[llvm] [RISCV][VLOPT] Add support for Vector Fixed-Point Arithmetic Instructions (PR #126483)

via llvm-commits llvm-commits at lists.llvm.org
Mon Feb 10 00:32:16 PST 2025


llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

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

Author: Luke Lau (lukel97)

<details>
<summary>Changes</summary>

This patch adds the remaining support for fixed-point arithmetic instructions (we previously had support for averaging adds and subtracts).

For saturating adds/subs/multiplies/clips, we can't change `vl` if `vxsat` if used, since changing `vl` may change its value. So this patch checks to see if it's dead before considering it a candidate.

---

Patch is 25.34 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/126483.diff


3 Files Affected:

- (modified) llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp (+37-1) 
- (modified) llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll (+508) 
- (modified) llvm/test/CodeGen/RISCV/rvv/vl-opt.mir (+20) 


``````````diff
diff --git a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
index d4829bced247091..2d0e2a423bc4816 100644
--- a/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
+++ b/llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp
@@ -976,6 +976,17 @@ static bool isSupportedInstr(const MachineInstr &MI) {
   case RISCV::VMV_V_I:
   case RISCV::VMV_V_X:
   case RISCV::VMV_V_V:
+  // Vector Single-Width Saturating Add and Subtract
+  case RISCV::VSADDU_VV:
+  case RISCV::VSADDU_VX:
+  case RISCV::VSADDU_VI:
+  case RISCV::VSADD_VV:
+  case RISCV::VSADD_VX:
+  case RISCV::VSADD_VI:
+  case RISCV::VSSUBU_VV:
+  case RISCV::VSSUBU_VX:
+  case RISCV::VSSUB_VV:
+  case RISCV::VSSUB_VX:
   // Vector Single-Width Averaging Add and Subtract
   case RISCV::VAADDU_VV:
   case RISCV::VAADDU_VX:
@@ -985,6 +996,23 @@ static bool isSupportedInstr(const MachineInstr &MI) {
   case RISCV::VASUBU_VX:
   case RISCV::VASUB_VV:
   case RISCV::VASUB_VX:
+  // Vector Single-Width Fractional Multiply with Rounding and Saturation
+  case RISCV::VSMUL_VV:
+  case RISCV::VSMUL_VX:
+  // Vector Single-Width Scaling Shift Instructions
+  case RISCV::VSSRL_VV:
+  case RISCV::VSSRL_VX:
+  case RISCV::VSSRL_VI:
+  case RISCV::VSSRA_VV:
+  case RISCV::VSSRA_VX:
+  case RISCV::VSSRA_VI:
+  // Vector Narrowing Fixed-Point Clip Instructions
+  case RISCV::VNCLIPU_WV:
+  case RISCV::VNCLIPU_WX:
+  case RISCV::VNCLIPU_WI:
+  case RISCV::VNCLIP_WV:
+  case RISCV::VNCLIP_WX:
+  case RISCV::VNCLIP_WI:
 
   // Vector Crypto
   case RISCV::VWSLL_VI:
@@ -1173,8 +1201,16 @@ bool RISCVVLOptimizer::isCandidate(const MachineInstr &MI) const {
   const MCInstrDesc &Desc = MI.getDesc();
   if (!RISCVII::hasVLOp(Desc.TSFlags) || !RISCVII::hasSEWOp(Desc.TSFlags))
     return false;
-  if (MI.getNumDefs() != 1)
+
+  if (MI.getNumExplicitDefs() != 1)
+    return false;
+
+  // Some instructions have implicit defs e.g. $vxsat. If they might be read
+  // later then we can't reduce VL.
+  if (!MI.allImplicitDefsAreDead()) {
+    LLVM_DEBUG(dbgs() << "Not a candidate because has non-dead implicit def\n");
     return false;
+  }
 
   if (MI.mayRaiseFPException()) {
     LLVM_DEBUG(dbgs() << "Not a candidate because may raise FP exception\n");
diff --git a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
index 053f1209cf21464..2411e5359b40347 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll
@@ -2511,6 +2511,514 @@ define <vscale x 4 x i32> @vwmaccus_vx(<vscale x 4 x i32> %a, i16 %b, <vscale x
   ret <vscale x 4 x i32> %2
 }
 
+define <vscale x 4 x i32> @vsaddu_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
+; NOVLOPT-LABEL: vsaddu_vv:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vsaddu.vv v10, v8, v10
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vsaddu_vv:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vsaddu.vv v10, v8, v10
+; VLOPT-NEXT:    vadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vsaddu(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vsaddu_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
+; NOVLOPT-LABEL: vsaddu_vx:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a2, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vsaddu.vx v10, v8, a0
+; NOVLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vsaddu_vx:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; VLOPT-NEXT:    vsaddu.vx v10, v8, a0
+; VLOPT-NEXT:    vadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vsaddu(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, i32 %b, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vsaddu_vi(<vscale x 4 x i32> %a, iXLen %vl) {
+; NOVLOPT-LABEL: vsaddu_vi:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vsaddu.vi v10, v8, 5
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vsaddu_vi:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vsaddu.vi v10, v8, 5
+; VLOPT-NEXT:    vadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vsaddu(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, i32 5, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vsadd_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
+; NOVLOPT-LABEL: vsadd_vv:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vsadd.vv v10, v8, v10
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vsadd_vv:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vsadd.vv v10, v8, v10
