[InstCombine] Fold X * (2^N + 1) >> N -> X + X >> N, or directly to X if X >> N is 0

AZero13 · AZero13 · commit 009f86315f1e · 2024-05-26T17:33:54.000-04:00
Alive2 Proofs: https://alive2.llvm.org/ce/z/eSinJY https://alive2.llvm.org/ce/z/sweDgc https://alive2.llvm.org/ce/z/-2dXZi https://alive2.llvm.org/ce/z/kx2PhF https://alive2.llvm.org/ce/z/e9QjM6
diff --git a/llvm/lib/Analysis/InstructionSimplify.cpp b/llvm/lib/Analysis/InstructionSimplify.cpp
@@ -1479,6 +1479,29 @@ static Value *simplifyLShrInst(Value *Op0, Value *Op1, bool IsExact,
   if (Q.IIQ.UseInstrInfo && match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
     return X;
 
+  // Look for a "splat" mul pattern - it replicates bits across each half
+  // of a value, so a right shift is just a mask of the low bits:
+  const APInt *MulC;
+  const APInt *ShAmt;
+  if (Q.IIQ.UseInstrInfo && match(Op0, m_NUWMul(m_Value(X), m_APInt(MulC))) &&
+      match(Op1, m_APInt(ShAmt))) {
+    unsigned ShAmtC = ShAmt->getZExtValue();
+    unsigned BitWidth = ShAmt->getBitWidth();
+    if (BitWidth > 2 && (*MulC - 1).isPowerOf2() &&
+        MulC->logBase2() == ShAmtC) {
+      // FIXME: This condition should be covered by the computeKnownBits, but
+      // for some reason it is not, so keep this in for now. This has no
+      // negative effects, but KnownBits should be able to infer a number of
+      // leading bits based on 2^N + 1 not wrapping, as that means 2^N must not
+      // wrap either, which means the top N bits of X must be 0.
+      if (ShAmtC * 2 == BitWidth)
+        return X;
+      const KnownBits XKnown = computeKnownBits(X, /* Depth */ 0, Q);
+      if (XKnown.countMaxActiveBits() <= ShAmtC)
+        return X;
+    }
+  }
+
   // ((X << A) | Y) >> A -> X  if effective width of Y is not larger than A.
   // We can return X as we do in the above case since OR alters no bits in X.
   // SimplifyDemandedBits in InstCombine can do more general optimization for
@@ -1523,6 +1546,22 @@ static Value *simplifyAShrInst(Value *Op0, Value *Op1, bool IsExact,
   if (Q.IIQ.UseInstrInfo && match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
     return X;
 
+  const APInt *MulC;
+  const APInt *ShAmt;
+  if (Q.IIQ.UseInstrInfo && match(Op0, m_NUWMul(m_Value(X), m_APInt(MulC))) &&
+      match(Op1, m_APInt(ShAmt)) &&
+      cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap()) {
+    unsigned ShAmtC = ShAmt->getZExtValue();
+    unsigned BitWidth = ShAmt->getBitWidth();
+    if (BitWidth > 2 && (*MulC - 1).isPowerOf2() &&
+        MulC->logBase2() == ShAmtC &&
+        ShAmtC < BitWidth - 1) /* Minus 1 for the sign bit */ {
+      KnownBits KnownX = computeKnownBits(X, /* Depth */ 0, Q);
+      if (KnownX.countMaxActiveBits() <= ShAmtC)
+        return X;
+    }
+  }
+
   // Arithmetic shifting an all-sign-bit value is a no-op.
   unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
   if (NumSignBits == Op0->getType()->getScalarSizeInBits())
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp b/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp
@@ -1456,41 +1456,42 @@ Instruction *InstCombinerImpl::visitLShr(BinaryOperator &I) {
     }
 
     const APInt *MulC;
-    if (match(Op0, m_NUWMul(m_Value(X), m_APInt(MulC)))) {
+    if (match(Op0, m_OneUse(m_NUWMul(m_Value(X), m_APInt(MulC))))) {
       if (BitWidth > 2 && (*MulC - 1).isPowerOf2() &&
           MulC->logBase2() == ShAmtC) {
-        // Look for a "splat" mul pattern - it replicates bits across each half
-        // of a value, so a right shift is just a mask of the low bits:
-        // lshr i[2N] (mul nuw X, (2^N)+1), N --> and iN X, (2^N)-1
-        if (ShAmtC * 2 == BitWidth)
-          return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, *MulC - 2));
 
