[Thumb] Teach ISel how to lower compares of AND bitmasks efficiently

This is essentially a recommit of r285893, but with a correctness fix. The
problem of the original commit was that this:

bic r5, r7, #31
cbz r5, .LBB2_10

got rewritten into:

lsrs  r5, r7, #5
beq .LBB2_10

The result in destination register r5 is not the same and this is incorrect
when r5 is not dead. So this fix includes checking the uses of the AND
destination register. And also, compared to the original commit, some regression
tests didn't need changing anymore because of this extra check.

For completeness, this was the original commit message:

For the common pattern (CMPZ (AND x, #bitmask), #0), we can do some more
efficient instruction selection if the bitmask is one consecutive sequence of
set bits (32 - clz(bm) - ctz(bm) == popcount(bm)).

1) If the bitmask touches the LSB, then we can remove all the upper bits and
set the flags by doing one LSLS.
2) If the bitmask touches the MSB, then we can remove all the lower bits and
set the flags with one LSRS.
3) If the bitmask has popcount == 1 (only one set bit), we can shift that bit
into the sign bit with one LSLS and change the condition query from NE/EQ to
MI/PL (we could also implement this by shifting into the carry bit and
branching on BCC/BCS).
4) Otherwise, we can emit a sequence of LSLS+LSRS to remove the upper and lower
zero bits of the mask.

1-3 require only one 16-bit instruction and can elide the CMP. 4 requires two
16-bit instructions but can elide the CMP and doesn't require materializing a
complex immediate, so is also a win.

Differential Revision: https://reviews.llvm.org/D27761


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289794 91177308-0d34-0410-b5e6-96231b3b80d8

Author: Sjoerd Meijer

lib/Target/ARM/ARMBaseInstrInfo.cpp

@@ -2544,7 +2544,11 @@ bool ARMBaseInstrInfo::optimizeCompareInstr(
   case ARM::EORrr:
   case ARM::EORri:
   case ARM::t2EORrr:
-  case ARM::t2EORri: {
+  case ARM::t2EORri:
+  case ARM::t2LSRri:
+  case ARM::t2LSRrr:
+  case ARM::t2LSLri:
+  case ARM::t2LSLrr: {
     // Scan forward for the use of CPSR
     // When checking against MI: if it's a conditional code that requires
     // checking of the V bit or C bit, then this is not safe to do.

lib/Target/ARM/ARMISelDAGToDAG.cpp

@@ -245,6 +245,7 @@ class ARMDAGToDAGISel : public SelectionDAGISel {
   bool tryInlineAsm(SDNode *N);
 
   void SelectConcatVector(SDNode *N);
+  void SelectCMPZ(SDNode *N, bool &SwitchEQNEToPLMI);
 
   bool trySMLAWSMULW(SDNode *N);
 
@@ -2730,6 +2731,87 @@ void ARMDAGToDAGISel::SelectConcatVector(SDNode *N) {
   ReplaceNode(N, createDRegPairNode(VT, N->getOperand(0), N->getOperand(1)));
 }
 
