[PATCH] D108643: Introduce _BitInt, deprecate _ExtInt

[Erich Keane via Phabricator via cfe-commits]

erichkeane added a comment.

In D108643#3026550 <https://reviews.llvm.org/D108643#3026550>, @rjmccall wrote:

> I think it would be better to provide a generic ABI specification that is designed to "lower" `_BitInt` types into more basic concepts that ABIs typically already have rules for.  It would be appropriate to note at the top that this is only a generic specification and that the rules for any particular platform should be codified in a platform-specific ABI.  But we should make that easy so that platform owners who don't have strong opinions about the ABI can just say "as described in version 1.4 of the generic ABI at https://clang.llvm.org/..."
>
> Now, that is all independent of what we actually decide the ABI should be.  But I do think that if we're going to make this available on all targets, we should strongly consider doing a "legalizing" lowering in the frontend: at ABI-exposed points (call arguments/parameters and memory accesses), `_BitInt` types should be translated into types that a naive backend can be trusted to handle in a way that matches the documented ABI.  Individual targets can then opt in to using the natural lowerings if they promise to match the ABI, but that becomes a decision that they make to enable better optimization and code generation, not something that's necessary for correctness.

Don't we already have that?  We work the same way a 'bool' does, that is, we zero/sign extend for parameters and return values?  The backends then reliably up-scale these to the power-of-two/multiple-of-pointer.

Or am I missing something here?  The only difference here that I can see vs bool is that we don't do anything for in-function variables:

  ; Function Attrs: mustprogress noinline nounwind optnone
  define dso_local signext i5 @_Z3BarDB5_(i5 signext %b) #0 {
  entry:
   %b.addr = alloca i5, align 1
   %c = alloca i5, align 1
    store i5 %b, i5* %b.addr, align 1
    %0 = load i5, i5* %b.addr, align 1
    store i5 %0, i5* %c, align 1
    %1 = load i5, i5* %c, align 1
    ret i5 %1
  }
  
  ; Function Attrs: mustprogress noinline nounwind optnone
  define dso_local zeroext i5 @_Z3BazDU5_(i5 zeroext %b) #0 {
  entry:
    %b.addr = alloca i5, align 1
    %c = alloca i5, align 1
    store i5 %b, i5* %b.addr, align 1
    %0 = load i5, i5* %b.addr, align 1
    store i5 %0, i5* %c, align 1
    %1 = load i5, i5* %c, align 1
    ret i5 %1
  }


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D108643/new/

https://reviews.llvm.org/D108643