Unexpected codegen for __builtin_elementwise_bitreverse on unsigned char and short

[shafik]

Given the following: https://godbolt.org/z/M1hEqs9q5

unsigned char f(unsigned char a) {
  return __builtin_elementwise_bitreverse(a);
}

unsigned short f(unsigned short a) {
  return __builtin_elementwise_bitreverse(a);
}

signed char f3(signed char a) {
  return __builtin_elementwise_bitreverse(a);
}

signed short f3(signed short a) {
  return __builtin_elementwise_bitreverse(a);
}

for the unsigned char and unsigned short cases it automatically returns 0 w/ any level of optimization e.g.

define dso_local noundef zeroext i8 @f(unsigned char)(i8 noundef zeroext %a) local_unnamed_addr {
entry:
  tail call void @llvm.dbg.value(metadata i8 %a, metadata !16, metadata !DIExpression())
  ret i8 0
}

Based on the documentation: https://clang.llvm.org/docs/LanguageExtensions.html

This is not expected behavior.

[topperc]

Seems like its zext the input and calling the i32 bitreverse.

I thought the __builtin_elementwise* were for vectors, not scalars?

[shafik]

Seems like its zext the input and calling the i32 bitreverse.

I thought the __builtin_elementwise* were for vectors, not scalars?

Documentation says:

For scalar types, consider the operation applied to a vector with a single element.

[llvmbot]

@llvm/issue-subscribers-clang-codegen

Author: Shafik Yaghmour (shafik)

Given the following: https://godbolt.org/z/M1hEqs9q5

unsigned char f(unsigned char a) {
  return __builtin_elementwise_bitreverse(a);
}

unsigned short f(unsigned short a) {
  return __builtin_elementwise_bitreverse(a);
}

signed char f3(signed char a) {
  return __builtin_elementwise_bitreverse(a);
}

signed short f3(signed short a) {
  return __builtin_elementwise_bitreverse(a);
}

for the unsigned char and unsigned short cases it automatically returns 0 w/ any level of optimization e.g.

define dso_local noundef zeroext i8 @<!-- -->f(unsigned char)(i8 noundef zeroext %a) local_unnamed_addr {
entry:
  tail call void @<!-- -->llvm.dbg.value(metadata i8 %a, metadata !16, metadata !DIExpression())
  ret i8 0
}

Based on the documentation: https://clang.llvm.org/docs/LanguageExtensions.html

This is not expected behavior.

[shafik]

If I look closer the documentation does say the implementation is incomplete but does not specify:

Note: The implementation of vector builtins is work-in-progress and incomplete.

In addition to the operators mentioned above, Clang provides a set of builtins to perform additional operations on certain scalar and vector types.

Let T be one of the following types:

an integer type (as in C23 6.2.5p22), but excluding enumerated types and bool

...
For scalar types, consider the operation applied to a vector with a single element.

but nothing indicates it should not work w/ unsigned char or unsigned short

[shafik]

CC @bob80905 @arsenm

[RKSimon]

CC @frasercrmck who hit this in #119423