[llvm] [AMDGPU] Add the 3-dword image_gather4 variant for packed D16 + TFE (PR #215972)

[Arseniy Obolenskiy via llvm-commits]

================
@@ -0,0 +1,19 @@
+; RUN: llc < %s -mtriple=amdgpu9.08 | FileCheck -check-prefix=GCN %s
+
+; GCN-LABEL: {{^}}image_gather4_b_2d_v4f16_tfe_agpr:
+; GCN: image_gather4_b v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], s[0:7], s[8:11] dmask:0x4 tfe d16{{$}}
----------------
aobolensk wrote:

Turns out an AGPR result is not reachable here: gfx90a+ do not support image TFE at all, but gfx908 has FeatureImageGather4D16Bug so it selects V5, not V3

https://github.com/llvm/llvm-project/pull/215972