Argument limit for avx2_gather is too weak

[KanRobert]

https://www.godbolt.org/z/9T1eGPWjY

llvm/include/llvm/IR/IntrinsicsX86.td

  def int_x86_avx2_gather_q_q_256 : ClangBuiltin<"__builtin_ia32_gatherq_q256">,
      DefaultAttrsIntrinsic<[llvm_v4i64_ty],
        [llvm_v4i64_ty, llvm_ptr_ty, llvm_v4i64_ty, llvm_v4i64_ty, llvm_i8_ty],
        [IntrReadMem, ImmArg<ArgIndex<4>>]>;

The i8 operand is a scale, and should be 1, 2, 4 or 8. But now the middle end only knows it's 8-bit immediate. So llvm-reduce may assign 0 to the operand and then produce incorrect IR.

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Shengchen Kan (KanRobert)

https://www.godbolt.org/z/9T1eGPWjY llvm/include/llvm/IR/IntrinsicsX86.td ``` def int_x86_avx2_gather_q_q_256 : ClangBuiltin<"__builtin_ia32_gatherq_q256">, DefaultAttrsIntrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty, llvm_ptr_ty, llvm_v4i64_ty, llvm_v4i64_ty, llvm_i8_ty], [IntrReadMem, ImmArg<ArgIndex<4>>]>; ``` The i8 operand is a scale, and should be 1, 2, 4 or 8. But now the middle end only knows it's 8-bit immediate. So llvm-reduce may assign 0 to the operand and then produce incorrect IR.

[KanRobert]

CC @FreddyLeaf @RKSimon @phoebewang @topperc @e-kud @goldsteinn

[FreddyLeaf]

I noticed front end can throw error for wrong scale: https://gcc.godbolt.org/z/699YP5TTq

[KanRobert]

I noticed front end can throw error for wrong scale: https://gcc.godbolt.org/z/699YP5TTq

Yes. But I am talking about the middle-end and backend to make llvm-reduce work correctly.

[autoquery]

For target intrinsics, the front end check is enough. I think llvm-reduce should provide some options to control the reduction rule with target intrinsics.

[KanRobert]

For target intrinsics, the front end check is enough. I think llvm-reduce should provide some options to control the reduction rule with target intrinsics.

But that option would definitely degrade the ability of llvm-reduce to reduce the test. Could we have sth like ImmArg<ArgIndex<4>, [1, 2, 4, 8]>?

[RKSimon]

Why don't we just remove the x86_avx2_gather* intrinsics and auto upgrade them to a gep + masked.gather intrinsic?

[autoquery]

For target intrinsics, the front end check is enough. I think llvm-reduce should provide some options to control the reduction rule with target intrinsics.

But that option would definitely degrade the ability of llvm-reduce to reduce the test. Could we have sth like ImmArg<ArgIndex<4>, [1, 2, 4, 8]>?

Still, this is something should be improved in llvm-reduce instead of middle-end/backend.