[SelectionDAG] Optimize BSWAP yet again once more

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

@@ -9899,17 +9899,28 @@ SDValue TargetLowering::expandBSWAP(SDNode *N, SelectionDAG &DAG) const {
     // Use a rotate by 8. This can be further expanded if necessary.
     return DAG.getNode(ISD::ROTL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
   case MVT::i32:
-    // This is meant for ARM speficially, which has ROTR but no ROTL.
+    // This is meant for ARM specifically, which has ROTR but no ROTL.
     if (isOperationLegalOrCustom(ISD::ROTR, VT)) {
-      SDValue Mask = DAG.getConstant(0x00FF00FF, dl, VT);
-      // (x & 0x00FF00FF) rotr 8 | (x rotl 8) & 0x00FF00FF
-      SDValue And = DAG.getNode(ISD::AND, dl, VT, Op, Mask);
-      SDValue Rotr =
-          DAG.getNode(ISD::ROTR, dl, VT, And, DAG.getConstant(8, dl, SHVT));
-      SDValue Rotl =
-          DAG.getNode(ISD::ROTR, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
-      SDValue And2 = DAG.getNode(ISD::AND, dl, VT, Rotl, Mask);
-      return DAG.getNode(ISD::OR, dl, VT, Rotr, And2);
+      // eor r3, r0, r0, ror #16
+      SDValue Ror16 =
+          DAG.getNode(ISD::ROTR, dl, VT, Op, DAG.getConstant(16, dl, SHVT));
+      SDValue Xor1 = DAG.getNode(ISD::XOR, dl, VT, Op, Ror16);
+
+      // lsr r3, r3, #8
+      SDValue Lsr8 = DAG.getNode(ISD::SRL, dl, VT, Xor1,
+                                 DAG.getConstant(8, dl, SHVT));
+
+      // bic r3, r3, #65280  (0xFF00)
+      // So we need the negated value: ~0x0000FF00 = 0xFFFF00FF
+      SDValue Mask = DAG.getConstant(0xFFFF00FFu, dl, VT);
+      SDValue BicResult = DAG.getNode(ISD::AND, dl, VT, Lsr8, Mask);
+
+      // ror r0, r0, #8
+      SDValue Ror8 =
+          DAG.getNode(ISD::ROTR, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
+
+      // eor r0, r3, r0, ror #8
+      return DAG.getNode(ISD::XOR, dl, VT, BicResult, Ror8);
     }
     Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
     Tmp3 = DAG.getNode(ISD::AND, dl, VT, Op,

llvm/lib/Target/ARM/ARMISelLowering.cpp

@@ -118,6 +118,7 @@ using namespace llvm;
 #define DEBUG_TYPE "arm-isel"
 
 STATISTIC(NumTailCalls, "Number of tail calls");
+STATISTIC(NumOptimizedImms, "Number of times immediates were optimized");
 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
 STATISTIC(NumConstpoolPromoted,
@@ -142,6 +143,12 @@ static cl::opt<unsigned> ConstpoolPromotionMaxTotal(
     cl::desc("Maximum size of ALL constants to promote into a constant pool"),
     cl::init(128));
 
+static cl::opt<bool>
+    EnableOptimizeLogicalImm("arm-enable-logical-imm", cl::Hidden,
+                             cl::desc("Enable ARM logical imm instruction "
+                                      "optimization"),
+                             cl::init(true));
+
 cl::opt<unsigned>
 MVEMaxSupportedInterleaveFactor("mve-max-interleave-factor", cl::Hidden,
   cl::desc("Maximum interleave factor for MVE VLDn to generate."),
@@ -20138,6 +20145,16 @@ void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
   }
 }
 
+static bool isLegalLogicalImmediate(unsigned Imm,
+                                    const ARMSubtarget *Subtarget) {
+  if (!Subtarget->isThumb())
+    return ARM_AM::getSOImmVal(Imm) != -1;
+  if (Subtarget->isThumb2())
+    return ARM_AM::getT2SOImmVal(Imm) != -1;
+  // Thumb1 only has 8-bit unsigned immediate.
+  return Imm <= 255;
+}
+
 bool ARMTargetLowering::targetShrinkDemandedConstant(
     SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts,
     TargetLoweringOpt &TLO) const {
@@ -20146,8 +20163,7 @@ bool ARMTargetLowering::targetShrinkDemandedConstant(
   if (!TLO.LegalOps)
     return false;
 
-  // Only optimize AND for now.
-  if (Op.getOpcode() != ISD::AND)
+  if (!EnableOptimizeLogicalImm)
     return false;
 
   EVT VT = Op.getValueType();
@@ -20158,28 +20174,28 @@ bool ARMTargetLowering::targetShrinkDemandedConstant(
 
   assert(VT == MVT::i32 && "Unexpected integer type");
 
