[llvm] [llvm][CodeGen] Add a new software pipeliner 'Window Scheduler' (PR #84443)

[Hua Tian via llvm-commits]

================
@@ -0,0 +1,692 @@
+//======----------- WindowScheduler.cpp - window scheduler -------------======//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// An implementation of the Window Scheduling software pipelining algorithm.
+//
+// The fundamental concept of the window scheduling algorithm involves folding
+// the original MBB at a specific position, followed by list scheduling on the
+// folded MIs. The optimal scheduling result is then chosen from various folding
+// positions as the final scheduling outcome.
+//
+// The primary challenge in this algorithm lies in generating the folded MIs and
+// establishing their dependencies. We have innovatively employed a new MBB,
+// created by copying the original MBB three times, known as TripleMBB. This
+// TripleMBB enables the convenient implementation of MI folding and dependency
+// establishment. To facilitate the algorithm's implementation, we have also
+// devised data structures such as OriMIs, TriMIs, TriToOri, and OriToCycle.
+//
+// Another challenge in the algorithm is the scheduling of phis. Semantically,
+// it is difficult to place the phis in the window and perform list scheduling.
+// Therefore, we schedule these phis separately after each list scheduling.
+//
+// The provided implementation is designed for use before the Register Allocator
+// (RA). If the target requires implementation after RA, it is recommended to
+// reimplement analyseII(), schedulePhi(), and expand(). Additionally,
+// target-specific logic can be added in initialize(), preProcess(), and
+// postProcess().
+//
+// Lastly, it is worth mentioning that getSearchIndexes() is an important
+// function. We have experimented with more complex heuristics on downstream
+// target and achieved favorable results.
+//
+//===----------------------------------------------------------------------===//
+#include "llvm/CodeGen/WindowScheduler.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/LiveIntervals.h"
+#include "llvm/CodeGen/MachineLoopInfo.h"
+#include "llvm/CodeGen/MachinePipeliner.h"
+#include "llvm/CodeGen/ModuloSchedule.h"
+#include "llvm/CodeGen/TargetPassConfig.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/TimeProfiler.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "pipeliner"
+
+namespace {
+STATISTIC(NumTryWindowSchedule,
+          "Number of loops that we attempt to use window scheduling");
+STATISTIC(NumTryWindowSearch,
+          "Number of times that we run list schedule in the window scheduling");
+STATISTIC(NumWindowSchedule,
+          "Number of loops that we successfully use window scheduling");
+STATISTIC(NumFailAnalyseII,
+          "Window scheduling abort due to the failure of the II analysis");
+
+cl::opt<unsigned>
+    WindowSearchNum("window-search-num",
+                    cl::desc("The number of searches per loop in the window "
+                             "algorithm. 0 means no search number limit."),
+                    cl::Hidden, cl::init(6));
+
+cl::opt<unsigned> WindowSearchRatio(
+    "window-search-ratio",
+    cl::desc("The ratio of searches per loop in the window algorithm. 100 "
+             "means search all positions in the loop, while 0 means not "
+             "performing any search."),
+    cl::Hidden, cl::init(40));
+
+cl::opt<unsigned> WindowIICoeff(
+    "window-ii-coeff",
+    cl::desc(
+        "The coefficient used when initializing II in the window algorithm."),
+    cl::Hidden, cl::init(5));
+
+cl::opt<unsigned> WindowRegionLimit(
+    "window-region-limit",
+    cl::desc(
+        "The lower limit of the scheduling region in the window algorithm."),
+    cl::Hidden, cl::init(3));
+
+cl::opt<unsigned> WindowDiffLimit(
+    "window-diff-limit",
+    cl::desc("The lower limit of the difference between best II and base II in "
+             "the window algorithm. If the difference is smaller than "
+             "this lower limit, window scheduling will not be performed."),
+    cl::Hidden, cl::init(2));
+} // namespace
+
+// WindowIILimit serves as an indicator of abnormal scheduling results and could
+// potentially be referenced by the derived target window scheduler.
+cl::opt<unsigned>
+    WindowIILimit("window-ii-limit",
+                  cl::desc("The upper limit of II in the window algorithm."),
+                  cl::Hidden, cl::init(1000));
+
+WindowScheduler::WindowScheduler(MachineSchedContext *C, MachineLoop &ML)
+    : Context(C), MF(C->MF), MBB(ML.getHeader()), Loop(ML) {
+  Subtarget = &(MF->getSubtarget());
----------------
huaatian wrote:

Updated

https://github.com/llvm/llvm-project/pull/84443