[llvm] [AMDGPU] Classify FLAT instructions as VMEM (PR #137148)

[Jay Foad via llvm-commits]

================
@@ -630,20 +630,29 @@ body: |
     ; CHECK-LABEL: name: flat_global_load
     ; CHECK: liveins: $vgpr0_vgpr1
     ; CHECK-NEXT: {{  $}}
-    ; CHECK-NEXT: $vgpr2 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
-    ; CHECK-NEXT: $vgpr3 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 4, 0, implicit $exec, implicit $flat_scr
+    ; CHECK-NEXT: BUNDLE implicit-def $vgpr2, implicit-def $vgpr2_lo16, implicit-def $vgpr2_hi16, implicit-def $vgpr3, implicit-def $vgpr3_lo16, implicit-def $vgpr3_hi16, implicit $vgpr0_vgpr1, implicit $exec, implicit $flat_scr {
+    ; CHECK-NEXT:   S_CLAUSE 1
+    ; CHECK-NEXT:   $vgpr2 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
+    ; CHECK-NEXT:   $vgpr3 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 4, 0, implicit $exec, implicit $flat_scr
+    ; CHECK-NEXT: }
----------------
jayfoad wrote:

This is bad. FLAT and GLOBAL instructions can't be mixed in a clause (at least by GFX10 rules).

https://github.com/llvm/llvm-project/pull/137148