[MachineLICM] Re-enable hoisting of constant stores

[syzaara]

Summary

This patch fixes an issue exposed on the SystemZ build bots when committing https://reviews.llvm.org/rL327856. The hoisting was temporarily disabled with an option. This patch now re-enables hoisting and checks that we only hoist a store instruction when all its operands are either constant caller preserved registers or immediates.

[jonpa]

• The test could perhaps be freed of alignment, tbaa operands and dso_local specifiers, but maybe it doesn't matter. Adding Uli as reviewer since the test is in SystemZ.
• Using isImm() but not isCImm() or isFPImm() (I don't see any method like isAnyImm() available)? It seems there may be other constant operands also, like isSymbol(), but not sure if it's useful.

Since it seems only PowerPC has implemented isCallerPreservedPhysReg() and are therefore the only user of this patch, I leave the review to the power-llvm-team. (Side question: Could the SP handling be put in common code?)

I was actually expecting to not only look for function-constant registers via isCallerPreservedPhysReg(), but to have a set of loop-invariant phys-regs for each loop (and preheader terminators). I suppose this is not needed given that the LLVM I/R LICM pass has done most of the work, or?

[uweigand]

In D45286#1058138, @jonpa wrote:

The test could perhaps be freed of alignment, tbaa operands and dso_local specifiers, but maybe it doesn't matter. Adding Uli as reviewer since the test is in SystemZ.

I agree it doesn't really matter. The test is fine with me as is.

[syzaara]

In D45286#1058138, @jonpa wrote:

• The test could perhaps be freed of alignment, tbaa operands and dso_local specifiers, but maybe it doesn't matter. Adding Uli as reviewer since the test is in SystemZ.
• Using isImm() but not isCImm() or isFPImm() (I don't see any method like isAnyImm() available)? It seems there may be other constant operands also, like isSymbol(), but not sure if it's useful.

Since it seems only PowerPC has implemented isCallerPreservedPhysReg() and are therefore the only user of this patch, I leave the review to the power-llvm-team. (Side question: Could the SP handling be put in common code?)

I was actually expecting to not only look for function-constant registers via isCallerPreservedPhysReg(), but to have a set of loop-invariant phys-regs for each loop (and preheader terminators). I suppose this is not needed given that the LLVM I/R LICM pass has done most of the work, or?

Thanks for the comments. The original intention of this patch was to start hoisting store instructions which are constant throughout the function, since till now there was no handling of store instructions at all. I will commit this now that the test case looks good.

[jonpa]

In D45286#1059761, @syzaara wrote:

In D45286#1058138, @jonpa wrote:

• The test could perhaps be freed of alignment, tbaa operands and dso_local specifiers, but maybe it doesn't matter. Adding Uli as reviewer since the test is in SystemZ.
• Using isImm() but not isCImm() or isFPImm() (I don't see any method like isAnyImm() available)? It seems there may be other constant operands also, like isSymbol(), but not sure if it's useful.

Since it seems only PowerPC has implemented isCallerPreservedPhysReg() and are therefore the only user of this patch, I leave the review to the power-llvm-team. (Side question: Could the SP handling be put in common code?)

I was actually expecting to not only look for function-constant registers via isCallerPreservedPhysReg(), but to have a set of loop-invariant phys-regs for each loop (and preheader terminators). I suppose this is not needed given that the LLVM I/R LICM pass has done most of the work, or?

Thanks for the comments. The original intention of this patch was to start hoisting store instructions which are constant throughout the function, since till now there was no handling of store instructions at all. I will commit this now that the test case looks good.

I see. Perhaps put a comment as well that explains that there is more to do near that function..?

lib/CodeGen/MachineLICM.cpp

 // checks that each register operand statisfies isCallerPreservedPhysReg.
 // This means, the value being stored and the address where it is being stored
 // is constant throughout the body of the function (not including prologue and
 // epilogue). When called with an MI that isn't a store, it returns false.
 static bool isInvariantStore(const MachineInstr &MI,
                              const TargetRegisterInfo *TRI,
                              const MachineRegisterInfo *MRI) {
 
+  bool FoundCallerPresReg = false;
   if (!MI.mayStore() || MI.hasUnmodeledSideEffects() ||
       (MI.getNumOperands() == 0))
     return false;
 
   // Check that all register operands are caller-preserved physical registers.
   for (const MachineOperand &MO : MI.operands()) {
     if (MO.isReg()) {
       unsigned Reg = MO.getReg();
       // If operand is a virtual register, check if it comes from a copy of a
       // physical register.
       if (TargetRegisterInfo::isVirtualRegister(Reg))
         Reg = TRI->lookThruCopyLike(MO.getReg(), MRI);
       if (TargetRegisterInfo::isVirtualRegister(Reg))
         return false;
       if (!TRI->isCallerPreservedPhysReg(Reg, *MI.getMF()))
         return false;
+      else
+        FoundCallerPresReg = true;
+    } else if (!MO.isImm()) {
+        return false;
     }
   }
-  return true;
+  return FoundCallerPresReg;
 }
 
