LSR: rewrite inner loops only.

Rewriting the entire loop nest now requires -enable-lsr-nested.
See PR11035 for some performance data.
A few unit tests specifically test nested LSR, and are now under a flag.

llvm-svn: 140762

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

@@ -78,6 +78,9 @@
 using namespace llvm;
 
 namespace llvm {
+cl::opt<bool> EnableNested(
+  "enable-lsr-nested", cl::Hidden, cl::desc("Enable LSR on nested loops"));
+
 cl::opt<bool> EnableRetry(
     "enable-lsr-retry", cl::Hidden, cl::desc("Enable LSR retry"));
 }
@@ -723,11 +726,14 @@ void Cost::RateRegister(const SCEV *Reg,
     if (AR->getLoop() == L)
       AddRecCost += 1; /// TODO: This should be a function of the stride.
 
-    // If this is an addrec for a loop that's already been visited by LSR,
-    // don't second-guess its addrec phi nodes. LSR isn't currently smart
-    // enough to reason about more than one loop at a time. Consider these
-    // registers free and leave them alone.
-    else if (L->contains(AR->getLoop()) ||
+    // If this is an addrec for another loop, don't second-guess its addrec phi
+    // nodes. LSR isn't currently smart enough to reason about more than one
+    // loop at a time. LSR has either already run on inner loops, will not run
+    // on other loops, and cannot be expected to change sibling loops. If the
+    // AddRec exists, consider it's register free and leave it alone. Otherwise,
+    // do not consider this formula at all.
+    // FIXME: why do we need to generate such fomulae?
+    else if (!EnableNested || L->contains(AR->getLoop()) ||
              (!AR->getLoop()->contains(L) &&
               DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
       for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
@@ -738,6 +744,10 @@ void Cost::RateRegister(const SCEV *Reg,
             SE.getSCEV(PN) == AR)
           return;
       }
+      if (!EnableNested) {
+        Loose();
+        return;
+      }
       // If this isn't one of the addrecs that the loop already has, it
       // would require a costly new phi and add. TODO: This isn't
       // precisely modeled right now.
@@ -3801,6 +3811,12 @@ LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
   // If loop preparation eliminates all interesting IV users, bail.
   if (IU.empty()) return;
 
+  // Skip nested loops until we can model them better with forulae.
+  if (!EnableNested && !L->empty()) {
+    DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
+    return false;
+  }
+
   // Start collecting data and preparing for the solver.
   CollectInterestingTypesAndFactors();
   CollectFixupsAndInitialFormulae();

llvm/test/CodeGen/ARM/lsr-on-unrolled-loops.ll

@@ -1,4 +1,4 @@
-; RUN: llc -mtriple=thumbv7-apple-darwin10 -mcpu=cortex-a8 < %s | FileCheck %s
+; RUN: llc -mtriple=thumbv7-apple-darwin10 -mcpu=cortex-a8  -enable-lsr-nested < %s | FileCheck %s
 
 ; LSR should recognize that this is an unrolled loop which can use
 ; constant offset addressing, so that each of the following stores
@@ -8,6 +8,9 @@
 ; CHECK: vstr.32 s{{.*}}, [{{(r[0-9]+)|(lr)}}, #64]
 ; CHECK: vstr.32 s{{.*}}, [{{(r[0-9]+)|(lr)}}, #96]
 
+; We can also save a register in the outer loop, but that requires
+; performing LSR on the outer loop.
+
 target datalayout = "e-p:32:32:32-i1:8:32-i8:8:32-i16:16:32-i32:32:32-i64:32:32-f32:32:32-f64:32:32-v64:64:64-v128:128:128-a0:0:32-n32"
 
 %0 = type { %1*, %3*, %6*, i8*, i32, i32, %8*, i32, i32, i32, i32, i32, i32, i32, double, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i8**, i32, i32, i32, i32, i32, [64 x i32]*, [4 x %9*], [4 x %10*], [4 x %10*], i32, %11*, i32, i32, [16 x i8], [16 x i8], [16 x i8], i32, i32, i8, i8, i8, i16, i16, i32, i8, i32, %12*, i32, i32, i32, i32, i8*, i32, [4 x %11*], i32, i32, i32, [10 x i32], i32, i32, i32, i32, i32, %13*, %14*, %15*, %16*, %17*, %18*, %19*, %20*, %21*, %22*, %23* }

llvm/test/CodeGen/X86/change-compare-stride-0.ll

@@ -1,4 +1,7 @@
-; RUN: llc < %s -march=x86 | FileCheck %s
+; RUN: llc < %s -march=x86 -enable-lsr-nested | FileCheck %s
+;
+; Nested LSR is required to optimize this case.
+; We do not expect to see this form of IR without -enable-iv-rewrite.
 
 define void @borf(i8* nocapture %in, i8* nocapture %out) nounwind {
 ; CHECK: borf:

llvm/test/CodeGen/X86/change-compare-stride-1.ll

@@ -1,4 +1,7 @@
-; RUN: llc < %s -march=x86-64 | FileCheck %s
+; RUN: llc < %s -march=x86-64 -enable-lsr-nested | FileCheck %s
+;
+; Nested LSR is required to optimize this case.
+; We do not expect to see this form of IR without -enable-iv-rewrite.
 
 define void @borf(i8* nocapture %in, i8* nocapture %out) nounwind {
 ; CHECK: borf:

llvm/test/CodeGen/X86/iv-users-in-other-loops.ll

@@ -1,4 +1,4 @@
-; RUN: llc < %s -march=x86-64 -o %t
+; RUN: llc < %s -march=x86-64 -enable-lsr-nested -o %t
 ; RUN: not grep inc %t
 ; RUN: grep dec %t | count 2
 ; RUN: grep addq %t | count 12
@@ -11,6 +11,10 @@
 ; to insert new induction variables. Previously it would create a
 ; flood of new induction variables.
 ; Also, the loop reversal should kick in once.
+;
+; In this example, performing LSR on the entire loop nest,
+; as opposed to only the inner loop can further reduce induction variables,
+; and their related instructions and registers.
 
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128"
 target triple = "x86_64-unknown-linux-gnu"

llvm/test/CodeGen/X86/loop-strength-reduce3.ll

@@ -1,5 +1,5 @@
-; RUN: llc < %s -march=x86 | grep cmp | grep 240
-; RUN: llc < %s -march=x86 | grep inc | count 1
+; RUN: llc < %s -march=x86 -enable-lsr-nested | grep cmp | grep 240
+; RUN: llc < %s -march=x86 -enable-lsr-nested | grep inc | count 1
 
 define i32 @foo(i32 %A, i32 %B, i32 %C, i32 %D) nounwind {
 entry:

llvm/test/CodeGen/X86/pr3495.ll

@@ -1,7 +1,9 @@
-; RUN: llc < %s -march=x86 -stats -regalloc=linearscan |& grep {Number of loads added} | grep 2
-; RUN: llc < %s -march=x86 -stats -regalloc=linearscan |& grep {Number of spill slots allocated} | grep 1
-; RUN: llc < %s -march=x86 -stats -regalloc=linearscan |& grep {Number of machine instrs printed} | grep 34
+; RUN: llc < %s -march=x86 -stats -regalloc=linearscan -enable-lsr-nested |& grep {Number of loads added} | grep 2
+; RUN: llc < %s -march=x86 -stats -regalloc=linearscan -enable-lsr-nested |& grep {Number of spill slots allocated} | grep 1
+; RUN: llc < %s -march=x86 -stats -regalloc=linearscan -enable-lsr-nested |& grep {Number of machine instrs printed} | grep 34
 ; PR3495
+;
+; Note: this should not spill at all with either good LSR or good regalloc.
 
 target triple = "i386-pc-linux-gnu"
 @x = external global [8 x i32], align 32		; <[8 x i32]*> [#uses=1]