UCS/SYS: Disable "always inline" when -Og flag is present by shasson5 · Pull Request #11251 · openucx/ucx

shasson5 · 2026-03-11T18:48:11Z

What?

Disable "always inline" when -Og flag is present

Why?

Support -Og optimization level

tvegas1 · 2026-03-12T09:16:26Z

        uct_pack_callback_t pack_cb, void *arg, const uct_iov_t *iov,
        size_t iovcnt, unsigned flags)
 {
+    uint8_t elem_flags = 0;


i think it is not needed?

tvegas1 · 2026-03-12T09:22:43Z

+#
+# Check if -Og flag was supplied
+#
+AS_IF([echo " $BASE_CFLAGS $CFLAGS " | grep -q -- ' -Og '], 


maybe enhance --enable-compiler-opt to support 0,1,2,3,g since as far as I understand they are all mutually exclusive? and maybe AC_DEFINE HAVE_OG_OPTIMIZATION for --enable-compiler-opt=g?

notice that -Og can be passed from CFLAGS which comes from env (see example in CI job).
I think that the existing block should remain as is because it has a single well-defined output (BASE_CFLAGS).

BTW, which functions fail to compile with -Og ?
Maybe we could convert them to non-forced inline w/o performance impact

Those are the failed functions:

ucp_proto_request_bcopy_complete_success

ucp_proto_put_offload_bcopy_send_func

ucp_proto_rndv_put_common_flush_send

ucp_proto_rndv_put_common_data_sent

ucp_proto_rndv_put_mtype_send_func

how would you suggest converting them?

or did you mean to convert only them? @yosefe

but then if we don't recognize -O3 the impact will not be so bad, worse case some functions may be non inline which may impact the performance a little. The impact of mis detection will not fail compilation.

Notice that built-in gcc flags __ OPTIMIZE __ and __ NO_INLINE __ are set for all opt levels
except O0, so it's unuseful for us.

So basically the plan is:

Add a flag in complier.m4 to identify high opt level (which includes O2 and O3).

if this flag is set, we define UCS_F_INLINE_OPTIMIZED in complier_def.h

apply UCS_F_INLINE_OPTIMIZED to list of functions above

@yosefe is this correct?

Notice that built-in gcc flags __ OPTIMIZE __ and __ NO_INLINE __ are set for all opt levels
except O0, so it's unuseful for us.

Right

Regarding the plan:

It's enough to do it as part of compiler-opt logic.

If we set optimization level to 2 or higher as part of this macro - define UCS_F_INLINE_OPTIMIZED to always_inline. Otherwise, define it as regular inline

Yes

It's enough to do it as part of compiler-opt logic.

without considering CFLAGS=-Og?

Yes, just setting CFLAGS=-Og without compiler-opt will keep UCS_F_INLINE_OPTIMIZED as "inline"

tvegas1 · 2026-03-17T10:25:39Z

regarding -Og passed as CFLAGS, if it is only about the line below, maybe we could fix this use-case to pass it using --enable-compiler-opt=g and avoid option detection, there could still be new users of always inline?

./buildlib/tools/builds.sh:	build_gcc CFLAGS=-Og CXXFLAGS=-Og

shasson5 · 2026-03-18T10:23:34Z

regarding -Og passed as CFLAGS, if it is only about the line below, maybe we could fix this use-case to pass it using --enable-compiler-opt=g and avoid option detection

so I guess the question is how you want handle a contradiction between CFLAGS and --enable-compiler-opt.
You can either decide to always give precedence to CFLAGS and ignore --enable-compliler-opt or treat it as error usecase (and then throw responsibility to the user or display error message)

there could still be new users of always inline?

can you explain this question?

tvegas1 · 2026-03-18T10:57:18Z

so I guess the question is how you want handle a contradiction between CFLAGS and --enable-compiler-opt. You can either decide to always give precedence to CFLAGS and ignore --enable-compliler-opt or treat it as error usecase (and then throw responsibility to the user or display error message)

maybe we can handle --enable-compiler-opt=g only as i think by default add -O3 anyways, which could already conflict with an -O specified from CFLAGS?

there could still be new users of always inline?

can you explain this question?

I understand the -Og can cause build failure for always inline. now we are able to remove the failures by moving to always inline to inline. But we would still need to have a macro set when -Og, to downgrade always inline to inline.

shasson5 · 2026-03-18T11:21:46Z

maybe we can handle --enable-compiler-opt=g only as i think by default add -O3 anyways, which could already conflict with an -O specified from CFLAGS?

won't it cause CFLAGS=-Og to fail compilation?

