[RISCV] Add LLVM IR intrinsics and codegen for XSfvfexp* and XSfvfexpa*

llvm/include/llvm/IR/IntrinsicsRISCVXsf.td

@@ -162,6 +162,10 @@ let TargetPrefix = "riscv" in {
   defm "" : RISCVSFCustomVC_XVV<["x", "i", "v", "f"]>;
   defm "" : RISCVSFCustomVC_XVW<["x", "i", "v", "f"]>;
 
+  // XSfvfexp* and XSfvfexpa*
+  defm sf_vfexp : RISCVUnaryAA;
+  defm sf_vfexpa : RISCVUnaryAA;
+
   // XSfvqmaccdod
   def int_riscv_sf_vqmaccu_2x8x2  : RISCVSFCustomVMACC;
   def int_riscv_sf_vqmacc_2x8x2   : RISCVSFCustomVMACC;

llvm/lib/Target/RISCV/RISCVInstrInfoXSf.td

@@ -218,11 +218,13 @@ let Predicates = [HasVendorXSfvcp], mayLoad = 0, mayStore = 0,
 }
 
 let Predicates = [HasVendorXSfvfexpAny], DecoderNamespace = "XSfvector" in {
-  def SF_VFEXP_V : VALUVs2<0b010011, 0b00111, OPFVV, "sf.vfexp.v">;
+  def SF_VFEXP_V : VALUVs2<0b010011, 0b00111, OPFVV, "sf.vfexp.v">,
+                   SchedUnaryMC<"WriteSF_VFExp", "ReadSF_VFExp">;
 }
 
 let Predicates = [HasVendorXSfvfexpa], DecoderNamespace = "XSfvector" in {
-  def SF_VFEXPA_V : VALUVs2<0b010011, 0b00110, OPFVV, "sf.vfexpa.v">;
+  def SF_VFEXPA_V : VALUVs2<0b010011, 0b00110, OPFVV, "sf.vfexpa.v">,
+                    SchedUnaryMC<"WriteSF_VFExpa", "ReadSF_VFExpa">;
 }
 
 let Predicates = [HasVendorXSfvqmaccdod], DecoderNamespace = "XSfvector",
@@ -487,6 +489,48 @@ let Predicates = [HasVendorXSfvfnrclipxfqf] in {
   defm SF_VFNRCLIP_X_F_QF : VPseudoSiFiveVFNRCLIP;
 }
 
