Merge lp:~sumit-semwal/pocl/arm-chrome-port into lp:~pocl/pocl/trunk

[Sumit Semwal]

ARM chromebook port: some minor build changes needed to get a pocl working set under ubuntu 13.04, with 3.3 llvm.

[Kalle Raiskila (kraiskil)]

This causes all tests (except AMD's radix sort!) to fail with:

+WARNING: Linking two modules of different target triples: /home/buildbot/buildslave/lp_pocl-ARM/build/lib/kernel/arm/kernel-armv7.bc: 'armv7--' and 'armv4t--'

This on Ubuntu 12.04, ARM CortexA9

review: Needs Fixing

[Pekka Jääskeläinen (pekka-jaaskelainen)]

Probably because the kernel binary is compiled with different clang parameters (in devices/common.c) than the kernel library.

[Kalle Raiskila (kraiskil)]

Looks like cause is the line(s) 61/62 in the diff. Is that there by mistake?

[lp:~sumit-semwal/pocl/arm-chrome-port]

723. By Erik Schnetter on 2013-05-06

Re-format tables in test_convert_type.cl to make them easier to check by eye

724. By Erik Schnetter on 2013-05-06

Merge from upstream

725. By Erik Schnetter on 2013-05-06

Correct C++11 issue in Khronos header files

[Sumit Semwal (sumit-semwal)]

Yes, it is due to line(s) 61/62, and yes, it is there by mistake! Sending out a new merge request to test!

review: Needs Resubmitting

[lp:~sumit-semwal/pocl/arm-chrome-port]

726. By Pekka Jääskeläinen on 2013-05-07

Clang++ 3.3 fails to compile the libstdc++ headers from gcc 4.7 in the std=gnu++11 mode. c++11 works.

727. By Pekka Jääskeläinen on 2013-05-07

Fixed the test suite status' for i686/Ubuntu 12.04LTS (Pentium 4).

728. By Pekka Jääskeläinen on 2013-05-07

The rotate case seems to fail on LLVM 3.2. XFAIL back until figured out why.

729. By Pekka Jääskeläinen on 2013-05-07

egrep instead of grep -e seems to be more compatible.

730. By Kalle Raiskila on 2013-05-08

Don't choke on doubles in test_rotate

731. By Kalle Raiskila on 2013-05-08

Add --disable-tce option to configure

732. By Pekka Jääskeläinen on 2013-05-08

Reverted r725 because it caused all tests using cl.hpp to segfault on CentOS x86_64 and ARM. Added cl.hpp to CLEANFILES.

733. By Kalle Raiskila on 2013-05-08

Handle LLVM 3.4 gracefully.

734. By Kalle Raiskila on 2013-05-08

Unmark convert type: it started to pass on PPC

735. By Kalle Raiskila on 2013-05-08

Fix 729 - x86 only should fail.

736. By Kalle Raiskila on 2013-05-08

Fix compiler flags for ARM gnueabihf (i.e. Ubuntu 13.04).

Patch by Sumit Semwal. Thanks!

737. By Kalle Raiskila on 2013-05-08

Fix 729 and 735 some more.

738. By Pekka Jääskeläinen on 2013-05-10

Wrong path to the Khronos ICD loader patch.

739. By Pekka Jääskeläinen on 2013-05-10

Vectors as value args is still broken.

740. By Pekka Jääskeläinen on 2013-05-10

Added a kernel compiler test case for a b-loop with two latches.

741. By Kalle Raiskila on 2013-05-10

Add test that shows a (yet unreported) bug with get_*_id() when kernel dim(X)<dim(Y).

742. By Pekka Jääskeläinen on 2013-05-10

Direct URL to CL registry in the Khronos web page.

743. By Pekka Jääskeläinen on 2013-05-10

Fix for get_local_id(0) when dim(x) < 2.

744. By Kalle Raiskila on 2013-05-10

AMD ludecomposition nolonger fails after 743.

745. By Pekka Jääskeläinen on 2013-05-12

* Marked the XFAIL status for LLVM 3.3+VML on Debian 7.0 x86_64.
* The AMD SDK was not enabled if testing from a separate build tree.

746. By Pekka Jääskeläinen on 2013-05-12

Minimal test case for the min() max() issue.

747. By Kalle Raiskila on 2013-05-13

Mark AMD ludecomposition as XFAIL on platforms lacking double.

748. By Kalle Raiskila on 2013-05-13

Mark XFAILs on LLVM 3.4 :(

749. By Pekka Jääskeläinen on 2013-05-13

Added the flaky case to the test suite. Now it seems to pass when running through the test suite and using LLVM 3.2. Strange one.

