[llvm-branch-commits] [llvm] AMDGPU/GlobalISel: RBLegalize rules for load (PR #112882)

[Matt Arsenault via llvm-branch-commits]

================
@@ -293,7 +506,87 @@ RegBankLegalizeRules::RegBankLegalizeRules(const GCNSubtarget &_ST,
       .Any({{UniS64, S32}, {{Sgpr64}, {Sgpr32}, Ext32To64}})
       .Any({{DivS64, S32}, {{Vgpr64}, {Vgpr32}, Ext32To64}});
 
-  addRulesForGOpcs({G_LOAD}).Any({{DivS32, DivP1}, {{Vgpr32}, {VgprP1}}});
+  bool hasUnAlignedLoads = ST->getGeneration() >= AMDGPUSubtarget::GFX12;
+  bool hasSMRDSmall = ST->hasScalarSubwordLoads();
+
+  Predicate isAlign16([](const MachineInstr &MI) -> bool {
+    return (*MI.memoperands_begin())->getAlign() >= Align(16);
+  });
+
+  Predicate isAlign4([](const MachineInstr &MI) -> bool {
+    return (*MI.memoperands_begin())->getAlign() >= Align(4);
+  });
+
+  Predicate isAtomicMMO([](const MachineInstr &MI) -> bool {
+    return (*MI.memoperands_begin())->isAtomic();
+  });
+
+  Predicate isUniMMO([](const MachineInstr &MI) -> bool {
+    return AMDGPUInstrInfo::isUniformMMO(*MI.memoperands_begin());
+  });
+
+  Predicate isConst([](const MachineInstr &MI) -> bool {
+    // This is wierd. Can AS in MMO be different then AS on pointer?
----------------
arsenm wrote:

Typo wierd.

Yes, the verifier doesn't enforce this (although maybe it should). In the cases where the addrspace comes from an underlying IR value, it definitely can mismatch 

https://github.com/llvm/llvm-project/pull/112882