[InstCombine] recognizeBSwapOrBitReverseIdiom - support for 'partial' bswap patterns...

If we're bswap'ing some bytes and zero'ing the remainder we can perform this as a bswap+mask which helps us match 'partial' bswaps as a first step towards folding into a more complex bswap pattern.

Reapplied with early-out if recognizeBSwapOrBitReverseIdiom collects a source wider than the result type.

Differential Revision: https://reviews.llvm.org/D88578

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
index cbc3f5a..edb2dc8 100644 (file)
--- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
@@ -2046,29 +2046,18 @@ Instruction *InstCombinerImpl::matchBSwap(BinaryOperator &Or) {
     Op1 = Ext->getOperand(0);
 
   // (A | B) | C  and  A | (B | C)                  -> bswap if possible.
-  bool OrOfOrs = match(Op0, m_Or(m_Value(), m_Value())) ||
-                 match(Op1, m_Or(m_Value(), m_Value()));
-
-  // (A >> B) | (C << D)  and  (A << B) | (B >> C)  -> bswap if possible.
-  bool OrOfShifts = match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
-                    match(Op1, m_LogicalShift(m_Value(), m_Value()));
-
-  // (A & B) | (C & D)                              -> bswap if possible.
-  bool OrOfAnds = match(Op0, m_And(m_Value(), m_Value())) &&
-                  match(Op1, m_And(m_Value(), m_Value()));
-
-  // (A << B) | (C & D)                              -> bswap if possible.
-  // The bigger pattern here is ((A & C1) << C2) | ((B >> C2) & C1), which is a
-  // part of the bswap idiom for specific values of C1, C2 (e.g. C1 = 16711935,
-  // C2 = 8 for i32).
-  // This pattern can occur when the operands of the 'or' are not canonicalized
-  // for some reason (not having only one use, for example).
-  bool OrOfAndAndSh = (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
-                       match(Op1, m_And(m_Value(), m_Value()))) ||
-                      (match(Op0, m_And(m_Value(), m_Value())) &&
-                       match(Op1, m_LogicalShift(m_Value(), m_Value())));
-
-  if (!OrOfOrs && !OrOfShifts && !OrOfAnds && !OrOfAndAndSh)
+  bool OrWithOrs = match(Op0, m_Or(m_Value(), m_Value())) ||
+                   match(Op1, m_Or(m_Value(), m_Value()));
+
+  // (A >> B) | C  and  (A << B) | C                -> bswap if possible.
+  bool OrWithShifts = match(Op0, m_LogicalShift(m_Value(), m_Value())) ||
+                      match(Op1, m_LogicalShift(m_Value(), m_Value()));
+
+  // (A & B) | C  and  A | (B & C)                  -> bswap if possible.
+  bool OrWithAnds = match(Op0, m_And(m_Value(), m_Value())) ||
+                    match(Op1, m_And(m_Value(), m_Value()));
+
+  if (!OrWithOrs && !OrWithShifts && !OrWithAnds)
     return nullptr;
 
   SmallVector<Instruction*, 4> Insts;

diff --git a/llvm/lib/Transforms/Utils/Local.cpp b/llvm/lib/Transforms/Utils/Local.cpp
index 0fd0dfa..1c4cbc7 100644 (file)
--- a/llvm/lib/Transforms/Utils/Local.cpp
+++ b/llvm/lib/Transforms/Utils/Local.cpp
@@ -2940,6 +2940,24 @@ collectBitParts(Value *V, bool MatchBSwaps, bool MatchBitReversals,
       return Result;
     }
 
+    // BSWAP - most likely due to us previous matching a partial bswap.
+    if (match(V, m_BSwap(m_Value(X)))) {
+      const auto &Res =
+          collectBitParts(X, MatchBSwaps, MatchBitReversals, BPS, Depth + 1);
+      if (!Res)
+        return Result;
+
+      unsigned ByteWidth = BitWidth / 8;
+      Result = BitPart(Res->Provider, BitWidth);
+      for (unsigned ByteIdx = 0; ByteIdx < ByteWidth; ++ByteIdx) {
+        unsigned ByteBitOfs = ByteIdx * 8;
+        for (unsigned BitIdx = 0; BitIdx < 8; ++BitIdx)
+          Result->Provenance[(BitWidth - 8 - ByteBitOfs) + BitIdx] =
+              Res->Provenance[ByteBitOfs + BitIdx];
+      }
+      return Result;
+    }
+
     // Funnel 'double' shifts take 3 operands, 2 inputs and the shift
     // amount (modulo).
     // fshl(X,Y,Z): (X << (Z % BW)) | (Y >> (BW - (Z % BW)))
@@ -3029,13 +3047,22 @@ bool llvm::recognizeBSwapOrBitReverseIdiom(
     DemandedTy = IntegerType::get(I->getContext(), BitProvenance.size());
   }
 