+class VFExpSchedSEWSet<string mx, bit IsBF16, bit IsApprox> {
+  defvar BaseSet = SchedSEWSet<mx, isF=1>.val;
+  list<int> val = !if(IsBF16, !listremove(BaseSet, [32, 64]),
+                      !if(IsApprox, BaseSet, !listremove(BaseSet, [64])));
+}
+multiclass VPseudoVFExp_V<bit IsBF16 = false, bit IsApprox = false> {
+  defvar SchedSuffix = !if(IsApprox, "VFExpa", "VFExp");
+
+  foreach m = MxListF in {
+    defvar mx = m.MX;
+    foreach e = VFExpSchedSEWSet<mx, IsBF16, IsApprox>.val in {
+      let VLMul = m.value in {
+        def "_V_" # mx # "_E" # e
+            : VPseudoUnaryNoMask<m.vrclass, m.vrclass>,
+              SchedUnary<"WriteSF_" # SchedSuffix, "ReadSF_" # SchedSuffix,
+                         mx, e, forcePassthruRead=true>;
+        def "_V_" # mx # "_E" # e # "_MASK"
+            : VPseudoUnaryMask<m.vrclass, m.vrclass>,
+              RISCVMaskedPseudo<MaskIdx = 2>,
+              SchedUnary<"WriteSF_" # SchedSuffix, "ReadSF_" # SchedSuffix,
+                         mx, e, forcePassthruRead=true>;
+      }
+    }
+  }
+}
+
+let Predicates = [HasVendorXSfvfbfexp16e], hasSideEffects = 0 in {
+  let AltFmtType = IS_ALTFMT in {
+    defm PseudoSF_VFEXP_ALT : VPseudoVFExp_V<IsBF16=true>;
+  }
+}
+
+let Predicates = [HasVendorXSfvfexpAnyFloat], hasSideEffects = 0 in {
+  let AltFmtType = IS_NOT_ALTFMT in {
+    defm PseudoSF_VFEXP : VPseudoVFExp_V;
+  }
+}
+
+let Predicates = [HasVendorXSfvfexpa], AltFmtType = IS_NOT_ALTFMT in {
+  defm PseudoSF_VFEXPA : VPseudoVFExp_V<IsApprox=true>;
+}
+
 // SDNode
 def SDT_SF_VC_V_X : SDTypeProfile<1, 4, [SDTCisVec<0>,
                                          SDTCisVT<1, XLenVT>,
@@ -893,3 +937,36 @@ let Predicates = [HasVendorXSfcease] in {
     let rs2 = 0b00101;
 }
 }
+
+let Predicates = [HasVendorXSfvfbfexp16e] in {
+  defm : VPatUnaryV_V<"int_riscv_sf_vfexp", "PseudoSF_VFEXP_ALT",
+                      AllBF16Vectors,
+                      isSEWAware=1>;
+}
+
+let Predicates = [HasVendorXSfvfexp16e] in {
+  defm : VPatUnaryV_V<"int_riscv_sf_vfexp", "PseudoSF_VFEXP",
+                      [VF16MF4, VF16MF2, VF16M1, VF16M2, VF16M4, VF16M8],
+                      isSEWAware=1>;
+}
+
+let Predicates = [HasVendorXSfvfexp32e] in {
+  defm : VPatUnaryV_V<"int_riscv_sf_vfexp", "PseudoSF_VFEXP",
+                      [VF32MF2, VF32M1, VF32M2, VF32M4, VF32M8], isSEWAware=1>;
+}
+
+let Predicates = [HasVendorXSfvfexpa] in {
+  defm : VPatUnaryV_V<"int_riscv_sf_vfexpa", "PseudoSF_VFEXPA",
+                      [VF32MF2, VF32M1, VF32M2, VF32M4, VF32M8], isSEWAware=1>;
+}
+
+let Predicates = [HasVendorXSfvfexpa, HasVInstructionsF16] in {
+  defm : VPatUnaryV_V<"int_riscv_sf_vfexpa", "PseudoSF_VFEXPA",
+                      [VF16MF4, VF16MF2, VF16M1, VF16M2, VF16M4, VF16M8],
+                      isSEWAware=1>;
+}
+
+let Predicates = [HasVendorXSfvfexpa64e] in {
+  defm : VPatUnaryV_V<"int_riscv_sf_vfexpa", "PseudoSF_VFEXPA",
+                      [VF64M1, VF64M2, VF64M4, VF64M8], isSEWAware=1>;
+}

llvm/lib/Target/RISCV/RISCVSchedSiFive7.td

@@ -1588,6 +1588,10 @@ multiclass SiFive7SchedResources<int vlen, bit dualVALU,
   //===----------------------------------------------------------------------===//
   // Unsupported extensions
   defm : UnsupportedSchedQ;
+  // TODO: scheduling info of XSfvfexp* and XSfvfexpa*
+  // for SiFive7 will be added in follow-up patches.
+  defm : UnsupportedSchedXSfvfexp;
+  defm : UnsupportedSchedXSfvfexpa;
   defm : UnsupportedSchedZabha;
   defm : UnsupportedSchedZbc;
   defm : UnsupportedSchedZbkb;

