[LoongArch] [CodeGen] Add options for Clang to generate LoongArch-spe…

…cific frecipe & frsqrte instructions (llvm#109917)

Two options: `-mfrecipe` & `-mno-frecipe`.
Enable or Disable frecipe.{s/d} and frsqrte.{s/d} instructions. 
The default is `-mno-frecipe`.

clang/include/clang/Driver/Options.td

@@ -5387,6 +5387,10 @@ def mno_lasx : Flag<["-"], "mno-lasx">, Group<m_loongarch_Features_Group>,
 let Flags = [TargetSpecific] in {
 def msimd_EQ : Joined<["-"], "msimd=">, Group<m_loongarch_Features_Group>,
   HelpText<"Select the SIMD extension(s) to be enabled in LoongArch either 'none', 'lsx', 'lasx'.">;
+def mfrecipe : Flag<["-"], "mfrecipe">, Group<m_loongarch_Features_Group>,
+  HelpText<"Enable frecipe.{s/d} and frsqrte.{s/d}">;
+def mno_frecipe : Flag<["-"], "mno-frecipe">, Group<m_loongarch_Features_Group>,
+  HelpText<"Disable frecipe.{s/d} and frsqrte.{s/d}">;
 def mannotate_tablejump : Flag<["-"], "mannotate-tablejump">, Group<m_loongarch_Features_Group>,
   HelpText<"Enable annotate table jump instruction to correlate it with the jump table.">;
 def mno_annotate_tablejump : Flag<["-"], "mno-annotate-tablejump">, Group<m_loongarch_Features_Group>,

clang/lib/Driver/ToolChains/Arch/LoongArch.cpp

@@ -251,6 +251,15 @@ void loongarch::getLoongArchTargetFeatures(const Driver &D,
     } else /*-mno-lasx*/
       Features.push_back("-lasx");
   }
+
+  // Select frecipe feature determined by -m[no-]frecipe.
+  if (const Arg *A =
+          Args.getLastArg(options::OPT_mfrecipe, options::OPT_mno_frecipe)) {
+    if (A->getOption().matches(options::OPT_mfrecipe))
+      Features.push_back("+frecipe");
+    else
+      Features.push_back("-frecipe");
+  }
 }
 
 std::string loongarch::postProcessTargetCPUString(const std::string &CPU,

clang/test/Driver/loongarch-mfrecipe.c

@@ -0,0 +1,30 @@
+/// Test -m[no]frecipe options.
+
+// RUN: %clang --target=loongarch64 -mfrecipe -fsyntax-only %s -### 2>&1 | \
+// RUN:     FileCheck %s --check-prefix=CC1-FRECIPE
+// RUN: %clang --target=loongarch64 -mno-frecipe -fsyntax-only %s -### 2>&1 | \
+// RUN:     FileCheck %s --check-prefix=CC1-NO-FRECIPE
+// RUN: %clang --target=loongarch64 -mno-frecipe -mfrecipe -fsyntax-only %s -### 2>&1 | \
+// RUN:     FileCheck %s --check-prefix=CC1-FRECIPE
+// RUN: %clang --target=loongarch64  -mfrecipe -mno-frecipe -fsyntax-only %s -### 2>&1 | \
+// RUN:     FileCheck %s --check-prefix=CC1-NO-FRECIPE
+
+// RUN: %clang --target=loongarch64 -mfrecipe -S -emit-llvm %s -o - | \
+// RUN: FileCheck %s --check-prefix=IR-FRECIPE
+// RUN: %clang --target=loongarch64 -mno-frecipe -S -emit-llvm %s -o - | \
+// RUN: FileCheck %s --check-prefix=IR-NO-FRECIPE
+// RUN: %clang --target=loongarch64 -mno-frecipe -mfrecipe -S -emit-llvm %s -o - | \
+// RUN: FileCheck %s --check-prefix=IR-FRECIPE
+// RUN: %clang --target=loongarch64 -mfrecipe -mno-frecipe -S -emit-llvm %s -o - | \
+// RUN: FileCheck %s --check-prefix=IR-NO-FRECIPE
+
+
+// CC1-FRECIPE: "-target-feature" "+frecipe"
+// CC1-NO-FRECIPE: "-target-feature" "-frecipe"
+
+// IR-FRECIPE: attributes #[[#]] ={{.*}}"target-features"="{{(.*,)?}}+frecipe{{(,.*)?}}"
+// IR-NO-FRECIPE: attributes #[[#]] ={{.*}}"target-features"="{{(.*,)?}}-frecipe{{(,.*)?}}"
+
+int foo(void) {
+  return 42;
+}

