From: Michael Berg Date: Fri, 14 Jun 2019 23:30:52 +0000 (+0000) Subject: adding more fmf propagation for selects plus tests X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=aad9d5e5dd9bccf64421d544db81a8b55108667f;p=llvm adding more fmf propagation for selects plus tests git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363474 91177308-0d34-0410-b5e6-96231b3b80d8 --- diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp index 477a5136b70..39ad5fe15a1 100644 --- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp +++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp @@ -2091,7 +2091,9 @@ SDValue DAGCombiner::foldBinOpIntoSelect(SDNode *BO) { !isConstantFPBuildVectorOrConstantFP(NewCF)) return SDValue(); - return DAG.getSelect(DL, VT, Sel.getOperand(0), NewCT, NewCF); + SDValue SelectOp = DAG.getSelect(DL, VT, Sel.getOperand(0), NewCT, NewCF); + SelectOp->setFlags(BO->getFlags()); + return SelectOp; } static SDValue foldAddSubBoolOfMaskedVal(SDNode *N, SelectionDAG &DAG) { @@ -7997,6 +7999,7 @@ SDValue DAGCombiner::visitSELECT(SDNode *N) { EVT VT = N->getValueType(0); EVT VT0 = N0.getValueType(); SDLoc DL(N); + SDNodeFlags Flags = N->getFlags(); if (SDValue V = DAG.simplifySelect(N0, N1, N2)) return V; @@ -8047,10 +8050,10 @@ SDValue DAGCombiner::visitSELECT(SDNode *N) { SDValue Cond0 = N0->getOperand(0); SDValue Cond1 = N0->getOperand(1); SDValue InnerSelect = - DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Cond1, N1, N2); + DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Cond1, N1, N2, Flags); if (normalizeToSequence || !InnerSelect.use_empty()) return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Cond0, - InnerSelect, N2); + InnerSelect, N2, Flags); // Cleanup on failure. if (InnerSelect.use_empty()) recursivelyDeleteUnusedNodes(InnerSelect.getNode()); @@ -8059,11 +8062,11 @@ SDValue DAGCombiner::visitSELECT(SDNode *N) { if (N0->getOpcode() == ISD::OR && N0->hasOneUse()) { SDValue Cond0 = N0->getOperand(0); SDValue Cond1 = N0->getOperand(1); - SDValue InnerSelect = - DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Cond1, N1, N2); + SDValue InnerSelect = DAG.getNode(ISD::SELECT, DL, N1.getValueType(), + Cond1, N1, N2, Flags); if (normalizeToSequence || !InnerSelect.use_empty()) return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Cond0, N1, - InnerSelect); + InnerSelect, Flags); // Cleanup on failure. if (InnerSelect.use_empty()) recursivelyDeleteUnusedNodes(InnerSelect.getNode()); @@ -8078,12 +8081,14 @@ SDValue DAGCombiner::visitSELECT(SDNode *N) { // Create the actual and node if we can generate good code for it. if (!normalizeToSequence) { SDValue And = DAG.getNode(ISD::AND, DL, N0.getValueType(), N0, N1_0); - return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), And, N1_1, N2); + return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), And, N1_1, + N2, Flags); } // Otherwise see if we can optimize the "and" to a better pattern. - if (SDValue Combined = visitANDLike(N0, N1_0, N)) + if (SDValue Combined = visitANDLike(N0, N1_0, N)) { return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Combined, N1_1, - N2); + N2, Flags); + } } } // select Cond0, X, (select Cond1, X, Y) -> select (or Cond0, Cond1), X, Y @@ -8095,19 +8100,23 @@ SDValue DAGCombiner::visitSELECT(SDNode *N) { // Create the actual or node if we can generate good code for it. if (!normalizeToSequence) { SDValue Or = DAG.getNode(ISD::OR, DL, N0.getValueType(), N0, N2_0); - return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Or, N1, N2_2); + return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Or, N1, + N2_2, Flags); } // Otherwise see if we can optimize to a better pattern. if (SDValue Combined = visitORLike(N0, N2_0, N)) return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Combined, N1, - N2_2); + N2_2, Flags); } } } // select (not Cond), N1, N2 -> select Cond, N2, N1 - if (SDValue F = extractBooleanFlip(N0, TLI)) - return DAG.getSelect(DL, VT, F, N2, N1); + if (SDValue F = extractBooleanFlip(N0, TLI)) { + SDValue SelectOp = DAG.getSelect(DL, VT, F, N2, N1); + SelectOp->setFlags(Flags); + return SelectOp; + } // Fold selects based on a setcc into other things, such as min/max/abs. if (N0.getOpcode() == ISD::SETCC) { @@ -8157,7 +8166,7 @@ SDValue DAGCombiner::visitSELECT(SDNode *N) { TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))) { // Any flags available in a select/setcc fold will be on the setcc as they // migrated from fcmp - const SDNodeFlags Flags = N0.getNode()->getFlags(); + Flags = N0.getNode()->getFlags(); SDValue SelectNode = DAG.getNode(ISD::SELECT_CC, DL, VT, Cond0, Cond1, N1, N2, N0.getOperand(2)); SelectNode->setFlags(Flags); @@ -8743,9 +8752,11 @@ SDValue DAGCombiner::visitSELECT_CC(SDNode *N) { return N2; } else if (SCC.getOpcode() == ISD::SETCC) { // Fold to a simpler select_cc - return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(), - SCC.getOperand(0), SCC.getOperand(1), N2, N3, - SCC.getOperand(2)); + SDValue SelectOp = DAG.getNode( + ISD::SELECT_CC, SDLoc(N), N2.getValueType(), SCC.getOperand(0), + SCC.getOperand(1), N2, N3, SCC.getOperand(2)); + SelectOp->setFlags(SCC->getFlags()); + return SelectOp; } } @@ -19412,13 +19423,16 @@ SDValue DAGCombiner::SimplifySelect(const SDLoc &DL, SDValue N0, SDValue N1, // Check to see if we got a select_cc back (to turn into setcc/select). // Otherwise, just return whatever node we got back, like fabs. if (SCC.getOpcode() == ISD::SELECT_CC) { + const SDNodeFlags Flags = N0.getNode()->getFlags(); SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0), N0.getValueType(), SCC.getOperand(0), SCC.getOperand(1), - SCC.getOperand(4)); + SCC.getOperand(4), Flags); AddToWorklist(SETCC.getNode()); - return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC, - SCC.getOperand(2), SCC.getOperand(3)); + SDValue SelectNode = DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC, + SCC.getOperand(2), SCC.getOperand(3)); + SelectNode->setFlags(Flags); + return SelectNode; } return SCC; diff --git a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp index cb64b7a81cd..1ef90445994 100644 --- a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp +++ b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp @@ -3591,6 +3591,7 @@ bool SelectionDAGLegalize::ExpandNode(SDNode *Node) { // Use the new condition code and swap true and false Legalized = true; Tmp1 = DAG.getSelectCC(dl, Tmp1, Tmp2, Tmp4, Tmp3, InvCC); + Tmp1->setFlags(Node->getFlags()); } else { // If The inverse is not legal, then try to swap the arguments using // the inverse condition code. @@ -3600,6 +3601,7 @@ bool SelectionDAGLegalize::ExpandNode(SDNode *Node) { // lhs and rhs. Legalized = true; Tmp1 = DAG.getSelectCC(dl, Tmp2, Tmp1, Tmp4, Tmp3, SwapInvCC); + Tmp1->setFlags(Node->getFlags()); } } @@ -3626,6 +3628,7 @@ bool SelectionDAGLegalize::ExpandNode(SDNode *Node) { Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0), Tmp1, Tmp2, Tmp3, Tmp4, CC); } + Tmp1->setFlags(Node->getFlags()); } Results.push_back(Tmp1); break; diff --git a/test/CodeGen/AArch64/arm64-fmax.ll b/test/CodeGen/AArch64/arm64-fmax.ll index 8337d299ea5..3488defb2e1 100644 --- a/test/CodeGen/AArch64/arm64-fmax.ll +++ b/test/CodeGen/AArch64/arm64-fmax.ll @@ -1,4 +1,5 @@ -; RUN: llc < %s -mtriple=arm64-eabi -enable-no-nans-fp-math | FileCheck %s +; RUN: llc < %s -mtriple=arm64-eabi -enable-no-nans-fp-math | FileCheck %s --check-prefix=CHECK +; RUN: llc < %s -mtriple=arm64-eabi -enable-no-nans-fp-math -debug-only=isel -o /dev/null 2>&1 | FileCheck %s --check-prefix=FMFDEBUG define double @test_direct(float %in) { ; CHECK-LABEL: test_direct: @@ -20,6 +21,7 @@ define double @test_cross(float %in) { ; CHECK: fmin } + ; Same as previous, but with ordered comparison; ; can't be converted in safe-math mode. define double @test_cross_fail_nan(float %in) { @@ -32,6 +34,10 @@ define double @test_cross_fail_nan(float %in) { ; CHECK: fmin } +; FMFDEBUG-LABEL: Optimized lowered selection DAG: %bb.0 'test_cross_fail:' +; FMFDEBUG: select_cc nnan {{t[0-9]+}}, {{t[0-9]+}}, {{t[0-9]+}}, {{t[0-9]+}}, setne:ch +; FMFDEBUG: Type-legalized selection DAG: %bb.0 'test_cross_fail:' + ; This isn't a min or a max, but passes the first condition for swapping the ; results. Make sure they're put back before we resort to the normal fcsel. define float @test_cross_fail(float %lhs, float %rhs) { diff --git a/test/CodeGen/AArch64/select_fmf.ll b/test/CodeGen/AArch64/select_fmf.ll new file mode 100644 index 00000000000..f152423f4b3 --- /dev/null +++ b/test/CodeGen/AArch64/select_fmf.ll @@ -0,0 +1,75 @@ +; RUN: llc < %s -mtriple=arm64-- -debug-only=isel -o /dev/null 2>&1 | FileCheck %s --check-prefix=FMFDEBUG + +; This test provides fmf coverage for DAG combining of selects + +; FMFDEBUG-LABEL: Optimized lowered selection DAG: %bb.0 'select_select_fold_select_and:' +; FMFDEBUG: [[AND:t[0-9]+]]: i1 = and {{t[0-9]+}}, {{t[0-9]+}} +; FMFDEBUG: [[FMAX:t[0-9]+]]: f32 = fmaxnum nnan ninf nsz arcp contract afn reassoc {{t[0-9]+}}, {{t[0-9]+}} +; FMFDEBUG-NEXT: select nnan ninf nsz arcp contract afn reassoc [[AND]], [[FMAX]], {{t[0-9]+}} +; FMFDEBUG: Type-legalized selection DAG: %bb.0 'select_select_fold_select_and:' + +; select Cond0, (select Cond1, X, Y), Y -> select (and Cond0, Cond1), X, Y +define float @select_select_fold_select_and(float %w, float %x, float %y, float %z) { + %tmp21 = fcmp fast olt float %x, %y + %tmp22 = select fast i1 %tmp21, float %x, float %y + %tmp24 = fcmp fast ogt float %tmp22, %w + %tmp78 = fcmp fast ogt float %w, %z + %select0 = select fast i1 %tmp78, float %w, float %z + %select1 = select fast i1 %tmp21, float %select0, float %w + %select2 = select fast i1 %tmp24, float %select1, float %w + %tmp82 = fadd fast float %w, 5.000000e-01 + %tmp102 = fadd fast float %tmp82, %select2 + %cmp.i155.i.i = fcmp fast ogt float %tmp102, %tmp82 + br i1 %cmp.i155.i.i, label %if.then.i157.i.i, label %if.end.i159.i.i + +if.then.i157.i.i: ; preds = %0 + %add.i156.i.i = fadd fast float %select2, 1.000000e+00 + br label %exit + +if.end.i159.i.i: ; preds = %0 + %sub.i158.i.i = fadd fast float %w, 0xBFD99999A0000000 + %sub15.i.i.i = fadd fast float %z, 