         // lshr (mul nuw (X, 2^N + 1)), N -> add nuw (X, lshr(X, N))
-        if (Op0->hasOneUse()) {
-          auto *NewAdd = BinaryOperator::CreateNUWAdd(
-              X, Builder.CreateLShr(X, ConstantInt::get(Ty, ShAmtC), "",
-                                    I.isExact()));
-          NewAdd->setHasNoSignedWrap(
-              cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap());
-          return NewAdd;
-        }
+        auto *NewAdd = BinaryOperator::CreateNUWAdd(
+            X, Builder.CreateLShr(X, ConstantInt::get(Ty, ShAmtC), "",
+                                  I.isExact()));
+        NewAdd->setHasNoSignedWrap(
+            cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap());
+        return NewAdd;
       }
 
       // The one-use check is not strictly necessary, but codegen may not be
       // able to invert the transform and perf may suffer with an extra mul
       // instruction.
-      if (Op0->hasOneUse()) {
-        APInt NewMulC = MulC->lshr(ShAmtC);
-        // if c is divisible by (1 << ShAmtC):
-        // lshr (mul nuw x, MulC), ShAmtC -> mul nuw nsw x, (MulC >> ShAmtC)
-        if (MulC->eq(NewMulC.shl(ShAmtC))) {
-          auto *NewMul =
-              BinaryOperator::CreateNUWMul(X, ConstantInt::get(Ty, NewMulC));
-          assert(ShAmtC != 0 &&
-                 "lshr X, 0 should be handled by simplifyLShrInst.");
-          NewMul->setHasNoSignedWrap(true);
-          return NewMul;
-        }
+      APInt NewMulC = MulC->lshr(ShAmtC);
+      // if c is divisible by (1 << ShAmtC):
+      // lshr (mul nuw x, MulC), ShAmtC -> mul nuw nsw x, (MulC >> ShAmtC)
+      if (MulC->eq(NewMulC.shl(ShAmtC))) {
+        auto *NewMul =
+            BinaryOperator::CreateNUWMul(X, ConstantInt::get(Ty, NewMulC));
+        assert(ShAmtC != 0 &&
+               "lshr X, 0 should be handled by simplifyLShrInst.");
+        NewMul->setHasNoSignedWrap(true);
+        return NewMul;
+      }
+    }
+
+    // lshr (mul nsw (X, 2^N + 1)), N -> add nsw (X, lshr(X, N))
+    if (match(Op0, m_OneUse(m_NSWMul(m_Value(X), m_APInt(MulC))))) {
+      if (BitWidth > 2 && (*MulC - 1).isPowerOf2() &&
+          MulC->logBase2() == ShAmtC) {
+        return BinaryOperator::CreateNSWAdd(
+            X, Builder.CreateLShr(X, ConstantInt::get(Ty, ShAmtC), "",
+                                  I.isExact()));
       }
     }
 
diff --git a/llvm/test/Transforms/InstCombine/ashr-lshr.ll b/llvm/test/Transforms/InstCombine/ashr-lshr.ll
@@ -862,4 +862,26 @@ define i32 @ashr_mul_times_5_div_4_exact_2(i32 %x) {
   ret i32 %ashr
 }
 
