SSA in the Front End

Project Infrastructure LLVM Dev List Archives
1

Hi,

I have been looking into the code that generates the LLVM assembly in the LLVM front end, but I am
not very sure if at the time that the llvm_c_expand_body_1 function is called, the SSA form was
already constructed (each definition dominates all the uses). Can somebody please tell me?

Thanks

2

Ricardo wrote:

Hi,

I have been looking into the code that generates the LLVM assembly in the LLVM front end, but I am
not very sure if at the time that the llvm_c_expand_body_1 function is called, the SSA form was
already constructed (each definition dominates all the uses). Can somebody please tell me?

The LLVM GCC frontend does not translate variables directly into LLVM virtual regsiters (which must be in SSA form).

Instead, the LLVM GCC frontend simply uses a memory location for each variable (e.g. a local variable is allocated via the LLVM alloca instruction). If you use llvm-gcc -S -o file.ll file.c, you can see the code that llvm-gcc generates.

Next, gccas (after assembling the LLVM assembly file) runs a pass called mem2reg. This pass finds memory locations that can be promoted to LLVM virtual registers and promotes them into registers. The following commands will allow you to see the change:

llvm-as -o file.bc file.ll
opt -mem2reg -o newfile.bc file.bc
llvm-dis -o newfile.ll newfile.bc

If you compare newfile.ll to file.ll, you'll see what the mem2reg transformation does.

Using this technique, language frontends don't need to worry about maintaining SSA for variables that it generates. It simply generates them as memory locations and the LLVM optimizer fixes it up into more efficient code.

-- John T.

3

Thanks for the explanation. It's more clear now

The only thing that seems strange is that in the function llvm_expand_shortcircuit_truth_expr in
the front end, there is the creation of a PHI instruction. If there is no SSA yet, why do you do
that?

Thanks in advance

4

Thanks for the explanation. It's more clear now

The only thing that seems strange is that in the function llvm_expand_shortcircuit_truth_expr in
the front end, there is the creation of a PHI instruction. If there is no SSA yet, why do you do
that?

That is just a shortcut. We could introduce an alloca for every shortcircuit expression, but as a minor compile time efficiency issue, we don't.

-Chris