750. By Pekka Jääskeläinen on 2013-05-13

Added hint for an autogen.sh error when pkg-config is missing.

751. By Pekka Jääskeläinen on 2013-05-13

clGetCommandQueueInfo was not exported due to a missing POsym() call.

Unmerged revisions

751. By Pekka Jääskeläinen on 2013-05-13

clGetCommandQueueInfo was not exported due to a missing POsym() call.

750. By Pekka Jääskeläinen on 2013-05-13

Added hint for an autogen.sh error when pkg-config is missing.

749. By Pekka Jääskeläinen on 2013-05-13

Added the flaky case to the test suite. Now it seems to pass when running through the test suite and using LLVM 3.2. Strange one.

748. By Kalle Raiskila on 2013-05-13

Mark XFAILs on LLVM 3.4 :(

747. By Kalle Raiskila on 2013-05-13

Mark AMD ludecomposition as XFAIL on platforms lacking double.

746. By Pekka Jääskeläinen on 2013-05-12

Minimal test case for the min() max() issue.

745. By Pekka Jääskeläinen on 2013-05-12

* Marked the XFAIL status for LLVM 3.3+VML on Debian 7.0 x86_64.
* The AMD SDK was not enabled if testing from a separate build tree.

744. By Kalle Raiskila on 2013-05-10

AMD ludecomposition nolonger fails after 743.

743. By Pekka Jääskeläinen on 2013-05-10

Fix for get_local_id(0) when dim(x) < 2.

742. By Pekka Jääskeläinen on 2013-05-10

Direct URL to CL registry in the Khronos web page.

=== modified file 'config/xclang'
--- config/xclang	2012-11-21 10:53:42 +0000
+++ config/xclang	2013-05-06 16:06:27 +0000
@@ -52,5 +52,13 @@
 fi
 
 $CLANG -target $host -c -emit-llvm -o $tmpdir/xclang.bc $@
-$LLC -mtriple=$build -o $tmpdir/xclang.s $tmpdir/xclang.bc
+
+# Adjust $LLC call for hard-float in case the build is that
+case "$build" in
+arm*gnueabihf)
+   $LLC -mtriple=$build -float-abi=hard -o $tmpdir/xclang.s $tmpdir/xclang.bc;;
+*)
+   $LLC -mtriple=$build -o $tmpdir/xclang.s $tmpdir/xclang.bc;;
+esac
+
 exec $CLANG -target $build $output $tmpdir/xclang.s

=== modified file 'configure.ac'
--- configure.ac	2013-04-30 15:30:29 +0000
+++ configure.ac	2013-05-06 16:06:27 +0000
@@ -591,15 +591,22 @@
     ;;  
 esac
 AC_SUBST([ICD_LD_FLAGS])
-
 case $host_cpu in
   armv7l)
     AC_MSG_NOTICE([using the ARM optimized kernel lib for the native device])
-    #armv7l seems a popular host_cpu in distros. Clang uses "armv7", though
-    OCL_KERNEL_TARGET="armv7"
+    # check for gnueabihf, and set clang target accordingly
+    case $host in
+      armv7l*gnueabihf)
+         OCL_KERNEL_TARGET="armv7l-unknown-linux-gnueabihf"
+         HOST_LLC_FLAGS="$HOST_LLC_FLAGS -float-abi=hard";;
+      *)
+         #armv7l seems a popular host_cpu in distros. Clang uses "armv7", though
+         OCL_KERNEL_TARGET="armv7";;
+    esac
+
     # Assume -lm should not be used. TODO: add an embedded math lib
     # for compiling kernels for ARM envs without usable math libs.
-    HOST_CLANG_FLAGS="$HOST_CLANG_FLAGS -target $OCL_KERNEL_TARGET -mfloat-abi=hard -fPIC"
+    HOST_CLANG_FLAGS="$HOST_CLANG_FLAGS -target $OCL_KERNEL_TARGET -mfloat-abi=hard"
     kernel_dir="arm"
     CL_DEVICE_ADDRESS_BITS=32
     ;;

=== modified file 'scripts/pocl-build.in'
--- scripts/pocl-build.in	2013-04-25 13:51:09 +0000
+++ scripts/pocl-build.in	2013-05-06 16:06:27 +0000
@@ -45,6 +45,8 @@
 
 target_dir=${target}
 case $target in
+  arm*gnueabihf)    target_dir="arm"
+                    target="armv7l-unknown-linux-gnueabihf";;
   arm*)        target_dir="arm"
                target="armv7";;
   cellspu-*)   target_dir="cellspu";;