[bpf-next] tools/bpf: turn on llvm alu32 attribute by default
diff mbox series

Message ID 20191022023226.2255076-1-yhs@fb.com
State Changes Requested
Delegated to: BPF Maintainers
Headers show
Series
  • [bpf-next] tools/bpf: turn on llvm alu32 attribute by default
Related show

Commit Message

Yonghong Song Oct. 22, 2019, 2:32 a.m. UTC
llvm alu32 was introduced in llvm7:
  https://reviews.llvm.org/rL325987
  https://reviews.llvm.org/rL325989
Experiments showed that in general performance
is better with alu32 enabled:
  https://lwn.net/Articles/775316/

This patch turned on alu32 with no-flavor test_progs
which is tested most often. The flavor test at
no_alu32/test_progs can be used to test without
alu32 enabled. The Makefile check for whether
llvm supports '-mattr=+alu32 -mcpu=v3' is
removed as llvm7 should be available for recent
distributions and also latest llvm is preferred
to run bpf selftests.

Note that jmp32 is checked by -mcpu=probe and
will be enabled if the host kernel supports it.

Cc: Jiong Wang <jiong.wang@netronome.com>
Cc: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Yonghong Song <yhs@fb.com>
---
 tools/testing/selftests/bpf/Makefile | 26 ++++++++------------------
 1 file changed, 8 insertions(+), 18 deletions(-)

Comments

Andrii Nakryiko Oct. 22, 2019, 3:01 a.m. UTC | #1
On Mon, Oct 21, 2019 at 7:32 PM Yonghong Song <yhs@fb.com> wrote:
>
> llvm alu32 was introduced in llvm7:
>   https://reviews.llvm.org/rL325987
>   https://reviews.llvm.org/rL325989
> Experiments showed that in general performance
> is better with alu32 enabled:
>   https://lwn.net/Articles/775316/
>
> This patch turned on alu32 with no-flavor test_progs
> which is tested most often. The flavor test at
> no_alu32/test_progs can be used to test without
> alu32 enabled. The Makefile check for whether
> llvm supports '-mattr=+alu32 -mcpu=v3' is
> removed as llvm7 should be available for recent
> distributions and also latest llvm is preferred
> to run bpf selftests.
>
> Note that jmp32 is checked by -mcpu=probe and
> will be enabled if the host kernel supports it.
>
> Cc: Jiong Wang <jiong.wang@netronome.com>
> Cc: Andrii Nakryiko <andriin@fb.com>
> Signed-off-by: Yonghong Song <yhs@fb.com>
> ---

Sounds good to me, see minor nit below.

Acked-by: Andrii Nakryiko <andriin@fb.com>

>  tools/testing/selftests/bpf/Makefile | 26 ++++++++------------------
>  1 file changed, 8 insertions(+), 18 deletions(-)
>
> diff --git a/tools/testing/selftests/bpf/Makefile b/tools/testing/selftests/bpf/Makefile
> index 4ff5f4aada08..5a0bca2802fe 100644
> --- a/tools/testing/selftests/bpf/Makefile
> +++ b/tools/testing/selftests/bpf/Makefile
> @@ -32,15 +32,7 @@ TEST_GEN_PROGS = test_verifier test_tag test_maps test_lru_map test_lpm_map test
>         test_netcnt test_tcpnotify_user test_sock_fields test_sysctl test_hashmap \
>         test_cgroup_attach xdping
>
> -# Also test sub-register code-gen if LLVM has eBPF v3 processor support which
> -# contains both ALU32 and JMP32 instructions.
> -SUBREG_CODEGEN := $(shell echo "int cal(int a) { return a > 0; }" | \
> -                       $(CLANG) -target bpf -O2 -emit-llvm -S -x c - -o - | \
> -                       $(LLC) -mattr=+alu32 -mcpu=v3 2>&1 | \
> -                       grep 'if w')
> -ifneq ($(SUBREG_CODEGEN),)
> -TEST_GEN_PROGS += test_progs-alu32
> -endif
> +TEST_GEN_PROGS += test_progs-no_alu32

combine this with TEST_GEN_PROGS list above, it's not conditional anymore?