+static Optional<std::pair<unsigned, unsigned>>
+getContiguousRangeOfSetBits(const APInt &A) {
+  unsigned FirstOne = A.getBitWidth() - A.countLeadingZeros() - 1;
+  unsigned LastOne = A.countTrailingZeros();
+  if (A.countPopulation() != (FirstOne - LastOne + 1))
+    return Optional<std::pair<unsigned,unsigned>>();
+  return std::make_pair(FirstOne, LastOne);
+}
+
+void ARMDAGToDAGISel::SelectCMPZ(SDNode *N, bool &SwitchEQNEToPLMI) {
+  assert(N->getOpcode() == ARMISD::CMPZ);
+  SwitchEQNEToPLMI = false;
+
+  if (!Subtarget->isThumb())
+    // FIXME: Work out whether it is profitable to do this in A32 mode - LSL and
+    // LSR don't exist as standalone instructions - they need the barrel shifter.
+    return;
+
+  // select (cmpz (and X, C), #0) -> (LSLS X) or (LSRS X) or (LSRS (LSLS X))
+  SDValue And = N->getOperand(0);
+  if (!And->hasOneUse())
+    return;
+
+  SDValue Zero = N->getOperand(1);
+  if (!isa<ConstantSDNode>(Zero) || !cast<ConstantSDNode>(Zero)->isNullValue() ||
+      And->getOpcode() != ISD::AND)
+    return;
+  SDValue X = And.getOperand(0);
+  auto C = dyn_cast<ConstantSDNode>(And.getOperand(1));
+
+  if (!C || !X->hasOneUse())
+    return;
+  auto Range = getContiguousRangeOfSetBits(C->getAPIntValue());
+  if (!Range)
+    return;
+
+  // There are several ways to lower this:
+  SDNode *NewN;
+  SDLoc dl(N);
+
+  auto EmitShift = [&](unsigned Opc, SDValue Src, unsigned Imm) -> SDNode* {
+    if (Subtarget->isThumb2()) {
+      Opc = (Opc == ARM::tLSLri) ? ARM::t2LSLri : ARM::t2LSRri;
+      SDValue Ops[] = { Src, CurDAG->getTargetConstant(Imm, dl, MVT::i32),
+                        getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32),
+                        CurDAG->getRegister(0, MVT::i32) };
+      return CurDAG->getMachineNode(Opc, dl, MVT::i32, Ops);
+    } else {
+      SDValue Ops[] = {CurDAG->getRegister(ARM::CPSR, MVT::i32), Src,
+                       CurDAG->getTargetConstant(Imm, dl, MVT::i32),
+                       getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32)};
+      return CurDAG->getMachineNode(Opc, dl, MVT::i32, Ops);
+    }
+  };
+  
+  if (Range->second == 0) {
+    //  1. Mask includes the LSB -> Simply shift the top N bits off
+    NewN = EmitShift(ARM::tLSLri, X, 31 - Range->first);
+    ReplaceNode(And.getNode(), NewN);
+  } else if (Range->first == 31) {
+    //  2. Mask includes the MSB -> Simply shift the bottom N bits off
+    NewN = EmitShift(ARM::tLSRri, X, Range->second);
+    ReplaceNode(And.getNode(), NewN);
+  } else if (Range->first == Range->second) {
+    //  3. Only one bit is set. We can shift this into the sign bit and use a
+    //     PL/MI comparison.
+    NewN = EmitShift(ARM::tLSLri, X, 31 - Range->first);
+    ReplaceNode(And.getNode(), NewN);
+
+    SwitchEQNEToPLMI = true;
+  } else if (!Subtarget->hasV6T2Ops()) {
+    //  4. Do a double shift to clear bottom and top bits, but only in
+    //     thumb-1 mode as in thumb-2 we can use UBFX.
+    NewN = EmitShift(ARM::tLSLri, X, 31 - Range->first);
+    NewN = EmitShift(ARM::tLSRri, SDValue(NewN, 0),
+                     Range->second + (31 - Range->first));
+    ReplaceNode(And.getNode(), NewN);
+  }
+
+}
+
 void ARMDAGToDAGISel::Select(SDNode *N) {
   SDLoc dl(N);
 
@@ -2957,6 +3039,7 @@ void ARMDAGToDAGISel::Select(SDNode *N) {
         return;
       }
     }
+
     break;
   }
   case ARMISD::VMOVRRD:
@@ -3148,9 +3231,27 @@ void ARMDAGToDAGISel::Select(SDNode *N) {
     assert(N2.getOpcode() == ISD::Constant);
     assert(N3.getOpcode() == ISD::Register);
 
-    SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
-                               cast<ConstantSDNode>(N2)->getZExtValue()), dl,
-                               MVT::i32);
+    unsigned CC = (unsigned) cast<ConstantSDNode>(N2)->getZExtValue();
+    
+    if (InFlag.getOpcode() == ARMISD::CMPZ) {
+      bool SwitchEQNEToPLMI;
+      SelectCMPZ(InFlag.getNode(), SwitchEQNEToPLMI);
+      InFlag = N->getOperand(4);
+
+      if (SwitchEQNEToPLMI) {
+        switch ((ARMCC::CondCodes)CC) {
+        default: llvm_unreachable("CMPZ must be either NE or EQ!");
+        case ARMCC::NE:
+          CC = (unsigned)ARMCC::MI;
+          break;
+        case ARMCC::EQ:
+          CC = (unsigned)ARMCC::PL;
+          break;
+        }
+      }
+    }
+
+    SDValue Tmp2 = CurDAG->getTargetConstant(CC, dl, MVT::i32);
     SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
     SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
                                              MVT::Glue, Ops);
@@ -3205,6 +3306,38 @@ void ARMDAGToDAGISel::Select(SDNode *N) {
     // Other cases are autogenerated.
     break;
   }
+
+  case ARMISD::CMOV: {
+    SDValue InFlag = N->getOperand(4);
+
+    if (InFlag.getOpcode() == ARMISD::CMPZ) {
+      bool SwitchEQNEToPLMI;
+      SelectCMPZ(InFlag.getNode(), SwitchEQNEToPLMI);
+
+      if (SwitchEQNEToPLMI) {
+        SDValue ARMcc = N->getOperand(2);
+        ARMCC::CondCodes CC =
+          (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
+
+        switch (CC) {
+        default: llvm_unreachable("CMPZ must be either NE or EQ!");
+        case ARMCC::NE:
+          CC = ARMCC::MI;
+          break;
+        case ARMCC::EQ:
+          CC = ARMCC::PL;
+          break;
+        }
+        SDValue NewARMcc = CurDAG->getConstant((unsigned)CC, dl, MVT::i32);
+        SDValue Ops[] = {N->getOperand(0), N->getOperand(1), NewARMcc,
+                         N->getOperand(3), N->getOperand(4)};
+        CurDAG->MorphNodeTo(N, ARMISD::CMOV, N->getVTList(), Ops);
+      }
+
+    }
+    // Other cases are autogenerated.
+    break;
+  }
 
