diff mbox series

[RE-SEND] openjdk: Add ARC HS3x/4x support via "Zero Assembly Project"

Message ID 20211021115915.32447-1-abrodkin@synopsys.com
State Accepted
Headers show
Series [RE-SEND] openjdk: Add ARC HS3x/4x support via "Zero Assembly Project" | expand

Commit Message

Alexey Brodkin Oct. 21, 2021, 11:59 a.m. UTC 
This allows building and running a full-scale JVM in purely
interpretive mode on ARCv2 processors.

Once JIT'ed version is available for ARC we'll obviously switch
to it to gain a faster execution.

Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Cc: Fabrice Fontaine <fontaine.fabrice@gmail.com>
Cc: Yann E. MORIN <yann.morin.1998@free.fr>
Cc: Adam Duskett <aduskett@gmail.com>
---

Erroneously sent to buildroot@lists.buildroot.org initially,
now sending to buildroot@busybox.net.

Sorry for the noise!

---

 ...-Add-ARCv2-ISA-processors-support-to-Zero.patch | 109 +++++++++++++++++++++
 package/openjdk/Config.in                          |   2 +
 2 files changed, 111 insertions(+)
 create mode 100644 package/openjdk/0001-Add-ARCv2-ISA-processors-support-to-Zero.patch

Comments

Thomas Petazzoni Dec. 30, 2021, 8:14 p.m. UTC | #1 
Hello Alexey,

On Thu, 21 Oct 2021 04:59:15 -0700
Alexey Brodkin via buildroot <buildroot@buildroot.org> wrote:

> This allows building and running a full-scale JVM in purely
> interpretive mode on ARCv2 processors.
> 
> Once JIT'ed version is available for ARC we'll obviously switch
> to it to gain a faster execution.
> 
> Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
> Cc: Fabrice Fontaine <fontaine.fabrice@gmail.com>
> Cc: Yann E. MORIN <yann.morin.1998@free.fr>
> Cc: Adam Duskett <aduskett@gmail.com>

I have applied your patch (finally you could say!) but there was a
mistake in it: you've added your patch in package/openjdk/, but we
support two versions of OpenJDK (11 and 17), and your patch only
applies to version 17. So I moved it to a sub-folder + ensured that
version 11 couldn't be selected for ARC. This also raised the issue
that version 17 has an additional requirement on host gcc >= 4.9, so if
a package ever did "select BR2_PACKAGE_OPENJDK", they would have to be
careful that specifically on ARC, it's only available if host gcc >=
4.9. To avoid this, I've moved the host gcc >= 4.9 dependency to
BR2_PACKAGE_OPENJDK itself, even if not strictly required for OpenJDK
11.

However, there is one thing I didn't address as I wasn't sure: do you
support only the "zero" variant on ARC? If so, doesn't this:

config BR2_PACKAGE_OPENJDK_JVM_VARIANT_CLIENT
        bool "client"
        depends on !BR2_powerpc
        help
          Quick loading, but slower run-time performance.

config BR2_PACKAGE_OPENJDK_JVM_VARIANT_SERVER
        bool "server"
        depends on !BR2_powerpc
        help
          Slower loading, but faster run-time performance.

config BR2_PACKAGE_OPENJDK_JVM_VARIANT_ZERO
        bool "zero"
        select BR2_PACKAGE_LIBFFI
        help
          A non-assembler variant with wide arch support, however
          performance is less then client/server.

          http://openjdk.java.net/projects/zero

also needs some depends on !BR2_arc for the client and server variants?

Thanks a lot,

Thomas
Alexey Brodkin Dec. 30, 2021, 9:04 p.m. UTC | #2 
Hi Thomas,

> I have applied your patch (finally you could say!) but there was a
> mistake in it: you've added your patch in package/openjdk/, but we
> support two versions of OpenJDK (11 and 17), and your patch only
> applies to version 17.

Sure, thanks for taking care of this one - much appreciated especially
assuming all the extra troubles you've been through here ;)

As for version 11 I'm not really sure if it makes sense to kinda retrofit
ARC support in the older version so yeah I think only v17 should be OK.
And indeed it was not immediately clear that there're 2 versions which co-exist
at the same tim - so that's my bad for sure.

> So I moved it to a sub-folder + ensured that
> version 11 couldn't be selected for ARC. This also raised the issue
> that version 17 has an additional requirement on host gcc >= 4.9, so if
> a package ever did "select BR2_PACKAGE_OPENJDK", they would have to be
> careful that specifically on ARC, it's only available if host gcc >=
> 4.9. To avoid this, I've moved the host gcc >= 4.9 dependency to
> BR2_PACKAGE_OPENJDK itself, even if not strictly required for OpenJDK
> 11.

I hope it won't hurt indeed. But we'll see...