+define i32 @mul_splat_fold_known_active_bits(i32 %x) {
+; CHECK-LABEL: @mul_splat_fold_known_active_bits(
+; CHECK-NEXT:    [[XX:%.*]] = and i32 [[X:%.*]], 360
+; CHECK-NEXT:    ret i32 [[XX]]
+;
+  %xx = and i32 %x, 360
+  %m = mul nuw i32 %xx, 65537
+  %t = ashr i32 %m, 16
+  ret i32 %t
+}
+
+define i32 @mul_splat_fold_no_known_active_bits(i32 %x) {
+; CHECK-LABEL: @mul_splat_fold_no_known_active_bits(
+; CHECK-NEXT:    [[TMP1:%.*]] = ashr i32 [[X:%.*]], 16
+; CHECK-NEXT:    [[T:%.*]] = add nsw i32 [[TMP1]], [[X]]
+; CHECK-NEXT:    ret i32 [[T]]
+;
+  %m = mul nsw i32 %x, 65537
+  %t = ashr i32 %m, 16
+  ret i32 %t
+}
+
 declare void @use(i32)
diff --git a/llvm/test/Transforms/InstCombine/lshr.ll b/llvm/test/Transforms/InstCombine/lshr.ll
@@ -348,22 +348,31 @@ define <2 x i32> @narrow_lshr_constant(<2 x i8> %x, <2 x i8> %y) {
 
 define i32 @mul_splat_fold(i32 %x) {
 ; CHECK-LABEL: @mul_splat_fold(
-; CHECK-NEXT:    [[T:%.*]] = and i32 [[X:%.*]], 65535
-; CHECK-NEXT:    ret i32 [[T]]
+; CHECK-NEXT:    ret i32 [[X:%.*]]
 ;
   %m = mul nuw i32 %x, 65537
   %t = lshr i32 %m, 16
   ret i32 %t
 }
 
+define i32 @mul_splat_fold_known_zeros(i32 %x) {
+; CHECK-LABEL: @mul_splat_fold_known_zeros(
+; CHECK-NEXT:    [[XX:%.*]] = and i32 [[X:%.*]], 360
+; CHECK-NEXT:    ret i32 [[XX]]
+;
+  %xx = and i32 %x, 360
+  %m = mul nuw i32 %xx, 65537
+  %t = lshr i32 %m, 16
+  ret i32 %t
+}
+
 ; Vector type, extra use, weird types are all ok.
 
 define <3 x i14> @mul_splat_fold_vec(<3 x i14> %x) {
 ; CHECK-LABEL: @mul_splat_fold_vec(
 ; CHECK-NEXT:    [[M:%.*]] = mul nuw <3 x i14> [[X:%.*]], <i14 129, i14 129, i14 129>
 ; CHECK-NEXT:    call void @usevec(<3 x i14> [[M]])
-; CHECK-NEXT:    [[T:%.*]] = and <3 x i14> [[X]], <i14 127, i14 127, i14 127>
-; CHECK-NEXT:    ret <3 x i14> [[T]]
+; CHECK-NEXT:    ret <3 x i14> [[X]]
 ;
   %m = mul nuw <3 x i14> %x, <i14 129, i14 129, i14 129>
   call void @usevec(<3 x i14> %m)
@@ -628,8 +637,6 @@ define i32 @mul_splat_fold_wrong_lshr_const(i32 %x) {
   ret i32 %t
 }
 
-; Negative test (but simplifies into a different transform)
-
 define i32 @mul_splat_fold_no_nuw(i32 %x) {
 ; CHECK-LABEL: @mul_splat_fold_no_nuw(
 ; CHECK-NEXT:    [[TMP1:%.*]] = lshr i32 [[X:%.*]], 16
@@ -641,7 +648,7 @@ define i32 @mul_splat_fold_no_nuw(i32 %x) {
   ret i32 %t
 }
 
-; Negative test 
+; Negative test
 
 define i32 @mul_splat_fold_no_flags(i32 %x) {
 ; CHECK-LABEL: @mul_splat_fold_no_flags(