+  // Check BitProvenance hasn't found a source larger than the result type.
+  unsigned DemandedBW = DemandedTy->getBitWidth();
+  if (DemandedBW > ITy->getBitWidth())
+    return false;
+
   // Now, is the bit permutation correct for a bswap or a bitreverse? We can
   // only byteswap values with an even number of bytes.
-  unsigned DemandedBW = DemandedTy->getBitWidth();
+  APInt DemandedMask = APInt::getAllOnesValue(DemandedBW);
   bool OKForBSwap = MatchBSwaps && (DemandedBW % 16) == 0;
   bool OKForBitReverse = MatchBitReversals;
   for (unsigned BitIdx = 0;
        (BitIdx < DemandedBW) && (OKForBSwap || OKForBitReverse); ++BitIdx) {
+    if (BitProvenance[BitIdx] == BitPart::Unset) {
+      DemandedMask.clearBit(BitIdx);
+      continue;
+    }
     OKForBSwap &= bitTransformIsCorrectForBSwap(BitProvenance[BitIdx], BitIdx,
                                                 DemandedBW);
     OKForBitReverse &= bitTransformIsCorrectForBitReverse(BitProvenance[BitIdx],
@@ -3061,12 +3088,18 @@ bool llvm::recognizeBSwapOrBitReverseIdiom(
     Provider = Trunc;
   }
 
-  auto *CI = CallInst::Create(F, Provider, "rev", I);
-  InsertedInsts.push_back(CI);
+  Instruction *Result = CallInst::Create(F, Provider, "rev", I);
+  InsertedInsts.push_back(Result);
+
+  if (!DemandedMask.isAllOnesValue()) {
+    auto *Mask = ConstantInt::get(DemandedTy, DemandedMask);
+    Result = BinaryOperator::Create(Instruction::And, Result, Mask, "mask", I);
+    InsertedInsts.push_back(Result);
+  }
 
   // We may need to zeroextend back to the result type.
-  if (ITy != CI->getType()) {
-    auto *ExtInst = CastInst::Create(Instruction::ZExt, CI, ITy, "zext", I);
+  if (ITy != Result->getType()) {
+    auto *ExtInst = CastInst::Create(Instruction::ZExt, Result, ITy, "zext", I);
     InsertedInsts.push_back(ExtInst);
   }
 

diff --git a/llvm/test/Transforms/InstCombine/bswap.ll b/llvm/test/Transforms/InstCombine/bswap.ll
index aac3417..d6f0792 100644 (file)
--- a/llvm/test/Transforms/InstCombine/bswap.ll
+++ b/llvm/test/Transforms/InstCombine/bswap.ll
@@ -534,14 +534,8 @@ define i8 @PR39793_bswap_u32_as_u16_trunc(i32 %0) {
 
 define i32 @partial_bswap(i32 %x) {
 ; CHECK-LABEL: @partial_bswap(
-; CHECK-NEXT:    [[X3:%.*]] = shl i32 [[X:%.*]], 24
-; CHECK-NEXT:    [[A2:%.*]] = shl i32 [[X]], 8
-; CHECK-NEXT:    [[X2:%.*]] = and i32 [[A2]], 16711680
-; CHECK-NEXT:    [[X32:%.*]] = or i32 [[X3]], [[X2]]
-; CHECK-NEXT:    [[T1:%.*]] = and i32 [[X]], -65536
-; CHECK-NEXT:    [[T2:%.*]] = call i32 @llvm.bswap.i32(i32 [[T1]])
-; CHECK-NEXT:    [[R:%.*]] = or i32 [[X32]], [[T2]]
-; CHECK-NEXT:    ret i32 [[R]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @llvm.bswap.i32(i32 [[X:%.*]])
+; CHECK-NEXT:    ret i32 [[TMP1]]
 ;
   %x3 = shl i32 %x, 24
   %a2 = shl i32 %x, 8
@@ -578,10 +572,9 @@ declare <2 x i32> @llvm.bswap.v2i32(<2 x i32>)
 
