Inefficient code for passing an alloca as argument to a function

[llvmbot]

Bugzilla Link	10432
Version	trunk
OS	Linux
Reporter	LLVM Bugzilla Contributor
CC	@asl,@lattner,@stoklund

Extended Description

Given

extern bool ToNumberSlow(double *dp);
void Interpret() {
double l, r;
for (;;) {
ToNumberSlow(&l);
ToNumberSlow(&r);
}
}

We produce what is probably the best IL:

define void @​_Z9Interpretv() noreturn nounwind uwtable {
entry:
%l = alloca double, align 8
%r = alloca double, align 8
br label %for.cond
for.cond:
%call = call zeroext i1 @​_Z12ToNumberSlowPd(double* %l) nounwind
%call1 = call zeroext i1 @​_Z12ToNumberSlowPd(double* %r) nounwind
br label %for.cond
}
declare zeroext i1 @​_Z12ToNumberSlowPd(double*)

but that then codegens to:

    pushq   %r14
    pushq   %rbx
    subq    $24, %rsp
    leaq    16(%rsp), %r14
    leaq    8(%rsp), %rbx
    .align  16, 0x90

.LBB0_1:
movq %r14, %rdi
callq _Z12ToNumberSlowPd
movq %rbx, %rdi
callq _Z12ToNumberSlowPd
jmp .LBB0_1

gcc produces

    subq    $24, %rsp

.L2:
movq %rsp, %rdi
call _Z12ToNumberSlowPd
leaq 8(%rsp), %rdi
call _Z12ToNumberSlowPd
jmp .L2

[lattner]

It seems like an lea from the stack like this should be considered "as cheap as a move"?

[stoklund]

Yes.

We need to split live ranges to eliminate copies. That isn't implemented yet, but it is on my list.

Currently, we only split live ranges to avoid spilling.

[llvmbot]

Do you have some pointers to what needs to be done? I might have time to implement this.

[stoklund]

You can start by figuring out why the coalescer doesn't do the remat with -join-physregs:

32L %vreg2 = LEA64r <fi#1>, 1, %noreg, 0, %noreg; GR64:%vreg2
144L %RDI = COPY %vreg2; GR64:%vreg2
Considering merging %vreg2 with %RDI
RHS = %vreg2 = [32d,208L:0) 0@32d
LHS = %RDI,inf = [80d,96d:0)[96d,96S:1)[144d,160d:2)[160d,160S:3) 0@80d 1@96d 2@144d 3@160d
Interference!