[RISCV] Add more patten for FMA instruction to eliminate the fneg

[ChunyuLiao]

We use fli+fneg to generate negative float, eliminate the fneg for fma.
Fold fma to vfnmsac.vf,vfnmsub.vf, vfnmacc.vf, vfnmadd.vf

[llvmbot]

@llvm/pr-subscribers-backend-risc-v

Author: Liao Chunyu (ChunyuLiao)

Changes

We use fli+fneg to generate negative float, eliminate the fneg for fma.
Fold fma to vfnmsac.vf,vfnmsub.vf, vfnmacc.vf, vfnmadd.vf

---

Full diff: https://github.com/llvm/llvm-project/pull/173808.diff

2 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVInstrInfoVSDPatterns.td (+16)
• (modified) llvm/test/CodeGen/RISCV/rvv/vsplats-zfa.ll (+46)

diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoVSDPatterns.td b/llvm/lib/Target/RISCV/RISCVInstrInfoVSDPatterns.td
index b3cc33d31761d..552432c4b156a 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoVSDPatterns.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoVSDPatterns.td
@@ -1326,6 +1326,22 @@ foreach fvti = AllFloatVectors in {
                    // RISCVInsertReadWriteCSR
                    FRM_DYN,
                    fvti.AVL, fvti.Log2SEW, TAIL_AGNOSTIC)>;
+    def : Pat<(fvti.Vector (any_fma (SplatFPOp (fneg fvti.ScalarRegClass:$rs1)),
+                                    fvti.RegClass:$rd, (fneg fvti.RegClass:$rs2))),
+              (!cast<Instruction>("PseudoVFNMADD_V" # fvti.ScalarSuffix # "_" # suffix)
+                   fvti.RegClass:$rd, fvti.ScalarRegClass:$rs1, fvti.RegClass:$rs2,
+                   // Value to indicate no rounding mode change in
+                   // RISCVInsertReadWriteCSR
+                   FRM_DYN,
+                   fvti.AVL, fvti.Log2SEW, TAIL_AGNOSTIC)>;
+    def : Pat<(fvti.Vector (any_fma (SplatFPOp (fneg fvti.ScalarRegClass:$rs1)),
+                                    fvti.RegClass:$rd, fvti.RegClass:$rs2)),
+              (!cast<Instruction>("PseudoVFNMSUB_V" # fvti.ScalarSuffix # "_" # suffix)
+                   fvti.RegClass:$rd, fvti.ScalarRegClass:$rs1, fvti.RegClass:$rs2,
+                   // Value to indicate no rounding mode change in
+                   // RISCVInsertReadWriteCSR
+                   FRM_DYN,
+                   fvti.AVL, fvti.Log2SEW, TAIL_AGNOSTIC)>;
   }
 }
 
diff --git a/llvm/test/CodeGen/RISCV/rvv/vsplats-zfa.ll b/llvm/test/CodeGen/RISCV/rvv/vsplats-zfa.ll
index 1047860ec8db6..6f997082a3d35 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vsplats-zfa.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vsplats-zfa.ll
@@ -33,3 +33,49 @@ define <vscale x 8 x double> @vsplat_f64_neg1() {
 ; CHECK-NEXT:    ret
   ret <vscale x 8 x double> splat (double -1.0)
 }
+
+define <vscale x 4 x float> @vfnmsac(<vscale x 4 x float> %va, <vscale x 4 x float> %vb) {
+; CHECK-LABEL: vfnmsac:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    fli.s fa5, 2.0
+; CHECK-NEXT:    vsetvli a0, zero, e32, m2, ta, ma
+; CHECK-NEXT:    vfnmsac.vf v8, fa5, v10
+; CHECK-NEXT:    ret
+  %vd = tail call <vscale x 4 x float> @llvm.fmuladd.nxv4f32(<vscale x 4 x float> %vb, <vscale x 4 x float> splat (float -2.000000e+00), <vscale x 4 x float> %va)
+  ret <vscale x 4 x float> %vd
+}
+
+define <vscale x 4 x float> @vfnmsub(<vscale x 4 x float> %va, <vscale x 4 x float> %vb) {
+; CHECK-LABEL: vfnmsub:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    fli.s fa5, 2.0
+; CHECK-NEXT:    vsetvli a0, zero, e32, m2, ta, ma
+; CHECK-NEXT:    vfnmsub.vf v8, fa5, v10
+; CHECK-NEXT:    ret
+  %vd = tail call <vscale x 4 x float> @llvm.fmuladd.nxv4f32(<vscale x 4 x float> splat (float -2.000000e+00), <vscale x 4 x float> %va, <vscale x 4 x float> %vb)
+  ret <vscale x 4 x float> %vd
+}
+
+define <vscale x 8 x float> @vfnmacc(<vscale x 8 x float> %va, <vscale x 8 x float> %vb) {
+; CHECK-LABEL: vfnmacc:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    fli.s fa5, 2.0
+; CHECK-NEXT:    vsetvli a0, zero, e32, m4, ta, ma
+; CHECK-NEXT:    vfnmacc.vf v8, fa5, v12
+; CHECK-NEXT:    ret
+  %neg = fneg <vscale x 8 x float> %va
+  %vd = call <vscale x 8 x float> @llvm.fma.v8f32(<vscale x 8 x float> %vb, <vscale x 8 x float> splat (float -2.000000e+00), <vscale x 8 x float> %neg)
+  ret <vscale x 8 x float> %vd
+}
+
+define <vscale x 8 x float> @vfnmadd(<vscale x 8 x float> %va, <vscale x 8 x float> %vb) {
+; CHECK-LABEL: vfnmadd:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    fli.s fa5, 2.0
+; CHECK-NEXT:    vsetvli a0, zero, e32, m4, ta, ma
+; CHECK-NEXT:    vfnmadd.vf v8, fa5, v12
+; CHECK-NEXT:    ret
+  %neg = fneg <vscale x 8 x float> %vb
+  %vd = call <vscale x 8 x float> @llvm.fma.v8f32(<vscale x 8 x float> %va, <vscale x 8 x float> splat (float -2.000000e+00), <vscale x 8 x float> %neg)
+  ret <vscale x 8 x float> %vd
+}