> However, there is one thing I didn't address as I wasn't sure: do you
> support only the "zero" variant on ARC? If so, doesn't this:
> 
> config BR2_PACKAGE_OPENJDK_JVM_VARIANT_CLIENT
>         bool "client"
>         depends on !BR2_powerpc
>         help
>           Quick loading, but slower run-time performance.
> 
> config BR2_PACKAGE_OPENJDK_JVM_VARIANT_SERVER
>         bool "server"
>         depends on !BR2_powerpc
>         help
>           Slower loading, but faster run-time performance.
> 
> config BR2_PACKAGE_OPENJDK_JVM_VARIANT_ZERO
>         bool "zero"
>         select BR2_PACKAGE_LIBFFI
>         help
>           A non-assembler variant with wide arch support, however
>           performance is less then client/server.
> 
>           http://openjdk.java.net/projects/zero/
> 
> also needs some depends on !BR2_arc for the client and server variants?

Hm, indeed we need to add "!BR2_arc" here as ARC at this point supports only
interpretive mode (that's what "zero" mode stands for), while both "client" and
"server" modes are 2 flavors or tunes of a true HotSpot, which is by definition
is JIT'ed. I must have been in a real hurry to make all these stupid mistakes.

Do you want me to send a separate patch which adds these "!BR2_arc" dependencies
for both BR2_PACKAGE_OPENJDK_JVM_VARIANT_CLIENT &
BR2_PACKAGE_OPENJDK_JVM_VARIANT_SERVER or you prefer to do a fast fix-up yourself?

Thanks,
Alexey
Thomas Petazzoni Dec. 30, 2021, 9:32 p.m. UTC | #3 
On Thu, 30 Dec 2021 21:04:21 +0000
Alexey Brodkin <Alexey.Brodkin@synopsys.com> wrote:

> Do you want me to send a separate patch which adds these "!BR2_arc" dependencies
> for both BR2_PACKAGE_OPENJDK_JVM_VARIANT_CLIENT &
> BR2_PACKAGE_OPENJDK_JVM_VARIANT_SERVER or you prefer to do a fast fix-up yourself?

Please send a patch :-)

Thanks a lot!