llvm/lib/Target/LoongArch/LoongArchFloat32InstrInfo.td

@@ -19,12 +19,16 @@ def SDT_LoongArchMOVGR2FR_W_LA64
 def SDT_LoongArchMOVFR2GR_S_LA64
     : SDTypeProfile<1, 1, [SDTCisVT<0, i64>, SDTCisVT<1, f32>]>;
 def SDT_LoongArchFTINT : SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisFP<1>]>;
+def SDT_LoongArchFRECIPE : SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisFP<1>]>;
+def SDT_LoongArchFRSQRTE : SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisFP<1>]>;
 
 def loongarch_movgr2fr_w_la64
     : SDNode<"LoongArchISD::MOVGR2FR_W_LA64", SDT_LoongArchMOVGR2FR_W_LA64>;
 def loongarch_movfr2gr_s_la64
     : SDNode<"LoongArchISD::MOVFR2GR_S_LA64", SDT_LoongArchMOVFR2GR_S_LA64>;
 def loongarch_ftint : SDNode<"LoongArchISD::FTINT", SDT_LoongArchFTINT>;
+def loongarch_frecipe : SDNode<"LoongArchISD::FRECIPE", SDT_LoongArchFRECIPE>;
+def loongarch_frsqrte : SDNode<"LoongArchISD::FRSQRTE", SDT_LoongArchFRSQRTE>;
 
 //===----------------------------------------------------------------------===//
 // Instructions
@@ -286,6 +290,8 @@ let Predicates = [HasFrecipe] in {
 // FP approximate reciprocal operation
 def : Pat<(int_loongarch_frecipe_s FPR32:$src), (FRECIPE_S FPR32:$src)>;
 def : Pat<(int_loongarch_frsqrte_s FPR32:$src), (FRSQRTE_S FPR32:$src)>;
+def : Pat<(loongarch_frecipe FPR32:$src), (FRECIPE_S FPR32:$src)>;
+def : Pat<(loongarch_frsqrte FPR32:$src), (FRSQRTE_S FPR32:$src)>;
 }
 
 // fmadd.s: fj * fk + fa

llvm/lib/Target/LoongArch/LoongArchFloat64InstrInfo.td

@@ -253,6 +253,8 @@ let Predicates = [HasFrecipe] in {
 // FP approximate reciprocal operation
 def : Pat<(int_loongarch_frecipe_d FPR64:$src), (FRECIPE_D FPR64:$src)>;
 def : Pat<(int_loongarch_frsqrte_d FPR64:$src), (FRSQRTE_D FPR64:$src)>;
+def : Pat<(loongarch_frecipe FPR64:$src), (FRECIPE_D FPR64:$src)>;
+def : Pat<(loongarch_frsqrte FPR64:$src), (FRSQRTE_D FPR64:$src)>;
 }
 
 // fmadd.d: fj * fk + fa

llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp

@@ -4697,6 +4697,8 @@ const char *LoongArchTargetLowering::getTargetNodeName(unsigned Opcode) const {
     NODE_NAME_CASE(VANY_ZERO)
     NODE_NAME_CASE(VALL_NONZERO)
     NODE_NAME_CASE(VANY_NONZERO)
+    NODE_NAME_CASE(FRECIPE)
+    NODE_NAME_CASE(FRSQRTE)
   }
 #undef NODE_NAME_CASE
   return nullptr;
@@ -5900,6 +5902,71 @@ Register LoongArchTargetLowering::getExceptionSelectorRegister(
   return LoongArch::R5;
 }
 