   case ARMISD::VZIP: {
     unsigned Opc = 0;

test/CodeGen/ARM/and-cmpz.ll

@@ -0,0 +1,71 @@
+; RUN: llc -mtriple=thumbv7m-linux-gnu < %s | FileCheck %s --check-prefix=CHECK --check-prefix=T2
+; RUN: llc -mtriple=thumbv6m-linux-gnu < %s | FileCheck %s --check-prefix=CHECK --check-prefix=T1
+
+; CHECK-LABEL: single_bit:
+; CHECK: lsls r0, r0, #23
+; T2-NEXT: mov
+; T2-NEXT: it
+; T1-NEXT: bmi
+define i32 @single_bit(i32 %p) {
+  %a = and i32 %p, 256
+  %b = icmp eq i32 %a, 0
+  br i1 %b, label %true, label %false
+
+true:
+  ret i32 1
+
+false:
+  ret i32 2
+}
+
+; CHECK-LABEL: multi_bit_lsb_ubfx:
+; CHECK: lsls r0, r0, #24
+; T2-NEXT: mov
+; T2-NEXT: it
+; T1-NEXT: beq
+define i32 @multi_bit_lsb_ubfx(i32 %p) {
+  %a = and i32 %p, 255
+  %b = icmp eq i32 %a, 0
+  br i1 %b, label %true, label %false
+
+true:
+  ret i32 1
+
+false:
+  ret i32 2
+}
+
+; CHECK-LABEL: multi_bit_msb:
+; CHECK: lsrs r0, r0, #24
+; T2-NEXT: mov
+; T2-NEXT: it
+; T1-NEXT: beq
+define i32 @multi_bit_msb(i32 %p) {
+  %a = and i32 %p, 4278190080  ; 0xff000000
+  %b = icmp eq i32 %a, 0
+  br i1 %b, label %true, label %false
+
+true:
+  ret i32 1
+
+false:
+  ret i32 2
+}
+
+; CHECK-LABEL: multi_bit_nosb:
+; T1: lsls r0, r0, #8
+; T1-NEXT: lsrs r0, r0, #24
+; T2: tst.w
+; T2-NEXT: it
+; T1-NEXT: beq
+define i32 @multi_bit_nosb(i32 %p) {
+  %a = and i32 %p, 16711680 ; 0x00ff0000
+  %b = icmp eq i32 %a, 0
+  br i1 %b, label %true, label %false
+
+true:
+  ret i32 1
+
+false:
+  ret i32 2
+}

test/CodeGen/ARM/arm-and-tst-peephole.ll

@@ -93,7 +93,7 @@ entry:
   %1 = load i8, i8* %0, align 1
   %2 = zext i8 %1 to i32
 ; ARM: ands
-; THUMB: ands
+; THUMB: ands 
 ; T2: ands
 ; V8: ands
 ; V8-NEXT: beq
@@ -150,10 +150,9 @@ define i32 @test_tst_assessment(i1 %lhs, i1 %rhs) {
   %rhs32 = zext i1 %rhs to i32
   %diff = sub nsw i32 %lhs32, %rhs32
 ; ARM: tst r1, #1
-; THUMB: movs [[RTMP:r[0-9]+]], #1
-; THUMB: tst r1, [[RTMP]]
-; T2: tst.w r1, #1
-; V8: tst.w r1, #1
+; THUMB: lsls r1, r1, #31
+; T2: lsls r1, r1, #31
+; V8: lsls r1, r1, #31
   ret i32 %diff
 }
 