llvm/lib/Target/RISCV/RISCVSchedule.td

@@ -523,6 +523,8 @@ include "RISCVScheduleZvk.td"
 // Vendor Extensions
 multiclass UnsupportedSchedXsf {
   defm : UnsupportedSchedXsfvcp;
+  defm : UnsupportedSchedXSfvfexp;
+  defm : UnsupportedSchedXSfvfexpa;
   defm : UnsupportedSchedXSfvfnrclipxfqf;
   defm : UnsupportedSchedXSfvfwmaccqqq;
   defm : UnsupportedSchedXSfvqmaccdod;

llvm/lib/Target/RISCV/RISCVScheduleXSf.td

@@ -99,3 +99,23 @@ defm : LMULWriteRes<"WriteSF_VFWMACC_QQQ", []>;
 defm : LMULReadAdvance<"ReadSF_VFWMACC_QQQ", 0>;
 } // Unsupported = true
 }
+
+defm "" : LMULSEWSchedWritesF<"WriteSF_VFExp">;
+defm "" : LMULSEWSchedReadsF<"ReadSF_VFExp">;
+
+multiclass UnsupportedSchedXSfvfexp {
+let Unsupported = true in {
+defm : LMULSEWWriteResF<"WriteSF_VFExp", []>;
+defm : LMULSEWReadAdvanceF<"ReadSF_VFExp", 0>;
+} // Unsupported = true
+}
+
+defm "" : LMULSEWSchedWritesF<"WriteSF_VFExpa">;
+defm "" : LMULSEWSchedReadsF<"ReadSF_VFExpa">;
+
+multiclass UnsupportedSchedXSfvfexpa {
+let Unsupported = true in {
+defm : LMULSEWWriteResF<"WriteSF_VFExpa", []>;
+defm : LMULSEWReadAdvanceF<"ReadSF_VFExpa", 0>;
+} // Unsupported = true
+}