+//===----------------------------------------------------------------------===//
+// Target Optimization Hooks
+//===----------------------------------------------------------------------===//
+
+static int getEstimateRefinementSteps(EVT VT,
+                                      const LoongArchSubtarget &Subtarget) {
+  // Feature FRECIPE instrucions relative accuracy is 2^-14.
+  // IEEE float has 23 digits and double has 52 digits.
+  int RefinementSteps = VT.getScalarType() == MVT::f64 ? 2 : 1;
+  return RefinementSteps;
+}
+
+SDValue LoongArchTargetLowering::getSqrtEstimate(SDValue Operand,
+                                                 SelectionDAG &DAG, int Enabled,
+                                                 int &RefinementSteps,
+                                                 bool &UseOneConstNR,
+                                                 bool Reciprocal) const {
+  if (Subtarget.hasFrecipe()) {
+    SDLoc DL(Operand);
+    EVT VT = Operand.getValueType();
+
+    if (VT == MVT::f32 || (VT == MVT::f64 && Subtarget.hasBasicD()) ||
+        (VT == MVT::v4f32 && Subtarget.hasExtLSX()) ||
+        (VT == MVT::v2f64 && Subtarget.hasExtLSX()) ||
+        (VT == MVT::v8f32 && Subtarget.hasExtLASX()) ||
+        (VT == MVT::v4f64 && Subtarget.hasExtLASX())) {
+
+      if (RefinementSteps == ReciprocalEstimate::Unspecified)
+        RefinementSteps = getEstimateRefinementSteps(VT, Subtarget);
+
+      SDValue Estimate = DAG.getNode(LoongArchISD::FRSQRTE, DL, VT, Operand);
+      if (Reciprocal)
+        Estimate = DAG.getNode(ISD::FMUL, DL, VT, Operand, Estimate);
+
+      return Estimate;
+    }
+  }
+
+  return SDValue();
+}
+
+SDValue LoongArchTargetLowering::getRecipEstimate(SDValue Operand,
+                                                  SelectionDAG &DAG,
+                                                  int Enabled,
+                                                  int &RefinementSteps) const {
+  if (Subtarget.hasFrecipe()) {
+    SDLoc DL(Operand);
+    EVT VT = Operand.getValueType();
+
+    if (VT == MVT::f32 || (VT == MVT::f64 && Subtarget.hasBasicD()) ||
+        (VT == MVT::v4f32 && Subtarget.hasExtLSX()) ||
+        (VT == MVT::v2f64 && Subtarget.hasExtLSX()) ||
+        (VT == MVT::v8f32 && Subtarget.hasExtLASX()) ||
+        (VT == MVT::v4f64 && Subtarget.hasExtLASX())) {
+
+      if (RefinementSteps == ReciprocalEstimate::Unspecified)
+        RefinementSteps = getEstimateRefinementSteps(VT, Subtarget);
+
+      return DAG.getNode(LoongArchISD::FRECIPE, DL, VT, Operand);
+    }
+  }
+
+  return SDValue();
+}
+
 //===----------------------------------------------------------------------===//
 //                           LoongArch Inline Assembly Support
 //===----------------------------------------------------------------------===//

llvm/lib/Target/LoongArch/LoongArchISelLowering.h

@@ -141,6 +141,10 @@ enum NodeType : unsigned {
   VALL_NONZERO,
   VANY_NONZERO,
 
+  // Floating point approximate reciprocal operation
+  FRECIPE,
+  FRSQRTE
+
   // Intrinsic operations end =============================================
 };
 } // end namespace LoongArchISD
@@ -216,6 +220,17 @@ class LoongArchTargetLowering : public TargetLowering {
   Register
   getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
 
+  bool isFsqrtCheap(SDValue Operand, SelectionDAG &DAG) const override {
+    return true;
+  }
+
+  SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
+                          int &RefinementSteps, bool &UseOneConstNR,
+                          bool Reciprocal) const override;
+
+  SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
+                           int &RefinementSteps) const override;
+
   ISD::NodeType getExtendForAtomicOps() const override {
     return ISD::SIGN_EXTEND;
   }

llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td

@@ -10,6 +10,7 @@
 //
 //===----------------------------------------------------------------------===//
 
+// Target nodes.
 def loongarch_xvpermi: SDNode<"LoongArchISD::XVPERMI", SDT_LoongArchV1RUimm>;
 
 def lasxsplati8
@@ -2094,6 +2095,15 @@ foreach Inst = ["XVFRECIPE_S", "XVFRSQRTE_S"] in
 foreach Inst = ["XVFRECIPE_D", "XVFRSQRTE_D"] in
   def : Pat<(deriveLASXIntrinsic<Inst>.ret (v4f64 LASX256:$xj)),
             (!cast<LAInst>(Inst) LASX256:$xj)>;
+
+def : Pat<(loongarch_vfrecipe v8f32:$src), 
+          (XVFRECIPE_S v8f32:$src)>;
+def : Pat<(loongarch_vfrecipe v4f64:$src), 
+          (XVFRECIPE_D v4f64:$src)>;
+def : Pat<(loongarch_vfrsqrte v8f32:$src), 
+          (XVFRSQRTE_S v8f32:$src)>;
+def : Pat<(loongarch_vfrsqrte v4f64:$src), 
+          (XVFRSQRTE_D v4f64:$src)>;
 }
 
