[GlobalISel] LegalizationArtifactCombiner: Elide redundant G_AND

The legalizer currently generates lots of G_AND artifacts.
For example between boolean uses and defs there is always a G_AND with a mask of 1, but when the target uses ZeroOrOneBooleanContents, this is unnecessary.
Currently these artifacts have to be removed using post-legalize combines.
Omitting these artifacts at their source in the artifact combiner has a few advantages:
- We know that the emitted G_AND is very likely to be useless, so our KnownBits call is likely worth it.
- The G_AND and G_CONSTANT can interrupt e.g. G_UADDE/... sequences generated during legalization of wide adds which makes it harder to detect these sequences in the instruction selector (e.g. useful to prevent unnecessary reloading of AArch64 NZCV register).
- This cleans up a lot of legalizer output and even improves compilation-times.
AArch64 CTMark geomean: `O0` -5.6% size..text; `O0` and `O3` ~-0.9% compilation-time (instruction count).

Since this introduces KnownBits into code-paths used by `O0`, I reduced the default recursion depth.
This doesn't seem to make a difference in CTMark, but should prevent excessive recursive calls in the worst case.

Reviewed By: aemerson

Differential Revision: https://reviews.llvm.org/D159140

Author: tobias-stadler

llvm/include/llvm/CodeGen/GlobalISel/GISelKnownBits.h

@@ -118,11 +118,7 @@ class GISelKnownBitsAnalysis : public MachineFunctionPass {
   GISelKnownBitsAnalysis() : MachineFunctionPass(ID) {
     initializeGISelKnownBitsAnalysisPass(*PassRegistry::getPassRegistry());
   }
-  GISelKnownBits &get(MachineFunction &MF) {
-    if (!Info)
-      Info = std::make_unique<GISelKnownBits>(MF);
-    return *Info.get();
-  }
+  GISelKnownBits &get(MachineFunction &MF);
   void getAnalysisUsage(AnalysisUsage &AU) const override;
   bool runOnMachineFunction(MachineFunction &MF) override;
   void releaseMemory() override { Info.reset(); }

llvm/include/llvm/CodeGen/GlobalISel/LegalizationArtifactCombiner.h

@@ -35,6 +35,7 @@ class LegalizationArtifactCombiner {
   MachineIRBuilder &Builder;
   MachineRegisterInfo &MRI;
   const LegalizerInfo &LI;
+  GISelKnownBits *KB;
 
   static bool isArtifactCast(unsigned Opc) {
     switch (Opc) {
@@ -50,8 +51,9 @@ class LegalizationArtifactCombiner {
 
 public:
   LegalizationArtifactCombiner(MachineIRBuilder &B, MachineRegisterInfo &MRI,
-                    const LegalizerInfo &LI)
-      : Builder(B), MRI(MRI), LI(LI) {}
+                               const LegalizerInfo &LI,
+                               GISelKnownBits *KB = nullptr)
+      : Builder(B), MRI(MRI), LI(LI), KB(KB) {}
 
   bool tryCombineAnyExt(MachineInstr &MI,
                         SmallVectorImpl<MachineInstr *> &DeadInsts,
@@ -131,13 +133,26 @@ class LegalizationArtifactCombiner {
       LLVM_DEBUG(dbgs() << ".. Combine MI: " << MI;);
       LLT SrcTy = MRI.getType(SrcReg);
       APInt MaskVal = APInt::getAllOnes(SrcTy.getScalarSizeInBits());
-      auto Mask = Builder.buildConstant(
-        DstTy, MaskVal.zext(DstTy.getScalarSizeInBits()));
       if (SextSrc && (DstTy != MRI.getType(SextSrc)))
         SextSrc = Builder.buildSExtOrTrunc(DstTy, SextSrc).getReg(0);
       if (TruncSrc && (DstTy != MRI.getType(TruncSrc)))
         TruncSrc = Builder.buildAnyExtOrTrunc(DstTy, TruncSrc).getReg(0);
-      Builder.buildAnd(DstReg, SextSrc ? SextSrc : TruncSrc, Mask);
+      APInt ExtMaskVal = MaskVal.zext(DstTy.getScalarSizeInBits());
+      Register AndSrc = SextSrc ? SextSrc : TruncSrc;
+      // Elide G_AND and mask constant if possible.
+      // The G_AND would also be removed by the post-legalize redundant_and
+      // combine, but in this very common case, eliding early and regardless of
+      // OptLevel results in significant compile-time and O0 code-size
+      // improvements. Inserting unnecessary instructions between a boolean def
+      // and use can also hinder ISel to detect e.g. that reloading a flags
+      // register is unnecessary.
+      if (KB && (KB->getKnownZeroes(AndSrc) | ExtMaskVal).isAllOnes()) {
+        replaceRegOrBuildCopy(DstReg, AndSrc, MRI, Builder, UpdatedDefs,
+                              Observer);
+      } else {
+        auto Mask = Builder.buildConstant(DstTy, ExtMaskVal);
+        Builder.buildAnd(DstReg, AndSrc, Mask);
+      }
       markInstAndDefDead(MI, *MRI.getVRegDef(SrcReg), DeadInsts);
       return true;
     }

llvm/lib/CodeGen/GlobalISel/GISelKnownBits.cpp

@@ -19,6 +19,7 @@
 #include "llvm/CodeGen/TargetLowering.h"
 #include "llvm/CodeGen/TargetOpcodes.h"
 #include "llvm/IR/Module.h"
+#include "llvm/Target/TargetMachine.h"
 
 #define DEBUG_TYPE "gisel-known-bits"
 
@@ -773,3 +774,12 @@ void GISelKnownBitsAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
 bool GISelKnownBitsAnalysis::runOnMachineFunction(MachineFunction &MF) {
   return false;
 }
+
+GISelKnownBits &GISelKnownBitsAnalysis::get(MachineFunction &MF) {
+  if (!Info) {
+    unsigned MaxDepth =
+        MF.getTarget().getOptLevel() == CodeGenOptLevel::None ? 2 : 6;
+    Info = std::make_unique<GISelKnownBits>(MF, MaxDepth);
+  }
+  return *Info.get();
+}

llvm/lib/CodeGen/GlobalISel/Legalizer.cpp

@@ -218,7 +218,7 @@ Legalizer::legalizeMachineFunction(MachineFunction &MF, const LegalizerInfo &LI,
   // This will keep all the observers notified about new insertions/deletions.
   RAIIMFObsDelInstaller Installer(MF, WrapperObserver);
   LegalizerHelper Helper(MF, LI, WrapperObserver, MIRBuilder, KB);
-  LegalizationArtifactCombiner ArtCombiner(MIRBuilder, MRI, LI);
+  LegalizationArtifactCombiner ArtCombiner(MIRBuilder, MRI, LI, KB);
   bool Changed = false;
   SmallVector<MachineInstr *, 128> RetryList;
   do {