llvm/test/CodeGen/RISCV/rvv/sf_vfbfexp16e.ll

@@ -0,0 +1,191 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: sed 's/iXLen/i32/g' %s | llc -mtriple=riscv32 -mattr=+v,+zvfbfmin,+xsfvfbfexp16e \
+; RUN:   -verify-machineinstrs | FileCheck %s --check-prefixes=CHECK
+; RUN: sed 's/iXLen/i64/g' %s | llc -mtriple=riscv64 -mattr=+v,+zvfbfmin,+xsfvfbfexp16e \
+; RUN:   -verify-machineinstrs | FileCheck %s --check-prefixes=CHECK
+
+define <vscale x 1 x bfloat>  @intrinsic_sf_vfexp_v_nxv1bf16(<vscale x 1 x bfloat> %0, iXLen %1) nounwind {
+; CHECK-LABEL: intrinsic_sf_vfexp_v_nxv1bf16:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vsetvli zero, a0, e16alt, mf4, ta, ma
+; CHECK-NEXT:    sf.vfexp.v v8, v8
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 1 x bfloat> @llvm.riscv.sf.vfexp.nxv1bf16(
+    <vscale x 1 x bfloat> undef,
+    <vscale x 1 x bfloat> %0,
+    iXLen %1)
+
+  ret <vscale x 1 x bfloat> %a
+}
+
+define <vscale x 2 x bfloat>  @intrinsic_sf_vfexp_v_nxv2bf16(<vscale x 2 x bfloat> %0, iXLen %1) nounwind {
+; CHECK-LABEL: intrinsic_sf_vfexp_v_nxv2bf16:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vsetvli zero, a0, e16alt, mf2, ta, ma
+; CHECK-NEXT:    sf.vfexp.v v8, v8
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 2 x bfloat> @llvm.riscv.sf.vfexp.nxv2bf16(
+    <vscale x 2 x bfloat> undef,
+    <vscale x 2 x bfloat> %0,
+    iXLen %1)
+
+  ret <vscale x 2 x bfloat> %a
+}
+
+define <vscale x 4 x bfloat>  @intrinsic_sf_vfexp_v_nxv4bf16(<vscale x 4 x bfloat> %0, iXLen %1) nounwind {
+; CHECK-LABEL: intrinsic_sf_vfexp_v_nxv4bf16:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vsetvli zero, a0, e16alt, m1, ta, ma
+; CHECK-NEXT:    sf.vfexp.v v8, v8
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 4 x bfloat> @llvm.riscv.sf.vfexp.nxv4bf16(
+    <vscale x 4 x bfloat> undef,
+    <vscale x 4 x bfloat> %0,
+    iXLen %1)
+
+  ret <vscale x 4 x bfloat> %a
+}
+
+define <vscale x 8 x bfloat>  @intrinsic_sf_vfexp_v_nxv8bf16(<vscale x 8 x bfloat> %0, iXLen %1) nounwind {
+; CHECK-LABEL: intrinsic_sf_vfexp_v_nxv8bf16:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vsetvli zero, a0, e16alt, m2, ta, ma
+; CHECK-NEXT:    sf.vfexp.v v8, v8
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 8 x bfloat> @llvm.riscv.sf.vfexp.nxv8bf16(
+    <vscale x 8 x bfloat> undef,
+    <vscale x 8 x bfloat> %0,
+    iXLen %1)
+
+  ret <vscale x 8 x bfloat> %a
+}
+
+define <vscale x 16 x bfloat>  @intrinsic_sf_vfexp_v_nxv16bf16(<vscale x 16 x bfloat> %0, iXLen %1) nounwind {
+; CHECK-LABEL: intrinsic_sf_vfexp_v_nxv16bf16:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vsetvli zero, a0, e16alt, m4, ta, ma
+; CHECK-NEXT:    sf.vfexp.v v8, v8
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 16 x bfloat> @llvm.riscv.sf.vfexp.nxv16bf16(
+    <vscale x 16 x bfloat> undef,
+    <vscale x 16 x bfloat> %0,
+    iXLen %1)
+
+  ret <vscale x 16 x bfloat> %a
+}
+
+define <vscale x 32 x bfloat>  @intrinsic_sf_vfexp_v_nxv32bf16(<vscale x 32 x bfloat> %0, iXLen %1) nounwind {
+; CHECK-LABEL: intrinsic_sf_vfexp_v_nxv32bf16:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vsetvli zero, a0, e16alt, m8, ta, ma
+; CHECK-NEXT:    sf.vfexp.v v8, v8
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 32 x bfloat> @llvm.riscv.sf.vfexp.nxv32bf16(
+    <vscale x 32 x bfloat> undef,
+    <vscale x 32 x bfloat> %0,
+    iXLen %1)
+
+  ret <vscale x 32 x bfloat> %a
+}
+
+define <vscale x 1 x bfloat>  @intrinsic_sf_vfexp_mask_v_nxv1bf16(<vscale x 1 x bfloat> %0, <vscale x 1 x bfloat> %1, <vscale x 1 x i1> %m, iXLen %2) nounwind {
