Skip to contentSkip to conversation

Few perf fixes

PATCH In review Mark Hatle 4 participants 11 messages Nov 21, 2013

lore
Share & tools
lore

Mark Hatle

author
#0 lore raw

cover Few perf fixes

Konrad Scherer (1):
perf: Disable -Werror flag

Randy MacLeod (1):
perf: disallow debug optimization.

 meta/recipes-kernel/perf/perf.bb | 16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)
Signature · 1 line
1.8.1.2.545.g2f19ada
author
#1 lore raw

PATCH 1/2 perf: disallow debug optimization.

From: Randy MacLeod <[email protected]>

perf fails to compile if someone tries to compile an entire image as -O0 so in this case, force to use -O2 and give a note about it.

---
 meta/recipes-kernel/perf/perf.bb | 11 +++++++++++
 1 file changed, 11 insertions(+)
meta/recipes-kernel/perf/perf.bb +11 -0
index 269069f..138595d 100644
@@ -60,6 +60,17 @@ B = "${WORKDIR}/${BPN}-${PV}"
 SCRIPTING_DEFINES = "${@perf_feature_enabled('perf-scripting', '', 'NO_LIBPERL=1 NO_LIBPYTHON=1',d)}"
 TUI_DEFINES = "${@perf_feature_enabled('perf-tui', '', 'NO_NEWT=1',d)}"
 
+# perf can't be built without optimization, if someone tries to compile an
+# entire image as -O0, we override it with -O2 here and give a note about it.
+def get_optimization(d):
+    selected_optimization = d.getVar("SELECTED_OPTIMIZATION", True)
+    if base_contains("SELECTED_OPTIMIZATION", "-O0", "x", "", d) == "x":
+        bb.note("eglibc can't be built with -O0, -O2 will be used instead.")
+        return selected_optimization.replace("-O0", "-O2")
+    return selected_optimization
+
+SELECTED_OPTIMIZATION := "${@get_optimization(d)}"
+
 # The LDFLAGS is required or some old kernels fails due missing
 # symbols and this is preferred than requiring patches to every old
 # supported kernel.
Signature · 1 line
1.8.1.2.545.g2f19ada
#2 lore raw

Re: [master][dora][PATCH 1/2] perf: disallow debug optimization.

meta/recipes-kernel/perf/perf.bb B = "${WORKDIR}/${BPN}-${PV}" · line 68

Quoted 4 lines · Mark Hatle
meta/recipes-kernel/perf/perf.bb +4 -0
+def get_optimization(d):
+    selected_optimization = d.getVar("SELECTED_OPTIMIZATION", True)
+    if base_contains("SELECTED_OPTIMIZATION", "-O0", "x", "", d) == "x":
+        bb.note("eglibc can't be built with -O0, -O2 will be used instead.")

Although the text of that warning is correct, users might find the reference to eglibc slightly confusing if it's perf that they're trying to build.

Also, as I mentioned in a different thread not all that long ago when someone submitted a similar patch for gcc-runtime, the proliferation of parse-time python functions to bash SELECTED_OPTIMIZATION around doesn't seem like all that good a thing: this will cause extra overhead for everyone, even those who are not using -O0 and have no interest in perf.

And, finally, it remains slightly unclear to me that this is really a problem that the metadata needs to be solving. I haven't seen any particularly convincing explanation of why this can't or shouldn't just be fixed in the distro configuration.

p.

#3 lore raw

Re: [master][dora][PATCH 1/2] perf: disallow debug optimization.

Quoted 16 lines · Phil Blundell

meta/recipes-kernel/perf/perf.bb B = "${WORKDIR}/${BPN}-${PV}" · line 68

Quoted 4 lines · Mark Hatle
meta/recipes-kernel/perf/perf.bb +4 -0
+def get_optimization(d):
+    selected_optimization = d.getVar("SELECTED_OPTIMIZATION", True)
+    if base_contains("SELECTED_OPTIMIZATION", "-O0", "x", "", d) == "x":
+        bb.note("eglibc can't be built with -O0, -O2 will be used instead.")

Although the text of that warning is correct, users might find the reference to eglibc slightly confusing if it's perf that they're trying to build.

Also, as I mentioned in a different thread not all that long ago when someone submitted a similar patch for gcc-runtime, the proliferation of parse-time python functions to bash SELECTED_OPTIMIZATION around doesn't seem like all that good a thing: this will cause extra overhead for everyone, even those who are not using -O0 and have no interest in perf.

Quoted 4 lines

And, finally, it remains slightly unclear to me that this is really a problem that the metadata needs to be solving. I haven't seen any particularly convincing explanation of why this can't or shouldn't just be fixed in the distro configuration.