+  // Exit early if we demand all bits.
+  if (DemandedBits.popcount() == 32)
+    return false;
+
+  // Only optimize AND for now.
+  if (Op.getOpcode() != ISD::AND)
+    return false;
+
   // Make sure the RHS really is a constant.
   ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
   if (!C)
     return false;
 
   unsigned Mask = C->getZExtValue();
 
+  if (Mask == 0 || Mask == ~0U)
+    return false;
+
   unsigned Demanded = DemandedBits.getZExtValue();
   unsigned ShrunkMask = Mask & Demanded;
   unsigned ExpandedMask = Mask | ~Demanded;
 
-  // If the mask is all zeros, let the target-independent code replace the
-  // result with zero.
-  if (ShrunkMask == 0)
-    return false;
-
-  // If the mask is all ones, erase the AND. (Currently, the target-independent
-  // code won't do this, so we have to do it explicitly to avoid an infinite
-  // loop in obscure cases.)
-  if (ExpandedMask == ~0U)
-    return TLO.CombineTo(Op, Op.getOperand(0));
-
   auto IsLegalMask = [ShrunkMask, ExpandedMask](unsigned Mask) -> bool {
     return (ShrunkMask & Mask) == ShrunkMask && (~ExpandedMask & Mask) == 0;
   };
@@ -20192,30 +20208,61 @@ bool ARMTargetLowering::targetShrinkDemandedConstant(
     return TLO.CombineTo(Op, NewOp);
   };
 
-  // Prefer uxtb mask.
-  if (IsLegalMask(0xFF))
-    return UseMask(0xFF);
+  // If the mask is all zeros, let the target-independent code replace the
+  // result with zero.
+  if (ShrunkMask == 0) {
+    ++NumOptimizedImms;
+    return UseMask(ShrunkMask);
+  }
 
-  // Prefer uxth mask.
-  if (IsLegalMask(0xFFFF))
-    return UseMask(0xFFFF);
+  // If the mask is all ones, erase the AND. (Currently, the target-independent
+  // code won't do this, so we have to do it explicitly to avoid an infinite
+  // loop in obscure cases.)
+  if (ExpandedMask == ~0U) {
+    ++NumOptimizedImms;
+    return UseMask(ExpandedMask);
+  }
 
-  // [1, 255] is Thumb1 movs+ands, legal immediate for ARM/Thumb2.
-  // FIXME: Prefer a contiguous sequence of bits for other optimizations.
-  if (ShrunkMask < 256)
+  // If thumb, check for uxth and uxtb masks first and foremost.
+  if (Subtarget->isThumb1Only() && Subtarget->hasV6Ops()) {
+    if (IsLegalMask(0xFF)) {
+      ++NumOptimizedImms;
+      return UseMask(0xFF);
+    }
+
+    if (IsLegalMask(0xFFFF)) {
+      ++NumOptimizedImms;
+      return UseMask(0xFFFF);
+    }
+  }
+
+  // Don't optimize if it is legal already.
+  if (isLegalLogicalImmediate(Mask, Subtarget))
+    return false;
+
+  if (isLegalLogicalImmediate(ShrunkMask, Subtarget)) {
+    ++NumOptimizedImms;
     return UseMask(ShrunkMask);
+  }
 
-  // [-256, -2] is Thumb1 movs+bics, legal immediate for ARM/Thumb2.
-  // FIXME: Prefer a contiguous sequence of bits for other optimizations.
-  if ((int)ExpandedMask <= -2 && (int)ExpandedMask >= -256)
+  // FIXME: The check for v6 is because this interferes with some ubfx
+  // optimizations
+  if (!Subtarget->hasV6Ops() &&
+      isLegalLogicalImmediate(~ExpandedMask, Subtarget)) {
+    ++NumOptimizedImms;
     return UseMask(ExpandedMask);
+  }
+
+  if ((~ExpandedMask) < 256) {
+    ++NumOptimizedImms;
+    return UseMask(ExpandedMask);
+  }
 
   // Potential improvements:
   //
   // We could try to recognize lsls+lsrs or lsrs+lsls pairs here.
   // We could try to prefer Thumb1 immediates which can be lowered to a
   // two-instruction sequence.
-  // We could try to recognize more legal ARM/Thumb2 immediates here.
 
   return false;
 }

llvm/lib/Target/ARM/README.txt

@@ -606,32 +606,6 @@ constant which was already loaded).  Not sure what's necessary to do that.
 
