Index: llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp =================================================================== --- llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp +++ llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp @@ -2403,6 +2403,43 @@ return &Sel; } + // A select of a "select shuffle" with a common operand can be rearranged + // to select followed by "select shuffle". Because of poison, this only works + // in the case of a shuffle with no undefined mask elements. + Value *Cond = Sel.getCondition(); + Value *TVal = Sel.getTrueValue(); + Value *FVal = Sel.getFalseValue(); + Value *X, *Y; + ArrayRef Mask; + if (match(TVal, m_OneUse(m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask)))) && + !is_contained(Mask, UndefMaskElem) && + ShuffleVectorInst::isSelectMask(Mask)) { + if (X == FVal) { + // select Cond, (shuf_sel X, Y), X --> shuf_sel X, (select Cond, Y, X) + Value *NewSel = Builder.CreateSelect(Cond, Y, X, "sel", &Sel); + return new ShuffleVectorInst(X, NewSel, Mask); + } + if (Y == FVal) { + // select Cond, (shuf_sel X, Y), Y --> shuf_sel (select Cond, X, Y), Y + Value *NewSel = Builder.CreateSelect(Cond, X, Y, "sel", &Sel); + return new ShuffleVectorInst(NewSel, Y, Mask); + } + } + if (match(FVal, m_OneUse(m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask)))) && + !is_contained(Mask, UndefMaskElem) && + ShuffleVectorInst::isSelectMask(Mask)) { + if (X == TVal) { + // select Cond, X, (shuf_sel X, Y) --> shuf_sel X, (select Cond, X, Y) + Value *NewSel = Builder.CreateSelect(Cond, X, Y, "sel", &Sel); + return new ShuffleVectorInst(X, NewSel, Mask); + } + if (Y == TVal) { + // select Cond, Y, (shuf_sel X, Y) --> shuf_sel (select Cond, Y, X), Y + Value *NewSel = Builder.CreateSelect(Cond, Y, X, "sel", &Sel); + return new ShuffleVectorInst(NewSel, Y, Mask); + } + } + return nullptr; } Index: llvm/test/Transforms/InstCombine/select-select.ll =================================================================== --- llvm/test/Transforms/InstCombine/select-select.ll +++ llvm/test/Transforms/InstCombine/select-select.ll @@ -44,10 +44,12 @@ ret <2 x i32> %sel1 } +; Four variations of (select (select-shuffle)) with a common operand. + define <4 x i8> @sel_shuf_commute0(<4 x i8> %x, <4 x i8> %y, <4 x i1> %cmp) { ; CHECK-LABEL: @sel_shuf_commute0( -; CHECK-NEXT: [[BLEND:%.*]] = shufflevector <4 x i8> [[X:%.*]], <4 x i8> [[Y:%.*]], <4 x i32> -; CHECK-NEXT: [[R:%.*]] = select <4 x i1> [[CMP:%.*]], <4 x i8> [[BLEND]], <4 x i8> [[X]] +; CHECK-NEXT: [[SEL:%.*]] = select <4 x i1> [[CMP:%.*]], <4 x i8> [[Y:%.*]], <4 x i8> [[X:%.*]] +; CHECK-NEXT: [[R:%.*]] = shufflevector <4 x i8> [[X]], <4 x i8> [[SEL]], <4 x i32> ; CHECK-NEXT: ret <4 x i8> [[R]] ; %blend = shufflevector <4 x i8> %x, <4 x i8> %y, <4 x i32> @@ -55,10 +57,12 @@ ret <4 x i8> %r } +; Weird types are ok. + define <5 x i9> @sel_shuf_commute1(<5 x i9> %x, <5 x i9> %y, <5 x i1> %cmp) { ; CHECK-LABEL: @sel_shuf_commute1( -; CHECK-NEXT: [[BLEND:%.