[llvm] [AMDGPU] Lazily emit waitcnts on function entry (PR #73122)

[Jay Foad via llvm-commits]

================
@@ -171,15 +171,15 @@ define <2 x i16> @v_add_v2i16_neg_inline_imm_splat(<2 x i16> %a) {
 ;
 ; GFX9-LABEL: v_add_v2i16_neg_inline_imm_splat:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v1, 0xffc0ffc0
----------------
jayfoad wrote:

Stepping back a bit, it might make more sense to change the ABI to say that expcnt must be zero at a function call boundary. I suspect there aren't any realistic cases where it would be useful to make a function call with an un-completed export in flight. But I don't want to change the ABI in this patch.

https://github.com/llvm/llvm-project/pull/73122