[llvm] [RISCV][ISEL] Lowering to load-acquire/store-release for RISCV Zalasr (PR #82914)

[Wang Pengcheng via llvm-commits]

================
@@ -57,3 +57,47 @@ let Predicates = [HasStdExtZalasr, IsRV64] in {
 defm LD : LAQ_r_aq_rl<0b011, "ld">;
 defm SD : SRL_r_aq_rl<0b011, "sd">;
 } // Predicates = [HasStdExtZalasr, IsRV64]
+
+//===----------------------------------------------------------------------===//
+// Pseudo-instructions and codegen patterns
+//===----------------------------------------------------------------------===//
+
+class PatLAQ<SDPatternOperator OpNode, RVInst Inst, ValueType vt = XLenVT>
+    : Pat<(vt (OpNode (vt GPRMemZeroOffset:$rs1))), (Inst GPRMemZeroOffset:$rs1)>;
+
+class PatSRL<SDPatternOperator OpNode, RVInst Inst, ValueType vt = XLenVT>
+    : Pat<(OpNode (vt GPR:$rs2), (vt GPRMemZeroOffset:$rs1)),
+          (Inst GPRMemZeroOffset:$rs1, GPR:$rs2)>; // n.b. this switches order of arguments
+                                                   // to deal with the fact that SRL has addr, data
+                                                   // while atomic_store has data, addr
+                              
+
+let Predicates = [HasStdExtZalasr] in {
+  def : PatLAQ<acquiring_load<atomic_load_8>, LB_AQ>;
+  def : PatLAQ<seq_cst_load<atomic_load_8>, LB_AQ>; // the sequentially consistent loads use
+                                                    // .aq instead of .aqrl to match the psABI/A.7
+
+  def : PatLAQ<acquiring_load<atomic_load_16>, LH_AQ>;
+  def : PatLAQ<seq_cst_load<atomic_load_16>, LH_AQ>;
+
+  def : PatLAQ<acquiring_load<atomic_load_32>, LW_AQ>;
+  def : PatLAQ<seq_cst_load<atomic_load_32>, LW_AQ>;
+
+  def : PatSRL<releasing_store<atomic_store_8>, SB_RL>;
+  def : PatSRL<seq_cst_store<atomic_store_8>, SB_RL>; // the sequentially consistent stores use
----------------
wangpc-pp wrote:

Ditto.

https://github.com/llvm/llvm-project/pull/82914