*]] = shufflevector <5 x i9> [[X:%.*]], <5 x i9> [[Y:%.*]], <5 x i32> -; CHECK-NEXT: [[R:%.*]] = select <5 x i1> [[CMP:%.*]], <5 x i9> [[BLEND]], <5 x i9> [[Y]] +; CHECK-NEXT: [[SEL:%.*]] = select <5 x i1> [[CMP:%.*]], <5 x i9> [[X:%.*]], <5 x i9> [[Y:%.*]] +; CHECK-NEXT: [[R:%.*]] = shufflevector <5 x i9> [[SEL]], <5 x i9> [[Y]], <5 x i32> ; CHECK-NEXT: ret <5 x i9> [[R]] ; %blend = shufflevector <5 x i9> %x, <5 x i9> %y, <5 x i32> @@ -68,8 +72,8 @@ define <4 x float> @sel_shuf_commute2(<4 x float> %x, <4 x float> %y, <4 x i1> %cmp) { ; CHECK-LABEL: @sel_shuf_commute2( -; CHECK-NEXT: [[BLEND:%.*]] = shufflevector <4 x float> [[X:%.*]], <4 x float> [[Y:%.*]], <4 x i32> -; CHECK-NEXT: [[R:%.*]] = select <4 x i1> [[CMP:%.*]], <4 x float> [[X]], <4 x float> [[BLEND]] +; CHECK-NEXT: [[SEL:%.*]] = select <4 x i1> [[CMP:%.*]], <4 x float> [[X:%.*]], <4 x float> [[Y:%.*]] +; CHECK-NEXT: [[R:%.*]] = shufflevector <4 x float> [[X]], <4 x float> [[SEL]], <4 x i32> ; CHECK-NEXT: ret <4 x float> [[R]] ; %blend = shufflevector <4 x float> %x, <4 x float> %y, <4 x i32> @@ -77,10 +81,12 @@ ret <4 x float> %r } +; Scalar condition is ok. + define <4 x i8> @sel_shuf_commute3(<4 x i8> %x, <4 x i8> %y, i1 %cmp) { ; CHECK-LABEL: @sel_shuf_commute3( -; CHECK-NEXT: [[BLEND:%.*]] = shufflevector <4 x i8> [[X:%.*]], <4 x i8> [[Y:%.*]], <4 x i32> -; CHECK-NEXT: [[R:%.*]] = select i1 [[CMP:%.*]], <4 x i8> [[Y]], <4 x i8> [[BLEND]] +; CHECK-NEXT: [[SEL:%.*]] = select i1 [[CMP:%.*]], <4 x i8> [[Y:%.*]], <4 x i8> [[X:%.*]] +; CHECK-NEXT: [[R:%.*]] = shufflevector <4 x i8> [[SEL]], <4 x i8> [[Y]], <4 x i32> ; CHECK-NEXT: ret <4 x i8> [[R]] ; %blend = shufflevector <4 x i8> %x, <4 x i8> %y, <4 x i32> @@ -90,6 +96,8 @@ declare void @use(<4 x i8>) +; Negative test - extra use would require another instruction. + define <4 x i8> @sel_shuf_use(<4 x i8> %x, <4 x i8> %y, <4 x i1> %cmp) { ; CHECK-LABEL: @sel_shuf_use( ; CHECK-NEXT: [[BLEND:%.*]] = shufflevector <4 x i8> [[X:%.*]], <4 x i8> [[Y:%.*]], <4 x i32> @@ -103,6 +111,8 @@ ret <4 x i8> %r } +; Negative test - undef in shuffle mask prevents transform. + define <4 x i8> @sel_shuf_undef(<4 x i8> %x, <4 x i8> %y, <4 x i1> %cmp) { ; CHECK-LABEL: @sel_shuf_undef( ; CHECK-NEXT: [[BLEND:%.*]] = shufflevector <4 x i8> [[X:%.*]], <4 x i8> [[Y:%.*]], <4 x i32> @@ -114,6 +124,8 @@ ret <4 x i8> %r } +; Negative test - not a "select shuffle" + define <4 x i8> @sel_shuf_not(<4 x i8> %x, <4 x i8> %y, <4 x i1> %cmp) { ; CHECK-LABEL: @sel_shuf_not( ; CHECK-NEXT: [[NOTBLEND:%.*]] = shufflevector <4 x i8> [[X:%.*]], <4 x i8> [[Y:%.*]], <4 x i32> @@ -125,6 +137,8 @@ ret <4 x i8> %r } +; Negative test - must shuffle one of the select operands + define <4 x i8> @sel_shuf_no_common_operand(<4 x i8> %x, <4 x i8> %y, <4 x i1> %cmp, <4 x i8> %z) { ; CHECK-LABEL: @sel_shuf_no_common_operand( ; CHECK-NEXT: [[BLEND:%.