Enable querying constants and value kinds #936

tbennun · 2023-04-19T01:42:42Z

This PR adds bindings for obtaining whether a value is constant, what its ValueKind is, and its exact constant value in certain cases (integer, floating point; in other cases the IR string is returned instead).

This is useful for inspecting constant values in Python code.

tbennun · 2023-04-19T01:44:28Z

This PR also fixes #727

apmasell

Thanks! This is a useful feature. Please add unit tests.

apmasell · 2023-04-24T14:28:49Z

ffi/value.cpp

@@ -309,6 +309,34 @@ LLVMPY_DisposeOperandsIter(LLVMOperandsIteratorRef GI) {
    delete llvm::unwrap(GI);
 }

+API_EXPORT(LLVMBool)


Please use bool rather than LLVMBool since that's that's the type preferred by the Python API, even though they should be the same.

Fixed, thank you!

tbennun · 2023-04-26T01:53:23Z

@apmasell Removed LLVMBool from Python-facing API, added tests, and fixed formatting.

apmasell

Looks good. Thank you!

sklam

I have some additional concerns mainly the lack of docs for the new code, and the lack of test coverage for some cases.

llvmlite/binding/value.py

sklam · 2023-04-26T16:15:07Z

llvmlite/binding/value.py

+            value = ffi.lib.LLVMPY_GetConstantFPValue(self,
+                                                      byref(accuracy_loss))
+            if accuracy_loss.value:
+                warnings.warn(


Needs to add test for the warning

@sklam For some reason, despite how much I try, I can't seem to trigger the losesInfo boolean with a single constant (maybe this appears during passes). Not sure how to proceed here, any suggestions?

Sorry for the delay.

After some digging into the LLVM source, the floating-point constants are encoded as double-precision so there will not be any information loss here. To force info loss, we will need to use bigger types like the long-double. For example: %const = fadd fp128 0xLF3CB1CCF26FBC178452FB4EC7F91DEAD, 0xLF3CB1CCF26FBC178452FB4EC7F91973F. This will require converting the 128-bit float to 64-bit float thus triggering the warning.

However, this brings to my attention that a warning is not appropriate here. This API is not giving user a choice to read float bigger than double-precision. In the LLVM OCAML binding, the similar function will return None (https://github.com/llvm/llvm-project/blob/1defa781243f9d0bc66719465e4de33e9fb7a243/llvm/bindings/ocaml/llvm/llvm_ocaml.c#L1030-L1032). I think the Python API should raise an exception.

Also, maybe it is better to have separate functions for int constant and float constant, so that the float version can provide more control on what to do if information is lost. For instance, if user is okay with the info loss, they can force the downcast.

That's great, I agree and already modified the constant float test. Only thing I'm not sure about is the separate functions for int and float constants. One of the main benefits of Python is that you can return multiple types. Maybe this should be another keyword argument, and keyword args should act like preferences?

I am already extending the behavior to support other types of constants (e.g. structs), and it was beneficial to be able to call the same method on the nested data: tbennun@3992ea1#diff-27ab06b9fa73e854c37e5ccf261269588cd8e02fadbd71c788102bc46c4a2f1bR390

Maybe this should be another keyword argument, and keyword args should act like preferences?

Yes, that works too

I am already extending the behavior to support other types of constants (e.g. structs), and it was beneficial to be able to call the same method on the nested data: tbennun@3992ea1#diff-27ab06b9fa73e854c37e5ccf261269588cd8e02fadbd71c788102bc46c4a2f1bR390

👍

@sklam all changes are complete and the PR is ready for re(re)review.

llvmlite/binding/value.py

tbennun · 2023-04-27T20:59:19Z

@sklam Added documentation and more tests (apart from testing the warning, which I couldn't reproduce but is part of the LLVM interface, see above comment). I also noticed that adding one argument to get_constant_value significantly improves its usage for signed integers, which was apparent in one of the tests.

tbennun · 2023-06-08T18:48:47Z

Any updates on this @sklam ? This is not waiting for me as of two months ago.

esc · 2023-07-05T11:14:41Z

@tbennun sorry, seems like we have dropped the ball on this one. I am changing the label now and will ping @sklam again. It is now scheduled for inclusion in the 0.41 release to be tagged during the week starting 7th of August.

sklam

Thanks for adding to the docs. Those looks great.

