LTO aggressively optimizing out function despite explicit calls in logical decoding

Started by Arkady Skvorcov4 months ago2 messages
Jump to latest
#1Arkady Skvorcov
arkashaskv@gmail.com

Hi contributors,

I'm working on a logical decoding improvement and running into a persistent
build issue where LTO (-flto=thin) is optimizing out my LogicalWaitForWal
function, even though there are explicit calls to it.

The situation:

Function LogicalWaitForWal() is defined in
src/backend/replication/logical/logical.c

It's called directly from logical_read_xlog_page_cv() in the same file

logical_read_xlog_page_cv is assigned as the page_read callback via
XL_ROUTINE in multiple places in logical.c

The function appears in the object file (nm logical.o shows it)

But it's completely absent from the final binary (nm postgres shows nothing)

What I've tried so far:

Direct calls: The function is called explicitly from
logical_read_xlog_page_cv

Attributes: __attribute__((used)), visibility("default"), retain

Global references:

c
void *_force_keep __attribute__((used)) = (void *)LogicalWaitForWal;
Makefile overrides: Attempted to disable LTO for just logical.o

Various force-linkage patterns: Global function pointers, used arrays, etc.

The function signature:
c
XLogRecPtr LogicalWaitForWal(XLogRecPtr targetLSN);
It's called here in logical_read_xlog_page_cv:
c
flushptr = LogicalWaitForWal(targetPagePtr + reqLen);
And logical_read_xlog_page_cv is assigned via:
c
XL_ROUTINE(.page_read = logical_read_xlog_page_cv, ...)
The puzzling part is that both functions are in the same compilation unit,
there's an explicit call, yet LTO still removes LogicalWaitForWal.

Is there a PostgreSQL-specific pattern or macro I should be using to ensure
functions aren't optimized out when they're called via function pointers
assigned in structs?

Thank you for your time!

Thanks,
Arkady Skvorcov

#2Andres Freund
andres@anarazel.de
In reply to: Arkady Skvorcov (#1)
Re: LTO aggressively optimizing out function despite explicit calls in logical decoding

Hi,

On 2025-11-06 11:39:03 +0400, Arkady Skvorcov wrote:

I'm working on a logical decoding improvement and running into a persistent
build issue where LTO (-flto=thin) is optimizing out my LogicalWaitForWal
function, even though there are explicit calls to it.

The situation:

Function LogicalWaitForWal() is defined in
src/backend/replication/logical/logical.c

It's called directly from logical_read_xlog_page_cv() in the same file

logical_read_xlog_page_cv is assigned as the page_read callback via
XL_ROUTINE in multiple places in logical.c

The function appears in the object file (nm logical.o shows it)

But it's completely absent from the final binary (nm postgres shows nothing)

Presumably what is happening is that it is inlined into the caller(s)?

Is there a PostgreSQL-specific pattern or macro I should be using to ensure
functions aren't optimized out when they're called via function pointers
assigned in structs?

What problem is this actually causing for you?

Greetings,

Andres Freund