Turns out this is needed!

Craig Topper authored Aug 18, 2021 and rferrer committed Nov 08, 2022

This fixes a bug in the extension of integers in which a multiplication
with a large integer might overflow and needs to happen within i32.

This is a problem caused by some instructions expecting v0. Remove that
and rewrite the insert pattern that expects be able to do this.

Closes #338

We may end introducing unexpected broadcasts.

We were enabling early clobber for a number of instructions that don't
need it.

We forgot to do that for that case.

This is in preparation of building documentation while we build the
compiler

Daniel Trujillo Viedma authored Jul 13, 2022 and rferrer committed Sep 28, 2022

Implement parser/encoder and printer

Update the error message now printed in the invalid tests

Initial implementation of bfloat

Prevent use of 'this' and fix the comparison to include the alternative format bit

Continue the implementation of bfloat16. Stuck at the bitcast conversion from bf16 to i16

Promote bf16 to f32 when the target doesn't support it

This is modeled after the half-precision fp support. Two new nodes are
introduced for casting from and to bf16. Since casting from bf16 is a
simple operation I opted to always directly lower it to integer
arithmetic. The other way round is more complicated if you want to
preserve IEEE semantics, so it's handled by a new __truncsfbf2
compiler-rt builtin.

This is of course very bare bones, but sufficient to get a semi-softened
fadd on x86.

Possible future improvements:
 - Targets with bf16 conversion instructions can now make fp_to_bf16 legal
 - The software conversion to bf16 can be replaced by a trivial
   implementation under fast math.

Differential Revision: https://reviews.llvm.org/D126953

Implement a missing function after cherry-picking, and cleanup lowerSTORE

Fix can't legalize result type

Implement tests with b16 in vsetvli and fix an error with vfnmadd

Implement the alternative b16 format in EPI InstrInfo, and simplify a bit the management of widened type info

Delete X86 and PowerPC tests that don't pass now

Ideally we'd have a better way to deal with this but for now this should
do.

An infinite loop was fixed in EPI, so let's backport this change as
well.

This is an interim change.

When there are no NaNs the comparisons get folded to {max,min}num but
those were not recognised by IVDescriptor yet. Add a flag to avoid the
fold so the form is recognised.

Mostly register renamings.