[Hexagon] Gracefully handle reg class mismatch in HexagonLoopReschedule

llvm-svn: 276793

diff --git a/llvm/lib/Target/Hexagon/HexagonBitSimplify.cpp b/llvm/lib/Target/Hexagon/HexagonBitSimplify.cpp
index e6a03aa..c847fc1 100644 (file)
--- a/llvm/lib/Target/Hexagon/HexagonBitSimplify.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonBitSimplify.cpp
@@ -2382,8 +2382,8 @@ namespace {
     struct PhiInfo {
       PhiInfo(MachineInstr &P, MachineBasicBlock &B);
       unsigned DefR;
-      BitTracker::RegisterRef LR, PR;
-      MachineBasicBlock *LB, *PB;
+      BitTracker::RegisterRef LR, PR; // Loop Register, Preheader Register
+      MachineBasicBlock *LB, *PB;     // Loop Block, Preheader Block
     };
 
     static unsigned getDefReg(const MachineInstr *MI);
@@ -2742,31 +2742,37 @@ bool HexagonLoopRescheduling::processLoop(LoopCand &C) {
     auto F = std::find_if(Phis.begin(), Phis.end(), LoopInpEq);
     if (F == Phis.end())
       continue;
-    unsigned PredR = 0;
-    if (!isSameShuffle(G.Out.Reg, G.Inp.Reg, F->PR.Reg, PredR)) {
-      const MachineInstr *DefPredR = MRI->getVRegDef(F->PR.Reg);
-      unsigned Opc = DefPredR->getOpcode();
+    unsigned PrehR = 0;
+    if (!isSameShuffle(G.Out.Reg, G.Inp.Reg, F->PR.Reg, PrehR)) {
+      const MachineInstr *DefPrehR = MRI->getVRegDef(F->PR.Reg);
+      unsigned Opc = DefPrehR->getOpcode();
       if (Opc != Hexagon::A2_tfrsi && Opc != Hexagon::A2_tfrpi)
         continue;
-      if (!DefPredR->getOperand(1).isImm())
+      if (!DefPrehR->getOperand(1).isImm())
         continue;
-      if (DefPredR->getOperand(1).getImm() != 0)
+      if (DefPrehR->getOperand(1).getImm() != 0)
         continue;
       const TargetRegisterClass *RC = MRI->getRegClass(G.Inp.Reg);
       if (RC != MRI->getRegClass(F->PR.Reg)) {
-        PredR = MRI->createVirtualRegister(RC);
+        PrehR = MRI->createVirtualRegister(RC);
         unsigned TfrI = (RC == &Hexagon::IntRegsRegClass) ? Hexagon::A2_tfrsi
                                                           : Hexagon::A2_tfrpi;
         auto T = C.PB->getFirstTerminator();
         DebugLoc DL = (T != C.PB->end()) ? T->getDebugLoc() : DebugLoc();
-        BuildMI(*C.PB, T, DL, HII->get(TfrI), PredR)
+        BuildMI(*C.PB, T, DL, HII->get(TfrI), PrehR)
           .addImm(0);
       } else {
-        PredR = F->PR.Reg;
+        PrehR = F->PR.Reg;
       }
     }
-    assert(MRI->getRegClass(PredR) == MRI->getRegClass(G.Inp.Reg));
-    moveGroup(G, *F->LB, *F->PB, F->LB->getFirstNonPHI(), F->DefR, PredR);
+    // isSameShuffle could match with PrehR being of a wider class than
+    // G.Inp.Reg, for example if G shuffles the low 32 bits of its input,
+    // it would match for the input being a 32-bit register, and PrehR
+    // being a 64-bit register (where the low 32 bits match). This could
+    // be handled, but for now skip these cases.
+    if (MRI->getRegClass(PrehR) != MRI->getRegClass(G.Inp.Reg))
+      continue;
+    moveGroup(G, *F->LB, *F->PB, F->LB->getFirstNonPHI(), F->DefR, PrehR);
     Changed = true;
   }
 

diff --git a/llvm/test/CodeGen/Hexagon/bit-loop-rc-mismatch.ll b/llvm/test/CodeGen/Hexagon/bit-loop-rc-mismatch.ll
new file mode 100644 (file)
index 0000000..db57998
--- /dev/null
+++ b/llvm/test/CodeGen/Hexagon/bit-loop-rc-mismatch.ll
@@ -0,0 +1,30 @@
+; RUN: llc -march=hexagon < %s
+; REQUIRES: asserts
+
+target datalayout = "e-m:e-p:32:32:32-a:0-n16:32-i64:64:64-i32:32:32-i16:16:16-i1:8:8-f32:32:32-f64:64:64-v32:32:32-v64:64:64-v512:512:512-v1024:1024:1024-v2048:2048:2048"
+target triple = "hexagon"
+
+define weak_odr hidden i32 @fred(i32* %this, i32* nocapture readonly dereferenceable(4) %__k) #0 align 2 {
+entry:
+  %call = tail call i64 @danny(i32* %this, i32* nonnull dereferenceable(4) %__k) #2
+  %__p.sroa.0.0.extract.trunc = trunc i64 %call to i32
+  br i1 undef, label %for.end, label %for.body
+
+for.body:                                         ; preds = %for.body, %entry
+  %__p.sroa.0.018 = phi i32 [ %call8, %for.body ], [ %__p.sroa.0.0.extract.trunc, %entry ]
+  %call8 = tail call i32 @sammy(i32* %this, i32 %__p.sroa.0.018) #2
+  %0 = inttoptr i32 %call8 to i32*
+  %lnot.i = icmp eq i32* %0, undef
+  br i1 %lnot.i, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.body, %entry
+  ret i32 0
+}
+
+declare hidden i64 @danny(i32*, i32* nocapture readonly dereferenceable(4)) #1 align 2
+declare hidden i32 @sammy(i32* nocapture, i32) #0 align 2
+
+attributes #0 = { nounwind optsize "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="hexagonv60" "target-features"="+hvx,-hvx-double" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #1 = { nounwind optsize readonly "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="hexagonv60" "target-features"="+hvx,-hvx-double" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #2 = { optsize }
+