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.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@199891 91177308-0d34-0410-b5e6-96231b3b80d8

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
+        UseCP = (Val > 255 &&                                  // MOV
-                 ~Val > 255 &&                         // MOV + MVN
+                 ~Val > 255 &&                                 // MOV + MVN
-                 !ARM_AM::isThumbImmShiftedVal(Val));  // MOV + LSL
+                 !ARM_AM::isThumbImmShiftedVal(Val) &&         // MOV + LSL
+                 !(Subtarget->hasV6T2Ops() && Val <= 0xffff)); // MOVW
       } else
-        UseCP = (ARM_AM::getSOImmVal(Val) == -1 &&     // MOV
+        UseCP = (ARM_AM::getSOImmVal(Val) == -1 &&             // MOV
-                 ARM_AM::getSOImmVal(~Val) == -1 &&    // MVN
+                 ARM_AM::getSOImmVal(~Val) == -1 &&            // MVN
-                 !ARM_AM::isSOImmTwoPartVal(Val));     // two instrs.
+                 !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() };

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());
 }]>;

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.

test/CodeGen/ARM/minsize-imms.ll

@@ -0,0 +1,57 @@
+; RUN: llc -mtriple=thumbv7m-macho -o - -show-mc-encoding %s | FileCheck %s
+; RUN: llc -mtriple=thumbv6m-macho -o - -show-mc-encoding %s | FileCheck %s --check-prefix=CHECK-V6M
+; RUN: llc -mtriple=armv6-macho -o - -show-mc-encoding %s | FileCheck %s --check-prefix=CHECK-ARM
+define i32 @test_mov() minsize {
+; CHECK-LABEL: test_mov:
+; CHECK: movs r0, #255 @ encoding: [0xff,0x20]
+
+  ret i32 255
+}
+
+define i32 @test_mov_mvn() minsize {
+; CHECK-LABEL: test_mov_mvn:
+; CHECK: mvn r0, #203 @ encoding: [0x6f,0xf0,0xcb,0x00]
+
+; CHECK-V6M-LABEL: test_mov_mvn:
+; CHECK-V6M: movs [[TMP:r[0-7]]], #203 @ encoding: [0xcb,0x20]
+; CHECK-V6M: mvns r0, [[TMP]] @ encoding: [0xc0,0x43]
+
+; CHECK-ARM-LABEL: test_mov_mvn:
+; CHECK-ARM: mvn r0, #203 @ encoding: [0xcb,0x00,0xe0,0xe3]
+  ret i32 4294967092
+}
+
+define i32 @test_mov_lsl() minsize {
+; CHECK-LABEL: test_mov_lsl:
+; CHECK: mov.w r0, #589824 @ encoding: [0x4f,0xf4,0x10,0x20]
+
+; CHECK-V6M-LABEL: test_mov_lsl:
+; CHECK-V6M: movs [[TMP:r[0-7]]], #9 @ encoding: [0x09,0x20]
+; CHECK-V6M: lsls r0, [[TMP]], #16 @ encoding: [0x00,0x04]
+
+; CHECK-ARM-LABEL: test_mov_lsl:
+; CHECK-ARM: mov r0, #589824 @ encoding: [0x09,0x08,0xa0,0xe3]
+  ret i32 589824
+}
+
+define i32 @test_movw() minsize {
+; CHECK-LABEL: test_movw:
+; CHECK: movw r0, #65535
+
+; CHECK-V6M-LABEL: test_movw:
+; CHECK-V6M: ldr r0, [[CONSTPOOL:LCPI[0-9]+_[0-9]+]] @ encoding: [A,0x48]
+; CHECK-V6M: [[CONSTPOOL]]:
+; CHECK-V6M-NEXT: .long 65535
+
+; CHECK-ARM-LABEL: test_movw:
+; CHECK-ARM: mov r0, #255 @ encoding: [0xff,0x00,0xa0,0xe3]
+; CHECK-ARM: orr r0, r0, #65280 @ encoding: [0xff,0x0c,0x80,0xe3]
+ ret i32 65535
+}
+
+define i32 @test_regress1() {
+; CHECK-ARM-LABEL: test_regress1:
+; CHECK-ARM: mov r0, #248 @ encoding: [0xf8,0x00,0xa0,0xe3]
+; CHECK-ARM: orr r0, r0, #16252928 @ encoding: [0x3e,0x07,0x80,0xe3]
+  ret i32 16253176
+}