[RISCV] Add patterns for vector widening floating-point multiply

[jacquesguan]

Summary

Add patterns for vector widening floating-point multiply

[craig.topper]

LGTM

llvm/lib/Target/RISCV/RISCVInstrInfoVSDPatterns.td

 defm : VPatWidenBinaryFPSDNode_VV_VF_WV_WF<fadd, "PseudoVFWADD">;
 defm : VPatWidenBinaryFPSDNode_VV_VF_WV_WF<fsub, "PseudoVFWSUB">;
 
 // 14.4. Vector Single-Width Floating-Point Multiply/Divide Instructions
 defm : VPatBinaryFPSDNode_VV_VF<fmul, "PseudoVFMUL">;
 defm : VPatBinaryFPSDNode_VV_VF<fdiv, "PseudoVFDIV">;
 defm : VPatBinaryFPSDNode_R_VF<fdiv, "PseudoVFRDIV">;
 
+// 14.5. Vector Widening Floating-Point Multiply Instructions
+defm : VPatWidenBinaryFPSDNode_VV_VF<fmul, "PseudoVFWMUL">;
+
 // 14.6 Vector Single-Width Floating-Point Fused Multiply-Add Instructions.
 foreach fvti = AllFloatVectors in {
   // NOTE: We choose VFMADD because it has the most commuting freedom. So it
   // works best with how TwoAddressInstructionPass tries commuting.
   defvar suffix = fvti.LMul.MX;
   def : Pat<(fvti.Vector (fma fvti.RegClass:$rs1, fvti.RegClass:$rd,
                               fvti.RegClass:$rs2)),
             (!cast<Instruction>("PseudoVFMADD_VV_"# suffix)

