From 4d8e47ca180557e9bb872da0b8a6ad0fc515c4d3 Mon Sep 17 00:00:00 2001 From: Roman Lebedev Date: Sat, 4 Jan 2020 16:31:18 +0300 Subject: [PATCH] [InstCombine] Sink sub into hands of select if one hand becomes zero (PR44426) This decreases use count of %Op1, makes one hand of select to be 0, and possibly exposes further folding potential. Name: sub (select %Cond, %Op1, %FalseVal), %Op1 -> select %Cond, 0, (sub %FalseVal, %Op1) %Op1 = %TrueVal %o = select i1 %Cond, i8 %Op1, i8 %FalseVal %r = sub i8 %o, %Op1 => %n = sub i8 %FalseVal, %Op1 %r = select i1 %Cond, i8 0, i8 %n Name: sub (select %Cond, %TrueVal, %Op1), %Op1 -> select %Cond, (sub %TrueVal, %Op1), 0 %Op1 = %FalseVal %o = select i1 %Cond, i8 %TrueVal, i8 %Op1 %r = sub i8 %o, %Op1 => %n = sub i8 %TrueVal, %Op1 %r = select i1 %Cond, i8 %n, i8 0 https://rise4fun.com/Alive/avL https://bugs.llvm.org/show_bug.cgi?id=44426 --- .../InstCombine/InstCombineAddSub.cpp | 27 +++++++++++++++++++ .../test/Transforms/InstCombine/sub-minmax.ll | 26 +++++++----------- .../subtract-of-one-hand-of-select.ll | 12 ++++----- 3 files changed, 43 insertions(+), 22 deletions(-) diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp index 9676c6abef00..05d13aeabe52 100644 --- a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp +++ b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp @@ -1898,6 +1898,33 @@ Instruction *InstCombiner::visitSub(BinaryOperator &I) { } } + { + // If we are subtracting from select with one hand matching the value + // we are subtracting, sink subtraction into hands of select: + // sub (select %Cond, %TrueVal, %FalseVal), %Op1 + // -> + // select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1) + // This will result in select between new subtraction and 0. + Value *Cond, *TrueVal, *FalseVal; + if (match(Op0, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal), + m_Value(FalseVal)))) && + (Op1 == TrueVal || Op1 == FalseVal)) { + // While it is really tempting to just create two subtractions and let + // InstCombine fold one of those to 0, it isn't possible to do so + // because of worklist visitation order. So ugly it is. + bool SubtractingTrueVal = Op1 == TrueVal; + Value *NewSub = + Builder.CreateSub(SubtractingTrueVal ? FalseVal : TrueVal, Op1); + Constant *Zero = Constant::getNullValue(I.getType()); + SelectInst *NewSel = + SelectInst::Create(Cond, SubtractingTrueVal ? Zero : NewSub, + SubtractingTrueVal ? NewSub : Zero); + // Preserve prof metadata if any. + NewSel->copyMetadata(cast(*Op0)); + return NewSel; + } + } + if (Op1->hasOneUse()) { Value *X = nullptr, *Y = nullptr, *Z = nullptr; Constant *C = nullptr; diff --git a/llvm/test/Transforms/InstCombine/sub-minmax.ll b/llvm/test/Transforms/InstCombine/sub-minmax.ll index ccc3483ce95b..7fb5e31d81ed 100644 --- a/llvm/test/Transforms/InstCombine/sub-minmax.ll +++ b/llvm/test/Transforms/InstCombine/sub-minmax.ll @@ -4,10 +4,9 @@ define i32 @max_na_b_minux_na(i32 %A, i32 %B) { ; CHECK-LABEL: @max_na_b_minux_na( ; CHECK-NEXT: [[NOT:%.*]] = xor i32 [[A:%.*]], -1 -; CHECK-NEXT: [[L0:%.*]] = icmp ult i32 [[NOT]], [[B:%.*]] -; CHECK-NEXT: [[L1:%.*]] = select i1 [[L0]], i32 [[NOT]], i32 [[B]] -; CHECK-NEXT: [[X:%.*]] = sub i32 [[L1]], [[NOT]] -; CHECK-NEXT: ret i32 [[X]] +; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.usub.sat.i32(i32 [[NOT]], i32 [[B:%.*]]) +; CHECK-NEXT: [[TMP2:%.*]] = sub i32 0, [[TMP1]] +; CHECK-NEXT: ret i32 [[TMP2]] ; %not = xor i32 %A, -1 %l0 = icmp ult i32 %not, %B @@ -34,10 +33,9 @@ define i32 @na_minus_max_na_b(i32 %A, i32 %B) { define i32 @max_b_na_minus_na(i32 %A, i32 %B) { ; CHECK-LABEL: @max_b_na_minus_na( ; CHECK-NEXT: [[NOT:%.