test/CodeGen/ARM/arm-shrink-wrapping.ll

@@ -638,14 +638,13 @@ declare double @llvm.pow.f64(double, double)
 ; during PEI with shrink-wrapping enable.
 ; CHECK-LABEL: debug_info:
 ;
-; ENABLE: tst{{(\.w)?}}  r2, #1
+; ENABLE: {{tst  r2, #1|lsls r1, r2, #31}}
 ; ENABLE-NEXT: beq      [[BB13:LBB[0-9_]+]]
 ;
 ; CHECK: push
 ;
-; DISABLE: tst{{(\.w)?}}  r2, #1
-; DISABLE-NEXT: vst1.64
-; DISABLE-NEXT: beq      [[BB13:LBB[0-9_]+]]
+; DISABLE: {{tst  r2, #1|lsls r1, r2, #31}}
+; DISABLE: beq      [[BB13:LBB[0-9_]+]]
 ;
 ; CHECK: bl{{x?}} _pow
 ;

test/CodeGen/ARM/call-tc.ll

@@ -120,7 +120,7 @@ if.end:                                           ; preds = %entry
   br i1 %tobool2, label %if.end5, label %if.then3
 
 if.then3:                                         ; preds = %if.end
-; CHECKT2D: bne.w _b
+; CHECKT2D: bmi.w _b
   %call4 = tail call i32 @b(i32 %x) nounwind
   br label %return
 

test/CodeGen/ARM/debug-info-branch-folding.ll

@@ -3,8 +3,7 @@ target datalayout = "e-p:32:32:32-i1:8:32-i8:8:32-i16:16:32-i32:32:32-i64:32:32-
 target triple = "thumbv7-apple-macosx10.6.7"
 
 ;CHECK: 	vadd.f32	q4, q8, q8
-;CHECK:         LBB0_1:
-;CHECK-NOT:     beq LBB0_1
+;CHECK-NEXT: LBB0_1
 
 ;CHECK:         @DEBUG_VALUE: x <- %Q4{{$}}
 ;CHECK-NEXT:    @DEBUG_VALUE: y <- %Q4{{$}}

test/CodeGen/Thumb/thumb-shrink-wrapping.ll

@@ -650,11 +650,14 @@ define i1 @beq_to_bx(i32* %y, i32 %head) {
 
 ; CHECK: tst r3, r4
 ; ENABLE-NEXT: pop {r4}
-; ENABLE-NEXT: pop {r3}
-; ENABLE-NEXT: mov lr, r3
+; ENABLE-NEXT: mov r12, r{{.*}}
+; ENABLE-NEXT: pop {r0}
+; ENABLE-NEXT: mov lr, r0
+; ENABLE-NEXT: mov r0, r12
 ; CHECK-NEXT: beq [[EXIT_LABEL]]
 
 ; CHECK: str r1, [r2]
+; CHECK: str r3, [r2]
 ; CHECK-NEXT: movs r0, #0
 ; CHECK-NEXT: [[EXIT_LABEL]]: @ %cleanup
 ; ENABLE-NEXT: bx lr
@@ -675,6 +678,7 @@ if.end:
 
 if.end4:
   store i32 %head, i32* %y, align 4
+  store volatile i32 %z, i32* %y, align 4
   br label %cleanup
 
 cleanup:

test/CodeGen/Thumb2/float-ops.ll

@@ -259,9 +259,9 @@ define i64 @bitcast_d_to_i(double %a) {
 
 define float @select_f(float %a, float %b, i1 %c) {
 ; CHECK-LABEL: select_f:
-; NONE: tst.w   r2, #1
+; NONE: lsls    r2, r2, #31
 ; NONE: moveq   r0, r1
-; HARD: tst.w   r0, #1
+; HARD: lsls    r0, r0, #31
 ; VFP4-ALL: vmovne.f32      s1, s0
 ; VFP4-ALL: vmov.f32        s0, s1
 ; FP-ARMv8: vseleq.f32 s0, s1, s0
@@ -271,8 +271,8 @@ define float @select_f(float %a, float %b, i1 %c) {
 
 define double @select_d(double %a, double %b, i1 %c) {
 ; CHECK-LABEL: select_d:
-; NONE: ldr.w   [[REG:r[0-9]+]], [sp]
-; NONE: ands    [[REG]], [[REG]], #1
+; NONE: ldr{{(.w)?}}     [[REG:r[0-9]+]], [sp]
+; NONE  ands    [[REG]], [[REG]], #1
 ; NONE: moveq   r0, r2
 ; NONE: moveq   r1, r3
 ; SP: ands r0, r0, #1
@@ -282,7 +282,7 @@ define double @select_d(double %a, double %b, i1 %c) {
 ; SP-DAG: movne [[BLO]], [[ALO]]
 ; SP-DAG: movne [[BHI]], [[AHI]]
 ; SP: vmov d0, [[BLO]], [[BHI]]
-; DP: tst.w   r0, #1
+; DP: lsls   r0, r0, #31
 ; VFP4-DP: vmovne.f64      d1, d0
 ; VFP4-DP: vmov.f64        d0, d1
 ; FP-ARMV8: vseleq.f64      d0, d1, d0