[Hexagon] Allow redefinition with immediates for hw loop conversion

Normally, if the registers holding the induction variable's bounds
are redefined inside of the loop's body, the loop cannot be converted
to a hardware loop. However, if the redefining instruction is actually
loading an immediate value into the register, this conversion is both
possible and legal (since the immediate itself will be used in the
loop setup in the preheader).


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@316218 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/Hexagon/HexagonHardwareLoops.cpp

@@ -111,9 +111,7 @@ namespace {
   public:
     static char ID;
 
-    HexagonHardwareLoops() : MachineFunctionPass(ID) {
-      initializeHexagonHardwareLoopsPass(*PassRegistry::getPassRegistry());
-    }
+    HexagonHardwareLoops() : MachineFunctionPass(ID) {}
 
     bool runOnMachineFunction(MachineFunction &MF) override;
 
@@ -685,15 +683,21 @@ CountValue *HexagonHardwareLoops::getLoopTripCount(MachineLoop *L,
   if (InitialValue->isReg()) {
     unsigned R = InitialValue->getReg();
     MachineBasicBlock *DefBB = MRI->getVRegDef(R)->getParent();
-    if (!MDT->properlyDominates(DefBB, Header))
-      return nullptr;
+    if (!MDT->properlyDominates(DefBB, Header)) {
+      int64_t V;
+      if (!checkForImmediate(*InitialValue, V))
+        return nullptr;
+    }
     OldInsts.push_back(MRI->getVRegDef(R));
   }
   if (EndValue->isReg()) {
     unsigned R = EndValue->getReg();
     MachineBasicBlock *DefBB = MRI->getVRegDef(R)->getParent();
-    if (!MDT->properlyDominates(DefBB, Header))
-      return nullptr;
+    if (!MDT->properlyDominates(DefBB, Header)) {
+      int64_t V;
+      if (!checkForImmediate(*EndValue, V))
+        return nullptr;
+    }
     OldInsts.push_back(MRI->getVRegDef(R));
   }
 

lib/Target/Hexagon/HexagonTargetMachine.cpp

@@ -126,6 +126,7 @@ namespace llvm {
   void initializeHexagonEarlyIfConversionPass(PassRegistry&);
   void initializeHexagonExpandCondsetsPass(PassRegistry&);
   void initializeHexagonGenMuxPass(PassRegistry&);
+  void initializeHexagonHardwareLoopsPass(PassRegistry&);
   void initializeHexagonLoopIdiomRecognizePass(PassRegistry&);
   void initializeHexagonVectorLoopCarriedReusePass(PassRegistry&);
   void initializeHexagonNewValueJumpPass(PassRegistry&);
@@ -184,6 +185,7 @@ extern "C" void LLVMInitializeHexagonTarget() {
   initializeHexagonConstExtendersPass(PR);
   initializeHexagonEarlyIfConversionPass(PR);
   initializeHexagonGenMuxPass(PR);
+  initializeHexagonHardwareLoopsPass(PR);
   initializeHexagonLoopIdiomRecognizePass(PR);
   initializeHexagonVectorLoopCarriedReusePass(PR);
   initializeHexagonNewValueJumpPass(PR);

test/CodeGen/Hexagon/hwloop-redef-imm.mir

@@ -0,0 +1,63 @@
+# RUN: llc -march=hexagon -run-pass hwloops %s -o - | FileCheck %s
+
+# Normally, if the registers holding the induction variable's bounds
+# are redefined inside of the loop's body, the loop cannot be converted
+# to a hardware loop. However, if the redefining instruction is actually
+# loading an immediate value into the register, this conversion is both
+# possible and legal (since the immediate itself will be used in the
+# loop setup in the preheader).
+
+# CHECK:  [[R0:%[0-9]+]] = A2_tfrsi 1920
+# CHECK:  J2_loop0r %bb.1.b1, [[R0]]
+#
+# CHECK: bb.1.b1 (address-taken):
+# CHECK:   ENDLOOP0 %bb.1.b1
+
+
+--- |
+  define void @fred() {
+  b0:
+    br label %b1
+  b1:
+    br label %b2
+  b2:
+    ret void
+  }
+...
+
+---
+name: fred
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: intregs }
+  - { id: 1, class: intregs }
+  - { id: 2, class: intregs }
+  - { id: 3, class: intregs }
+  - { id: 4, class: intregs }
+  - { id: 5, class: intregs }
+  - { id: 6, class: intregs }
+  - { id: 7, class: intregs }
+  - { id: 8, class: predregs }
+body: |
+  bb.0.b0:
+    liveins: %r0
+    successors: %bb.1
+    %0 = A2_tfrsi 0
+    %1 = A2_tfrsi 0
+    %2 = COPY %r0
+
+  bb.1.b1:
+    successors: %bb.1, %bb.2
+    %3 = PHI %0, %bb.0, %6, %bb.1
+    %4 = PHI %1, %bb.0, %5, %bb.1
+    S4_storerh_rr %2, %4, 0, %3
+    %5 = A2_addi %4, 2
+    %6 = A2_addi %3, 1
+    ; This definition of %7 should not prevent conversion to hardware loop.
+    %7 = A2_tfrsi 3840
+    %8 = C2_cmpeq %5, %7
+    J2_jumpf %8, %bb.1, implicit-def %pc
+    J2_jump %bb.2, implicit-def %pc
+
+  bb.2.b2:
+...