I have to admit at this point, this may look better as an include file along the lines of:

SELECTED_OPTIMIZATION = "-O0"
SELECTED_OPTIMIZATION_pn-eglibc = "-O2"
SELECTED_OPTIMIZATION_pn-perf = "-O2"

since clutter the recipes with anonymous python fragments isn't particular desirable.

Cheers,

Richard

author
#4 lore raw

Re: [master][dora][PATCH 1/2] perf: disallow debug optimization.

Quoted 25 lines · Richard Purdie
Quoted 16 lines · Phil Blundell

meta/recipes-kernel/perf/perf.bb B = "${WORKDIR}/${BPN}-${PV}" · line 68

Quoted 4 lines · Mark Hatle
meta/recipes-kernel/perf/perf.bb +4 -0
+def get_optimization(d):
+    selected_optimization = d.getVar("SELECTED_OPTIMIZATION", True)
+    if base_contains("SELECTED_OPTIMIZATION", "-O0", "x", "", d) == "x":
+        bb.note("eglibc can't be built with -O0, -O2 will be used instead.")

Although the text of that warning is correct, users might find the reference to eglibc slightly confusing if it's perf that they're trying to build.

Also, as I mentioned in a different thread not all that long ago when someone submitted a similar patch for gcc-runtime, the proliferation of parse-time python functions to bash SELECTED_OPTIMIZATION around doesn't seem like all that good a thing: this will cause extra overhead for everyone, even those who are not using -O0 and have no interest in perf.

Quoted 4 lines

And, finally, it remains slightly unclear to me that this is really a problem that the metadata needs to be solving. I haven't seen any particularly convincing explanation of why this can't or shouldn't just be fixed in the distro configuration.

I have to admit at this point, this may look better as an include file along the lines of:

Quoted 6 lines

SELECTED_OPTIMIZATION = "-O0"
SELECTED_OPTIMIZATION_pn-eglibc = "-O2"
SELECTED_OPTIMIZATION_pn-perf = "-O2"

since clutter the recipes with anonymous python fragments isn't particular desirable.

Thats part of the problem. We only need to set -O2, when someone sets -O0. But if they set -O1 or -Os (or any other -O...) it appears to work properly...

So the python fragment is used to establish a known functional set for that item.

--Mark

Quoted 4 lines

Cheers,

Richard

#5 lore raw

Re: [master][dora][PATCH 1/2] perf: disallow debug optimization.

In reply to Richard Purdie
Quoted 9 lines · Richard Purdie

I have to admit at this point, this may look better as an include file along the lines of:

SELECTED_OPTIMIZATION = "-O0"
SELECTED_OPTIMIZATION_pn-eglibc = "-O2"
SELECTED_OPTIMIZATION_pn-perf = "-O2"

since clutter the recipes with anonymous python fragments isn't particular desirable.

Right, exactly. This is what I already do in my distro configuration to deal with a similar situation involving -fasynchronous-unwind-tables and it works fine.

p.

author
#6 lore raw

Re: [master][dora][PATCH 1/2] perf: disallow debug optimization.

In reply to Phil Blundell
Quoted 9 lines · Phil Blundell

meta/recipes-kernel/perf/perf.bb B = "${WORKDIR}/${BPN}-${PV}" · line 68

Quoted 4 lines · Mark Hatle
meta/recipes-kernel/perf/perf.bb +4 -0
+def get_optimization(d):
+    selected_optimization = d.getVar("SELECTED_OPTIMIZATION", True)
+    if base_contains("SELECTED_OPTIMIZATION", "-O0", "x", "", d) == "x":
+        bb.note("eglibc can't be built with -O0, -O2 will be used instead.")

Although the text of that warning is correct, users might find the reference to eglibc slightly confusing if it's perf that they're trying to build.

I'll get that fixed.

Quoted 5 lines

Also, as I mentioned in a different thread not all that long ago when someone submitted a similar patch for gcc-runtime, the proliferation of parse-time python functions to bash SELECTED_OPTIMIZATION around doesn't seem like all that good a thing: this will cause extra overhead for everyone, even those who are not using -O0 and have no interest in perf.

We have users who desire to build their system at different levels of optimizations for debug, size, profiling, etc. So they do change the default optimization levels from -O2 to -O0, etc. The python fragement is used to only adjust -O0, as -O1 (or -Os) work correctly.

--Mark

Quoted 8 lines

And, finally, it remains slightly unclear to me that this is really a problem that the metadata needs to be solving. I haven't seen any particularly convincing explanation of why this can't or shouldn't just be fixed in the distro configuration.

p.

