[X86] Add experimental option to separately tune alignment of innermost loops

[mkazantsev]

Summary

We already have an experimental option to tune loop alignment. Its impact
is very wide (and there is a suspicion that it's not always profitable). We want
to have something more narrow to play with. This patch adds similar option that
overrides preferred alignment for innermost loops. This is for experimental
purposes, default values do not change the existing behavior.

[pengfei]

llvm/lib/Target/X86/X86ISelLowering.cpp

51713

Should it be Align(1ULL << ExperimentalPrefLoopAlignment)?

llvm/test/CodeGen/X86/innermost-loop-alignment.ll

4

Should add a test for combination with option x86-experimental-pref-loop-alignment?

[mkazantsev]

Added test for combination of these options.

llvm/lib/Target/X86/X86ISelLowering.cpp

51713

This is already stored in PrefLoopAlignment field. See usages of this option in code. IMO usage of default implementation is better than code copy-paste.

[pengfei]

LGTM.

llvm/lib/Target/X86/X86ISelLowering.cpp

51713

I see. Thank you.

llvm/lib/Target/X86/X86ISelLowering.h

public:
     /// instructions/intrinsics.
     bool lowerInterleavedStore(StoreInst *SI, ShuffleVectorInst *SVI,
                                unsigned Factor) const override;
 
     SDValue expandIndirectJTBranch(const SDLoc& dl, SDValue Value,
                                    SDValue Addr, SelectionDAG &DAG)
                                    const override;
 
+    Align getPrefLoopAlignment(MachineLoop *ML) const override;
+
   protected:
     std::pair<const TargetRegisterClass *, uint8_t>
     findRepresentativeClass(const TargetRegisterInfo *TRI,
                             MVT VT) const override;
 
   private:
     /// Keep a reference to the X86Subtarget around so that we can
     /// make the right decision when generating code for different targets.

llvm/lib/Target/X86/X86ISelLowering.cpp

 #include "llvm/Analysis/EHPersonalities.h"
 #include "llvm/Analysis/ProfileSummaryInfo.h"
 #include "llvm/Analysis/VectorUtils.h"
 #include "llvm/CodeGen/IntrinsicLowering.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineJumpTableInfo.h"
+#include "llvm/CodeGen/MachineLoopInfo.h"
 #include "llvm/CodeGen/MachineModuleInfo.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/TargetLowering.h"
 #include "llvm/CodeGen/WinEHFuncInfo.h"
 #include "llvm/IR/CallingConv.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/DerivedTypes.h"
 #include "llvm/IR/DiagnosticInfo.h"

 static cl::opt<int> ExperimentalPrefLoopAlignment(
     "x86-experimental-pref-loop-alignment", cl::init(4),
     cl::desc(
         "Sets the preferable loop alignment for experiments (as log2 bytes)"
         "(the last x86-experimental-pref-loop-alignment bits"
         " of the loop header PC will be 0)."),
     cl::Hidden);
 
+static cl::opt<int> ExperimentalPrefInnermostLoopAlignment(
+    "x86-experimental-pref-innermost-loop-alignment", cl::init(4),
+    cl::desc(
+        "Sets the preferable loop alignment for experiments (as log2 bytes) "
+        "for innermost loops only. If specified, this option overrides "
+        "alignment set by x86-experimental-pref-loop-alignment."),
+    cl::Hidden);
+
 static cl::opt<bool> MulConstantOptimization(
     "mul-constant-optimization", cl::init(true),
     cl::desc("Replace 'mul x, Const' with more effective instructions like "
              "SHIFT, LEA, etc."),
     cl::Hidden);
 