I understand the -Og can cause build failure for always inline. now we are able to remove the failures by moving to always inline to inline. But we would still need to have a macro set when -Og, to downgrade always inline to inline.

right that's the idea

tvegas1 · 2026-03-19T11:10:57Z

+#
+# Define OPTIMIZE_HIGH for optimization levels -O2 or -O3
+#
+AS_IF([echo " $BASE_CFLAGS $CFLAGS " | grep -qE -- ' -O(2|3) '],


maybe i am getting confused, but the case where CI sets -Og would not be handled right, as below comes with -O3 and then -Og, and AFAIU, gcc will use -Og, where it should not add the inlines?

$ ./contrib/configure-devel CFLAGS=-Og configure: CFLAGS: -O3 -g -Wall ... -Og

shasson5 · 2026-03-19T21:04:11Z

@yosefe please review

yosefe · 2026-03-20T09:46:03Z

+opt_level=""
+for flag in $BASE_CFLAGS $CFLAGS; do
+    case $flag in -O*) opt_level=$flag;; esac
+done
+AS_IF([test "x$opt_level" = "x-O2" || test "x$opt_level" = "x-O3"],
+      [AC_DEFINE([OPTIMIZE_HIGH], 1, [Compiled with high optimization level])])


the approach is not bad. i would fine tune the implementation:

use shell pattern substitution to set "opt_level" to the exact optimization level that follows the -O argument. The -O argument can be a number or a string /(like "z", "fast", "g", etc)

define OPTIMIZATION_LEVEL macro to be the numeric optimization level. If the optimization is not a number, the macro is left undefined

in compiler_def.h we CANNOT use OPTIMIZATION_LEVEL macro, because it does not include config.h. need to use compiler.h

in compiler.h, need to check if OPTIMIZATION_LEVEL is defined and its level is >=2

) * UCS/SYS: Disable "always inline" when -Og flag is present * UCS/SYS: CR fixes * UCS/SYS: added complier.m4 fix * UCT/MM: reverted change * UCS/BUILD: CR fix * CONFIG: Claude CR fixes --------- Co-authored-by: Yossi Itigin <yosefe@nvidia.com>

UCS/SYS: Disable "always inline" when -Og flag is present

13358b1

tvegas1 reviewed Mar 12, 2026

View reviewed changes

shasson5 added 3 commits March 18, 2026 16:41

UCS/SYS: CR fixes

44f5480

UCS/SYS: added complier.m4 fix

11eb6e2

UCT/MM: reverted change

027a0ab

shasson5 requested review from tvegas1 and yosefe March 18, 2026 18:39

tvegas1 reviewed Mar 19, 2026

View reviewed changes

UCS/BUILD: CR fix

3733989

shasson5 requested a review from tvegas1 March 19, 2026 16:15

tvegas1 reviewed Mar 19, 2026

View reviewed changes

Comment thread config/m4/compiler.m4

shasson5 requested a review from tvegas1 March 19, 2026 21:06

yosefe reviewed Mar 20, 2026

View reviewed changes

CONFIG: Claude CR fixes

d4e1d2b

tvegas1 approved these changes Mar 20, 2026

View reviewed changes

yosefe enabled auto-merge (squash) March 21, 2026 09:26

yosefe merged commit 4f24dbe into openucx:master Mar 21, 2026
152 checks passed

shasson5 mentioned this pull request Mar 22, 2026

UCS/SYS: Disable "always inline" when -Og flag is present - v1.21.x #11281

Merged

Conversation

shasson5 commented Mar 11, 2026

What?

Why?

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

shasson5 Mar 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

shasson5 Mar 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

shasson5 Mar 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tvegas1 commented Mar 17, 2026

Uh oh!

shasson5 commented Mar 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

tvegas1 commented Mar 18, 2026

Uh oh!

shasson5 commented Mar 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

shasson5 commented Mar 19, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

shasson5 Mar 15, 2026 •

edited

Loading

shasson5 Mar 15, 2026 •

edited

Loading

shasson5 Mar 18, 2026 •

edited

Loading

shasson5 commented Mar 18, 2026 •

edited

Loading

shasson5 commented Mar 18, 2026 •

edited

Loading