 define i64 @bswap_and_mask_0(i64 %0) {
 ; CHECK-LABEL: @bswap_and_mask_0(
-; CHECK-NEXT:    [[TMP2:%.*]] = lshr i64 [[TMP0:%.*]], 56
-; CHECK-NEXT:    [[TMP3:%.*]] = shl i64 [[TMP0]], 56
-; CHECK-NEXT:    [[TMP4:%.*]] = or i64 [[TMP2]], [[TMP3]]
-; CHECK-NEXT:    ret i64 [[TMP4]]
+; CHECK-NEXT:    [[TMP2:%.*]] = and i64 [[TMP0:%.*]], -72057594037927681
+; CHECK-NEXT:    [[TMP3:%.*]] = call i64 @llvm.bswap.i64(i64 [[TMP2]])
+; CHECK-NEXT:    ret i64 [[TMP3]]
 ;
   %2 = lshr i64 %0, 56
   %3 = shl i64 %0, 56
@@ -606,13 +599,9 @@ define i64 @bswap_and_mask_1(i64 %0) {
 
 define i64 @bswap_and_mask_2(i64 %0) {
 ; CHECK-LABEL: @bswap_and_mask_2(
-; CHECK-NEXT:    [[TMP2:%.*]] = lshr i64 [[TMP0:%.*]], 56
-; CHECK-NEXT:    [[TMP3:%.*]] = shl i64 [[TMP0]], 56
-; CHECK-NEXT:    [[TMP4:%.*]] = or i64 [[TMP2]], [[TMP3]]
-; CHECK-NEXT:    [[TMP5:%.*]] = shl i64 [[TMP0]], 40
-; CHECK-NEXT:    [[TMP6:%.*]] = and i64 [[TMP5]], 71776119061217280
-; CHECK-NEXT:    [[TMP7:%.*]] = or i64 [[TMP4]], [[TMP6]]
-; CHECK-NEXT:    ret i64 [[TMP7]]
+; CHECK-NEXT:    [[TMP2:%.*]] = and i64 [[TMP0:%.*]], -72057594037862401
+; CHECK-NEXT:    [[TMP3:%.*]] = call i64 @llvm.bswap.i64(i64 [[TMP2]])
+; CHECK-NEXT:    ret i64 [[TMP3]]
 ;
   %2 = lshr i64 %0, 56
   %3 = shl i64 %0, 56
@@ -735,28 +724,8 @@ define i32 @funnel_binary(i32 %abcd) {
 
 define i64 @PR47191_problem1(i64 %0) {
 ; CHECK-LABEL: @PR47191_problem1(
-; CHECK-NEXT:    [[TMP2:%.*]] = lshr i64 [[TMP0:%.*]], 56
-; CHECK-NEXT:    [[TMP3:%.*]] = lshr i64 [[TMP0]], 40
-; CHECK-NEXT:    [[TMP4:%.*]] = and i64 [[TMP3]], 65280
-; CHECK-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP0]], 24
-; CHECK-NEXT:    [[TMP6:%.*]] = and i64 [[TMP5]], 16711680
-; CHECK-NEXT:    [[TMP7:%.*]] = lshr i64 [[TMP0]], 8
-; CHECK-NEXT:    [[TMP8:%.*]] = and i64 [[TMP7]], 4278190080
-; CHECK-NEXT:    [[TMP9:%.*]] = shl i64 [[TMP0]], 56
-; CHECK-NEXT:    [[TMP10:%.*]] = shl i64 [[TMP0]], 40
-; CHECK-NEXT:    [[TMP11:%.*]] = and i64 [[TMP10]], 71776119061217280
-; CHECK-NEXT:    [[TMP12:%.*]] = shl i64 [[TMP0]], 24
-; CHECK-NEXT:    [[TMP13:%.*]] = and i64 [[TMP12]], 280375465082880
-; CHECK-NEXT:    [[TMP14:%.*]] = or i64 [[TMP9]], [[TMP2]]
-; CHECK-NEXT:    [[TMP15:%.*]] = or i64 [[TMP14]], [[TMP4]]
-; CHECK-NEXT:    [[TMP16:%.*]] = or i64 [[TMP15]], [[TMP6]]
-; CHECK-NEXT:    [[TMP17:%.*]] = or i64 [[TMP16]], [[TMP8]]
-; CHECK-NEXT:    [[TMP18:%.*]] = or i64 [[TMP17]], [[TMP11]]
-; CHECK-NEXT:    [[TMP19:%.*]] = or i64 [[TMP18]], [[TMP13]]
-; CHECK-NEXT:    [[TMP20:%.*]] = shl i64 [[TMP0]], 8
-; CHECK-NEXT:    [[TMP21:%.*]] = and i64 [[TMP20]], 1095216660480
-; CHECK-NEXT:    [[TMP22:%.*]] = add i64 [[TMP19]], [[TMP21]]
-; CHECK-NEXT:    ret i64 [[TMP22]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.bswap.i64(i64 [[TMP0:%.*]])
+; CHECK-NEXT:    ret i64 [[TMP2]]
 ;
   %2 = lshr i64 %0, 56
   %3 = lshr i64 %0, 40
@@ -784,28 +753,8 @@ define i64 @PR47191_problem1(i64 %0) {
 
