[llvm-branch-commits] [llvm] [AMDGPU] Codegen support for constrained multi-dword sloads (PR #96163)

[Stanislav Mekhanoshin via llvm-branch-commits]

================
@@ -886,26 +977,17 @@ multiclass SMRD_Pattern <string Instr, ValueType vt, bit immci = true> {
   def : GCNPat <
     (smrd_load (SMRDSgpr i64:$sbase, i32:$soffset)),
     (vt (!cast<SM_Pseudo>(Instr#"_SGPR") $sbase, $soffset, 0))> {
-    let OtherPredicates = [isNotGFX9Plus];
-  }
-  def : GCNPat <
-    (smrd_load (SMRDSgpr i64:$sbase, i32:$soffset)),
-    (vt (!cast<SM_Pseudo>(Instr#"_SGPR_IMM") $sbase, $soffset, 0, 0))> {
-    let OtherPredicates = [isGFX9Plus];
+    let OtherPredicates = [isGFX6GFX7];
   }
 
-  // 4. SGPR+IMM offset
+  // 4. No offset
   def : GCNPat <
-    (smrd_load (SMRDSgprImm i64:$sbase, i32:$soffset, i32:$offset)),
-    (vt (!cast<SM_Pseudo>(Instr#"_SGPR_IMM") $sbase, $soffset, $offset, 0))> {
-    let OtherPredicates = [isGFX9Plus];
+    (vt (smrd_load (i64 SReg_64:$sbase))),
+    (vt (!cast<SM_Pseudo>(Instr#"_IMM") i64:$sbase, 0, 0))> {
+    let OtherPredicates = [isGFX6GFX7];
   }
 
-  // 5. No offset
-  def : GCNPat <
-    (vt (smrd_load (i64 SReg_64:$sbase))),
-    (vt (!cast<SM_Pseudo>(Instr#"_IMM") i64:$sbase, 0, 0))
-  >;
+  defm : SMRD_Align_Pattern<Instr, vt>;
----------------
rampitec wrote:

You can avoid duplicating patterns for aligned case, you just need to check if xnack is on (and it is off before gfx8).
I also do not see xnack checked anywhere.

https://github.com/llvm/llvm-project/pull/96163