*]] = shufflevector <4 x i8> [[X:%.*]], <4 x i8> [[Y:%.*]], <4 x i32> Index: llvm/test/Transforms/InstCombine/vec_demanded_elts.ll =================================================================== --- llvm/test/Transforms/InstCombine/vec_demanded_elts.ll +++ llvm/test/Transforms/InstCombine/vec_demanded_elts.ll @@ -677,9 +677,8 @@ define <4 x i8> @select_cond_with_eq_true_false_elts(<4 x i8> %x, <4 x i8> %y, <4 x i1> %cmp) { ; CHECK-LABEL: @select_cond_with_eq_true_false_elts( -; CHECK-NEXT: [[TVAL:%.*]] = shufflevector <4 x i8> [[X:%.*]], <4 x i8> [[Y:%.*]], <4 x i32> -; CHECK-NEXT: [[SPLAT:%.*]] = shufflevector <4 x i1> [[CMP:%.*]], <4 x i1> undef, <4 x i32> zeroinitializer -; CHECK-NEXT: [[R:%.*]] = select <4 x i1> [[SPLAT]], <4 x i8> [[TVAL]], <4 x i8> [[Y]] +; CHECK-NEXT: [[SEL:%.*]] = select <4 x i1> [[CMP:%.*]], <4 x i8> [[X:%.*]], <4 x i8> [[Y:%.*]] +; CHECK-NEXT: [[R:%.*]] = shufflevector <4 x i8> [[SEL]], <4 x i8> [[Y]], <4 x i32> ; CHECK-NEXT: ret <4 x i8> [[R]] ; %tval = shufflevector <4 x i8> %x, <4 x i8> %y, <4 x i32> @@ -692,9 +691,9 @@ define <4 x i8> @select_cond_with_eq_true_false_elts2(<4 x i8> %x, <4 x i8> %y, <4 x i1> %cmp) { ; CHECK-LABEL: @select_cond_with_eq_true_false_elts2( -; CHECK-NEXT: [[TVAL:%.*]] = shufflevector <4 x i8> [[X:%.*]], <4 x i8> [[Y:%.*]], <4 x i32> -; CHECK-NEXT: [[COND:%.*]] = shufflevector <4 x i1> [[CMP:%.*]], <4 x i1> undef, <4 x i32> -; CHECK-NEXT: [[R:%.*]] = select <4 x i1> [[COND]], <4 x i8> [[TVAL]], <4 x i8> [[X]] +; CHECK-NEXT: [[COND:%.*]] = shufflevector <4 x i1> [[CMP:%.*]], <4 x i1> undef, <4 x i32> +; CHECK-NEXT: [[SEL:%.*]] = select <4 x i1> [[COND]], <4 x i8> [[Y:%.*]], <4 x i8> [[X:%.*]] +; CHECK-NEXT: [[R:%.*]] = shufflevector <4 x i8> [[X]], <4 x i8> [[SEL]], <4 x i32> ; CHECK-NEXT: ret <4 x i8> [[R]] ; %tval = shufflevector <4 x i8> %x, <4 x i8> %y, <4 x i32> @@ -734,13 +733,12 @@ ret <4 x i8> %r } -; The insert can not be safely eliminated because cmp[0] might be poison. +; The insert can be safely eliminated because the shuffle blocks poison from cmp[0]. define <4 x i8> @select_cond_(<4 x i8> %x, <4 x i8> %min, <4 x i1> %cmp, i1 %poison_blocker) { ; CHECK-LABEL: @select_cond_( -; CHECK-NEXT: [[INS:%.*]] = insertelement <4 x i1> [[CMP:%.*]], i1 [[POISON_BLOCKER:%.*]], i32 0 -; CHECK-NEXT: [[VECINS:%.*]] = shufflevector <4 x i8> [[X:%.*]], <4 x i8> [[MIN:%.*]], <4 x i32> -; CHECK-NEXT: [[R:%.*]] = select <4 x i1> [[INS]], <4 x i8> [[VECINS]], <4 x i8> [[X]] +; CHECK-NEXT: [[SEL:%.*]] = select <4 x i1> [[CMP:%.*]], <4 x i8> [[MIN:%.*]], <4 x i8> [[X:%.*]] +; CHECK-NEXT: [[R:%.*]] = shufflevector <4 x i8> [[X]], <4 x i8> [[SEL]], <4 x i32> ; CHECK-NEXT: ret <4 x i8> [[R]] ; %ins = insertelement <4 x i1> %cmp, i1 %poison_blocker, i32 0