0xBFD6666660000000 + %tmp191 = fcmp fast ogt float %tmp82, 0.000000e+00 + %select3 = select fast i1 %tmp191, float %sub.i158.i.i, float %sub15.i.i.i + br label %exit + +exit: ; preds = %if.end.i159.i.i, %if.then.i157.i.i + %phi1 = phi float [ %add.i156.i.i, %if.then.i157.i.i ], [ %select3, %if.end.i159.i.i ] + ret float %phi1 +} + +; FMFDEBUG-LABEL: Optimized lowered selection DAG: %bb.0 'select_select_fold_select_or:' +; FMFDEBUG: [[OR:t[0-9]+]]: i1 = or {{t[0-9]+}}, {{t[0-9]+}} +; FMFDEBUG: [[FMAX:t[0-9]+]]: f32 = fmaxnum nnan ninf nsz arcp contract afn reassoc {{t[0-9]+}}, {{t[0-9]+}} +; FMFDEBUG-NEXT: select nnan ninf nsz arcp contract afn reassoc [[OR]], {{t[0-9]+}}, [[FMAX]] +; FMFDEBUG: Type-legalized selection DAG: %bb.0 'select_select_fold_select_or:' + +; select Cond0, X, (select Cond1, X, Y) -> select (or Cond0, Cond1), X, Y +define float @select_select_fold_select_or(float %w, float %x, float %y, float %z) { + %tmp21 = fcmp fast olt float %x, %y + %tmp22 = select fast i1 %tmp21, float %x, float %y + %tmp24 = fcmp fast ogt float %tmp22, %w + %tmp78 = fcmp fast ogt float %w, %z + %select0 = select fast i1 %tmp78, float %w, float %z + %select1 = select fast i1 %tmp21, float %w, float %select0 + %select2 = select fast i1 %tmp24, float %w, float %select1 + %tmp82 = fadd fast float %w, 5.000000e-01 + %tmp102 = fadd fast float %tmp82, %select2 + %cmp.i155.i.i = fcmp fast ogt float %tmp102, %tmp82 + br i1 %cmp.i155.i.i, label %if.then.i157.i.i, label %if.end.i159.i.i + +if.then.i157.i.i: ; preds = %0 + %add.i156.i.i = fadd fast float %select2, 1.000000e+00 + br label %exit + +if.end.i159.i.i: ; preds = %0 + %sub.i158.i.i = fadd fast float %w, 0xBFD99999A0000000 + %sub15.i.i.i = fadd fast float %z, 0xBFD6666660000000 + %tmp191 = fcmp fast ogt float %tmp82, 0.000000e+00 + %select3 = select fast i1 %tmp191, float %sub.i158.i.i, float %sub15.i.i.i + br label %exit + +exit: ; preds = %if.end.i159.i.i, %if.then.i157.i.i + %phi1 = phi float [ %add.i156.i.i, %if.then.i157.i.i ], [ %select3, %if.end.i159.i.i ] + ret float %phi1 +} diff --git a/test/CodeGen/X86/fdiv-combine.ll b/test/CodeGen/X86/fdiv-combine.ll index cd2b959ee0d..586c68f5afb 100644 --- a/test/CodeGen/X86/fdiv-combine.ll +++ b/test/CodeGen/X86/fdiv-combine.ll @@ -1,5 +1,6 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64 | FileCheck %s +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -debug-only=isel -o /dev/null 2>&1 | FileCheck %s --check-prefix=FMFDEBUG ; More than one 'arcp' division using a single divisor operand ; should be converted into a reciprocal and multiplication. @@ -96,6 +97,10 @@ define double @div3_arcp(double %x, double %y, double %z) { ret double %ret } +; FMFDEBUG-LABEL: Optimized lowered selection DAG: %bb.0 'div_select_constant_fold:' +; FMFDEBUG: select nnan {{t[0-9]+}}, ConstantFP:f32<2.500000e+00>, ConstantFP:f32<3.000000e+00> +; FMFDEBUG: Type-legalized selection DAG: %bb.0 'div_select_constant_fold:' + define float @div_select_constant_fold(i1 zeroext %arg) { ; CHECK-LABEL: div_select_constant_fold: ; CHECK: # %bb.0: @@ -108,7 +113,7 @@ define float @div_select_constant_fold(i1 zeroext %arg) { ; CHECK-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero ; CHECK-NEXT: retq %tmp = select i1 %arg, float 5.000000e+00, float 6.000000e+00 - %B2 = fdiv float %tmp, 1.000000e+00 + %B2 = fdiv nnan float %tmp, 2.000000e+00 ret float %B2 }