Skip to content

lower unbounded_shl/unbounded_shr without branching - #710

Open
nazar-pc wants to merge 1 commit into
Rust-GPU:mainfrom
nazar-pc:unbounded-shifts
Open

nazar-pc wants to merge 1 commit into
Rust-GPU:mainfrom
nazar-pc:unbounded-shifts

Conversation

@nazar-pc

Copy link
Copy Markdown
Contributor

This resolves #403.

Essentially reduces this:

%30 = OpULessThan %bool %29 %uint_32
      OpSelectionMerge %31 None
      OpBranchConditional %30 %32 %33
%33 = OpLabel
      OpBranch %31
%32 = OpLabel
%41 = OpShiftLeftLogical %uint %26 %29
      OpBranch %31
%31 = OpLabel
%42 = OpPhi %uint %41 %32 %uint_0 %33
      OpSelectionMerge %44 None
      OpBranchConditional %30 %45 %46
%46 = OpLabel
      OpBranch %44
%45 = OpLabel
%54 = OpShiftRightLogical %uint %26 %29
      OpBranch %44
%44 = OpLabel
%55 = OpPhi %uint %54 %45 %uint_0 %46
%56 = OpBitwiseXor %uint %42 %55
      OpSelectionMerge %58 None
      OpBranchConditional %30 %59 %60
%60 = OpLabel
%76 = OpShiftRightArithmetic %int %27 %uint_31
      OpBranch %58
%59 = OpLabel
%68 = OpShiftRightArithmetic %int %27 %29
      OpBranch %58
%58 = OpLabel
%77 = OpPhi %int %68 %59 %76 %60
%78 = OpULessThan %bool %29 %uint_64
      OpSelectionMerge %79 None
      OpBranchConditional %78 %80 %81
%81 = OpLabel
      OpBranch %79
%80 = OpLabel
%89 = OpUConvert %ulong %29
%90 = OpShiftLeftLogical %ulong %28 %89
      OpBranch %79
%79 = OpLabel
%91 = OpPhi %ulong %90 %80 %ulong_0 %81

Into this:

%28 = OpULessThan %bool %27 %uint_32
%29 = OpShiftLeftLogical %uint %24 %27
%30 = OpSelect %uint %28 %29 %uint_0
%32 = OpShiftRightLogical %uint %24 %27
%33 = OpSelect %uint %28 %32 %uint_0
%34 = OpBitwiseXor %uint %30 %33
%36 = OpSelect %uint %28 %27 %uint_31
%37 = OpShiftRightArithmetic %int %25 %36
%38 = OpULessThan %bool %27 %uint_64
%39 = OpShiftLeftLogical %ulong %26 %27
%40 = OpSelect %ulong %38 %39 %ulong_0

And hopefully optimizes better in a driver (though I have not benchmarked it).

Implemented by a clanker, but looks reasonable to me.

The `core` implementation of these methods branches on whether the shift
amount is less than the bit width, which turns into structured control flow
with an `OpPhi`. It also goes through `unchecked_shl`/`unchecked_shr`, whose
precondition check forces callers to be inlined because of the panic path.

Intercept calls to the integer `unbounded_shl`/`unbounded_shr` methods from
`core` and emit the shift followed by an `OpSelect` on whether the shift
amount is in range, since SPIR-V leaves shifts by the bit width or more
undefined. Signed `unbounded_shr` clamps the shift amount to `BITS - 1`
instead, which fills the result with the sign bit.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Implement unbounded_shl/unbounded_shr

1 participant