*]] = xor i32 [[A:%.*]], -1 -; CHECK-NEXT: [[L0:%.*]] = icmp ugt i32 [[NOT]], [[B:%.*]] -; CHECK-NEXT: [[L1:%.*]] = select i1 [[L0]], i32 [[B]], i32 [[NOT]] -; CHECK-NEXT: [[X:%.*]] = sub i32 [[L1]], [[NOT]] -; CHECK-NEXT: ret i32 [[X]] +; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.usub.sat.i32(i32 [[NOT]], i32 [[B:%.*]]) +; CHECK-NEXT: [[TMP2:%.*]] = sub i32 0, [[TMP1]] +; CHECK-NEXT: ret i32 [[TMP2]] ; %not = xor i32 %A, -1 %l0 = icmp ugt i32 %not, %B @@ -79,10 +77,8 @@ define i32 @max_na_bi_minux_na(i32 %A, i32 %Bi) { define i32 @na_minus_max_na_bi(i32 %A, i32 %Bi) { ; CHECK-LABEL: @na_minus_max_na_bi( -; CHECK-NEXT: [[TMP1:%.*]] = icmp ult i32 [[A:%.*]], [[BI:%.*]] -; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], i32 [[BI]], i32 [[A]] -; CHECK-NEXT: [[X:%.*]] = sub i32 [[TMP2]], [[A]] -; CHECK-NEXT: ret i32 [[X]] +; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.usub.sat.i32(i32 [[BI:%.*]], i32 [[A:%.*]]) +; CHECK-NEXT: ret i32 [[TMP1]] ; %B = xor i32 %Bi, -1 %not = xor i32 %A, -1 @@ -109,10 +105,8 @@ define i32 @max_bi_na_minus_na(i32 %A, i32 %Bi) { define i32 @na_minus_max_bi_na(i32 %A, i32 %Bi) { ; CHECK-LABEL: @na_minus_max_bi_na( -; CHECK-NEXT: [[TMP1:%.*]] = icmp ult i32 [[A:%.*]], [[BI:%.*]] -; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], i32 [[BI]], i32 [[A]] -; CHECK-NEXT: [[X:%.*]] = sub i32 [[TMP2]], [[A]] -; CHECK-NEXT: ret i32 [[X]] +; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.usub.sat.i32(i32 [[BI:%.*]], i32 [[A:%.*]]) +; CHECK-NEXT: ret i32 [[TMP1]] ; %B = xor i32 %Bi, -1 %not = xor i32 %A, -1 diff --git a/llvm/test/Transforms/InstCombine/subtract-of-one-hand-of-select.ll b/llvm/test/Transforms/InstCombine/subtract-of-one-hand-of-select.ll index d4d5e496952d..931a7e9d2c09 100644 --- a/llvm/test/Transforms/InstCombine/subtract-of-one-hand-of-select.ll +++ b/llvm/test/Transforms/InstCombine/subtract-of-one-hand-of-select.ll @@ -12,8 +12,8 @@ define i8 @t0_sub_of_trueval(i1 %c, i8 %Op1, i8 %FalseVal) { ; CHECK-LABEL: @t0_sub_of_trueval( -; CHECK-NEXT: [[O:%.*]] = select i1 [[C:%.*]], i8 [[OP1:%.*]], i8 [[FALSEVAL:%.*]], !prof !0 -; CHECK-NEXT: [[R:%.*]] = sub i8 [[O]], [[OP1]] +; CHECK-NEXT: [[TMP1:%.*]] = sub i8 [[FALSEVAL:%.*]], [[OP1:%.*]] +; CHECK-NEXT: [[R:%.*]] = select i1 [[C:%.*]], i8 0, i8 [[TMP1]], !prof !0 ; CHECK-NEXT: ret i8 [[R]] ; %o = select i1 %c, i8 %Op1, i8 %FalseVal, !prof !0 ; while there, ensure preservation of prof md @@ -22,8 +22,8 @@ define i8 @t0_sub_of_trueval(i1 %c, i8 %Op1, i8 %FalseVal) { } define i8 @t1_sub_of_falseval(i1 %c, i8 %TrueVal, i8 %Op1) { ; CHECK-LABEL: @t1_sub_of_falseval( -; CHECK-NEXT: [[O:%.*]] = select i1 [[C:%.*]], i8 [[TRUEVAL:%.*]], i8 [[OP1:%.*]], !prof !0 -; CHECK-NEXT: [[R:%.*]] = sub i8 [[O]], [[OP1]] +; CHECK-NEXT: [[TMP1:%.*]] = sub i8 [[TRUEVAL:%.*]], [[OP1:%.*]] +; CHECK-NEXT: [[R:%.*]] = select i1 [[C:%.*]], i8 [[TMP1]], i8 0, !prof !0 ; CHECK-NEXT: ret i8 [[R]] ; %o = select i1 %c, i8 %TrueVal, i8 %Op1, !prof !0 ; while there, ensure preservation of prof md @@ -35,8 +35,8 @@ define i8 @t1_sub_of_falseval(i1 %c, i8 %TrueVal, i8 %Op1) { define <2 x i8> @t2_vec(i1 %c, <2 x i8> %Op1, <2 x i8> %FalseVal) { ; CHECK-LABEL: @t2_vec( -; CHECK-NEXT: [[O:%.*]] = select i1 [[C:%.*]], <2 x i8> [[OP1:%.*]], <2 x i8> [[FALSEVAL:%.*]] -; CHECK-NEXT: [[R:%.*]] = sub <2 x i8> [[O]], [[OP1]] +; CHECK-NEXT: [[TMP1:%.*]] = sub <2 x i8> [[FALSEVAL:%.*]], [[OP1:%.*]] +; CHECK-NEXT: [[R:%.*]] = select i1 [[C:%.*]], <2 x i8> zeroinitializer, <2 x i8> [[TMP1]] ; CHECK-NEXT: ret <2 x i8> [[R]] ; %o = select i1 %c, <2 x i8> %Op1, <2 x i8> %FalseVal