[llvm-commits] [llvm] r114785 - in /llvm/trunk: lib/Transforms/Scalar/GVN.cpp test/Transforms/GVN/pre-single-pred.ll

[Owen Anderson]

Author: resistor
Date: Sat Sep 25 00:26:18 2010
New Revision: 114785

URL: http://llvm.org/viewvc/llvm-project?rev=114785&view=rev
Log:
LoadPRE was not properly checking that the load it was PRE'ing post-dominated the block it was being hoisted to.
Splitting critical edges at the merge point only addressed part of the issue; it is also possible for non-post-domination
to occur when the path from the load to the merge has branches in it.  Unfortunately, full anticipation analysis is
time-consuming, so for now approximate it.  This is strictly more conservative than real anticipation, so we will miss
some cases that real PRE would allow, but we also no longer insert loads into paths where they didn't exist before. :-)

This is a very slight net positive on SPEC for me (0.5% on average).  Most of the benchmarks are largely unaffected, but
when it pays off it pays off decently: 181.mcf improves by 4.5% on my machine.

Modified:
    llvm/trunk/lib/Transforms/Scalar/GVN.cpp
    llvm/trunk/test/Transforms/GVN/pre-single-pred.ll

Modified: llvm/trunk/lib/Transforms/Scalar/GVN.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/GVN.cpp?rev=114785&r1=114784&r2=114785&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/Scalar/GVN.cpp (original)
+++ llvm/trunk/lib/Transforms/Scalar/GVN.cpp Sat Sep 25 00:26:18 2010
@@ -1532,8 +1532,14 @@
       return false;
     if (Blockers.count(TmpBB))
       return false;
+    
+    // If any of these blocks has more than one successor (i.e. if the edge we
+    // just traversed was critical), then there are other paths through this 
+    // block along which the load may not be anticipated.  Hoisting the load 
+    // above this block would be adding the load to execution paths along
+    // which it was not previously executed.
     if (TmpBB->getTerminator()->getNumSuccessors() != 1)
-      allSingleSucc = false;
+      return false;
   }
 
   assert(TmpBB);

Modified: llvm/trunk/test/Transforms/GVN/pre-single-pred.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/GVN/pre-single-pred.ll?rev=114785&r1=114784&r2=114785&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/GVN/pre-single-pred.ll (original)
+++ llvm/trunk/test/Transforms/GVN/pre-single-pred.ll Sat Sep 25 00:26:18 2010
@@ -1,4 +1,13 @@
-; RUN: opt < %s -gvn -enable-load-pre -S | not grep {tmp3 = load}
+; RUN: opt < %s -gvn -enable-load-pre -S | FileCheck %s
+; This testcase assumed we'll PRE the load into %for.cond, but we don't actually
+; verify that doing so is safe.  If there didn't _happen_ to be a load in
+; %for.end, we would actually be lengthening the execution on some paths, and
+; we were never actually checking that case.  Now we actually do perform some
+; conservative checking to make sure we don't make paths longer, but we don't
+; currently get this case, which we got lucky on previously.
+;
+; Now that that faulty assumption is corrected, test that we DON'T incorrectly
+; hoist the load.  Doing the right thing for the wrong reasons is still a bug.
 
 @p = external global i32
 define i32 @f(i32 %n) nounwind {
@@ -13,6 +22,8 @@
 for.cond.for.end_crit_edge:		; preds = %for.cond
 	br label %for.end
 
+; CHECK: for.body:
+; CHECK-NEXT: %tmp3 = load i32* @p
 for.body:		; preds = %for.cond
 	%tmp3 = load i32* @p		; <i32> [#uses=1]
 	%dec = add i32 %tmp3, -1		; <i32> [#uses=2]
@@ -20,6 +31,7 @@
 	%cmp6 = icmp slt i32 %dec, 0		; <i1> [#uses=1]
 	br i1 %cmp6, label %for.body.for.end_crit_edge, label %for.inc
 
+; CHECK: for.body.for.end_crit_edge:
 for.body.for.end_crit_edge:		; preds = %for.body
 	br label %for.end