maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

Move the re-sort of invalidated NonLocalPointerDeps cache earlier 

so that all code paths get it.  PR4256 was about a case where the
phi translation loop would find all preds in the Visited cache, so
it could get by without re-sorting the NonLocalPointerDeps cache.
Fix this by resorting it earlier, there is no reason not to do this.

This patch inspired by Jakub Staszak's patch.

llvm-svn: 75476
This commit is contained in:
Chris Lattner 2009-07-13 17:14:23 +00:00
parent 9c3a0188cd
commit 2f0c1c44d5
2 changed files with 81 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
llvm/lib/Analysis/MemoryDependenceAnalysis.cpp
View File

@ -731,10 +731,22 @@ getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize,
// If we do need to do phi translation, then there are a bunch of different
// cases, because we have to find a Value* live in the predecessor block. We
// know that PtrInst is defined in this block at least.
// We may have added values to the cache list before this PHI translation.
// If so, we haven't done anything to ensure that the cache remains sorted.
// Sort it now (if needed) so that recursive invocations of
// getNonLocalPointerDepFromBB and other routines that could reuse the cache
// value will only see properly sorted cache arrays.
if (Cache && NumSortedEntries != Cache->size()) {
std::sort(Cache->begin(), Cache->end());
NumSortedEntries = Cache->size();
}
// If this is directly a PHI node, just use the incoming values for each
// pred as the phi translated version.
if (PHINode *PtrPHI = dyn_cast<PHINode>(PtrInst)) {
Cache = 0;
for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
BasicBlock *Pred = *PI;
Value *PredPtr = PtrPHI->getIncomingValueForBlock(Pred);
@ -759,15 +771,6 @@ getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize,
goto PredTranslationFailure;
}
// We may have added values to the cache list before this PHI
// translation. If so, we haven't done anything to ensure that the
// cache remains sorted. Sort it now (if needed) so that recursive
// invocations of getNonLocalPointerDepFromBB that could reuse the cache
// value will only see properly sorted cache arrays.
if (Cache && NumSortedEntries != Cache->size())
std::sort(Cache->begin(), Cache->end());
Cache = 0;
// FIXME: it is entirely possible that PHI translating will end up with
// the same value. Consider PHI translating something like:
// X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
@ -779,7 +782,7 @@ getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize,
Result, Visited))
goto PredTranslationFailure;
}
// Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
CacheInfo = &NonLocalPointerDeps[CacheKey];
Cache = &CacheInfo->second;
@ -806,11 +809,8 @@ getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize,
CacheInfo = &NonLocalPointerDeps[CacheKey];
Cache = &CacheInfo->second;
NumSortedEntries = Cache->size();
} else if (NumSortedEntries != Cache->size()) {
std::sort(Cache->begin(), Cache->end());
NumSortedEntries = Cache->size();
}
// Since we did phi translation, the "Cache" set won't contain all of the
// results for the query. This is ok (we can still use it to accelerate
// specific block queries) but we can't do the fastpath "return all

67
llvm/test/Transforms/GVN/2009-07-13-MemDepSortFail.ll Normal file
View File

@ -0,0 +1,67 @@
; RUN: llvm-as < %s | opt -gvn | llvm-dis
; PR4256
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-f80:32:32"
target triple = "i386-linux-gnu"
%llvm.dbg.anchor.type = type { i32, i32 }
%struct.cset = type { i8*, i8, i8, i32, i8* }
%struct.lmat = type { %struct.re_guts*, i32, %llvm.dbg.anchor.type*, i8*, i8*, i8*, i8*, i8**, i32, i8*, i8*, i8*, i8*, i8* }
%struct.re_guts = type { i32*, %struct.cset*, i8*, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i8*, i8*, i32, i32, i32, i32, [1 x i8] }
define i8* @lbackref(%struct.lmat* %m, i8* %start, i8* %stop, i32 %startst, i32 %stopst, i32 %lev, i32 %rec) nounwind {
entry:
br label %bb63
bb: ; preds = %bb63
switch i32 0, label %bb62 [
i32 268435456, label %bb2
i32 805306368, label %bb9
i32 -1610612736, label %bb51
]
bb2: ; preds = %bb
br label %bb62
bb9: ; preds = %bb
%0 = load i8* %sp.1, align 1 ; <i8> [#uses=0]
br label %bb62
bb51: ; preds = %bb
%1 = load i8* %sp.1, align 1 ; <i8> [#uses=0]
ret i8* null
bb62: ; preds = %bb9, %bb2, %bb
br label %bb63
bb63: ; preds = %bb84, %bb69, %bb62, %entry
%sp.1 = phi i8* [ null, %bb62 ], [ %sp.1.lcssa, %bb84 ], [ %start, %entry ], [ %sp.1.lcssa, %bb69 ] ; <i8*> [#uses=3]
br i1 false, label %bb, label %bb65
bb65: ; preds = %bb63
%sp.1.lcssa = phi i8* [ %sp.1, %bb63 ] ; <i8*> [#uses=4]
br i1 false, label %bb66, label %bb69
bb66: ; preds = %bb65
ret i8* null
bb69: ; preds = %bb65
switch i32 0, label %bb108.loopexit2.loopexit.loopexit [
i32 1342177280, label %bb63
i32 1476395008, label %bb84
i32 1879048192, label %bb104
i32 2013265920, label %bb93
]
bb84: ; preds = %bb69
%2 = tail call i8* @lbackref(%struct.lmat* %m, i8* %sp.1.lcssa, i8* %stop, i32 0, i32 %stopst, i32 0, i32 0) nounwind ; <i8*> [#uses=0]
br label %bb63
bb93: ; preds = %bb69
ret i8* null
bb104: ; preds = %bb69
%sp.1.lcssa.lcssa33 = phi i8* [ %sp.1.lcssa, %bb69 ] ; <i8*> [#uses=0]
unreachable
bb108.loopexit2.loopexit.loopexit: ; preds = %bb69
ret i8* null
}