[InstCombine] try to narrow more shifted bswap-of-zext

[Chenbing.Zheng]

Summary

Try to narrow more bswap, if the shift amount is less than the zext
(bswap (zext X)) >> C --> (zext (bswap X)) << C'

This is 'TODO' left in D122166.

https://alive2.llvm.org/ce/z/i7ddjn

[RKSimon]

The alive2 link appears to be broken?

[Chenbing.Zheng]

In D124598#3484668, @RKSimon wrote:

The alive2 link appears to be broken?

I have update it https://alive2.llvm.org/ce/z/i7ddjn

[RKSimon]

LGTM

llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp

if (match(Op0, m_NUWMul(m_Value(X), m_APInt(MulC)))) {
               BinaryOperator::CreateNUWMul(X, ConstantInt::get(Ty, NewMulC));
           BinaryOperator *OrigMul = cast<BinaryOperator>(Op0);
           NewMul->setHasNoSignedWrap(OrigMul->hasNoSignedWrap());
           return NewMul;
         }
       }
     }
 
-    // Try to narrow a bswap:
-    // (bswap (zext X)) >> C --> zext (bswap X >> C')
+    // Try to narrow bswap.
     // In the case where the shift amount equals the bitwidth difference, the
     // shift is eliminated.
     if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::bswap>(
                        m_OneUse(m_ZExt(m_Value(X))))))) {
-      // TODO: If the shift amount is less than the zext, we could shift left.
       unsigned SrcWidth = X->getType()->getScalarSizeInBits();
       unsigned WidthDiff = BitWidth - SrcWidth;
-      if (SrcWidth % 16 == 0 && ShAmtC >= WidthDiff) {
+      if (SrcWidth % 16 == 0) {
         Value *NarrowSwap = Builder.CreateUnaryIntrinsic(Intrinsic::bswap, X);
+        if (ShAmtC >= WidthDiff) {
+          // (bswap (zext X)) >> C --> zext (bswap X >> C')
           Value *NewShift = Builder.CreateLShr(NarrowSwap, ShAmtC - WidthDiff);
           return new ZExtInst(NewShift, Ty);
+        } else {
+          // (bswap (zext X)) >> C --> (zext (bswap X)) << C'
+          Value *NewZExt = Builder.CreateZExt(NarrowSwap, Ty);
+          Constant *ShiftDiff = ConstantInt::get(Ty, WidthDiff - ShAmtC);
+          return BinaryOperator::CreateShl(NewZExt, ShiftDiff);
+        }
       }
     }
 
     // If the shifted-out value is known-zero, then this is an exact shift.
     if (!I.isExact() &&
         MaskedValueIsZero(Op0, APInt::getLowBitsSet(BitWidth, ShAmtC), 0, &I)) {
       I.setIsExact();
       return &I;

llvm/test/Transforms/InstCombine/lshr.ll

 ; CHECK-NEXT:    ret i128 [[S]]
 ;
   %z = zext i16 %x to i128
   %b = call i128 @llvm.bswap.i128(i128 %z)
   %s = lshr i128 %b, 112
   ret i128 %s
 }
 
-; TODO: The bswap can be narrowed followed by shl.
-
 define i32 @narrow_bswap_undershift(i16 %x) {
 ; CHECK-LABEL: @narrow_bswap_undershift(
-; CHECK-NEXT:    [[Z:%.*]] = zext i16 [[X:%.*]] to i32
-; CHECK-NEXT:    [[B:%.*]] = call i32 @llvm.bswap.i32(i32 [[Z]])
-; CHECK-NEXT:    [[S:%.*]] = lshr exact i32 [[B]], 8
+; CHECK-NEXT:    [[TMP1:%.*]] = call i16 @llvm.bswap.i16(i16 [[X:%.*]])
+; CHECK-NEXT:    [[TMP2:%.*]] = zext i16 [[TMP1]] to i32
+; CHECK-NEXT:    [[S:%.*]] = shl nuw nsw i32 [[TMP2]], 7
 ; CHECK-NEXT:    ret i32 [[S]]
 ;
   %z = zext i16 %x to i32
   %b = call i32 @llvm.bswap.i32(i32 %z)
-  %s = lshr i32 %b, 8
+  %s = lshr i32 %b, 9
   ret i32 %s
 }
 
 define <2 x i64> @narrow_bswap_splat(<2 x i16> %x) {
 ; CHECK-LABEL: @narrow_bswap_splat(
 ; CHECK-NEXT:    [[TMP1:%.*]] = call <2 x i16> @llvm.bswap.v2i16(<2 x i16> [[X:%.*]])
 ; CHECK-NEXT:    [[S:%.*]] = zext <2 x i16> [[TMP1]] to <2 x i64>
 ; CHECK-NEXT:    ret <2 x i64> [[S]]