucial part of the test, but it's much easier to
miss it in lengthy script output. (Searching for it might be a good exercise
for the reader.) I'd also like to note that target audience of DR tests is
likely to be far less experienced in reading LLVM IR than regular users of the
script, so it's much less obvious for them to know which parts of the output to
pay attention to. On the other hand wildcards help hide uninteresting parts of
the codegen output, highlighting what's important.
This brings me to conclusion that hand-written style of codegen tests presented
above is a better fit for purposes of C++ defect report codegen tests.
CC @AaronBallman
https://github.com/llvm/llvm-project/pull/77637
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