 def : Pat<(int_loongarch_lasx_xvpickve_w_f v8f32:$xj, timm:$imm),

llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td

@@ -23,6 +23,8 @@ def SDT_LoongArchV2R : SDTypeProfile<1, 2, [SDTCisVec<0>,
                                      SDTCisSameAs<0, 1>, SDTCisSameAs<1, 2>]>;
 def SDT_LoongArchV1RUimm: SDTypeProfile<1, 2, [SDTCisVec<0>,
                                         SDTCisSameAs<0,1>, SDTCisVT<2, i64>]>;
+def SDT_LoongArchVFRECIPE : SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisVec<0>, SDTCisSameAs<0, 1>]>;
+def SDT_LoongArchVFRSQRTE : SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisVec<0>, SDTCisSameAs<0, 1>]>;
 
 // Target nodes.
 def loongarch_vreplve : SDNode<"LoongArchISD::VREPLVE", SDT_LoongArchVreplve>;
@@ -50,6 +52,8 @@ def loongarch_vilvh: SDNode<"LoongArchISD::VILVH", SDT_LoongArchV2R>;
 
 def loongarch_vshuf4i: SDNode<"LoongArchISD::VSHUF4I", SDT_LoongArchV1RUimm>;
 def loongarch_vreplvei: SDNode<"LoongArchISD::VREPLVEI", SDT_LoongArchV1RUimm>;
+def loongarch_vfrecipe: SDNode<"LoongArchISD::FRECIPE", SDT_LoongArchVFRECIPE>;
+def loongarch_vfrsqrte: SDNode<"LoongArchISD::FRSQRTE", SDT_LoongArchVFRSQRTE>;
 
 def immZExt1 : ImmLeaf<i64, [{return isUInt<1>(Imm);}]>;
 def immZExt2 : ImmLeaf<i64, [{return isUInt<2>(Imm);}]>;
@@ -2238,6 +2242,15 @@ foreach Inst = ["VFRECIPE_S", "VFRSQRTE_S"] in
 foreach Inst = ["VFRECIPE_D", "VFRSQRTE_D"] in
   def : Pat<(deriveLSXIntrinsic<Inst>.ret (v2f64 LSX128:$vj)),
             (!cast<LAInst>(Inst) LSX128:$vj)>;
+
+def : Pat<(loongarch_vfrecipe v4f32:$src), 
+          (VFRECIPE_S v4f32:$src)>;
+def : Pat<(loongarch_vfrecipe v2f64:$src), 
+          (VFRECIPE_D v2f64:$src)>;
+def : Pat<(loongarch_vfrsqrte v4f32:$src), 
+          (VFRSQRTE_S v4f32:$src)>;
+def : Pat<(loongarch_vfrsqrte v2f64:$src), 
+          (VFRSQRTE_D v2f64:$src)>;
 }
 