However, I have followed the rabbit hole and discovered more details to be considered for get_constant_value and the info loss issue. See below.

sklam · 2023-07-18T22:31:19Z

llvmlite/binding/value.py

+            value = ffi.lib.LLVMPY_GetConstantFPValue(self,
+                                                      byref(accuracy_loss))
+            if accuracy_loss.value:
+                warnings.warn(


Sorry for the delay.

After some digging into the LLVM source, the floating-point constants are encoded as double-precision so there will not be any information loss here. To force info loss, we will need to use bigger types like the long-double. For example: %const = fadd fp128 0xLF3CB1CCF26FBC178452FB4EC7F91DEAD, 0xLF3CB1CCF26FBC178452FB4EC7F91973F. This will require converting the 128-bit float to 64-bit float thus triggering the warning.

However, this brings to my attention that a warning is not appropriate here. This API is not giving user a choice to read float bigger than double-precision. In the LLVM OCAML binding, the similar function will return None (https://github.com/llvm/llvm-project/blob/1defa781243f9d0bc66719465e4de33e9fb7a243/llvm/bindings/ocaml/llvm/llvm_ocaml.c#L1030-L1032). I think the Python API should raise an exception.

Also, maybe it is better to have separate functions for int constant and float constant, so that the float version can provide more control on what to do if information is lost. For instance, if user is okay with the info loss, they can force the downcast.

Enable querying constants and value kinds

fd9c175

apmasell added the 3 - Ready for Review label Apr 19, 2023

apmasell reviewed Apr 24, 2023

View reviewed changes

apmasell self-assigned this Apr 24, 2023

apmasell added 4 - Waiting on author and removed 3 - Ready for Review labels Apr 24, 2023

tbennun added 4 commits April 25, 2023 16:27

Use preferred Python API types

0e8f535

Add tests

2c0d168

clang-format

1bc9047

Python formatting

bf3adad

tbennun requested a review from apmasell April 26, 2023 01:51

apmasell approved these changes Apr 26, 2023

View reviewed changes

apmasell added 5 - Ready to merge and removed 4 - Waiting on author labels Apr 26, 2023

sklam requested changes Apr 26, 2023

View reviewed changes

sklam added 4 - Waiting on author and removed 5 - Ready to merge labels Apr 26, 2023

tbennun added 3 commits April 27, 2023 10:00

Add signed integer option for constants

278178a

Add tests

70e4501

Documentation

8224437

tbennun requested a review from sklam April 27, 2023 22:02

esc added this to the v0.41.0 milestone Jun 27, 2023

esc added 4 - Waiting on reviewer and removed 4 - Waiting on author labels Jul 5, 2023

esc assigned sklam Jul 5, 2023

esc unassigned apmasell and sklam Jul 13, 2023

sklam requested changes Jul 18, 2023

View reviewed changes

tbennun added 3 commits July 18, 2023 18:00

Merge branch 'main' into constant-values

520eb6f

Apply review suggestions

1260f35

Add floating point rounding as an argument to get_constant_value

8daa56a

sklam approved these changes Jul 26, 2023

View reviewed changes

sklam added 5 - Ready to merge and removed 4 - Waiting on reviewer labels Jul 26, 2023

sklam merged commit add17f8 into numba:main Jul 26, 2023
18 checks passed

tbennun deleted the constant-values branch July 26, 2023 23:48

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Enable querying constants and value kinds #936

Enable querying constants and value kinds #936

tbennun commented Apr 19, 2023

tbennun commented Apr 19, 2023

apmasell left a comment

apmasell Apr 24, 2023

tbennun Apr 26, 2023

tbennun commented Apr 26, 2023

apmasell left a comment

sklam left a comment

sklam Apr 26, 2023

tbennun Apr 27, 2023

sklam Jul 18, 2023

tbennun Jul 19, 2023

sklam Jul 19, 2023

tbennun Jul 19, 2023

tbennun commented Apr 27, 2023

tbennun commented Jun 8, 2023

esc commented Jul 5, 2023

sklam left a comment

sklam Jul 18, 2023

Enable querying constants and value kinds #936

Enable querying constants and value kinds #936

Conversation

tbennun commented Apr 19, 2023

tbennun commented Apr 19, 2023

apmasell left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tbennun commented Apr 26, 2023

apmasell left a comment

Choose a reason for hiding this comment

sklam left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tbennun commented Apr 27, 2023

tbennun commented Jun 8, 2023

esc commented Jul 5, 2023

sklam left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment