Skip to contentSkip to conversation

Test compiling with Clang in the Travis-CI

PATCH In review Thomas Huth 4 participants 9 messages Jun 22, 2021 – Jun 23, 2021

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover Test compiling with Clang in the Travis-CI

Travis-CI recently changed their policy so that builds on the non-x86 build machines are possible without consuming any credits again. While we're already testing the non-x86 builds in the gitlab-CI with the GCC cross-compilers, we could still benefit from the non-x86 builders in the Travis-CI by compiling the code with Clang there, too (since there are AFAIK no Clang cross-compilers available in the usual distros on x86).

Thomas Huth (4):
configure: Add the possibility to specify additional cflags
powerpc: Probe whether the compiler understands -mabi=no-altivec lib/s390x: Fix the epsw inline assembly
Test compilation with Clang on aarch64, ppc64le and s390x in Travis-CI

 .travis.yml              | 44 ++++++++++++++++++++++++++++++++++++++++
 Makefile                 |  3 ---
 configure                | 10 +++++++--
 lib/s390x/asm/arch_def.h |  2 +-
 powerpc/Makefile.common  |  4 +++-
 5 files changed, 56 insertions(+), 7 deletions(-)
 create mode 100644 .travis.yml
Signature · 1 line
2.27.0
#1 lore raw

PATCH 1/4 configure: Add the possibility to specify additional cflags

For certain compilers or experiments, it might be necessary to specify additional CFLAGS for the build. Let's add an option to the configure script to specify such additional compiler flags.

---
 Makefile  |  3 ---
 configure | 10 ++++++++--
 2 files changed, 8 insertions(+), 5 deletions(-)
Makefile +0 -3
index 1afa90e..f7b9f28 100644
@@ -22,9 +22,6 @@ DESTDIR := $(PREFIX)/share/kvm-unit-tests/
 cc-option = $(shell if $(CC) -Werror $(1) -S -o /dev/null -xc /dev/null \
               > /dev/null 2>&1; then echo "$(1)"; else echo "$(2)"; fi ;)
 
-#make sure env CFLAGS variable is not used
-CFLAGS =
-
 libcflat := lib/libcflat.a
 cflatobjs := \
 	lib/argv.o \
configure +8 -2
index d21601f..c48ab3d 100755
@@ -8,6 +8,7 @@ fi
 srcdir=$(cd "$(dirname "$0")"; pwd)
 prefix=/usr/local
 cc=gcc
+cflags=
 ld=ld
 objcopy=objcopy
 objdump=objdump
@@ -38,8 +39,9 @@ usage() {
 	    --target=TARGET        target platform that the tests will be running on (qemu or
 	                           kvmtool, default is qemu) (arm/arm64 only)
 	    --cross-prefix=PREFIX  cross compiler prefix
-	    --cc=CC		   c compiler to use ($cc)
-	    --ld=LD		   ld linker to use ($ld)
+	    --cc=CC                c compiler to use ($cc)
+	    --cflags=FLAGS         extra options to be passed to the c compiler
+	    --ld=LD                ld linker to use ($ld)
 	    --prefix=PREFIX        where to install things ($prefix)
 	    --endian=ENDIAN        endianness to compile for (little or big, ppc64 only)
 	    --[enable|disable]-pretty-print-stacks
@@ -100,6 +102,9 @@ while [[ "$1" = -* ]]; do
 	--cc)
 	    cc="$arg"
 	    ;;
+	--cflags)
+	    cflags="$arg"
+	    ;;
 	--ld)
 	    ld="$arg"
 	    ;;
@@ -316,6 +321,7 @@ ARCH=$arch
 ARCH_NAME=$arch_name
 PROCESSOR=$processor
 CC=$cross_prefix$cc
+CFLAGS=$cflags
 LD=$cross_prefix$ld
 OBJCOPY=$cross_prefix$objcopy
 OBJDUMP=$cross_prefix$objdump
Signature · 1 line
2.27.0
#2 lore raw

PATCH 2/4 powerpc: Probe whether the compiler understands -mabi=no-altivec

In reply to Thomas Huth

Clang does not support "-mabi=no-altivec", so let's check whether this option is supported before using it.

---
 powerpc/Makefile.common | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
powerpc/Makefile.common +3 -1
index 4c3121a..12c280c 100644
@@ -17,9 +17,11 @@ all: directories $(TEST_DIR)/boot_rom.bin $(tests-all)
 
 ##################################################################
 
+mabi_no_altivec := $(call cc-option,-mabi=no-altivec,"")
+
 CFLAGS += -std=gnu99
 CFLAGS += -ffreestanding
-CFLAGS += -O2 -msoft-float -mabi=no-altivec -mno-altivec
+CFLAGS += -O2 -msoft-float -mno-altivec $(mabi_no_altivec)
 CFLAGS += -I $(SRCDIR)/lib -I $(SRCDIR)/lib/libfdt -I lib
 CFLAGS += -Wa,-mregnames
 
Signature · 1 line
2.27.0
#3 lore raw

PATCH 3/4 lib/s390x: Fix the epsw inline assembly

In reply to Thomas Huth

According to the Principles of Operation, the epsw instruction does not touch the second register if it is r0. With GCC we were lucky so far that it never tried to use r0 here, but when compiling the kvm-unit-tests with Clang, this indeed happens and leads to very weird crashes. Thus let's make sure to never use r0 for the second operand of the epsw instruction.

---
 lib/s390x/asm/arch_def.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
lib/s390x/asm/arch_def.h +1 -1
index 3aa5da9..15cf7d4 100644
@@ -265,7 +265,7 @@ static inline uint64_t extract_psw_mask(void)
 
 	asm volatile(
 		"	epsw	%0,%1\n"
-		: "+r" (mask_upper), "+r" (mask_lower) : : );
+		: "=r" (mask_upper), "=a" (mask_lower));
 
 	return (uint64_t) mask_upper << 32 | mask_lower;
 }
Signature · 1 line
2.27.0
Reviewed-by on v1 3/4
#4 lore raw

Re: [kvm-unit-tests PATCH 3/4] lib/s390x: Fix the epsw inline assembly

Quoted 8 lines · On Tue, 22 Jun 2021 15:55:16 +0200 Thomas Huth <[email protected]> wrote:

According to the Principles of Operation, the epsw instruction does not touch the second register if it is r0. With GCC we were lucky so far that it never tried to use r0 here, but when compiling the kvm-unit-tests with Clang, this indeed happens and leads to very weird crashes. Thus let's make sure to never use r0 for the second operand of the epsw instruction.

maybe also mention in the patch description why you changed + to =

Quoted 17 lines
---
 lib/s390x/asm/arch_def.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
lib/s390x/asm/arch_def.h +1 -1
index 3aa5da9..15cf7d4 100644
@@ -265,7 +265,7 @@ static inline uint64_t extract_psw_mask(void)
 
 	asm volatile(
 		"	epsw	%0,%1\n"
-		: "+r" (mask_upper), "+r" (mask_lower) : : );
+		: "=r" (mask_upper), "=a" (mask_lower));
 
 	return (uint64_t) mask_upper << 32 | mask_lower;
 }
#5 lore raw

Re: [kvm-unit-tests PATCH 3/4] lib/s390x: Fix the epsw inline assembly

Quoted 10 lines · Claudio Imbrenda
Quoted 7 lines · On Tue, 22 Jun 2021 15:55:16 +0200 Thomas Huth <[email protected]> wrote:

According to the Principles of Operation, the epsw instruction does not touch the second register if it is r0. With GCC we were lucky so far that it never tried to use r0 here, but when compiling the kvm-unit-tests with Clang, this indeed happens and leads to very weird crashes. Thus let's make sure to never use r0 for the second operand of the epsw instruction.

Quoted 5 lines
Quoted 1 line

maybe also mention in the patch description why you changed + to =

Yes, that makes sense. I'll add something like:

While we're at it, also change the constraint modifier from "+" to "=" since these are only output parameters, not input.

Thomas

Acked-by on v1 3/4
#6 lore raw

Re: [kvm-unit-tests PATCH 3/4] lib/s390x: Fix the epsw inline assembly

In reply to Thomas Huth
Quoted 18 lines · Thomas Huth

According to the Principles of Operation, the epsw instruction does not touch the second register if it is r0. With GCC we were lucky so far that it never tried to use r0 here, but when compiling the kvm-unit-tests with Clang, this indeed happens and leads to very weird crashes. Thus let's make sure to never use r0 for the second operand of the epsw instruction.

---
 lib/s390x/asm/arch_def.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
lib/s390x/asm/arch_def.h +0 -0
index 3aa5da9..15cf7d4 100644

lib/s390x/asm/arch_def.h static inline uint64_t extract_psw_mask(void) · line 271

Quoted 8 lines
lib/s390x/asm/arch_def.h +1 -1
@@ -265,7 +265,7 @@ static inline uint64_t extract_psw_mask(void)
 
 	asm volatile(
 		"	epsw	%0,%1\n"
-		: "+r" (mask_upper), "+r" (mask_lower) : : );
+		: "=r" (mask_upper), "=a" (mask_lower));
 
 	return (uint64_t) mask_upper << 32 | mask_lower;
 }

With Claudio's nits fixed:

#7 lore raw

PATCH 4/4 Test compilation with Clang on aarch64, ppc64le and s390x in Travis-CI

In reply to Thomas Huth

Travis-CI recently changed their policy so that builds on the non-x86 build machines are possible without consuming any credits again. We can use these systems to test compilation of the non-x86 code with Clang. Unfortunately, the qemu-system-s390x of Ubuntu 20.04 seems to be buggy, so that the s390x binaries cause that QEMU to crash. Thus we can only run the TCG tests for ppc64le and aarch64.

---
 .travis.yml | 44 ++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 44 insertions(+)
 create mode 100644 .travis.yml
.travis.yml new file +44 -0
new file mode 100644
index 0000000..4fcb687
@@ -0,0 +1,44 @@
+dist: focal
+language: c
+cache: ccache
+compiler: clang
+git:
+  submodules: false
+
+jobs:
+  include:
+
+    - arch: arm64
+      addons:
+        apt_packages: qemu-system-aarch64
+      env:
+      - CONFIG="--arch=arm64 --cc=clang"
+      - TESTS="cache gicv2-active gicv2-ipi gicv2-mmio gicv3-active gicv3-ipi
+          pci-test pmu-cycle-counter pmu-event-counter-config pmu-sw-incr
+          selftest-setup selftest-smp selftest-vectors-kernel
+          selftest-vectors-user timer"
+
+    - arch: ppc64le
+      addons:
+        apt_packages: clang-11 qemu-system-ppc
+      env:
+      - CONFIG="--arch=ppc64 --endian=little --cc=clang-11 --cflags=-no-integrated-as"
+      - TESTS="emulator rtas-get-time-of-day rtas-get-time-of-day-base
+          rtas-set-time-of-day selftest-setup spapr_hcall"
+
+    - arch: s390x
+      addons:
+        apt_packages: clang-11 qemu-system-s390x
+      env:
+      - CONFIG="--arch=s390x --cc=clang-11 --cflags=-no-integrated-as"
+      - TESTS=""
+
+before_script:
+  - mkdir -p build && cd build
+  - $TRAVIS_BUILD_DIR/configure $CONFIG
+script:
+  - make -j3
+  - if [ -n "$TESTS" ]; then
+      ACCEL="${ACCEL:-tcg}" ./run_tests.sh -v $TESTS | tee results.txt &&
+      grep -q PASS results.txt && ! grep -q FAIL results.txt ;
+    fi
Signature · 1 line
2.27.0
maintainer
#8 lore raw
In reply to Thomas Huth
Quoted 7 lines · Thomas Huth

Travis-CI recently changed their policy so that builds on the non-x86 build machines are possible without consuming any credits again. While we're already testing the non-x86 builds in the gitlab-CI with the GCC cross-compilers, we could still benefit from the non-x86 builders in the Travis-CI by compiling the code with Clang there, too (since there are AFAIK no Clang cross-compilers available in the usual distros on x86).

Looks good apart from Claudio's remark. Thanks very much!

Paolo