>
>  # Also test bpf-gcc, if present
>  ifneq ($(BPF_GCC),)
> @@ -179,7 +171,7 @@ endef
>  # $eval()) and pass control to DEFINE_TEST_RUNNER_RULES.
>  # Parameters:

[...]
Yonghong Song Oct. 22, 2019, 3:49 a.m. UTC | #2
On 10/21/19 8:01 PM, Andrii Nakryiko wrote:
> On Mon, Oct 21, 2019 at 7:32 PM Yonghong Song <yhs@fb.com> wrote:
>>
>> llvm alu32 was introduced in llvm7:
>>    https://urldefense.proofpoint.com/v2/url?u=https-3A__reviews.llvm.org_rL325987&d=DwIBaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=DA8e1B5r073vIqRrFz7MRA&m=NHAn2DX4M75upJno8E_Vhg5Wx0cj_CtKLDdMYJJHgVA&s=zlXKgOdNDd7V6cQk3lYECzrp8MYswrOnc-ApTPJ6Q5o&e=
>>    https://urldefense.proofpoint.com/v2/url?u=https-3A__reviews.llvm.org_rL325989&d=DwIBaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=DA8e1B5r073vIqRrFz7MRA&m=NHAn2DX4M75upJno8E_Vhg5Wx0cj_CtKLDdMYJJHgVA&s=NSHiry5gsEeL5ThRWj-kc0yfD0YpZCt-OhAsBIvwNZQ&e=
>> Experiments showed that in general performance
>> is better with alu32 enabled:
>>    https://urldefense.proofpoint.com/v2/url?u=https-3A__lwn.net_Articles_775316_&d=DwIBaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=DA8e1B5r073vIqRrFz7MRA&m=NHAn2DX4M75upJno8E_Vhg5Wx0cj_CtKLDdMYJJHgVA&s=FaOOW0RJNF7B2U4JUKG2dPlKk0c1oQtnq06-eFx8mBg&e=
>>
>> This patch turned on alu32 with no-flavor test_progs
>> which is tested most often. The flavor test at
>> no_alu32/test_progs can be used to test without
>> alu32 enabled. The Makefile check for whether
>> llvm supports '-mattr=+alu32 -mcpu=v3' is
>> removed as llvm7 should be available for recent
>> distributions and also latest llvm is preferred
>> to run bpf selftests.
>>
>> Note that jmp32 is checked by -mcpu=probe and
>> will be enabled if the host kernel supports it.
>>
>> Cc: Jiong Wang <jiong.wang@netronome.com>
>> Cc: Andrii Nakryiko <andriin@fb.com>
>> Signed-off-by: Yonghong Song <yhs@fb.com>
>> ---
> 
> Sounds good to me, see minor nit below.
> 
> Acked-by: Andrii Nakryiko <andriin@fb.com>
> 
>>   tools/testing/selftests/bpf/Makefile | 26 ++++++++------------------
>>   1 file changed, 8 insertions(+), 18 deletions(-)
>>
>> diff --git a/tools/testing/selftests/bpf/Makefile b/tools/testing/selftests/bpf/Makefile
>> index 4ff5f4aada08..5a0bca2802fe 100644
>> --- a/tools/testing/selftests/bpf/Makefile
>> +++ b/tools/testing/selftests/bpf/Makefile
>> @@ -32,15 +32,7 @@ TEST_GEN_PROGS = test_verifier test_tag test_maps test_lru_map test_lpm_map test
>>          test_netcnt test_tcpnotify_user test_sock_fields test_sysctl test_hashmap \
>>          test_cgroup_attach xdping
>>
>> -# Also test sub-register code-gen if LLVM has eBPF v3 processor support which
>> -# contains both ALU32 and JMP32 instructions.
>> -SUBREG_CODEGEN := $(shell echo "int cal(int a) { return a > 0; }" | \
>> -                       $(CLANG) -target bpf -O2 -emit-llvm -S -x c - -o - | \
>> -                       $(LLC) -mattr=+alu32 -mcpu=v3 2>&1 | \
>> -                       grep 'if w')
>> -ifneq ($(SUBREG_CODEGEN),)
>> -TEST_GEN_PROGS += test_progs-alu32
>> -endif
>> +TEST_GEN_PROGS += test_progs-no_alu32
> 
> combine this with TEST_GEN_PROGS list above, it's not conditional anymore?

Ya, make sense. Will respin.