 // load

llvm/test/CodeGen/LoongArch/fdiv-reciprocal-estimate.ll

@@ -0,0 +1,80 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc --mtriple=loongarch32 --mattr=+f,-d,-frecipe < %s | FileCheck %s --check-prefix=LA32F
+; RUN: llc --mtriple=loongarch32 --mattr=+f,-d,+frecipe < %s | FileCheck %s --check-prefix=LA32F-FRECIPE
+; RUN: llc --mtriple=loongarch64 --mattr=+d,-frecipe < %s | FileCheck %s --check-prefix=LA64D
+; RUN: llc --mtriple=loongarch64 --mattr=+d,+frecipe < %s | FileCheck %s --check-prefix=LA64D-FRECIPE
+
+;; Exercise the 'fdiv' LLVM IR: https://llvm.org/docs/LangRef.html#fdiv-instruction
+
+define float @fdiv_s(float %x, float %y) {
+; LA32F-LABEL: fdiv_s:
+; LA32F:       # %bb.0:
+; LA32F-NEXT:    fdiv.s	$fa0, $fa0, $fa1
+; LA32F-NEXT:    ret
+;
+; LA32F-FRECIPE-LABEL: fdiv_s:
+; LA32F-FRECIPE:       # %bb.0:
+; LA32F-FRECIPE-NEXT:    frecipe.s	$fa2, $fa1
+; LA32F-FRECIPE-NEXT:    fmul.s	$fa3, $fa0, $fa2
+; LA32F-FRECIPE-NEXT:    fnmsub.s	$fa0, $fa1, $fa3, $fa0
+; LA32F-FRECIPE-NEXT:    fmadd.s	$fa0, $fa2, $fa0, $fa3
+; LA32F-FRECIPE-NEXT:    ret
+;
+; LA64D-LABEL: fdiv_s:
+; LA64D:       # %bb.0:
+; LA64D-NEXT:    fdiv.s $fa0, $fa0, $fa1
+; LA64D-NEXT:    ret
+;
+; LA64D-FRECIPE-LABEL: fdiv_s:
+; LA64D-FRECIPE:       # %bb.0:
+; LA64D-FRECIPE-NEXT:    frecipe.s	$fa2, $fa1
+; LA64D-FRECIPE-NEXT:    fmul.s	$fa3, $fa0, $fa2
+; LA64D-FRECIPE-NEXT:    fnmsub.s	$fa0, $fa1, $fa3, $fa0
+; LA64D-FRECIPE-NEXT:    fmadd.s	$fa0, $fa2, $fa0, $fa3
+; LA64D-FRECIPE-NEXT:    ret
+  %div = fdiv fast float %x, %y
+  ret float %div
+}
+
+define double @fdiv_d(double %x, double %y) {
+; LA32F-LABEL: fdiv_d:
+; LA32F:       # %bb.0:
+; LA32F-NEXT:    addi.w	$sp, $sp, -16
+; LA32F-NEXT:    .cfi_def_cfa_offset 16
+; LA32F-NEXT:    st.w	$ra, $sp, 12                    # 4-byte Folded Spill
+; LA32F-NEXT:    .cfi_offset 1, -4
+; LA32F-NEXT:    bl	%plt(__divdf3)
+; LA32F-NEXT:    ld.w	$ra, $sp, 12                    # 4-byte Folded Reload
+; LA32F-NEXT:    addi.w	$sp, $sp, 16
+; LA32F-NEXT:    ret
+;
+; LA32F-FRECIPE-LABEL: fdiv_d:
+; LA32F-FRECIPE:       # %bb.0:
+; LA32F-FRECIPE-NEXT:    addi.w	$sp, $sp, -16
+; LA32F-FRECIPE-NEXT:    .cfi_def_cfa_offset 16
+; LA32F-FRECIPE-NEXT:    st.w	$ra, $sp, 12                    # 4-byte Folded Spill
+; LA32F-FRECIPE-NEXT:    .cfi_offset 1, -4
+; LA32F-FRECIPE-NEXT:    bl	%plt(__divdf3)
+; LA32F-FRECIPE-NEXT:    ld.w	$ra, $sp, 12                    # 4-byte Folded Reload
+; LA32F-FRECIPE-NEXT:    addi.w	$sp, $sp, 16
+; LA32F-FRECIPE-NEXT:    ret
+;
+; LA64D-LABEL: fdiv_d:
+; LA64D:       # %bb.0:
+; LA64D-NEXT:    fdiv.d $fa0, $fa0, $fa1
+; LA64D-NEXT:    ret
+;
+; LA64D-FRECIPE-LABEL: fdiv_d:
+; LA64D-FRECIPE:       # %bb.0:
+; LA64D-FRECIPE-NEXT:    pcalau12i	$a0, %pc_hi20(.LCPI1_0)
+; LA64D-FRECIPE-NEXT:    fld.d	$fa2, $a0, %pc_lo12(.LCPI1_0)
+; LA64D-FRECIPE-NEXT:    frecipe.d	$fa3, $fa1
+; LA64D-FRECIPE-NEXT:    fmadd.d	$fa2, $fa1, $fa3, $fa2
+; LA64D-FRECIPE-NEXT:    fnmsub.d	$fa2, $fa2, $fa3, $fa3
+; LA64D-FRECIPE-NEXT:    fmul.d	$fa3, $fa0, $fa2
+; LA64D-FRECIPE-NEXT:    fnmsub.d	$fa0, $fa1, $fa3, $fa0
+; LA64D-FRECIPE-NEXT:    fmadd.d	$fa0, $fa2, $fa0, $fa3
+; LA64D-FRECIPE-NEXT:    ret
+  %div = fdiv fast double %x, %y
+  ret double %div
+}