+; CHECK-LABEL: intrinsic_sf_vfexp_mask_v_nxv1bf16:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vsetvli zero, a0, e16alt, mf4, tu, mu
+; CHECK-NEXT:    sf.vfexp.v v8, v9, v0.t
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 1 x bfloat> @llvm.riscv.sf.vfexp.mask.nxv1bf16(
+    <vscale x 1 x bfloat> %0,
+    <vscale x 1 x bfloat> %1,
+    <vscale x 1 x i1> %m,
+    iXLen %2, iXLen 0)
+
+  ret <vscale x 1 x bfloat> %a
+}
+
+define <vscale x 2 x bfloat>  @intrinsic_sf_vfexp_mask_v_nxv2bf16(<vscale x 2 x bfloat> %0, <vscale x 2 x bfloat> %1, <vscale x 2 x i1> %m, iXLen %2) nounwind {
+; CHECK-LABEL: intrinsic_sf_vfexp_mask_v_nxv2bf16:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vsetvli zero, a0, e16alt, mf2, tu, mu
+; CHECK-NEXT:    sf.vfexp.v v8, v9, v0.t
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 2 x bfloat> @llvm.riscv.sf.vfexp.mask.nxv2bf16(
+    <vscale x 2 x bfloat> %0,
+    <vscale x 2 x bfloat> %1,
+    <vscale x 2 x i1> %m,
+    iXLen %2, iXLen 0)
+
+  ret <vscale x 2 x bfloat> %a
+}
+
+define <vscale x 4 x bfloat>  @intrinsic_sf_vfexp_mask_v_nxv4bf16(<vscale x 4 x bfloat> %0, <vscale x 4 x bfloat> %1, <vscale x 4 x i1> %m, iXLen %2) nounwind {
+; CHECK-LABEL: intrinsic_sf_vfexp_mask_v_nxv4bf16:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vsetvli zero, a0, e16alt, m1, tu, mu
+; CHECK-NEXT:    sf.vfexp.v v8, v9, v0.t
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 4 x bfloat> @llvm.riscv.sf.vfexp.mask.nxv4bf16(
+    <vscale x 4 x bfloat> %0,
+    <vscale x 4 x bfloat> %1,
+    <vscale x 4 x i1> %m,
+    iXLen %2, iXLen 0)
+
+  ret <vscale x 4 x bfloat> %a
+}
+
+define <vscale x 8 x bfloat>  @intrinsic_sf_vfexp_mask_v_nxv8bf16(<vscale x 8 x bfloat> %0, <vscale x 8 x bfloat> %1, <vscale x 8 x i1> %m, iXLen %2) nounwind {
+; CHECK-LABEL: intrinsic_sf_vfexp_mask_v_nxv8bf16:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vsetvli zero, a0, e16alt, m2, tu, mu
+; CHECK-NEXT:    sf.vfexp.v v8, v10, v0.t
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 8 x bfloat> @llvm.riscv.sf.vfexp.mask.nxv8bf16(
+    <vscale x 8 x bfloat> %0,
+    <vscale x 8 x bfloat> %1,
+    <vscale x 8 x i1> %m,
+    iXLen %2, iXLen 0)
+
+  ret <vscale x 8 x bfloat> %a
+}
+
+define <vscale x 16 x bfloat>  @intrinsic_sf_vfexp_mask_v_nxv16bf16(<vscale x 16 x bfloat> %0, <vscale x 16 x bfloat> %1, <vscale x 16 x i1> %m, iXLen %2) nounwind {
+; CHECK-LABEL: intrinsic_sf_vfexp_mask_v_nxv16bf16:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vsetvli zero, a0, e16alt, m4, tu, mu
+; CHECK-NEXT:    sf.vfexp.v v8, v12, v0.t
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 16 x bfloat> @llvm.riscv.sf.vfexp.mask.nxv16bf16(
+    <vscale x 16 x bfloat> %0,
+    <vscale x 16 x bfloat> %1,
+    <vscale x 16 x i1> %m,
+    iXLen %2, iXLen 0)
+
+  ret <vscale x 16 x bfloat> %a
+}
+
+define <vscale x 32 x bfloat>  @intrinsic_sf_vfexp_mask_v_nxv32bf16(<vscale x 32 x bfloat> %0, <vscale x 32 x bfloat> %1, <vscale x 32 x i1> %m, iXLen %2) nounwind {
+; CHECK-LABEL: intrinsic_sf_vfexp_mask_v_nxv32bf16:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vsetvli zero, a0, e16alt, m8, tu, mu
+; CHECK-NEXT:    sf.vfexp.v v8, v16, v0.t
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 32 x bfloat> @llvm.riscv.sf.vfexp.mask.nxv32bf16(
+    <vscale x 32 x bfloat> %0,
+    <vscale x 32 x bfloat> %1,
+    <vscale x 32 x i1> %m,
+    iXLen %2, iXLen 0)
+
+  ret <vscale x 32 x bfloat> %a
+}