 define i64 @PR47191_problem2(i64 %0) {
 ; CHECK-LABEL: @PR47191_problem2(
-; CHECK-NEXT:    [[TMP2:%.*]] = lshr i64 [[TMP0:%.*]], 56
-; CHECK-NEXT:    [[TMP3:%.*]] = lshr i64 [[TMP0]], 40
-; CHECK-NEXT:    [[TMP4:%.*]] = and i64 [[TMP3]], 65280
-; CHECK-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP0]], 24
-; CHECK-NEXT:    [[TMP6:%.*]] = and i64 [[TMP5]], 16711680
-; CHECK-NEXT:    [[TMP7:%.*]] = lshr i64 [[TMP0]], 8
-; CHECK-NEXT:    [[TMP8:%.*]] = and i64 [[TMP7]], 4278190080
-; CHECK-NEXT:    [[TMP9:%.*]] = shl i64 [[TMP0]], 56
-; CHECK-NEXT:    [[TMP10:%.*]] = shl i64 [[TMP0]], 40
-; CHECK-NEXT:    [[TMP11:%.*]] = and i64 [[TMP10]], 71776119061217280
-; CHECK-NEXT:    [[TMP12:%.*]] = or i64 [[TMP9]], [[TMP2]]
-; CHECK-NEXT:    [[TMP13:%.*]] = or i64 [[TMP12]], [[TMP4]]
-; CHECK-NEXT:    [[TMP14:%.*]] = or i64 [[TMP13]], [[TMP6]]
-; CHECK-NEXT:    [[TMP15:%.*]] = or i64 [[TMP14]], [[TMP8]]
-; CHECK-NEXT:    [[TMP16:%.*]] = or i64 [[TMP15]], [[TMP11]]
-; CHECK-NEXT:    [[TMP17:%.*]] = shl i64 [[TMP0]], 24
-; CHECK-NEXT:    [[TMP18:%.*]] = and i64 [[TMP17]], 280375465082880
-; CHECK-NEXT:    [[TMP19:%.*]] = shl i64 [[TMP0]], 8
-; CHECK-NEXT:    [[TMP20:%.*]] = and i64 [[TMP19]], 1095216660480
-; CHECK-NEXT:    [[TMP21:%.*]] = or i64 [[TMP20]], [[TMP18]]
-; CHECK-NEXT:    [[TMP22:%.*]] = xor i64 [[TMP21]], [[TMP16]]
-; CHECK-NEXT:    ret i64 [[TMP22]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.bswap.i64(i64 [[TMP0:%.*]])
+; CHECK-NEXT:    ret i64 [[TMP2]]
 ;
   %2 = lshr i64 %0, 56
   %3 = lshr i64 %0, 40
@@ -833,28 +782,8 @@ define i64 @PR47191_problem2(i64 %0) {
 
