CODEWITHSUNDEEP
assemblycompiler-constructionllvmcompiler-optimizationllvm-ir
Myk
Myk

Reputation: 1608

How does LLVM avoid generating redundant native code for the `br` IR instruction?

For the following C code

void foo() {
    int forty_two = 42;
    if (forty_two == 42) {
    }
}

clang -S -emit-llvm foo.c emits this IR code:

define dso_local void @foo() #0 {
  %1 = alloca i32, align 4
  store i32 42, i32* %1, align 4
  %2 = load i32, i32* %1, align 4
  %3 = icmp eq i32 %2, 42
  br i1 %3, label %4, label %5

4:                                                ; preds = %0
  br label %5

5:                                                ; preds = %4, %0
  ret void
}

And for the IR, LLVM (llc foo.ll) generates the x64 code below. It's abridged for readability.

foo:                                    # @foo
# %bb.0:
    pushq   %rbp
    movq    %rsp, %rbp
    movl    $42, -4(%rbp)
    cmpl    $42, -4(%rbp)
    jne .LBB0_2
# %bb.1:
    jmp .LBB0_2
.LBB0_2:
    popq    %rbp
    retq

In contrast to the native code emitted by LLVM, translating the IR code in a straightforward way would contain a number of redundant instructions. Something along these lines:

foo:
# %bb.0:
    pushq   %rbp
    movq    %rsp, %rbp
    movl    $42, -4(%rbp)
    cmpl    $42, -4(%rbp)

# create the i1 boolean in a register.
# (This instruction is redundant and LLVM doesn't emit is)
    sete    %al
# See whether the comparison's result was `true` or `false`.
# (This instruction is redundant and LLVM doesn't emit it)
    cmpb    $1, %al

    jne .LBB0_2
# %bb.1:
    jmp .LBB0_2
.LBB0_2:
    popq    %rbp
    retq

My question is: Where is the portion of LLVM code that makes sure these redundant instructions are not emitted? And how does it work?

I read an excellent post Life of an instruction in LLVM by @Eli Bendersky and looked at the code in SelectionDAGBuilder::visitICmp and SelectionDAGBuilder::visitBr. But didn't manage to figure out the answer on my own.

Upvotes: 3

Views: 492

Answers (2)

Myk
Myk

Reputation: 1608

TLDR: X86FastISel::X86SelectBranch

Someone on LLVM's discord told me about the -print-after-all flag of llc. (Actually, @arnt mentioned it in their answer even before I asked on discord and I have no idea why I didn't give the flag a try right away...)

That flag let me see that "X86 DAG->DAG Instruction Selection" was the first pass that didn't only transform IR but turned it into x86-specific Machine IR (MIR). (The corresponding class is X86DAGToDAGISel).
And from the MIR it emitted, it was clear the decision to emit or not SETCC/TEST instructions happens during the pass run.

Stepping through X86DAGToDAGISel::runOnMachineFunction eventually brought me to X86FastISel::X86SelectBranch. There,

  • in case br is the only user of icmp's result and the instructions are in the same basic block, the pass decides not to emit SETCC/TEST
  • in case icmp's result has other users or the two IR instructions aren't in the same basic block, the pass will actually emit SETCC/TEST.

So, for this C code:

void foo() {
    int forty_two = 42;
    int is_forty_two;
    if (is_forty_two = (forty_two == 42)) {
    }
}

clang -S -emit-llvm brcond.c produces the following IR:

define void @foo() #0 {
entry:
  %forty_two = alloca i32, align 4
  %is_forty_two = alloca i32, align 4
  store i32 42, i32* %forty_two, align 4
  %0 = load i32, i32* %forty_two, align 4
  %cmp = icmp eq i32 %0, 42
  %conv = zext i1 %cmp to i32
  store i32 %conv, i32* %is_forty_two, align 4
  br i1 %cmp, label %if.then, label %if.end

if.then:                                          ; preds = %entry
  br label %if.end

if.end:                                           ; preds = %if.then, %entry
  ret void
}

Obviously, %cmp has more than one user. So llc brcond.ll emits the assembly below (abridged a bit):

foo:                                    # @foo
# %bb.0:                                # %entry
    pushq   %rax
    movl    $42, (%rsp)
    cmpl    $42, (%rsp)
    sete    %al
    movb    %al, %cl
    andb    $1, %cl
    movzbl  %cl, %ecx
    movl    %ecx, 4(%rsp)
    testb   $1, %al
    jne .LBB1_1
    jmp .LBB1_2
.LBB1_1:                                # %if.then
    jmp .LBB1_2
.LBB1_2:                                # %if.end
    popq    %rax
    retq

Upvotes: 1

arnt
arnt

Reputation: 9675

LLVM runs passes that change the code in beneficial ways. Each pass decides what "beneficial" means. Am I right in assuming that you're more interested in a generic answer and are using that br as an example? If so, the -print-after-all flag, which instructs the compiler to print the IR after each of the passes, may be the what you want. There's also a -print-before-all and more specific flags.

Reading the output and seeing how it changes gfives you a good overview of which passes add/eliminate which warts.

Upvotes: 1

Related Questions