+; VLOPT-NEXT:    vadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vsadd(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vsadd_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
+; NOVLOPT-LABEL: vsadd_vx:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a2, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vsadd.vx v10, v8, a0
+; NOVLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vsadd_vx:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; VLOPT-NEXT:    vsadd.vx v10, v8, a0
+; VLOPT-NEXT:    vadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vsadd(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, i32 %b, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vsadd_vi(<vscale x 4 x i32> %a, iXLen %vl) {
+; NOVLOPT-LABEL: vsadd_vi:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vsadd.vi v10, v8, 5
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vsadd_vi:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vsadd.vi v10, v8, 5
+; VLOPT-NEXT:    vadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vsadd(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, i32 5, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vssubu_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
+; NOVLOPT-LABEL: vssubu_vv:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vssubu.vv v10, v8, v10
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vssubu_vv:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vssubu.vv v10, v8, v10
+; VLOPT-NEXT:    vadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vssubu(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vssubu_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
+; NOVLOPT-LABEL: vssubu_vx:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a2, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vssubu.vx v10, v8, a0
+; NOVLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vssubu_vx:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; VLOPT-NEXT:    vssubu.vx v10, v8, a0
+; VLOPT-NEXT:    vadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vssubu(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, i32 %b, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vssub_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
+; NOVLOPT-LABEL: vssub_vv:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vssub.vv v10, v8, v10
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vssub_vv:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vssub.vv v10, v8, v10
+; VLOPT-NEXT:    vadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vssub(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vssub_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
+; NOVLOPT-LABEL: vssub_vx:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    vsetvli a2, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vssub.vx v10, v8, a0
+; NOVLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vssub_vx:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; VLOPT-NEXT:    vssub.vx v10, v8, a0
+; VLOPT-NEXT:    vadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vssub(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, i32 %b, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vsmul_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
+; NOVLOPT-LABEL: vsmul_vv:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    csrwi vxrm, 0
+; NOVLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vsmul.vv v8, v8, v10
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v8, v10
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vsmul_vv:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    csrwi vxrm, 0
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vsmul.vv v8, v8, v10
+; VLOPT-NEXT:    vadd.vv v8, v8, v10
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vsmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen 0, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %b, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vsmul_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
+; NOVLOPT-LABEL: vsmul_vx:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    csrwi vxrm, 0
+; NOVLOPT-NEXT:    vsetvli a2, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vsmul.vx v10, v8, a0
+; NOVLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vsmul_vx:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    csrwi vxrm, 0
+; VLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; VLOPT-NEXT:    vsmul.vx v10, v8, a0
+; VLOPT-NEXT:    vadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vsmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, i32 %b, iXLen 0, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vssrl_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
+; NOVLOPT-LABEL: vssrl_vv:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    csrwi vxrm, 0
+; NOVLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vssrl.vv v8, v8, v10
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v8, v10
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vssrl_vv:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    csrwi vxrm, 0
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vssrl.vv v8, v8, v10
+; VLOPT-NEXT:    vadd.vv v8, v8, v10
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vssrl.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen 0, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %b, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vssrl_vx(<vscale x 4 x i32> %a, iXLen %b, iXLen %vl) {