 define i64 @PR47191_problem3(i64 %0) {
 ; CHECK-LABEL: @PR47191_problem3(
-; CHECK-NEXT:    [[TMP2:%.*]] = lshr i64 [[TMP0:%.*]], 56
-; CHECK-NEXT:    [[TMP3:%.*]] = lshr i64 [[TMP0]], 40
-; CHECK-NEXT:    [[TMP4:%.*]] = and i64 [[TMP3]], 65280
-; CHECK-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP0]], 24
-; CHECK-NEXT:    [[TMP6:%.*]] = and i64 [[TMP5]], 16711680
-; CHECK-NEXT:    [[TMP7:%.*]] = lshr i64 [[TMP0]], 8
-; CHECK-NEXT:    [[TMP8:%.*]] = and i64 [[TMP7]], 4278190080
-; CHECK-NEXT:    [[TMP9:%.*]] = shl i64 [[TMP0]], 56
-; CHECK-NEXT:    [[TMP10:%.*]] = shl i64 [[TMP0]], 40
-; CHECK-NEXT:    [[TMP11:%.*]] = and i64 [[TMP10]], 71776119061217280
-; CHECK-NEXT:    [[TMP12:%.*]] = or i64 [[TMP9]], [[TMP2]]
-; CHECK-NEXT:    [[TMP13:%.*]] = or i64 [[TMP12]], [[TMP4]]
-; CHECK-NEXT:    [[TMP14:%.*]] = or i64 [[TMP13]], [[TMP6]]
-; CHECK-NEXT:    [[TMP15:%.*]] = or i64 [[TMP14]], [[TMP8]]
-; CHECK-NEXT:    [[TMP16:%.*]] = or i64 [[TMP15]], [[TMP11]]
-; CHECK-NEXT:    [[TMP17:%.*]] = shl i64 [[TMP0]], 24
-; CHECK-NEXT:    [[TMP18:%.*]] = and i64 [[TMP17]], 280375465082880
-; CHECK-NEXT:    [[TMP19:%.*]] = shl i64 [[TMP0]], 8
-; CHECK-NEXT:    [[TMP20:%.*]] = and i64 [[TMP19]], 1095216660480
-; CHECK-NEXT:    [[TMP21:%.*]] = or i64 [[TMP20]], [[TMP18]]
-; CHECK-NEXT:    [[TMP22:%.*]] = xor i64 [[TMP21]], [[TMP16]]
-; CHECK-NEXT:    ret i64 [[TMP22]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.bswap.i64(i64 [[TMP0:%.*]])
+; CHECK-NEXT:    ret i64 [[TMP2]]
 ;
   %2 = lshr i64 %0, 56
   %3 = lshr i64 %0, 40
@@ -882,28 +811,8 @@ define i64 @PR47191_problem3(i64 %0) {
 
 define i64 @PR47191_problem4(i64 %0) {
 ; CHECK-LABEL: @PR47191_problem4(
-; CHECK-NEXT:    [[TMP2:%.*]] = lshr i64 [[TMP0:%.*]], 56
-; CHECK-NEXT:    [[TMP3:%.*]] = shl i64 [[TMP0]], 56
-; CHECK-NEXT:    [[TMP4:%.*]] = or i64 [[TMP2]], [[TMP3]]
-; CHECK-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP0]], 40
-; CHECK-NEXT:    [[TMP6:%.*]] = and i64 [[TMP5]], 65280
-; CHECK-NEXT:    [[TMP7:%.*]] = or i64 [[TMP4]], [[TMP6]]
-; CHECK-NEXT:    [[TMP8:%.*]] = shl i64 [[TMP0]], 40
-; CHECK-NEXT:    [[TMP9:%.*]] = and i64 [[TMP8]], 71776119061217280
-; CHECK-NEXT:    [[TMP10:%.*]] = or i64 [[TMP7]], [[TMP9]]
-; CHECK-NEXT:    [[TMP11:%.*]] = lshr i64 [[TMP0]], 24
-; CHECK-NEXT:    [[TMP12:%.*]] = and i64 [[TMP11]], 16711680
-; CHECK-NEXT:    [[TMP13:%.*]] = or i64 [[TMP10]], [[TMP12]]
-; CHECK-NEXT:    [[TMP14:%.*]] = shl i64 [[TMP0]], 24
-; CHECK-NEXT:    [[TMP15:%.*]] = and i64 [[TMP14]], 280375465082880
-; CHECK-NEXT:    [[TMP16:%.*]] = or i64 [[TMP13]], [[TMP15]]
-; CHECK-NEXT:    [[TMP17:%.*]] = lshr i64 [[TMP0]], 8
-; CHECK-NEXT:    [[TMP18:%.*]] = and i64 [[TMP17]], 4278190080
-; CHECK-NEXT:    [[TMP19:%.*]] = or i64 [[TMP16]], [[TMP18]]
-; CHECK-NEXT:    [[TMP20:%.*]] = shl i64 [[TMP0]], 8
-; CHECK-NEXT:    [[TMP21:%.*]] = and i64 [[TMP20]], 1095216660480
-; CHECK-NEXT:    [[TMP22:%.*]] = add i64 [[TMP19]], [[TMP21]]
-; CHECK-NEXT:    ret i64 [[TMP22]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.bswap.i64(i64 [[TMP0:%.*]])
+; CHECK-NEXT:    ret i64 [[TMP2]]
 ;
   %2 = lshr i64 %0, 56
   %3 = shl i64 %0, 56