Thomas
diff mbox series

Patch

diff --git a/package/openjdk/0001-Add-ARCv2-ISA-processors-support-to-Zero.patch b/package/openjdk/0001-Add-ARCv2-ISA-processors-support-to-Zero.patch
new file mode 100644
index 0000000000..1ef523a419
--- /dev/null
+++ b/package/openjdk/0001-Add-ARCv2-ISA-processors-support-to-Zero.patch
@@ -0,0 +1,109 @@ 
+From 9a3c2a0714420186ae8d5159d5416c6fa2009ad9 Mon Sep 17 00:00:00 2001
+From: Alexey Brodkin <abrodkin@synopsys.com>
+Date: Mon, 18 Oct 2021 05:20:05 -0700
+Subject: [PATCH] Add ARCv2 ISA processors support to Zero
+
+This adds ARCv2 processors support in OpenJDK via
+"Zero Assembly Project" (see https://openjdk.java.net/projects/zero).
+
+That' a purely interpretive mode, so likely not that fast
+as JIT'ed version, but for starters it's much better than nothing.
+
+Once all the logistical problems are solved hopefully this
+change will be accepted upstream.
+
+Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
+---
+ make/autoconf/libraries.m4        | 8 ++++++++
+ make/autoconf/platform.m4         | 8 ++++++++
+ src/hotspot/os/linux/os_linux.cpp | 8 +++++++-
+ 3 files changed, 23 insertions(+), 1 deletion(-)
+
+diff --git a/make/autoconf/libraries.m4 b/make/autoconf/libraries.m4
+index 5120918aed2..f7a99955874 100644
+--- a/make/autoconf/libraries.m4
++++ b/make/autoconf/libraries.m4
+@@ -122,6 +122,14 @@ AC_DEFUN_ONCE([LIB_SETUP_LIBRARIES],
+     BASIC_JVM_LIBS="$BASIC_JVM_LIBS -lpthread"
+   fi
+ 
++  # Libatomic library
++  # 32-bit ARC needs fallback library for 8-byte atomic ops
++  # Inspired by a fix for MIPS, see https://github.com/openjdk/jdk/commit/1b3aa3af
++  if test "x$OPENJDK_TARGET_OS" = xlinux &&
++      (test "x$OPENJDK_TARGET_CPU" = xarc); then
++    BASIC_JVM_LIBS="$BASIC_JVM_LIBS -latomic"
++  fi
++
+   # perfstat lib
+   if test "x$OPENJDK_TARGET_OS" = xaix; then
+     BASIC_JVM_LIBS="$BASIC_JVM_LIBS -lperfstat"
+diff --git a/make/autoconf/platform.m4 b/make/autoconf/platform.m4
+index c0f2446dbd7..dc416ece885 100644
+--- a/make/autoconf/platform.m4
++++ b/make/autoconf/platform.m4
+@@ -54,6 +54,12 @@ AC_DEFUN([PLATFORM_EXTRACT_VARS_FROM_CPU],
+       VAR_CPU_BITS=64
+       VAR_CPU_ENDIAN=little
+       ;;
++    arc)
++      VAR_CPU=arc
++      VAR_CPU_ARCH=arc
++      VAR_CPU_BITS=32
++      VAR_CPU_ENDIAN=little
++      ;;
+     arm*)
+       VAR_CPU=arm
+       VAR_CPU_ARCH=arm
+@@ -478,6 +484,8 @@ AC_DEFUN([PLATFORM_SETUP_LEGACY_VARS_HELPER],
+     HOTSPOT_$1_CPU_DEFINE=PPC64
+ 
+   # The cpu defines below are for zero, we don't support them directly.
++  elif test "x$OPENJDK_$1_CPU" = xarc; then
++    HOTSPOT_$1_CPU_DEFINE=ARC
+   elif test "x$OPENJDK_$1_CPU" = xsparc; then
+     HOTSPOT_$1_CPU_DEFINE=SPARC
+   elif test "x$OPENJDK_$1_CPU" = xppc; then
+diff --git a/src/hotspot/os/linux/os_linux.cpp b/src/hotspot/os/linux/os_linux.cpp
+index b08caf4d5d3..2bf084895ba 100644
+--- a/src/hotspot/os/linux/os_linux.cpp
++++ b/src/hotspot/os/linux/os_linux.cpp
+@@ -1858,6 +1858,9 @@ void * os::dll_load(const char *filename, char *ebuf, int ebuflen) {
+ #ifndef EM_AARCH64
+   #define EM_AARCH64    183               /* ARM AARCH64 */
+ #endif
++#ifndef EM_ARC_COMPACT2
++  #define EM_ARC_COMPACT2 195             /* ARC ARCv2 ISA */
++#endif
+ #ifndef EM_RISCV
+   #define EM_RISCV      243               /* RISC-V */
+ #endif
+@@ -1879,6 +1882,7 @@ void * os::dll_load(const char *filename, char *ebuf, int ebuflen) {
+     {EM_SH,          EM_SH,      ELFCLASS32, ELFDATA2MSB, (char*)"SuperH BE"},
+ #endif
+     {EM_ARM,         EM_ARM,     ELFCLASS32, ELFDATA2LSB, (char*)"ARM"},
++    {EM_ARC_COMPACT2, EM_ARC_COMPACT2, ELFCLASS32, ELFDATA2LSB, (char*)"ARC"},
+     // we only support 64 bit z architecture
+     {EM_S390,        EM_S390,    ELFCLASS64, ELFDATA2MSB, (char*)"IBM System/390"},
+     {EM_ALPHA,       EM_ALPHA,   ELFCLASS64, ELFDATA2LSB, (char*)"Alpha"},
+@@ -1906,6 +1910,8 @@ void * os::dll_load(const char *filename, char *ebuf, int ebuflen) {
+   static  Elf32_Half running_arch_code=EM_PPC;
+ #elif  (defined AARCH64)
+   static  Elf32_Half running_arch_code=EM_AARCH64;
++#elif  (defined ARC)
++  static  Elf32_Half running_arch_code=EM_ARC_COMPACT2;
+ #elif  (defined ARM)
+   static  Elf32_Half running_arch_code=EM_ARM;
+ #elif  (defined S390)
+@@ -1926,7 +1932,7 @@ void * os::dll_load(const char *filename, char *ebuf, int ebuflen) {
+   static  Elf32_Half running_arch_code=EM_RISCV;
+ #else
+     #error Method os::dll_load requires that one of following is defined:\
+-        AARCH64, ALPHA, ARM, AMD64, IA32, IA64, M68K, MIPS, MIPSEL, PARISC, __powerpc__, __powerpc64__, RISCV, S390, SH, __sparc
++        AARCH64, ALPHA, ARC, ARM, AMD64, IA32, IA64, M68K, MIPS, MIPSEL, PARISC, __powerpc__, __powerpc64__, RISCV, S390, SH, __sparc
+ #endif
+ 
+   // Identify compatibility class for VM's architecture and library's architecture
+-- 
+2.16.2
+
diff --git a/package/openjdk/Config.in b/package/openjdk/Config.in
index f2b7f5ffd7..c09d1c0162 100644
--- a/package/openjdk/Config.in
+++ b/package/openjdk/Config.in
@@ -6,6 +6,7 @@  config BR2_PACKAGE_HOST_OPENJDK_BIN_ARCH_SUPPORTS
 config BR2_PACKAGE_OPENJDK_ARCH_SUPPORTS
 	bool
 	default y if BR2_aarch64
+	default y if BR2_arc && !(BR2_arc750d || BR2_arc770d)
 	default y if BR2_arm
 	default y if BR2_i386 || BR2_x86_64
 	default y if BR2_m68k
@@ -121,6 +122,7 @@  choice
 	prompt "openjdk variant"
 	default BR2_PACKAGE_OPENJDK_JVM_VARIANT_SERVER if !BR2_powerpc
 	default BR2_PACKAGE_OPENJDK_JVM_VARIANT_ZERO if BR2_powerpc
+	default BR2_PACKAGE_OPENJDK_JVM_VARIANT_ZERO if BR2_arc
 
 config BR2_PACKAGE_OPENJDK_JVM_VARIANT_CLIENT
 	bool "client"