+; NOVLOPT-LABEL: vssrl_vx:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    csrwi vxrm, 0
+; NOVLOPT-NEXT:    vsetvli a2, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vssrl.vx v10, v8, a0
+; NOVLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vssrl_vx:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    csrwi vxrm, 0
+; VLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; VLOPT-NEXT:    vssrl.vx v10, v8, a0
+; VLOPT-NEXT:    vadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vssrl.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, iXLen %b, iXLen 0, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vssrl_vi(<vscale x 4 x i32> %a, iXLen %vl) {
+; NOVLOPT-LABEL: vssrl_vi:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    csrwi vxrm, 0
+; NOVLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vssrl.vi v10, v8, 5
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vssrl_vi:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    csrwi vxrm, 0
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vssrl.vi v10, v8, 5
+; VLOPT-NEXT:    vadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vssrl.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, iXLen 5, iXLen 0, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vssra_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
+; NOVLOPT-LABEL: vssra_vv:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    csrwi vxrm, 0
+; NOVLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vssra.vv v8, v8, v10
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v8, v10
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vssra_vv:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    csrwi vxrm, 0
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vssra.vv v8, v8, v10
+; VLOPT-NEXT:    vadd.vv v8, v8, v10
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vssra.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen 0, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %b, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vssra_vx(<vscale x 4 x i32> %a, iXLen %b, iXLen %vl) {
+; NOVLOPT-LABEL: vssra_vx:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    csrwi vxrm, 0
+; NOVLOPT-NEXT:    vsetvli a2, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vssra.vx v10, v8, a0
+; NOVLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vssra_vx:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    csrwi vxrm, 0
+; VLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; VLOPT-NEXT:    vssra.vx v10, v8, a0
+; VLOPT-NEXT:    vadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vssra.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, iXLen %b, iXLen 0, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vssra_vi(<vscale x 4 x i32> %a, iXLen %vl) {
+; NOVLOPT-LABEL: vssra_vi:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    csrwi vxrm, 0
+; NOVLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vssra.vi v10, v8, 5
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v10, v8
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vssra_vi:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    csrwi vxrm, 0
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vssra.vi v10, v8, 5
+; VLOPT-NEXT:    vadd.vv v8, v10, v8
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vssra.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, iXLen 5, iXLen 0, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vnclipu_vv(<vscale x 4 x i64> %a, <vscale x 4 x i32> %b, iXLen %vl) {
+; NOVLOPT-LABEL: vnclipu_vv:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    csrwi vxrm, 0
+; NOVLOPT-NEXT:    vsetvli a1, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vnclipu.wv v14, v8, v12
+; NOVLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v14, v14
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vnclipu_vv:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    csrwi vxrm, 0
+; VLOPT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
+; VLOPT-NEXT:    vnclipu.wv v14, v8, v12
+; VLOPT-NEXT:    vadd.vv v8, v14, v14
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vnclipu(<vscale x 4 x i32> poison, <vscale x 4 x i64> %a, <vscale x 4 x i32> %b, iXLen 0, iXLen -1)
+  %2 = call <vscale x 4 x i32> @llvm.riscv.vadd(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %1, iXLen %vl)
+  ret <vscale x 4 x i32> %2
+}
+
+define <vscale x 4 x i32> @vnclipu_vx(<vscale x 4 x i64> %a, iXLen %b, iXLen %vl) {
+; NOVLOPT-LABEL: vnclipu_vx:
+; NOVLOPT:       # %bb.0:
+; NOVLOPT-NEXT:    csrwi vxrm, 0
+; NOVLOPT-NEXT:    vsetvli a2, zero, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vnclipu.wx v12, v8, a0
+; NOVLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; NOVLOPT-NEXT:    vadd.vv v8, v12, v12
+; NOVLOPT-NEXT:    ret
+;
+; VLOPT-LABEL: vnclipu_vx:
+; VLOPT:       # %bb.0:
+; VLOPT-NEXT:    csrwi vxrm, 0
+; VLOPT-NEXT:    vsetvli zero, a1, e32, m2, ta, ma
+; VLOPT-NEXT:    vnclipu.wx v12, v8, a0
+; VLOPT-NEXT:    vadd.vv v8, v12, v12
+; VLOPT-NEXT:    ret
+  %1 = call <vscale x 4 x i32> @llvm.riscv.vnclipu(<vscale x 4 x i32> poison, <vs...
[truncated]

``````````

</details>


https://github.com/llvm/llvm-project/pull/126483


More information about the llvm-commits mailing list