#7 lore raw

Re: [master][dora][PATCH 1/2] perf: disallow debug optimization.

Quoted 4 lines · Mark Hatle

We have users who desire to build their system at different levels of optimizations for debug, size, profiling, etc. So they do change the default optimization levels from -O2 to -O0, etc. The python fragement is used to only adjust -O0, as -O1 (or -Os) work correctly.

Sure, I understand what the python is doing. The things I'm not quite so clear about are:

a) If the user asks to build with -O0, is it appropriate for the metadata to second-guess this and quietly switch to using -O2 instead when it thinks it knows best?

Personally I am inclined to say that attempting to use -O0 with packages that don't support it should just produce an error and the user should fix their configuration to not do that. And, if we're going to enable optimisation that the user hasn't asked for, shouldn't it be the minimum level consistent with getting the package to build rather than the full -O2 set?

b) If the answer to (a) is that the metadata should indeed be doing this, can it be made to do so in a way that doesn't involve running extra python fragments for all users every time the recipe is parsed?

c) If the answer to (b) is no, is the feature really so important that it's worth adding extra overhead to the parse for all users in order to benefit the (presumably tiny) minority who might actually be trying to build perf at -O0?

p.

#8 lore raw

Re: [master][dora][PATCH 1/2] perf: disallow debug optimization.

Quoted 6 lines · Phil Blundell
Quoted 4 lines · Mark Hatle

We have users who desire to build their system at different levels of optimizations for debug, size, profiling, etc. So they do change the default optimization levels from -O2 to -O0, etc. The python fragement is used to only adjust -O0, as -O1 (or -Os) work correctly.

Quoted 6 lines

Sure, I understand what the python is doing. The things I'm not quite so clear about are:

a) If the user asks to build with -O0, is it appropriate for the metadata to second-guess this and quietly switch to using -O2 instead when it thinks it knows best?

I suppose the other question is: why exactly does perf fail to build at -O0, and can we just patch it so that it works rather than forcing optimisation on? Even if it can't be fixed, it would be good for the commit message associated with any workaround to explain what the problem is.

p.

author
#9 lore raw

PATCH 2/2 perf: Disable -Werror flag

In reply to Mark Hatle

From: Konrad Scherer <[email protected]>

If the sed command does not run before make is invoked, the compile fails. Defining the environment variable is the proper way to disable warnings as errors build option and eliminates the race condition.

---
 meta/recipes-kernel/perf/perf.bb | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)
meta/recipes-kernel/perf/perf.bb +1 -4
index 138595d..1b889aa 100644
@@ -40,6 +40,7 @@ export STAGING_INCDIR
 export STAGING_LIBDIR
 export BUILD_SYS
 export HOST_SYS
+export WERROR = "0"
 
 do_populate_lic[depends] += "virtual/kernel:do_populate_sysroot"
 
@@ -118,10 +119,6 @@ do_install() {
 	fi
 }
 
-do_configure_prepend () {
-    sed -i 's,-Werror ,,' ${S}/tools/perf/Makefile
-}
-
 python do_package_prepend() {
     bb.data.setVar('PKGV', get_kernelversion('${S}').split("-")[0], d)
 }
Signature · 1 line
1.8.1.2.545.g2f19ada
#10 lore raw

Re: [master][dora][PATCH 2/2] perf: Disable -Werror flag

Quoted 25 lines · Mark Hatle

From: Konrad Scherer <[email protected]>

If the sed command does not run before make is invoked, the compile fails. Defining the environment variable is the proper way to disable warnings as errors build option and eliminates the race condition.

---
  meta/recipes-kernel/perf/perf.bb | 5 +----
  1 file changed, 1 insertion(+), 4 deletions(-)
meta/recipes-kernel/perf/perf.bb +1 -0
index 138595d..1b889aa 100644
@@ -40,6 +40,7 @@ export STAGING_INCDIR
  export STAGING_LIBDIR
  export BUILD_SYS
  export HOST_SYS
+export WERROR = "0"
 
  do_populate_lic[depends] += "virtual/kernel:do_populate_sysroot"
 

meta/recipes-kernel/perf/perf.bb do_install() { · line 124

Quoted 7 lines
meta/recipes-kernel/perf/perf.bb +0 -4
@@ -118,10 +119,6 @@ do_install() {
  	fi
  }
 
-do_configure_prepend () {
-    sed -i 's,-Werror ,,' ${S}/tools/perf/Makefile
-}
-

This is will not work for kernels < 3.1. I did submit another patch which works for all kernel versions.

Signature · 1 line
Konrad Scherer, MTS, Linux Products Group, Wind River