 //===---------------------------------------------------------------------===//
 
-The code generated for bswap on armv4/5 (CPUs without rev) is less than ideal:
-
-int a(int x) { return __builtin_bswap32(x); }
-
-a:
-	mov	r1, #255, 24
-	mov	r2, #255, 16
-	and	r1, r1, r0, lsr #8
-	and	r2, r2, r0, lsl #8
-	orr	r1, r1, r0, lsr #24
-	orr	r0, r2, r0, lsl #24
-	orr	r0, r0, r1
-	bx	lr
-
-Something like the following would be better (fewer instructions/registers):
-	eor     r1, r0, r0, ror #16
-	bic     r1, r1, #0xff0000
-	mov     r1, r1, lsr #8
-	eor     r0, r1, r0, ror #8
-	bx	lr
-
-A custom Thumb version would also be a slight improvement over the generic
-version.
-
-//===---------------------------------------------------------------------===//
-
 Consider the following simple C code:
 
 void foo(unsigned char *a, unsigned char *b, int *c) {

llvm/test/CodeGen/ARM/and-cmpz.ll

@@ -92,7 +92,7 @@ false:
 ; T1:      uxth    r0, r0
 ; T1-NEXT: lsrs    r0, r0, #9
 ; T1-NEXT: bne
-; T2:      uxth    r0, r0
+; T2:      and	r0, r0, #65024
 ; T2-NEXT: movs    r2, #0
 ; T2-NEXT: cmp.w   r2, r0, lsr #9
 define void @i16_cmpz(i16 %x, ptr %foo) {

llvm/test/CodeGen/ARM/fpenv.ll

@@ -41,8 +41,8 @@ define void @func_05() {
 ; CHECK-LABEL: func_05:
 ; CHECK:       @ %bb.0:
 ; CHECK-NEXT:    vmrs r0, fpscr
-; CHECK-NEXT:    bic r0, r0, #12582912
 ; CHECK-NEXT:    orr r0, r0, #4194304
+; CHECK-NEXT:    bic r0, r0, #8388608
 ; CHECK-NEXT:    vmsr fpscr, r0
 ; CHECK-NEXT:    mov pc, lr
   call void @llvm.set.rounding(i32 2)
@@ -53,8 +53,8 @@ define void @func_06() {
 ; CHECK-LABEL: func_06:
 ; CHECK:       @ %bb.0:
 ; CHECK-NEXT:    vmrs r0, fpscr
-; CHECK-NEXT:    bic r0, r0, #12582912
 ; CHECK-NEXT:    orr r0, r0, #8388608
+; CHECK-NEXT:    bic r0, r0, #4194304
 ; CHECK-NEXT:    vmsr fpscr, r0
 ; CHECK-NEXT:    mov pc, lr
   call void @llvm.set.rounding(i32 3)

llvm/test/CodeGen/ARM/funnel-shift-rot.ll

@@ -19,7 +19,7 @@ declare <4 x i32> @llvm.fshr.v4i32(<4 x i32>, <4 x i32>, <4 x i32>)
 define i8 @rotl_i8_const_shift(i8 %x) {
 ; CHECK-LABEL: rotl_i8_const_shift:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    uxtb r1, r0
+; CHECK-NEXT:    and r1, r0, #224
 ; CHECK-NEXT:    lsl r0, r0, #3
 ; CHECK-NEXT:    orr r0, r0, r1, lsr #5
 ; CHECK-NEXT:    bx lr
@@ -161,8 +161,7 @@ define <4 x i32> @rotl_v4i32_rotl_const_shift(<4 x i32> %x) {
 define i8 @rotr_i8_const_shift(i8 %x) {
 ; CHECK-LABEL: rotr_i8_const_shift:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    uxtb r1, r0
-; CHECK-NEXT:    lsr r1, r1, #3
+; CHECK-NEXT:    ubfx r1, r0, #3, #5
 ; CHECK-NEXT:    orr r0, r1, r0, lsl #5
 ; CHECK-NEXT:    bx lr
   %f = call i8 @llvm.fshr.i8(i8 %x, i8 %x, i8 3)

