[llvm] [RISCV][VLOPT] Add support for vfmerge.vfm and vfmv.v.f (PR #146692)

[Luke Lau via llvm-commits]

https://github.com/lukel97 created https://github.com/llvm/llvm-project/pull/146692

I noticed these were missing when seeing some extra vl toggles with EVL tail folding.




>From 35c518aaee21321e57425154028facb12704085d Mon Sep 17 00:00:00 2001
From: Luke Lau <luke at igalia.com>
Date: Wed, 2 Jul 2025 14:19:43 +0100
Subject: [PATCH 1/2] Precommit tests

---
 llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll | 42 ++++++++++++++++++++
 1 file changed, 42 insertions(+)

diff --git a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
index 1cbb980aebffc..74807534a1294 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
@@ -4540,6 +4540,48 @@ define <vscale x 4 x float> @vfsgnjx_vf(<vscale x 4 x float> %a, float %b, iXLen
   ret <vscale x 4 x float> %2
 }
 
+define <vscale x 4 x float> @vfmerge_vf(<vscale x 4 x float> %a, float %b, <vscale x 4 x i1> %c, iXLen %vl) {
+; NOVLOPT-LABEL: vfmerge_vf:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vfmerge.vfm v10, v8, fa0, v0
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vfadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vfmerge_vf:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; VLOPT-NEXT:    vfmerge.vfm v10, v8, fa0, v0
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vfadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x float> @llvm.riscv.vfmerge(<vscale x 4 x float> poison, <vscale x 4 x float> %a, float %b, <vscale x 4 x i1> %c, iXLen -1)
+  %2 = call <vscale x 4 x float> @llvm.riscv.vfadd.nxv4f32.nxv4f32(<vscale x 4 x float> poison, <vscale x 4 x float> %1, <vscale x 4 x float> %a, iXLen 7, iXLen %vl)
+  ret <vscale x 4 x float> %2
+}
+
+define <vscale x 4 x float> @vfmv_v_f(<vscale x 4 x float> %a, float %b, iXLen %vl) {
+; NOVLOPT-LABEL: vfmv_v_f:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vfmv.v.f v10, fa0
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vfadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vfmv_v_f:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; VLOPT-NEXT:    vfmv.v.f v10, fa0
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vfadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x float> @llvm.riscv.vfmv.v.f(<vscale x 4 x float> poison, float %b, iXLen -1)
+  %2 = call <vscale x 4 x float> @llvm.riscv.vfadd.nxv4f32.nxv4f32(<vscale x 4 x float> poison, <vscale x 4 x float> %1, <vscale x 4 x float> %a, iXLen 7, iXLen %vl)
+  ret <vscale x 4 x float> %2
+}
+
 define <vscale x 4 x float> @vfmacc_vv(<vscale x 4 x float> %a, <vscale x 4 x float> %b, <vscale x 4 x float> %c, iXLen %vl) {
 ; NOVLOPT-LABEL: vfmacc_vv:
 ; NOVLOPT:       # %bb.0:

>From f59af75c83b4c12083d933343fc8cd93cc024558 Mon Sep 17 00:00:00 2001
From: Luke Lau <luke at igalia.com>
Date: Wed, 2 Jul 2025 14:22:28 +0100
Subject: [PATCH 2/2] [RISCV][VLOPT] Add support for vfmerge.vfm and vfmv.v.f

I noticed these were missing when seeing some extra vl toggles with EVL tail folding.
---
 llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp   | 4 ++++
 llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll | 6 ++----
 2 files changed, 6 insertions(+), 4 deletions(-)

diff --git a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
index d7aa85048eeda..9852b6b755144 100644
--- a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
+++ b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
@@ -1119,6 +1119,10 @@ static bool isSupportedInstr(const MachineInstr &MI) {
   case RISCV::VMFLE_VV:
   case RISCV::VMFGT_VF:
   case RISCV::VMFGE_VF:
+  // Vector Floating-Point Merge Instruction
+  case RISCV::VFMERGE_VFM:
+  // Vector Floating-Point Move Instruction
+  case RISCV::VFMV_V_F:
   // Single-Width Floating-Point/Integer Type-Convert Instructions
   case RISCV::VFCVT_XU_F_V:
   case RISCV::VFCVT_X_F_V:
diff --git a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
index 74807534a1294..f4f194de644a0 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
@@ -4551,9 +4551,8 @@ define <vscale x 4 x float> @vfmerge_vf(<vscale x 4 x float> %a, float %b, <vsca
 ;
 ; VLOPT-LABEL: vfmerge_vf:
 ; VLOPT:       # %bb.0:
-; VLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
-; VLOPT-NEXT:    vfmerge.vfm v10, v8, fa0, v0
 ; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vfmerge.vfm v10, v8, fa0, v0
 ; VLOPT-NEXT:    vfadd.vv v8, v10, v8
 ; VLOPT-NEXT:    ret
   %1 = call <vscale x 4 x float> @llvm.riscv.vfmerge(<vscale x 4 x float> poison, <vscale x 4 x float> %a, float %b, <vscale x 4 x i1> %c, iXLen -1)
@@ -4572,9 +4571,8 @@ define <vscale x 4 x float> @vfmv_v_f(<vscale x 4 x float> %a, float %b, iXLen %
 ;
 ; VLOPT-LABEL: vfmv_v_f:
 ; VLOPT:       # %bb.0:
-; VLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
-; VLOPT-NEXT:    vfmv.v.f v10, fa0
 ; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vfmv.v.f v10, fa0
 ; VLOPT-NEXT:    vfadd.vv v8, v10, v8
 ; VLOPT-NEXT:    ret
   %1 = call <vscale x 4 x float> @llvm.riscv.vfmv.v.f(<vscale x 4 x float> poison, float %b, iXLen -1)