add workaround for clang on AMD with FMA

The clang bug which causes incorrect code to be generated on AMD
with FMA (Piledriver and Bulldozer) is in fact caused by an assembler
bug and can be worked around by switching to an external assembler.
This commit implements such the workaround for clang 3.1 and 3.2.
The bug was fix in clang version 3.2-svn-r173176:
http://llvm.org/bugs/show_bug.cgi?id=15040
Hence, no workaround is needed for clang 3.3 with AVX_128_FMA
acceleration.

This workaround is ineffective on Mac OS because there is no suitable
external assembler, but distinguishing between vanilla and Apple clang
requires more code than what's worth adding to cover the case of
AMD-based Hackintoshes (all Mac-s are Intel-based).

It is appropriate to amend only C and not C++ compilation for
release-4-6 branch, but when merging this to master branch, reconsider
the detail of the implementation.

Refs #1099

Change-Id: I835e0f2436ee33507514d3bc3980fcc227e2e36f

diff --git a/CMakeLists.txt b/CMakeLists.txt
index dd02712160ad006a27019ac188b6031e27032d27..9537782c5f8d1840b54db1855de0b764493133e9 100644 (file)
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -280,6 +280,10 @@ endif()
 include(gmxCFlags)
 gmx_c_flags()
 
+# This variable should be used for additional compiler flags which are not
+# generated in gmxCFlags nor are acceleration or MPI related.
+set(EXTRA_C_FLAGS "")
+
 # gcc 4.4.x is buggy and crashes when compiling some files with O3 and OpenMP on.
 # Detect here whether applying a workaround is needed and will apply it later
 # on the affected files.
@@ -299,6 +303,15 @@ if(COMPILER_IS_CLANG_3_0)
     message(FATAL_ERROR "Your compiler is clang version 3.0, which is known to be buggy for GROMACS. Use a different compiler.")
 endif()
 
+# clang <=3.2 contains a bug that causes incorrect code to be generated for the
+# vfmaddps instruction and therefore the bug is triggered with AVX_128_FMA.
+# (see: http://llvm.org/bugs/show_bug.cgi?id=15040).
+# We can work around this by not using the integrated assembler (except on OS X
+# which has an outdated assembler that does not support AVX instructions).
+if (${CMAKE_C_COMPILER_ID} MATCHES "Clang" AND C_COMPILER_VERSION VERSION_LESS "3.3")
+    set(GMX_USE_CLANG_FMA_BUG_WORKAROUND TRUE)
+endif()
+
 if (CMAKE_C_COMPILER_ID STREQUAL "PGI")
     message(WARNING "All tested PGI compiler versions (up to 12.9.0) generate binaries which produce incorrect results, or even fail to compile Gromacs. Highly recommended to use a different compiler. If you choose to use PGI, make sure to run the regressiontests.")
 endif()
@@ -784,9 +797,6 @@ elseif(${GMX_CPU_ACCELERATION} STREQUAL "AVX_128_FMA" OR ${GMX_CPU_ACCELERATION}
 
     # Set the FMA4 flags (MSVC doesn't require any)
     if(${GMX_CPU_ACCELERATION} STREQUAL "AVX_128_FMA" AND NOT MSVC)
-        if (${CMAKE_COMPILER_ID} MATCHES "Clang")
-            message(FATAL_ERROR "Clang up to at least version 3.2 produces incorrect code for AVX_128_FMA. Sorry, but you will have to select a different compiler or acceleration.")
-        endif()
         GMX_TEST_CFLAG(GNU_FMA_CFLAG "-mfma4" ACCELERATION_C_FLAGS)
         if (NOT GNU_FMA_CFLAG)
             message(WARNING "No C FMA4 flag found. Consider a newer compiler, or try SSE4.1 (lower performance).")
@@ -834,6 +844,20 @@ elseif(${GMX_CPU_ACCELERATION} STREQUAL "AVX_128_FMA" OR ${GMX_CPU_ACCELERATION}
         if (NOT ACCELERATION_QUIETLY)
           message(STATUS "Enabling 128-bit AVX Gromacs acceleration (with fused-multiply add), and it will help compiler optimization.")
         endif()
+
+        # We don't have the full compiler version string yet (BUILD_C_COMPILER),
+        # so we can't distinguish vanilla and Apple clang, but catering for AMD
+        # hackintoshes is not worth the effort.
+        if (APPLE AND ${CMAKE_C_COMPILER_ID} STREQUAL "Clang")
+            message(WARNING "Due to a known compiler bug, Clang up to version 3.2 (and Apple Clang up to version 4.1) produces incorrect code with AVX_128_FMA acceleration. As we can not work around this bug on OS X, you will have to select a different compiler or CPU acceleration.")
+        endif()
+
+        if (GMX_USE_CLANG_FMA_BUG_WORKAROUND)
+            # we assume that we have an external assembler that supports AVX
+            message(STATUS "Clang ${C_COMPILER_VERSION} detected, enabling FMA bug workaround")
+            set(EXTRA_C_FLAGS "${EXTRA_C_FLAGS} -no-integrated-as")
+        endif()
+
     else()
         # If we are not doing AVX_128, it must be AVX_256...
         set(GMX_CPU_ACCELERATION_X86_AVX_256 1)
@@ -1111,13 +1135,13 @@ endif(GMX_FAHCORE)
 # # # # # # # # # # NO MORE TESTS AFTER THIS LINE! # # # # # # # # # # #
 # these are set after everything else
 if (NOT GMX_SKIP_DEFAULT_CFLAGS)
-    set(CMAKE_C_FLAGS "${ACCELERATION_C_FLAGS} ${MPI_COMPILE_FLAGS} ${CMAKE_C_FLAGS}")
+    set(CMAKE_C_FLAGS "${ACCELERATION_C_FLAGS} ${MPI_COMPILE_FLAGS} ${EXTRA_C_FLAGS} ${CMAKE_C_FLAGS}")
     set(CMAKE_CXX_FLAGS "${ACCELERATION_CXX_FLAGS} ${MPI_COMPILE_FLAGS} ${CMAKE_CXX_FLAGS}")
     set(CMAKE_EXE_LINKER_FLAGS "${FFT_LINKER_FLAGS} ${MPI_LINKER_FLAGS} ${CMAKE_EXE_LINKER_FLAGS}")
     set(CMAKE_SHARED_LINKER_FLAGS "${MPI_LINKER_FLAGS} ${CMAKE_SHARED_LINKER_FLAGS}")
 else()
     message("Recommended flags which are not added because GMX_SKIP_DEFAULT_CFLAGS=yes:")
-    message("CMAKE_C_FLAGS: ${ACCELERATION_C_FLAGS} ${MPI_COMPILE_FLAGS} ${GMXC_CFLAGS}")
+    message("CMAKE_C_FLAGS: ${ACCELERATION_C_FLAGS} ${MPI_COMPILE_FLAGS} ${EXTRA_C_FLAGS} ${GMXC_CFLAGS}")
     message("CMAKE_C_FLAGS_RELEASE: ${GMXC_CFLAGS_RELEASE}")
     message("CMAKE_C_FLAGS_DEBUG: ${GMXC_CFLAGS_DEBUG}")
     if(CMAKE_CXX_COMPILER_LOADED)