[llvm] [AMDGPU] Add custom MachineValueType entries for buffer fat poiners (PR #127692)

[Krzysztof Drewniak via llvm-commits]

================
@@ -331,6 +331,10 @@ def i64x8     : ValueType<512,  231>;  // 8 Consecutive GPRs (AArch64)
 def aarch64svcount
               : ValueType<16,  232>;  // AArch64 predicate-as-counter
 def spirvbuiltin : ValueType<0, 233>; // SPIR-V's builtin type
+// AMDGPU buffer fat pointer, buffer rsrc + offset, rewritten before MIR translation
+def amdgpuBufferFatPointer : ValueType<160, 234>;
+// AMDGPU buffer strided pointer, buffer rsrc + index + offset, doesn't reach MIR
+def amdgpuBufferStridedPointer : ValueType<192, 235>;
----------------
krzysz00 wrote:

Got sucked into gfx12 bringup stuff, here's a branch that *just* add i160 and doesn't even try to touch fat pointers, observe 15 diverse crashes on the x86 and amdgpu test suite https://github.com/llvm/llvm-project/pull/127953 

I'm thinking that just making a custom type - whose size is implicitly 160 bits per the way tablegen works, makes sense, since the ARM forks added i64x8 (8 contiguous registers) as a custom type for similar reasons instead of going to MVT::i512

And ... this type never hits  MIR anyway so

https://github.com/llvm/llvm-project/pull/127692