llvm/test/CodeGen/RISCV/rvv/vfwmul-sdnode.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv32 -mattr=+d,+zfh,+experimental-v -target-abi=ilp32d \
+; RUN:     -verify-machineinstrs < %s | FileCheck %s
+; RUN: llc -mtriple=riscv64 -mattr=+d,+zfh,+experimental-v -target-abi=lp64d \
+; RUN:     -verify-machineinstrs < %s | FileCheck %s
+
+define <vscale x 1 x double> @vfwmul_vv_nxv1f64(<vscale x 1 x float> %va, <vscale x 1 x float> %vb) {
+; CHECK-LABEL: vfwmul_vv_nxv1f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, mf2, ta, mu
+; CHECK-NEXT:    vfwmul.vv v10, v8, v9
+; CHECK-NEXT:    vmv1r.v v8, v10
+; CHECK-NEXT:    ret
+  %vc = fpext <vscale x 1 x float> %va to <vscale x 1 x double>
+  %vd = fpext <vscale x 1 x float> %vb to <vscale x 1 x double>
+  %ve = fmul <vscale x 1 x double> %vc, %vd
+  ret <vscale x 1 x double> %ve
+}
+
+define <vscale x 1 x double> @vfwmul_vf_nxv1f64(<vscale x 1 x float> %va, float %b) {
+; CHECK-LABEL: vfwmul_vf_nxv1f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, mf2, ta, mu
+; CHECK-NEXT:    vfwmul.vf v9, v8, fa0
+; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    ret
+  %head = insertelement <vscale x 1 x float> undef, float %b, i32 0
+  %splat = shufflevector <vscale x 1 x float> %head, <vscale x 1 x float> undef, <vscale x 1 x i32> zeroinitializer
+  %vc = fpext <vscale x 1 x float> %va to <vscale x 1 x double>
+  %vd = fpext <vscale x 1 x float> %splat to <vscale x 1 x double>
+  %ve = fmul <vscale x 1 x double> %vc, %vd
+  ret <vscale x 1 x double> %ve
+}
+
+define <vscale x 2 x double> @vfwmul_vv_nxv2f64(<vscale x 2 x float> %va, <vscale x 2 x float> %vb) {
+; CHECK-LABEL: vfwmul_vv_nxv2f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m1, ta, mu
+; CHECK-NEXT:    vfwmul.vv v10, v8, v9
+; CHECK-NEXT:    vmv2r.v v8, v10
+; CHECK-NEXT:    ret
+  %vc = fpext <vscale x 2 x float> %va to <vscale x 2 x double>
+  %vd = fpext <vscale x 2 x float> %vb to <vscale x 2 x double>
+  %ve = fmul <vscale x 2 x double> %vc, %vd
+  ret <vscale x 2 x double> %ve
+}
+
+define <vscale x 2 x double> @vfwmul_vf_nxv2f64(<vscale x 2 x float> %va, float %b) {
+; CHECK-LABEL: vfwmul_vf_nxv2f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m1, ta, mu
+; CHECK-NEXT:    vfwmul.vf v10, v8, fa0
+; CHECK-NEXT:    vmv2r.v v8, v10
+; CHECK-NEXT:    ret
+  %head = insertelement <vscale x 2 x float> undef, float %b, i32 0
+  %splat = shufflevector <vscale x 2 x float> %head, <vscale x 2 x float> undef, <vscale x 2 x i32> zeroinitializer
+  %vc = fpext <vscale x 2 x float> %va to <vscale x 2 x double>
+  %vd = fpext <vscale x 2 x float> %splat to <vscale x 2 x double>
+  %ve = fmul <vscale x 2 x double> %vc, %vd
+  ret <vscale x 2 x double> %ve
+}
+
+define <vscale x 4 x double> @vfwmul_vv_nxv4f64(<vscale x 4 x float> %va, <vscale x 4 x float> %vb) {
+; CHECK-LABEL: vfwmul_vv_nxv4f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m2, ta, mu
+; CHECK-NEXT:    vfwmul.vv v12, v8, v10
+; CHECK-NEXT:    vmv4r.v v8, v12
+; CHECK-NEXT:    ret
+  %vc = fpext <vscale x 4 x float> %va to <vscale x 4 x double>
+  %vd = fpext <vscale x 4 x float> %vb to <vscale x 4 x double>
+  %ve = fmul <vscale x 4 x double> %vc, %vd
+  ret <vscale x 4 x double> %ve
+}
+
+define <vscale x 4 x double> @vfwmul_vf_nxv4f64(<vscale x 4 x float> %va, float %b) {
+; CHECK-LABEL: vfwmul_vf_nxv4f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m2, ta, mu
+; CHECK-NEXT:    vfwmul.vf v12, v8, fa0
+; CHECK-NEXT:    vmv4r.v v8, v12
+; CHECK-NEXT:    ret
+  %head = insertelement <vscale x 4 x float> undef, float %b, i32 0
+  %splat = shufflevector <vscale x 4 x float> %head, <vscale x 4 x float> undef, <vscale x 4 x i32> zeroinitializer
+  %vc = fpext <vscale x 4 x float> %va to <vscale x 4 x double>
+  %vd = fpext <vscale x 4 x float> %splat to <vscale x 4 x double>
+  %ve = fmul <vscale x 4 x double> %vc, %vd
+  ret <vscale x 4 x double> %ve
+}
+
+define <vscale x 8 x double> @vfwmul_vv_nxv8f64(<vscale x 8 x float> %va, <vscale x 8 x float> %vb) {
+; CHECK-LABEL: vfwmul_vv_nxv8f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m4, ta, mu
+; CHECK-NEXT:    vfwmul.vv v16, v8, v12
+; CHECK-NEXT:    vmv8r.v v8, v16
+; CHECK-NEXT:    ret
+  %vc = fpext <vscale x 8 x float> %va to <vscale x 8 x double>
+  %vd = fpext <vscale x 8 x float> %vb to <vscale x 8 x double>
+  %ve = fmul <vscale x 8 x double> %vc, %vd
+  ret <vscale x 8 x double> %ve
+}
+
+define <vscale x 8 x double> @vfwmul_vf_nxv8f64(<vscale x 8 x float> %va, float %b) {
+; CHECK-LABEL: vfwmul_vf_nxv8f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m4, ta, mu
+; CHECK-NEXT:    vfwmul.vf v16, v8, fa0
+; CHECK-NEXT:    vmv8r.v v8, v16
+; CHECK-NEXT:    ret
+  %head = insertelement <vscale x 8 x float> undef, float %b, i32 0
+  %splat = shufflevector <vscale x 8 x float> %head, <vscale x 8 x float> undef, <vscale x 8 x i32> zeroinitializer
+  %vc = fpext <vscale x 8 x float> %va to <vscale x 8 x double>
+  %vd = fpext <vscale x 8 x float> %splat to <vscale x 8 x double>
+  %ve = fmul <vscale x 8 x double> %vc, %vd
+  ret <vscale x 8 x double> %ve
+}