[llvm] [AMDGPU] Add GFX12 wave register names with WAVE_ prefix (PR #144352)

[Jay Foad via llvm-commits]

================
@@ -167,7 +167,14 @@ namespace Hwreg {
 // NOLINTBEGIN
 // clang-format off
 static constexpr CustomOperand Operands[] = {
-  {{""}},
+  // GFX12+ renamed registers
+  {{"HW_REG_WAVE_MODE"},      ID_MODE,      isGFX12Plus},
+  {{"HW_REG_WAVE_STATUS"},    ID_STATUS,    isGFX12Plus},
+  {{"HW_REG_WAVE_GPR_ALLOC"}, ID_GPR_ALLOC, isGFX12Plus},
+  {{"HW_REG_WAVE_LDS_ALLOC"}, ID_LDS_ALLOC, isGFX12Plus},
+  {{"HW_REG_WAVE_HW_ID1"},    ID_HW_ID1,    isGFX12Plus},
+  {{"HW_REG_WAVE_HW_ID2"},    ID_HW_ID2,    isGFX12Plus},
+
   {{"HW_REG_MODE"},          ID_MODE},
----------------
jayfoad wrote:

The original design of this table was that you look up ID values directly. That is why entry 0 is "", because we want "HW_REG_MODE" to be entry 1, because ID_MODE == 1.

Then we started getting cases where the same ID value meant different registers on different subtargets, but we still tried to keep _most_ entries in the correct position. That is why getNameFromOperandTable first tries direct lookup (fast case) and then falls back to a table search (slow case).

Now I am wondering if we should abandon the idea of direct lookup. Instead we could keep this table ordered by ID, and always do a binary search.

https://github.com/llvm/llvm-project/pull/144352