llvm MemorySSA def-use chains


I have a question about how llvm MemorySSA works, as seems I misunderstand something.

Consider following code snippet and corresponding IR with MemorySSA annotations (got with opt -print-memoryssa)

void foo(int* b) {

int a = 0;

int d = 12;

if (b) {

a = 42;

d = 32;


int c = a;

int e = d;


; Function Attrs: noinline nounwind optnone uwtable

define void @foo(i32* %b) #0 {


%b.addr = alloca i32*, align 8

%a = alloca i32, align 4

%d = alloca i32, align 4

%c = alloca i32, align 4

%e = alloca i32, align 4

; 1 = MemoryDef(liveOnEntry)

store i32* %b, i32** %b.addr, align 8

; 2 = MemoryDef(1)

store i32 0, i32* %a, align 4

; 3 = MemoryDef(2)

store i32 12, i32* %d, align 4

; MemoryUse(1)

%0 = load i32*, i32** %b.addr, align 8

%tobool = icmp ne i32* %0, null

br i1 %tobool, label %if.then, label %if.end

if.then: ; preds = %entry

; 4 = MemoryDef(3)

store i32 42, i32* %a, align 4

; 5 = MemoryDef(4)

store i32 32, i32* %d, align 4

br label %if.end

if.end: ; preds = %if.then, %entry

; 9 = MemoryPhi({entry,3},{if.then,5})

; MemoryUse(9)

%1 = load i32, i32* %a, align 4

; 6 = MemoryDef(9)

store i32 %1, i32* %c, align 4

; MemoryUse(9)

%2 = load i32, i32* %d, align 4

; 7 = MemoryDef(6)

store i32 %2, i32* %e, align 4

ret void


Tracking back def-use chain for %1 = load i32, i32* %a, align 4 first I get 9 = MemoryPhi({entry,3},{if.then,5}) and then as definitions MemoryDef(2) which corresponds to store i32 12, i32* %d, align 4 and MemoryDef(4) corresponding to store i32 32, i32* %d, align 4. However neither of these stores define a, they both define d. What I was expecting to get was MemoryDef(3) and MemoryDef(1), which are real definitions for a.

Could you please explain what’s the reason MemorySSA gives incorrect definitions, and is there any way to get the correct ones.



My understanding is that MemorySSA doesn’t know whether memory locations can alias or not.


IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.


try adding some alias analysis to the pipeline: see examples in
for example assume.ll has
; RUN: opt -aa-pipeline=basic-aa -passes='print<memoryssa>,verify<memoryssa>'

without basic-aa the memory ssa will safely assume that the stores to
%a and %d alias which may be proven disjoint by one of the alias

Hi Sebastian,

Thanks for the answer. I’ve tried to run memoryssa piping it with alias analysis, as you’ve suggested, however the output is still the same. As a def for use of a I’m still getting a def of d.

You can see that there’s a chain between def-def as well - meaning that the later def updates the former def. Hence the result is not incorrect, it is just inaccurate.

For brevity of memory ssa, LLVM (as well as GCC) is maintaining at most 1 ‘latest’ def per block. (https://llvm.org/docs/MemorySSA.html#precision , keyword ‘partition’ means something similar to the number of latest def) You can test GCC with -fno-tree-vectorize -fdump-tree-all-vops as well.
In your example, 3 is the latest def of block entry, and 5 is the latest def of block if.then.
Having more than one latest def can increase number of MemoryPhis. Verbose ssa generation is bad for for compilation time & memory footprint.

Best Regards,
Juneyoung Lee

Hi Lee,

Thanks for the explanation, now it’s clear how MemorySSA works and why I’m getting these results.