 static cl::opt<bool> ExperimentalUnorderedISEL(
     "x86-experimental-unordered-atomic-isel", cl::init(false),
X86TargetLowering::getStackProbeSize(MachineFunction &MF) const {
   unsigned StackProbeSize = 4096;
   const Function &Fn = MF.getFunction();
   if (Fn.hasFnAttribute("stack-probe-size"))
     Fn.getFnAttribute("stack-probe-size")
         .getValueAsString()
         .getAsInteger(0, StackProbeSize);
   return StackProbeSize;
 }
+
+Align X86TargetLowering::getPrefLoopAlignment(MachineLoop *ML) const {
+  if (ML->isInnermost() &&
+      ExperimentalPrefInnermostLoopAlignment.getNumOccurrences())
+    return Align(1ULL << ExperimentalPrefInnermostLoopAlignment);
+  return TargetLowering::getPrefLoopAlignment();
+}

llvm/test/CodeGen/X86/innermost-loop-alignment.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=i686-pc-linux-gnu | FileCheck %s -check-prefix=DEFAULT
+; RUN: llc < %s -mtriple=i686-pc-linux-gnu -x86-experimental-pref-innermost-loop-alignment=5 | FileCheck %s -check-prefix=ALIGN32
+; RUN: llc < %s -mtriple=i686-pc-linux-gnu -x86-experimental-pref-loop-alignment=5 -x86-experimental-pref-innermost-loop-alignment=6 | FileCheck %s -check-prefix=ALIGN64
+
+declare void @foo()
+
+define void @test(i32 %n, i32 %m) {
+; DEFAULT-LABEL: test:
+; DEFAULT:         .p2align 4, 0x90
+; DEFAULT-NEXT:  .LBB0_1: # %outer
+; DEFAULT-NEXT:    # =>This Loop Header: Depth=1
+; DEFAULT-NEXT:    # Child Loop BB0_2 Depth 2
+; DEFAULT:         .p2align 4, 0x90
+; DEFAULT-NEXT:  .LBB0_2: # %inner
+; DEFAULT-NEXT:    # Parent Loop BB0_1 Depth=1
+
+; ALIGN32-LABEL: test:
+; ALIGN32:         .p2align 4, 0x90
+; ALIGN32-NEXT:  .LBB0_1: # %outer
+; ALIGN32-NEXT:    # =>This Loop Header: Depth=1
+; ALIGN32-NEXT:    # Child Loop BB0_2 Depth 2
+; ALIGN32:         .p2align 5, 0x90
+; ALIGN32-NEXT:  .LBB0_2: # %inner
+; ALIGN32-NEXT:    # Parent Loop BB0_1 Depth=1
+; ALIGN32-NEXT:    # => This Inner Loop Header: Depth=2
+
+; ALIGN64-LABEL: test:
+; ALIGN64:         .p2align 5, 0x90
+; ALIGN64-NEXT:  .LBB0_1: # %outer
+; ALIGN64-NEXT:    # =>This Loop Header: Depth=1
+; ALIGN64-NEXT:    # Child Loop BB0_2 Depth 2
+; ALIGN64:         .p2align 6, 0x90
+; ALIGN64-NEXT:  .LBB0_2: # %inner
+; ALIGN64-NEXT:    # Parent Loop BB0_1 Depth=1
+; ALIGN64-NEXT:    # => This Inner Loop Header: Depth=2
+
+entry:
+  br label %outer
+
+outer:
+  %outer.iv = phi i32 [0, %entry], [%outer.iv.next, %outer_bb]
+  br label %inner
+
+inner:
+  %inner.iv = phi i32 [0, %outer], [%inner.iv.next, %inner]
+  call void @foo()
+  %inner.iv.next = add i32 %inner.iv, 1
+  %inner.cond = icmp ne i32 %inner.iv.next, %m
+  br i1 %inner.cond, label %inner, label %outer_bb
+
+outer_bb:
+  %outer.iv.next = add i32 %outer.iv, 1
+  %outer.cond = icmp ne i32 %outer.iv.next, %n
+  br i1 %outer.cond, label %outer, label %exit
+
+exit:
+  ret void
+}