> 
>>
>>   # Also test bpf-gcc, if present
>>   ifneq ($(BPF_GCC),)
>> @@ -179,7 +171,7 @@ endef
>>   # $eval()) and pass control to DEFINE_TEST_RUNNER_RULES.
>>   # Parameters:
> 
> [...]
>

Patch
diff mbox series

diff --git a/tools/testing/selftests/bpf/Makefile b/tools/testing/selftests/bpf/Makefile
index 4ff5f4aada08..5a0bca2802fe 100644
--- a/tools/testing/selftests/bpf/Makefile
+++ b/tools/testing/selftests/bpf/Makefile
@@ -32,15 +32,7 @@  TEST_GEN_PROGS = test_verifier test_tag test_maps test_lru_map test_lpm_map test
 	test_netcnt test_tcpnotify_user test_sock_fields test_sysctl test_hashmap \
 	test_cgroup_attach xdping
 
-# Also test sub-register code-gen if LLVM has eBPF v3 processor support which
-# contains both ALU32 and JMP32 instructions.
-SUBREG_CODEGEN := $(shell echo "int cal(int a) { return a > 0; }" | \
-			$(CLANG) -target bpf -O2 -emit-llvm -S -x c - -o - | \
-			$(LLC) -mattr=+alu32 -mcpu=v3 2>&1 | \
-			grep 'if w')
-ifneq ($(SUBREG_CODEGEN),)
-TEST_GEN_PROGS += test_progs-alu32
-endif
+TEST_GEN_PROGS += test_progs-no_alu32
 
 # Also test bpf-gcc, if present
 ifneq ($(BPF_GCC),)
@@ -179,7 +171,7 @@  endef
 # $eval()) and pass control to DEFINE_TEST_RUNNER_RULES.
 # Parameters:
 # $1 - test runner base binary name (e.g., test_progs)
-# $2 - test runner extra "flavor" (e.g., alu32, gcc-bpf, etc)
+# $2 - test runner extra "flavor" (e.g., no_alu32, gcc-bpf, etc)
 define DEFINE_TEST_RUNNER
 
 TRUNNER_OUTPUT := $(OUTPUT)$(if $2,/)$2
@@ -201,7 +193,7 @@  endef
 # Using TRUNNER_XXX variables, provided by callers of DEFINE_TEST_RUNNER and
 # set up by DEFINE_TEST_RUNNER itself, create test runner build rules with:
 # $1 - test runner base binary name (e.g., test_progs)
-# $2 - test runner extra "flavor" (e.g., alu32, gcc-bpf, etc)
+# $2 - test runner extra "flavor" (e.g., no_alu32, gcc-bpf, etc)
 define DEFINE_TEST_RUNNER_RULES
 
 ifeq ($($(TRUNNER_OUTPUT)-dir),)
@@ -272,14 +264,12 @@  TRUNNER_EXTRA_FILES := $(OUTPUT)/urandom_read				\
 		       $(wildcard progs/btf_dump_test_case_*.c)
 TRUNNER_BPF_BUILD_RULE := CLANG_BPF_BUILD_RULE
 TRUNNER_BPF_CFLAGS := -I. -I$(OUTPUT) $(BPF_CFLAGS) $(CLANG_CFLAGS)
-TRUNNER_BPF_LDFLAGS :=
+TRUNNER_BPF_LDFLAGS := -mattr=+alu32
 $(eval $(call DEFINE_TEST_RUNNER,test_progs))
 
-# Define test_progs-alu32 test runner.
-ifneq ($(SUBREG_CODEGEN),)
-TRUNNER_BPF_LDFLAGS += -mattr=+alu32
-$(eval $(call DEFINE_TEST_RUNNER,test_progs,alu32))
-endif
+# Define test_progs-no_alu32 test runner.
+TRUNNER_BPF_LDFLAGS :=
+$(eval $(call DEFINE_TEST_RUNNER,test_progs,no_alu32))
 
 # Define test_progs BPF-GCC-flavored test runner.
 ifneq ($(BPF_GCC),)
@@ -319,4 +309,4 @@  $(OUTPUT)/test_verifier: test_verifier.c verifier/tests.h $(BPFOBJ) | $(OUTPUT)
 
 EXTRA_CLEAN := $(TEST_CUSTOM_PROGS)					\
 	prog_tests/tests.h map_tests/tests.h verifier/tests.h		\
-	feature $(OUTPUT)/*.o $(OUTPUT)/alu32 $(OUTPUT)/bpf_gcc
+	feature $(OUTPUT)/*.o $(OUTPUT)/no_alu32 $(OUTPUT)/bpf_gcc