Revert the main portion of r31856. It was causing BranchFolding

to break up CFG diamonds by banishing one of the blocks to the end of
the function, which is bad for code density and branch size.

This does pessimize MultiSource/Benchmarks/Ptrdist/yacr2, the
benchmark cited as the reason for the change, however I've examined
the code and it looks more like a case of gaming a particular
branch than of being generally applicable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@84803 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/BranchFolding.cpp b/lib/CodeGen/BranchFolding.cpp
index f9abeacbdbb3d9a73368c4e11d61cf0cafcb7185..66c5aa5ff784e84b6a2648276ac9d63b52dd9ec3 100644 (file)
--- a/lib/CodeGen/BranchFolding.cpp
+++ b/lib/CodeGen/BranchFolding.cpp
@@ -945,15 +945,15 @@ bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) {
       }
     }
     
-    // If this block doesn't fall through (e.g. it ends with an uncond branch or
-    // has no successors) and if the pred falls through into this block, and if
-    // it would otherwise fall through into the block after this, move this
-    // block to the end of the function.
+    // If this block has no successors (e.g. it is a return block or ends with
+    // a call to a no-return function like abort or __cxa_throw) and if the pred
+    // falls through into this block, and if it would otherwise fall through
+    // into the block after this, move this block to the end of the function.
     //
     // We consider it more likely that execution will stay in the function (e.g.
     // due to loops) than it is to exit it.  This asserts in loops etc, moving
     // the assert condition out of the loop body.
-    if (!PriorCond.empty() && PriorFBB == 0 &&
+    if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 &&
         MachineFunction::iterator(PriorTBB) == FallThrough &&
         !CanFallThrough(MBB)) {
       bool DoTransform = true;

diff --git a/test/CodeGen/CellSPU/2009-01-01-BrCond.ll b/test/CodeGen/CellSPU/2009-01-01-BrCond.ll
index 35422311c57433c836143b6e4d1b93ae673f623b..58e3190454f89e4d054a30d6fba3665294cd4010 100644 (file)
--- a/test/CodeGen/CellSPU/2009-01-01-BrCond.ll
+++ b/test/CodeGen/CellSPU/2009-01-01-BrCond.ll
@@ -1,4 +1,4 @@
-; RUN: llc < %s -march=cellspu -o - | grep brz
+; RUN: llc < %s -march=cellspu -o - | grep brnz
 ; PR3274
 
 target datalayout = "E-p:32:32:128-i1:8:128-i8:8:128-i16:16:128-i32:32:128-i64:32:128-f32:32:128-f64:64:128-v64:64:64-v128:128:128-a0:0:128-s0:128:128"

diff --git a/test/CodeGen/Generic/switch-lower-feature-2.ll b/test/CodeGen/Generic/switch-lower-feature-2.ll
index d6e56471c364c07838fb1b51727ee23883233779..80e0618e05f5df6a4771437bbbb1fea1dc8a2c67 100644 (file)
--- a/test/CodeGen/Generic/switch-lower-feature-2.ll
+++ b/test/CodeGen/Generic/switch-lower-feature-2.ll
@@ -5,9 +5,9 @@
 ; RUN: grep 1023 %t | count 1
 ; RUN: grep 119  %t | count 1
 ; RUN: grep JTI %t | count 2
-; RUN: grep jg %t | count 1
+; RUN: grep jg %t | count 3
 ; RUN: grep ja %t | count 1
-; RUN: grep js %t | count 1
+; RUN: grep jns %t | count 1
 
 target triple = "i686-pc-linux-gnu"
 

diff --git a/test/CodeGen/X86/loop-blocks.ll b/test/CodeGen/X86/loop-blocks.ll
index af50bd95f3bde4bf930c9242ffca240638a05c5b..c0379d115220cd5e93aeb64dfdaa8054dc1d357a 100644 (file)
--- a/test/CodeGen/X86/loop-blocks.ll
+++ b/test/CodeGen/X86/loop-blocks.ll
@@ -74,16 +74,16 @@ exit:
 ; CHECK: yet_more_involved:
 ;      CHECK:   jmp .LBB3_1
 ; CHECK-NEXT:   align
-; CHECK-NEXT: .LBB3_7:
-; CHECK-NEXT:   call block_a_true_func
-; CHECK-NEXT:   jmp .LBB3_4
-; CHECK-NEXT: .LBB3_2:
+; CHECK-NEXT: .LBB3_3:
 ; CHECK-NEXT:   call bar99
 ; CHECK-NEXT:   call get
 ; CHECK-NEXT:   cmpl $2999, %eax
-; CHECK-NEXT:   jle .LBB3_7
+; CHECK-NEXT:   jg .LBB3_5
+; CHECK-NEXT:   call block_a_true_func
+; CHECK-NEXT:   jmp .LBB3_6
+; CHECK-NEXT: .LBB3_5:
 ; CHECK-NEXT:   call block_a_false_func
-; CHECK-NEXT: .LBB3_4:
+; CHECK-NEXT: .LBB3_6:
 ; CHECK-NEXT:   call block_a_merge_func
 ; CHECK-NEXT: .LBB3_1:
 ; CHECK-NEXT:   call body