 // Return true if the input MI is a copy instruction that feeds an invariant
 // store instruction. This means that the src of the copy has to satisfy
 // isCallerPreservedPhysReg and atleast one of it's users should satisfy
 // isInvariantStore.
 static bool isCopyFeedingInvariantStore(const MachineInstr &MI,
                                         const MachineRegisterInfo *MRI,

test/CodeGen/SystemZ/hoist-const-stores.ll

+; RUN: llc -verify-machineinstrs -mtriple=s390x-ibm-linux -mcpu=z13 -O3 -hoist-const-stores < %s | FileCheck %s
+
+@b = dso_local local_unnamed_addr global i32 15, align 4
+@e = dso_local local_unnamed_addr global i32 -1, align 4
+@f = common dso_local global i32 0, align 4
+@g = dso_local local_unnamed_addr global i32* @f, align 8
+@c = common dso_local local_unnamed_addr global i32 0, align 4
+@a = common dso_local local_unnamed_addr global [6 x i32] zeroinitializer, align 4
+@d = common dso_local local_unnamed_addr global i32 0, align 4
+@h = common dso_local local_unnamed_addr global i32 0, align 4
+@.str = private unnamed_addr constant [15 x i8] c"checksum = %X\0A\00", align 2
+
+; Function Attrs: nounwind
+define dso_local signext i32 @main()  {
+entry:
+  %i = alloca i32, align 4
+  %.pr = load i32, i32* @c, align 4, !tbaa !2
+  %cmp6 = icmp slt i32 %.pr, 6
+  br i1 %cmp6, label %for.body.preheader, label %for.end
+
+for.body.preheader:                               ; preds = %entry
+  store i32 5, i32* getelementptr inbounds ([6 x i32], [6 x i32]* @a, i64 0, i64 1), align 4, !tbaa !2
+  store i32 6, i32* @c, align 4, !tbaa !2
+  br label %for.end
+
+for.end:                                          ; preds = %for.body.preheader, %entry
+  %0 = bitcast i32* %i to i8*
+  call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %0)
+  store i32 14, i32* %i, align 4, !tbaa !2
+  %.pr2 = load i32, i32* @d, align 4, !tbaa !2
+  %cmp25 = icmp sgt i32 %.pr2, -1
+  br i1 %cmp25, label %for.cond4thread-pre-split.lr.ph, label %for.end.for.end11_crit_edge
+
+for.end.for.end11_crit_edge:                      ; preds = %for.end
+  %.pre10 = load i32, i32* @b, align 4, !tbaa !2
+  br label %for.end11
+
+; CHECK: # %for.cond4thread-pre-split.lr.ph
+; CHECK-NOT: mvhi    164(%r15), 0
+; CHECK: # %for.end9
+; CHECK: mvhi    164(%r15), 0 
+
+for.cond4thread-pre-split.lr.ph:                  ; preds = %for.end
+  %1 = ptrtoint i32* %i to i64
+  %2 = trunc i64 %1 to i32
+  %3 = load i32*, i32** @g, align 8
+  %.pr3.pre = load i32, i32* @e, align 4, !tbaa !2
+  br label %for.cond4thread-pre-split
+
+for.cond4thread-pre-split:                        ; preds = %for.cond4thread-pre-split.lr.ph, %for.end9
+  %4 = phi i32 [ %.pr2, %for.cond4thread-pre-split.lr.ph ], [ %dec, %for.end9 ]
+  %5 = phi i32 [ 14, %for.cond4thread-pre-split.lr.ph ], [ 0, %for.end9 ]
+  %.pr3 = phi i32 [ %.pr3.pre, %for.cond4thread-pre-split.lr.ph ], [ %.pr37, %for.end9 ]
+  %cmp54 = icmp slt i32 %.pr3, 1
+  br i1 %cmp54, label %for.body6.preheader, label %for.end9
+
+for.body6.preheader:                              ; preds = %for.cond4thread-pre-split
+  store i32 %5, i32* %3, align 4, !tbaa !2
+  %6 = load i32, i32* @e, align 4, !tbaa !2
+  %inc811 = add nsw i32 %6, 1
+  store i32 %inc811, i32* @e, align 4, !tbaa !2
+  %cmp512 = icmp slt i32 %6, 0
+  br i1 %cmp512, label %for.body6.for.body6_crit_edge, label %for.end9.loopexit
+
+for.body6.for.body6_crit_edge:                    ; preds = %for.body6.preheader, %for.body6.for.body6_crit_edge.3
+  %.pre = load i32, i32* %i, align 4, !tbaa !2
+  store i32 %.pre, i32* %3, align 4, !tbaa !2
+  %7 = load i32, i32* @e, align 4, !tbaa !2
+  %inc8 = add nsw i32 %7, 1
+  store i32 %inc8, i32* @e, align 4, !tbaa !2
+  %cmp5 = icmp slt i32 %7, 0
+  br i1 %cmp5, label %for.body6.for.body6_crit_edge.1, label %for.end9.loopexit
+
+for.end9.loopexit:                                ; preds = %for.body6.for.body6_crit_edge, %for.body6.for.body6_crit_edge.1, %for.body6.for.body6_crit_edge.2, %for.body6.for.body6_crit_edge.3, %for.body6.preheader
+  %inc8.lcssa = phi i32 [ %inc811, %for.body6.preheader ], [ %inc8, %for.body6.for.body6_crit_edge ], [ %inc8.1, %for.body6.for.body6_crit_edge.1 ], [ %inc8.2, %for.body6.for.body6_crit_edge.2 ], [ %inc8.3, %for.body6.for.body6_crit_edge.3 ]
+  %.pre9 = load i32, i32* @d, align 4, !tbaa !2
+  br label %for.end9
+
+for.end9:                                         ; preds = %for.end9.loopexit, %for.cond4thread-pre-split
+  %8 = phi i32 [ %.pre9, %for.end9.loopexit ], [ %4, %for.cond4thread-pre-split ]
+  %.pr37 = phi i32 [ %inc8.lcssa, %for.end9.loopexit ], [ %.pr3, %for.cond4thread-pre-split ]
+  store i32 %2, i32* @h, align 4, !tbaa !2
+  store i32 0, i32* %i, align 4, !tbaa !2
+  %9 = load i32, i32* @b, align 4, !tbaa !2
+  %10 = load i32, i32* @f, align 4, !tbaa !2
+  %xor = xor i32 %10, %9
+  %idxprom = sext i32 %xor to i64
+  %arrayidx = getelementptr inbounds [6 x i32], [6 x i32]* @a, i64 0, i64 %idxprom
+  %11 = load i32, i32* %arrayidx, align 4, !tbaa !2
+  store i32 %11, i32* @b, align 4, !tbaa !2
+  %dec = add nsw i32 %8, -1
+  store i32 %dec, i32* @d, align 4, !tbaa !2
+  %cmp2 = icmp sgt i32 %8, 0
+  br i1 %cmp2, label %for.cond4thread-pre-split, label %for.end11
+
+for.end11:                                        ; preds = %for.end9, %for.end.for.end11_crit_edge
+  %12 = phi i32 [ %.pre10, %for.end.for.end11_crit_edge ], [ %11, %for.end9 ]
+  %call = call signext i32 (i8*, ...) @printf(i8* getelementptr inbounds ([15 x i8], [15 x i8]* @.str, i64 0, i64 0), i32 signext %12)
+  call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %0)
+  ret i32 0
+
+for.body6.for.body6_crit_edge.1:                  ; preds = %for.body6.for.body6_crit_edge
+  %.pre.1 = load i32, i32* %i, align 4, !tbaa !2
+  store i32 %.pre.1, i32* %3, align 4, !tbaa !2
+  %13 = load i32, i32* @e, align 4, !tbaa !2
+  %inc8.1 = add nsw i32 %13, 1
+  store i32 %inc8.1, i32* @e, align 4, !tbaa !2
+  %cmp5.1 = icmp slt i32 %13, 0
+  br i1 %cmp5.1, label %for.body6.for.body6_crit_edge.2, label %for.end9.loopexit
+
+for.body6.for.body6_crit_edge.2:                  ; preds = %for.body6.for.body6_crit_edge.1
+  %.pre.2 = load i32, i32* %i, align 4, !tbaa !2
+  store i32 %.pre.2, i32* %3, align 4, !tbaa !2
+  %14 = load i32, i32* @e, align 4, !tbaa !2
+  %inc8.2 = add nsw i32 %14, 1
+  store i32 %inc8.2, i32* @e, align 4, !tbaa !2
+  %cmp5.2 = icmp slt i32 %14, 0
+  br i1 %cmp5.2, label %for.body6.for.body6_crit_edge.3, label %for.end9.loopexit
+
+for.body6.for.body6_crit_edge.3:                  ; preds = %for.body6.for.body6_crit_edge.2
+  %.pre.3 = load i32, i32* %i, align 4, !tbaa !2
+  store i32 %.pre.3, i32* %3, align 4, !tbaa !2
+  %15 = load i32, i32* @e, align 4, !tbaa !2
+  %inc8.3 = add nsw i32 %15, 1
+  store i32 %inc8.3, i32* @e, align 4, !tbaa !2
+  %cmp5.3 = icmp slt i32 %15, 0
+  br i1 %cmp5.3, label %for.body6.for.body6_crit_edge, label %for.end9.loopexit
+}
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.start.p0i8(i64, i8* nocapture)
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.end.p0i8(i64, i8* nocapture)
+
+; Function Attrs: nounwind
+declare dso_local signext i32 @printf(i8* nocapture readonly, ...)
+
+!llvm.module.flags = !{!0}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!2 = !{!3, !3, i64 0}
+!3 = !{!"int", !4, i64 0}
+!4 = !{!"omnipotent char", !5, i64 0}
+!5 = !{!"Simple C/C++ TBAA"}