ARM: use litpools for normal i32 imms when compiling minsize.

With constant-sharing, litpool loads consume 4 + N*2 bytes of code, but
movw/movt pairs consume 8*N. This means litpools are better than movw/movt even
with just one use. Other materialisation strategies can still be better though,
so the logic is a little odd.

llvm-svn: 199891

3 files changed, 13 insertions, 11 deletions

diff --git a/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp b/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
index 8e236e16fe9..d561db2f07e 100644
--- a/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -2472,19 +2472,21 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
   case ISD::Constant: {
     unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
     bool UseCP = true;
-    if (Subtarget->hasThumb2())
+    if (Subtarget->useMovt())
       // Thumb2-aware targets have the MOVT instruction, so all immediates can
       // be done with MOV + MOVT, at worst.
-      UseCP = 0;
+      UseCP = false;
     else {
       if (Subtarget->isThumb()) {
-        UseCP = (Val > 255 &&                          // MOV
-                 ~Val > 255 &&                         // MOV + MVN
-                 !ARM_AM::isThumbImmShiftedVal(Val));  // MOV + LSL
+        UseCP = (Val > 255 &&                                  // MOV
+                 ~Val > 255 &&                                 // MOV + MVN
+                 !ARM_AM::isThumbImmShiftedVal(Val) &&         // MOV + LSL
+                 !(Subtarget->hasV6T2Ops() && Val <= 0xffff)); // MOVW
       } else
-        UseCP = (ARM_AM::getSOImmVal(Val) == -1 &&     // MOV
-                 ARM_AM::getSOImmVal(~Val) == -1 &&    // MVN
-                 !ARM_AM::isSOImmTwoPartVal(Val));     // two instrs.
+        UseCP = (ARM_AM::getSOImmVal(Val) == -1 &&             // MOV
+                 ARM_AM::getSOImmVal(~Val) == -1 &&            // MVN
+                 !ARM_AM::isSOImmTwoPartVal(Val) &&            // two instrs.
+                 !(Subtarget->hasV6T2Ops() && Val <= 0xffff)); // MOVW
     }
 
     if (UseCP) {
@@ -2494,7 +2496,7 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
                                       getTargetLowering()->getPointerTy());
 
       SDNode *ResNode;
-      if (Subtarget->isThumb1Only()) {
+      if (Subtarget->isThumb()) {
         SDValue Pred = getAL(CurDAG);
         SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
         SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
diff --git a/llvm/lib/Target/ARM/ARMInstrInfo.td b/llvm/lib/Target/ARM/ARMInstrInfo.td
index 6b93b945df2..9faa4cd840e 100644
--- a/llvm/lib/Target/ARM/ARMInstrInfo.td
+++ b/llvm/lib/Target/ARM/ARMInstrInfo.td
@@ -593,7 +593,7 @@ def so_imm2part : PatLeaf<(imm), [{
 /// arm_i32imm - True for +V6T2, or true only if so_imm2part is true.
 ///
 def arm_i32imm : PatLeaf<(imm), [{
-  if (Subtarget->hasV6T2Ops())
+  if (Subtarget->useMovt())
     return true;
   return ARM_AM::isSOImmTwoPartVal((unsigned)N->getZExtValue());
 }]>;
diff --git a/llvm/lib/Target/ARM/ARMInstrThumb2.td b/llvm/lib/Target/ARM/ARMInstrThumb2.td
index 73dc071e773..dd53ff6ccf0 100644
--- a/llvm/lib/Target/ARM/ARMInstrThumb2.td
+++ b/llvm/lib/Target/ARM/ARMInstrThumb2.td
@@ -3781,7 +3781,7 @@ def t2SUBS_PC_LR : T2I <(outs), (ins imm0_255:$imm), NoItinerary,
 let isReMaterializable = 1, isMoveImm = 1 in
 def t2MOVi32imm : PseudoInst<(outs rGPR:$dst), (ins i32imm:$src), IIC_iMOVix2,
                             [(set rGPR:$dst, (i32 imm:$src))]>,
-                            Requires<[IsThumb, HasV6T2]>;
+                            Requires<[IsThumb, UseMovt]>;
 
 // Pseudo instruction that combines movw + movt + add pc (if pic).
 // It also makes it possible to rematerialize the instructions.