[llvm] [MachinePipeliner] Limit the number of stores in BB (PR #154940)

[Ryotaro Kasuga via llvm-commits]

https://github.com/kasuga-fj updated https://github.com/llvm/llvm-project/pull/154940

>From 67a44590b5ccae8c285d2ea618e69ea0ee9b1847 Mon Sep 17 00:00:00 2001
From: Ryotaro Kasuga <kasuga.ryotaro at fujitsu.com>
Date: Fri, 22 Aug 2025 13:16:15 +0000
Subject: [PATCH] [MachinePipeliner] Limit the number of stores in BB

---
 llvm/lib/CodeGen/MachinePipeliner.cpp | 25 +++++++++++++++++++++++++
 1 file changed, 25 insertions(+)

diff --git a/llvm/lib/CodeGen/MachinePipeliner.cpp b/llvm/lib/CodeGen/MachinePipeliner.cpp
index 3a9651c5cee04..89ed4da1fa532 100644
--- a/llvm/lib/CodeGen/MachinePipeliner.cpp
+++ b/llvm/lib/CodeGen/MachinePipeliner.cpp
@@ -110,6 +110,7 @@ STATISTIC(NumFailZeroMII, "Pipeliner abort due to zero MII");
 STATISTIC(NumFailNoSchedule, "Pipeliner abort due to no schedule found");
 STATISTIC(NumFailZeroStage, "Pipeliner abort due to zero stage");
 STATISTIC(NumFailLargeMaxStage, "Pipeliner abort due to too many stages");
+STATISTIC(NumFailTooManyStores, "Pipeliner abort due to too many stores");
 
 /// A command line option to turn software pipelining on or off.
 static cl::opt<bool> EnableSWP("enable-pipeliner", cl::Hidden, cl::init(true),
@@ -193,6 +194,13 @@ static cl::opt<bool>
     MVECodeGen("pipeliner-mve-cg", cl::Hidden, cl::init(false),
                cl::desc("Use the MVE code generator for software pipelining"));
 
+/// A command line argument to limit the number of store instructions in the
+/// target basic block.
+static cl::opt<unsigned> SwpMaxNumStores(
+    "pipeliner-max-num-stores",
+    cl::desc("Maximum number of stores allwed in the target loop."), cl::Hidden,
+    cl::init(200));
+
 namespace llvm {
 
 // A command line option to enable the CopyToPhi DAG mutation.
@@ -544,6 +552,23 @@ bool MachinePipeliner::canPipelineLoop(MachineLoop &L) {
     return false;
   }
 
+  unsigned NumStores = 0;
+  for (MachineInstr &MI : *L.getHeader())
+    if (MI.mayStore())
+      ++NumStores;
+  if (NumStores > SwpMaxNumStores) {
+    LLVM_DEBUG(dbgs() << "Too many stores\n");
+    NumFailTooManyStores++;
+    ORE->emit([&]() {
+      return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop",
+                                               L.getStartLoc(), L.getHeader())
+             << "Too many store instructions in the loop: "
+             << ore::NV("NumStores", NumStores) << " > "
+             << ore::NV("SwpMaxNumStores", SwpMaxNumStores) << ".";
+    });
+    return false;
+  }
+
   // Remove any subregisters from inputs to phi nodes.
   preprocessPhiNodes(*L.getHeader());
   return true;