llvm/test/CodeGen/ARM/hoist-and-by-const-from-lshr-in-eqcmp-zero.ll

@@ -21,9 +21,9 @@ define i1 @scalar_i8_signbit_eq(i8 %x, i8 %y) nounwind {
 ; ARM-LABEL: scalar_i8_signbit_eq:
 ; ARM:       @ %bb.0:
 ; ARM-NEXT:    uxtb r1, r1
-; ARM-NEXT:    lsl r0, r0, r1
+; ARM-NEXT:    mov r2, #128
+; ARM-NEXT:    and r0, r2, r0, lsl r1
 ; ARM-NEXT:    mov r1, #1
-; ARM-NEXT:    uxtb r0, r0
 ; ARM-NEXT:    eor r0, r1, r0, lsr #7
 ; ARM-NEXT:    bx lr
 ;
@@ -42,7 +42,7 @@ define i1 @scalar_i8_signbit_eq(i8 %x, i8 %y) nounwind {
 ; THUMB78-NEXT:    uxtb r1, r1
 ; THUMB78-NEXT:    lsls r0, r1
 ; THUMB78-NEXT:    movs r1, #1
-; THUMB78-NEXT:    uxtb r0, r0
+; THUMB78-NEXT:    and r0, r0, #128
 ; THUMB78-NEXT:    eor.w r0, r1, r0, lsr #7
 ; THUMB78-NEXT:    bx lr
   %t0 = lshr i8 128, %y
@@ -122,9 +122,9 @@ define i1 @scalar_i16_signbit_eq(i16 %x, i16 %y) nounwind {
 ; ARM-LABEL: scalar_i16_signbit_eq:
 ; ARM:       @ %bb.0:
 ; ARM-NEXT:    uxth r1, r1
-; ARM-NEXT:    lsl r0, r0, r1
+; ARM-NEXT:    mov r2, #32768
+; ARM-NEXT:    and r0, r2, r0, lsl r1
 ; ARM-NEXT:    mov r1, #1
-; ARM-NEXT:    uxth r0, r0
 ; ARM-NEXT:    eor r0, r1, r0, lsr #15
 ; ARM-NEXT:    bx lr
 ;
@@ -144,7 +144,7 @@ define i1 @scalar_i16_signbit_eq(i16 %x, i16 %y) nounwind {
 ; THUMB78-NEXT:    uxth r1, r1
 ; THUMB78-NEXT:    lsls r0, r1
 ; THUMB78-NEXT:    movs r1, #1
-; THUMB78-NEXT:    uxth r0, r0
+; THUMB78-NEXT:    and r0, r0, #32768
 ; THUMB78-NEXT:    eor.w r0, r1, r0, lsr #15
 ; THUMB78-NEXT:    bx lr
   %t0 = lshr i16 32768, %y
@@ -862,21 +862,35 @@ define <4 x i1> @vec_4xi32_nonsplat_undef2_eq(<4 x i32> %x, <4 x i32> %y) nounwi
 ;------------------------------------------------------------------------------;
 
 define i1 @scalar_i8_signbit_ne(i8 %x, i8 %y) nounwind {
-; ARM-LABEL: scalar_i8_signbit_ne:
-; ARM:       @ %bb.0:
-; ARM-NEXT:    uxtb r1, r1
-; ARM-NEXT:    lsl r0, r0, r1
-; ARM-NEXT:    uxtb r0, r0
-; ARM-NEXT:    lsr r0, r0, #7
-; ARM-NEXT:    bx lr
+; ARM6-LABEL: scalar_i8_signbit_ne:
+; ARM6:       @ %bb.0:
+; ARM6-NEXT:    uxtb r1, r1
+; ARM6-NEXT:    mov r2, #128
+; ARM6-NEXT:    and r0, r2, r0, lsl r1
+; ARM6-NEXT:    lsr r0, r0, #7
+; ARM6-NEXT:    bx lr
 ;
-; THUMB-LABEL: scalar_i8_signbit_ne:
-; THUMB:       @ %bb.0:
-; THUMB-NEXT:    uxtb r1, r1
-; THUMB-NEXT:    lsls r0, r1
-; THUMB-NEXT:    uxtb r0, r0
-; THUMB-NEXT:    lsrs r0, r0, #7
-; THUMB-NEXT:    bx lr
+; ARM78-LABEL: scalar_i8_signbit_ne:
+; ARM78:       @ %bb.0:
+; ARM78-NEXT:    uxtb r1, r1
+; ARM78-NEXT:    lsl r0, r0, r1
+; ARM78-NEXT:    ubfx r0, r0, #7, #1
+; ARM78-NEXT:    bx lr
+;
+; THUMB6-LABEL: scalar_i8_signbit_ne:
+; THUMB6:       @ %bb.0:
+; THUMB6-NEXT:    uxtb r1, r1
+; THUMB6-NEXT:    lsls r0, r1
+; THUMB6-NEXT:    uxtb r0, r0
+; THUMB6-NEXT:    lsrs r0, r0, #7
+; THUMB6-NEXT:    bx lr
+;
+; THUMB78-LABEL: scalar_i8_signbit_ne:
+; THUMB78:       @ %bb.0:
+; THUMB78-NEXT:    uxtb r1, r1
+; THUMB78-NEXT:    lsls r0, r1
+; THUMB78-NEXT:    ubfx r0, r0, #7, #1
+; THUMB78-NEXT:    bx lr
   %t0 = lshr i8 128, %y
   %t1 = and i8 %t0, %x
   %res = icmp ne i8 %t1, 0 ;  we are perfectly happy with 'ne' predicate
@@ -1051,3 +1065,5 @@ define i1 @scalar_i8_signbit_eq_with_nonzero(i8 %x, i8 %y) nounwind {
   %res = icmp eq i8 %t1, 1 ; should be comparing with 0
   ret i1 %res
 }
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; THUMB: {{.*}}