I am trying to use the results of LLVMs built-in MemoryDependenceAnalysis (MDA) in a custom LLVM pass that I'm working on. Given some instruction which reads from memory (a load, for example), i'd like MDA to tell me all of the previous instructions which may have Defined or Clobbered it. If my understanding of the MDA documentation serves me correctly, MDA should be able to give me this info. However, I'm having a hard time getting the precision that I need out of it. Here's the relevant snippet of a simple test program i have been toying around with:
%1 = alloca i32, align 4
%result = alloca i32, align 4
%x = alloca i32, align 4
%xp = alloca i32*, align 8
store i32 0, i32* %1
store i32 5, i32* %result, align 4, !dbg !14
store i32 7, i32* %x, align 4, !dbg !16
store i32* %x, i32** %xp, align 8, !dbg !19
%2 = load i32* %x, align 4, !dbg !20
%3 = icmp eq i32 %2, 4, !dbg !20
br i1 %3, label %4, label %7, !dbg !22, !dataware.bbuid !23
; <label>:4 ; preds = %0
%5 = load i32** %xp, align 8, !dbg !24
%6 = load i32* %5, align 4, !dbg !26
store i32 %6, i32* %result, align 4, !dbg !27
br label %8, !dbg !28, !dataware.bbuid !29
; <label>:7 ; preds = %0
store i32 42, i32* %result, align 4, !dbg !30
br label %8, !dataware.bbuid !32
The command im using to run the analysis:
opt-3.6 -enable-tbaa -tbaa -basicaa -libcall-aa -scev-aa -globalsmodref-aa -domtree -memdep -print-memdeps -gvn -analyze test.bc
For some instructions, I am getting results as expected. For example, It tells me that %5 = load i32** %xp is dependent on store i32* %x, i32** %xp. However, it doesn't determine that %6 = load i32* %5 is dependent on store i32 7, i32* %x. Here's a snippet of the output:
Def in block %0 from: store i32* %x, i32** %xp, align 8, !dbg !19
%5 = load i32** %xp, align 8, !dbg !24
Unknown in block %4
%6 = load i32* %5, align 4, !dbg !26
The latter case (which it apparently doesn't know how to reason about) seems like it should be easy enough to detect, even with sub-optimal alias analysis. How do I go about investigating why the results of MDA are sub-optimal? And do you have suggestions for getting more precise results? Are there some additional analysis passes I can add to cause MDA to work better? Im using opt 3.6.0 - maybe MDA has been improved since this release?
Thanks.
I got the same results as you obtained. You might use a simpler pointer analysis result to infer that %5 and %x mayalias which in turn infer that the store to %x is potentially loaded from %5.
opt -basicaa -aa-eval -print-all-alias-modref-info test.bc -disable-output
MayAlias: i32* %5, i32* %x