[PowerPC] Use 'sync; ld; cmp; bc; isync' for atomic load seq-cst on 32-bit platform (#75905)

`cmp; bc; isync` is more performant than `lwsync` theoretically.

64-bit platform already features it, now implement it for 32-bit
platform.

Author: Kai Luo

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

@@ -10984,17 +10984,18 @@ SDValue PPCTargetLowering::LowerINTRINSIC_VOID(SDValue Op,
   switch (cast<ConstantSDNode>(Op.getOperand(ArgStart))->getZExtValue()) {
   case Intrinsic::ppc_cfence: {
     assert(ArgStart == 1 && "llvm.ppc.cfence must carry a chain argument.");
-    assert(Subtarget.isPPC64() && "Only 64-bit is supported for now.");
     SDValue Val = Op.getOperand(ArgStart + 1);
     EVT Ty = Val.getValueType();
     if (Ty == MVT::i128) {
       // FIXME: Testing one of two paired registers is sufficient to guarantee
       // ordering?
       Val = DAG.getNode(ISD::TRUNCATE, DL, MVT::i64, Val);
     }
+    unsigned Opcode = Subtarget.isPPC64() ? PPC::CFENCE8 : PPC::CFENCE;
+    EVT FTy = Subtarget.isPPC64() ? MVT::i64 : MVT::i32;
     return SDValue(
-        DAG.getMachineNode(PPC::CFENCE8, DL, MVT::Other,
-                           DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Val),
+        DAG.getMachineNode(Opcode, DL, MVT::Other,
+                           DAG.getNode(ISD::ANY_EXTEND, DL, FTy, Val),
                            Op.getOperand(0)),
         0);
   }
@@ -11825,7 +11826,7 @@ Instruction *PPCTargetLowering::emitTrailingFence(IRBuilderBase &Builder,
     // See http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html and
     // http://www.rdrop.com/users/paulmck/scalability/paper/N2745r.2011.03.04a.html
     // and http://www.cl.cam.ac.uk/~pes20/cppppc/ for justification.
-    if (isa<LoadInst>(Inst) && Subtarget.isPPC64())
+    if (isa<LoadInst>(Inst))
       return Builder.CreateCall(
           Intrinsic::getDeclaration(
               Builder.GetInsertBlock()->getParent()->getParent(),

llvm/lib/Target/PowerPC/PPCInstrInfo.cpp

@@ -3177,9 +3177,11 @@ bool PPCInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
   }
 
     // FIXME: Maybe we can expand it in 'PowerPC Expand Atomic' pass.
+  case PPC::CFENCE:
   case PPC::CFENCE8: {
     auto Val = MI.getOperand(0).getReg();
-    BuildMI(MBB, MI, DL, get(PPC::CMPD), PPC::CR7).addReg(Val).addReg(Val);
+    unsigned CmpOp = Subtarget.isPPC64() ? PPC::CMPD : PPC::CMPW;
+    BuildMI(MBB, MI, DL, get(CmpOp), PPC::CR7).addReg(Val).addReg(Val);
     BuildMI(MBB, MI, DL, get(PPC::CTRL_DEP))
         .addImm(PPC::PRED_NE_MINUS)
         .addReg(PPC::CR7)

llvm/lib/Target/PowerPC/PPCInstrInfo.td

@@ -5260,6 +5260,9 @@ def HASHCHKP : XForm_XD6_RA5_RB5<31, 690, (outs),
                                  "hashchkp $RB, $addr", IIC_IntGeneral, []>;
 }
 
+let Defs = [CR7], Itinerary = IIC_LdStSync in
+def CFENCE : PPCPostRAExpPseudo<(outs), (ins gprc:$cr), "#CFENCE", []>;
+
 // Now both high word and low word are reversed, next
 // swap the high word and low word.
 def : Pat<(i64 (bitreverse i64:$A)),

llvm/test/CodeGen/PowerPC/atomics-indexed.ll

@@ -15,7 +15,9 @@ define i8 @load_x_i8_seq_cst(ptr %mem) {
 ; PPC32-NEXT:    sync
 ; PPC32-NEXT:    ori r4, r4, 24464
 ; PPC32-NEXT:    lbzx r3, r3, r4
-; PPC32-NEXT:    lwsync
+; PPC32-NEXT:    cmpw cr7, r3, r3
+; PPC32-NEXT:    bne- cr7, .+4
+; PPC32-NEXT:    isync
 ; PPC32-NEXT:    blr
 ;
 ; PPC64-LABEL: load_x_i8_seq_cst:
@@ -38,7 +40,9 @@ define i16 @load_x_i16_acquire(ptr %mem) {
 ; PPC32-NEXT:    lis r4, 2
 ; PPC32-NEXT:    ori r4, r4, 48928
 ; PPC32-NEXT:    lhzx r3, r3, r4
-; PPC32-NEXT:    lwsync
+; PPC32-NEXT:    cmpw cr7, r3, r3
+; PPC32-NEXT:    bne- cr7, .+4
+; PPC32-NEXT:    isync
 ; PPC32-NEXT:    blr
 ;
 ; PPC64-LABEL: load_x_i16_acquire:

llvm/test/CodeGen/PowerPC/atomics.ll

@@ -29,7 +29,9 @@ define i32 @load_i32_acquire(ptr %mem) {
 ; PPC32-LABEL: load_i32_acquire:
 ; PPC32:       # %bb.0:
 ; PPC32-NEXT:    lwz r3, 0(r3)
-; PPC32-NEXT:    lwsync
+; PPC32-NEXT:    cmpw cr7, r3, r3
+; PPC32-NEXT:    bne- cr7, .+4
+; PPC32-NEXT:    isync
 ; PPC32-NEXT:    blr
 ;
 ; PPC64-LABEL: load_i32_acquire: