Revert "Disable codegen prepare critical edge splitting. Machine instruction passes now"

This reverts revision 114633. It was breaking llvm-gcc-i386-linux-selfhost.

It seems there is a downstream bug that is exposed by
-cgp-critical-edge-splitting=0. When that bug is fixed, this patch can go back
in.

Note that the changes to tailcallfp2.ll are not reverted. They were good are
required.

llvm-svn: 114859

lib/Transforms/Scalar/CodeGenPrepare.cpp

@@ -31,7 +31,6 @@
 #include "llvm/Transforms/Utils/BuildLibCalls.h"
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/ADT/SmallSet.h"
-#include "llvm/ADT/Statistic.h"
 #include "llvm/Assembly/Writer.h"
 #include "llvm/Support/CallSite.h"
 #include "llvm/Support/CommandLine.h"
@@ -43,12 +42,10 @@
 using namespace llvm;
 using namespace llvm::PatternMatch;
 
-STATISTIC(NumElim,  "Number of blocks eliminated");
-
 static cl::opt<bool>
 CriticalEdgeSplit("cgp-critical-edge-splitting",
                   cl::desc("Split critical edges during codegen prepare"),
-                  cl::init(false), cl::Hidden);
+                  cl::init(true), cl::Hidden);
 
 namespace {
   class CodeGenPrepare : public FunctionPass {
@@ -305,7 +302,6 @@ void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
     PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
   }
   BB->eraseFromParent();
-  ++NumElim;
 
   DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
 }

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

@@ -624,12 +624,12 @@ bb23:                                             ; preds = %bb22, %bb20, %bb9,
 bb24:                                             ; preds = %bb23
 
 ; LSR should use count-down iteration to avoid requiring the trip count
-; in a register.
+; in a register, and it shouldn't require any reloads here.
 
 ;      CHECK: @ %bb24
 ; CHECK-NEXT: @   in Loop: Header=BB1_1 Depth=1
-; CHECK: subs [[REGISTER:(r[0-9]+)|(lr)]], #1
+; CHECK-NEXT: sub{{.*}} [[REGISTER:(r[0-9]+)|(lr)]], #1
-; CHECK: bne.w
+; CHECK-NEXT: bne.w
 
   %92 = icmp eq i32 %tmp81, %indvar78             ; <i1> [#uses=1]
   %indvar.next79 = add i32 %indvar78, 1           ; <i32> [#uses=1]

test/CodeGen/Thumb/2010-07-15-debugOrdering.ll

@@ -10,7 +10,7 @@
 define void @_Z19getClosestDiagonal3ii(%0* noalias sret, i32, i32) nounwind {
 ; CHECK: blx ___muldf3
 ; CHECK: blx ___muldf3
-; CHECK: beq LBB0_7
+; CHECK: beq LBB0_8
 ; CHECK: blx ___muldf3
 ; <label>:3
   switch i32 %1, label %4 [

test/CodeGen/X86/2008-02-18-TailMergingBug.ll

@@ -1,4 +1,4 @@
-; RUN: llc < %s -march=x86 -mcpu=yonah -stats |& grep {Number of block tails merged} | grep 16
+; RUN: llc < %s -march=x86 -mcpu=yonah -stats |& grep {Number of block tails merged} | grep 9
 ; PR1909
 
 @.str = internal constant [48 x i8] c"transformed bounds: (%.2f, %.2f), (%.2f, %.2f)\0A\00"		; <[48 x i8]*> [#uses=1]

test/CodeGen/X86/lsr-reuse.ll

@@ -389,7 +389,7 @@ return:
 ; rdar://7657764
 
 ; CHECK: asd:
-; CHECK: BB9_4:
+; CHECK: BB9_5:
 ; CHECK-NEXT: addl  (%r{{[^,]*}},%rdi,4), %e
 ; CHECK-NEXT: incq  %rdi
 ; CHECK-NEXT: cmpq  %rdi, %r{{[^,]*}}
@@ -464,7 +464,7 @@ bb5:                                              ; preds = %bb3, %entry
 
 ; And the one at %bb68, where we want to be sure to use superhero mode:
 
-; CHECK:      BB10_7:
+; CHECK:      BB10_9:
 ; CHECK-NEXT:   movaps  48(%r{{[^,]*}}), %xmm{{.*}}
 ; CHECK-NEXT:   mulps   %xmm{{.*}}, %xmm{{.*}}
 ; CHECK-NEXT:   movaps  32(%r{{[^,]*}}), %xmm{{.*}}
@@ -484,6 +484,7 @@ bb5:                                              ; preds = %bb3, %entry
 ; CHECK-NEXT:   addq    $64, %r{{.*}}
 ; CHECK-NEXT:   addq    $64, %r{{.*}}
 ; CHECK-NEXT:   addq    $-16, %r{{.*}}
+; CHECK-NEXT: BB10_10:
 ; CHECK-NEXT:   cmpq    $15, %r{{.*}}
 ; CHECK-NEXT:   jg
 

test/CodeGen/X86/phi-immediate-factoring.ll

@@ -1,5 +1,5 @@
-; RUN: llc < %s -march=x86 -stats |& grep {Number of blocks eliminated} | grep 6
 ; PR1296
+; RUN: llc < %s -march=x86 | grep {movl	\$1} | count 1
 
 target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64"
 target triple = "i686-apple-darwin8"

test/CodeGen/X86/phys_subreg_coalesce-2.ll

@@ -1,4 +1,4 @@
-; RUN: llc < %s -march=x86 | grep mov | count 4
+; RUN: llc < %s -march=x86 | grep mov | count 5
 ; PR2659
 
 define i32 @binomial(i32 %n, i32 %k) nounwind {

test/CodeGen/X86/postra-licm.ll

@@ -68,7 +68,7 @@ bb26.preheader:                                   ; preds = %imix_test.exit
 
 bb23:                                             ; preds = %imix_test.exit
   unreachable
-; X86-32: %bb26.preheader
+; X86-32: %bb26.preheader.bb28_crit_edge
 ; X86-32: movl -16(%ebp),
 ; X86-32-NEXT: .align 4
 ; X86-32-NEXT: %bb28

test/CodeGen/X86/pr2659.ll

@@ -1,4 +1,4 @@
-; RUN: llc < %s -march=x86 -mtriple=i686-apple-darwin9.4.0 | grep movl | count 4
+; RUN: llc < %s -march=x86 -mtriple=i686-apple-darwin9.4.0 | grep movl | count 5
 ; RUN: llc < %s -march=x86 -mtriple=i686-apple-darwin9.4.0 | FileCheck %s
 ; PR2659
 
@@ -14,11 +14,10 @@ forcond.preheader:              ; preds = %entry
   %cmp44 = icmp eq i32 %k, 0            ; <i1> [#uses=1]
   br i1 %cmp44, label %afterfor, label %forbody
 
-; CHECK: %forcond.preheader
+; CHECK: %forcond.preheader.forbody_crit_edge
 ; CHECK: movl $1
 ; CHECK-NOT: xorl
-; CHECK-NOT: movl
+; CHECK-NEXT: movl
-; CHECK-NEXT: je
 
 ifthen:         ; preds = %entry
   ret i32 0

test/CodeGen/X86/tail-opts.ll

@@ -153,16 +153,19 @@ bb30:
 ; an unconditional jump to complete a two-way conditional branch.
 
 ; CHECK: c_expand_expr_stmt:
-; CHECK:        jmp .LBB3_11
+; CHECK:        jmp .LBB3_7
-; CHECK-NEXT: .LBB3_9:
+; CHECK-NEXT: .LBB3_12:
 ; CHECK-NEXT:   movq 8(%rax), %rax
-; CHECK-NEXT:   xorb %dl, %dl
 ; CHECK-NEXT:   movb 16(%rax), %al
 ; CHECK-NEXT:   cmpb $16, %al
-; CHECK-NEXT:   je .LBB3_11
+; CHECK-NEXT:   je .LBB3_6
 ; CHECK-NEXT:   cmpb $23, %al
-; CHECK-NEXT:   jne .LBB3_14
+; CHECK-NEXT:   je .LBB3_6
-; CHECK-NEXT: .LBB3_11:
+; CHECK-NEXT:   jmp .LBB3_15
+; CHECK-NEXT: .LBB3_14:
+; CHECK-NEXT:   cmpb $23, %bl
+; CHECK-NEXT:   jne .LBB3_15
+; CHECK-NEXT: .LBB3_15:
 
 %0 = type { %struct.rtx_def* }
 %struct.lang_decl = type opaque