diff --git a/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp b/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp --- a/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp +++ b/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp @@ -1173,6 +1173,21 @@ MulC->logBase2() == ShAmtC) return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, *MulC - 2)); + // Try to narrow a bswap: + // (bswap (zext X)) >> C --> zext (bswap X >> C') + // In the case where the shift amount equals the bitwidth difference, the + // shift is eliminated. + if (match(Op0, m_OneUse(m_Intrinsic( + m_OneUse(m_ZExt(m_Value(X))))))) { + // TODO: If the shift amount is less than the zext, we could shift left. + unsigned WidthDiff = BitWidth - X->getType()->getScalarSizeInBits(); + if (ShAmtC >= WidthDiff) { + Value *NarrowSwap = Builder.CreateUnaryIntrinsic(Intrinsic::bswap, X); + Value *NewShift = Builder.CreateLShr(NarrowSwap, ShAmtC - WidthDiff); + return new ZExtInst(NewShift, Ty); + } + } + // 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)) { diff --git a/llvm/test/Transforms/InstCombine/lshr.ll b/llvm/test/Transforms/InstCombine/lshr.ll --- a/llvm/test/Transforms/InstCombine/lshr.ll +++ b/llvm/test/Transforms/InstCombine/lshr.ll @@ -831,9 +831,8 @@ define i32 @narrow_bswap(i16 %x) { ; CHECK-LABEL: @narrow_bswap( -; 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]], 16 +; CHECK-NEXT: [[TMP1:%.*]] = call i16 @llvm.bswap.i16(i16 [[X:%.*]]) +; CHECK-NEXT: [[S:%.*]] = zext i16 [[TMP1]] to i32 ; CHECK-NEXT: ret i32 [[S]] ; %z = zext i16 %x to i32 @@ -844,9 +843,8 @@ define i128 @narrow_bswap_extra_wide(i16 %x) { ; CHECK-LABEL: @narrow_bswap_extra_wide( -; CHECK-NEXT: [[Z:%.*]] = zext i16 [[X:%.*]] to i128 -; CHECK-NEXT: [[B:%.*]] = call i128 @llvm.bswap.i128(i128 [[Z]]) -; CHECK-NEXT: [[S:%.*]] = lshr exact i128 [[B]], 112 +; CHECK-NEXT: [[TMP1:%.*]] = call i16 @llvm.bswap.i16(i16 [[X:%.*]]) +; CHECK-NEXT: [[S:%.*]] = zext i16 [[TMP1]] to i128 ; CHECK-NEXT: ret i128 [[S]] ; %z = zext i16 %x to i128 @@ -855,6 +853,8 @@ 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 @@ -870,9 +870,8 @@ define <2 x i64> @narrow_bswap_splat(<2 x i16> %x) { ; CHECK-LABEL: @narrow_bswap_splat( -; CHECK-NEXT: [[Z:%.*]] = zext <2 x i16> [[X:%.*]] to <2 x i64> -; CHECK-NEXT: [[B:%.*]] = call <2 x i64> @llvm.bswap.v2i64(<2 x i64> [[Z]]) -; CHECK-NEXT: [[S:%.*]] = lshr exact <2 x i64> [[B]], +; 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]] ; %z = zext <2 x i16> %x to <2 x i64> @@ -881,6 +880,8 @@ ret <2 x i64> %s } +; TODO: poison/undef in the shift amount is ok to propagate. + define <2 x i64> @narrow_bswap_splat_poison_elt(<2 x i16> %x) { ; CHECK-LABEL: @narrow_bswap_splat_poison_elt( ; CHECK-NEXT: [[Z:%.*]] = zext <2 x i16> [[X:%.*]] to <2 x i64> @@ -896,9 +897,9 @@ define <2 x i64> @narrow_bswap_overshift(<2 x i32> %x) { ; CHECK-LABEL: @narrow_bswap_overshift( -; CHECK-NEXT: [[Z:%.*]] = zext <2 x i32> [[X:%.*]] to <2 x i64> -; CHECK-NEXT: [[B:%.*]] = call <2 x i64> @llvm.bswap.v2i64(<2 x i64> [[Z]]) -; CHECK-NEXT: [[S:%.*]] = lshr <2 x i64> [[B]], +; CHECK-NEXT: [[TMP1:%.*]] = call <2 x i32> @llvm.bswap.v2i32(<2 x i32> [[X:%.*]]) +; CHECK-NEXT: [[TMP2:%.*]] = lshr <2 x i32> [[TMP1]], +; CHECK-NEXT: [[S:%.*]] = zext <2 x i32> [[TMP2]] to <2 x i64> ; CHECK-NEXT: ret <2 x i64> [[S]] ; %z = zext <2 x i32> %x to <2 x i64> @@ -909,9 +910,9 @@ define i128 @narrow_bswap_overshift2(i96 %x) { ; CHECK-LABEL: @narrow_bswap_overshift2( -; CHECK-NEXT: [[Z:%.*]] = zext i96 [[X:%.*]] to i128 -; CHECK-NEXT: [[B:%.*]] = call i128 @llvm.bswap.i128(i128 [[Z]]) -; CHECK-NEXT: [[S:%.*]] = lshr i128 [[B]], 61 +; CHECK-NEXT: [[TMP1:%.*]] = call i96 @llvm.bswap.i96(i96 [[X:%.*]]) +; CHECK-NEXT: [[TMP2:%.*]] = lshr i96 [[TMP1]], 29 +; CHECK-NEXT: [[S:%.*]] = zext i96 [[TMP2]] to i128 ; CHECK-NEXT: ret i128 [[S]] ; %z = zext i96 %x to i128