Record the sub-architecture (CPU features enabled in the compiler)

If the system's compiler has extra features on top of the base
settings enabled or if the mkspec or $CXXFLAGS variable included some,
record them.

This will allow us to choose whether or not to use our own special
compilers, based on whether the system default compiler contains it or
not.

Change-Id: I87cada9fab4cfa58846a831d0a7c7b50d8fa87fd
Reviewed-by: Oswald Buddenhagen <oswald.buddenhagen@nokia.com>

config.tests/arch/arch.cpp

@@ -76,7 +76,169 @@ const char msg[] = "==Qt=magic=Qt== Architecture:"
 #endif
                    ;
 
+extern const char msg2[];
+const char msg2[] = "==Qt=magic=Qt== Sub-architecture:"
+        // This is the list of features found in GCC or MSVC
+        // We don't use all of them, but this is ready for future expansion
+
+// -- x86 --
+#ifdef __3dNOW__
+// 3dNow!, introduced with the AMD K6-2, discontinued after 2010
+" 3dnow"
+#endif
+#ifdef __3dNOW_A__
+// Athlon
+" 3dnow-a"
+#endif
+#ifdef __ABM__
+// Advanced Bit Manipulation, AMD Barcelona (family 10h)
+" abm"
+#endif
+#ifdef __AES__
+// AES New Instructions, Intel Core-i7 second generation ("Sandy Bridge")
+" aes"
+#endif
+#ifdef __AVX__
+// Advanced Vector Extensions, Intel Core-i7 second generation ("Sandy Bridge")
+" avx"
+#endif
+#ifdef __AVX2__
+// AVX 2, Intel codename "Haswell"
+" avx2"
+#endif
+#ifdef __BMI__
+// Bit Manipulation Instructions 1, Intel codename "Haswell", AMD "Bulldozer 2"
+" bmi"
+#endif
+#ifdef __BMI2__
+// Bit Manipulation Instructions 2, Intel codename "Haswell"
+" bmi2"
+#endif
+#ifdef __GCC_HAVE_SYNC_COMPARE_AND_SWAP_16
+// cmpxchg16b instruction, Intel Pentium 4 64-bit ("Nocona"), AMD Barcelona (family 10h)
+// Notably, this instruction is missing on earlier AMD Athlon 64
+" cx16"
+#endif
+#ifdef __F16C__
+// 16-bit floating point conversion, Intel codename "Ivy Bridge"
+" f16c"
+#endif
+#ifdef __FMA__
+// Fused Multiply-Add with 3 arguments, Intel codename "Haswell", AMD "Bulldozer 2"
+// a.k.a. "FMA3"
+" fma"
+#endif
+#ifdef __FMA4__
+// Fused Multiply-Add with 4 arguments, AMD "Bulldozer"
+" fma4"
+#endif
+#ifdef __FSGSBASE__
+// rdfsgsbase, wrfsgsbase, Intel codename "Ivy Bridge"
+" fsgsbase"
+#endif
+#ifdef __LWP__
+// LWP instructions, AMD "Bulldozer"
+" lwp"
+#endif
+#ifdef __LZCNT__
+// Leading-Zero bit count, Intel codename "Haswell"
+" lzcnt"
+#endif
+#ifdef __MMX__
+// Multimedia Extensions, Pentium MMX, AMD K6-2
+" mmx"
+#endif
+#ifdef __MOVBE__
+// Move Big Endian, Intel Atom & "Haswell"
+" movbe"
+#endif
+#ifdef __NO_SAHF__
+// missing SAHF instruction in 64-bit, up to Intel Pentium 4 64-bit ("Nocona"), AMD Athlon FX
+// Note: the macro is not defined, so this will never show up
+" no-sahf"
+#endif
+#ifdef __PCLMUL__
+// (Packed) Carry-less multiplication, Intel Core-i7 second generation ("Sandy Bridge")
+" pclmul"
+#endif
+#ifdef __POPCNT__
+// Population Count (count of set bits), Intel Core-i7 second generation ("Sandy Bridge")
+" popcnt"
+#endif
+#ifdef __RDRND__
+// Random number generator, Intel codename "Ivy Bridge"
+" rdrnd"
+#endif
+#if defined(__SSE__) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1) || defined(_M_X64)
+// Streaming SIMD Extensions, Intel Pentium III, AMD Athlon
+" sse"
+#endif
+#if defined(__SSE2__) || (defined(_M_IX86_FP) && _M_IX86_FP >= 2) || defined(_M_X64)
+// SSE2, Intel Pentium-M, Intel Pentium 4, AMD Opteron and Athlon 64
+" sse2"
+#endif
+#ifdef __SSE3__
+// SSE3, Intel Pentium 4 "Prescott", AMD Athlon 64 rev E
+" sse3"
+#endif
+#ifdef __SSSE3__
+// Supplemental SSE3, Intel Core 2, AMD "Bulldozer"
+" ssse3"
+#endif
+#ifdef __SSE4A__
+// SSE4a, AMD Barcelona
+" sse4a"
+#endif
+#ifdef __SSE4_1__
+// SSE 4.1, Intel Core-i7 ("Nehalem"), AMD "Bulldozer"
+" sse4.1"
+#endif
+#ifdef __SSE4_2__
+// SSE 4.2, Intel Core-i7 ("Nehalem"), AMD "Bulldozer"
+" sse4.2"
+// Since no processor supports SSE4.2 without 4.1 and since no Intel processor
+// supports SSE4a, define "sse4" to indicate SSE4"
+" sse4"
+#endif
+#ifdef __TBM__
+// TBM, AMD "Bulldozer"
+" tbm"
+#endif
+#ifdef __XOP__
+// XOP, AMD "Bulldozer"
+" xop"
+#endif
+
+// -- ARM --
+#ifdef __ARM_NEON__
+" neon"
+#endif
+#ifdef __IWMMXT__
+" iwmmxt"
+#endif
+
+// -- SPARC --
+#ifdef __VIS__
+" vis"
+# if __VIS__ >= 0x200
+" vis2"
+# endif
+# if __VIS__ >= 0x300
+" vis3"
+# endif
+#endif
+
+// -- MIPS --
+// Wikipedia says there are extensions, but GCC installs no headers
+
+// -- POWER, PowerPC --
+#ifdef __ALTIVEC__
+" altivec"
+#endif
+                   "";
+
 int main()
 {
     puts(msg);
+    puts(msg2);
 }

config.tests/unix/arch.test

@@ -5,6 +5,7 @@ VERBOSE=$2
 SRCDIR=$3
 OUTDIR=$4
 RESULTFILE=$5
+VARPREFIX=$6
 
 # debuggery
 [ "$VERBOSE" = "yes" ] && echo "Determining architecture... ($*)"
@@ -28,14 +29,15 @@ else
     exit 2
 fi
 
-ARCH=`strings - $binary 2>/dev/null | awk -F: '/==Qt=magic=Qt==/ { print $2 }'` ||
+if strings - $binary 2>/dev/null | \
-{
+    awk -F: '/==Qt=magic=Qt== Architecture/ { print "'$VARPREFIX'_ARCH=\"" $2 "\"" }
+             /==Qt=magic=Qt== Sub-architecture/ { print "'$VARPREFIX'_CPUFEATURES=\"" $2 "\"" }' > "$RESULTFILE"
+then
+    [ "$VERBOSE" = "yes" ] && echo "    Found architecture in binary" && \
+        cat "$RESULTFILE"
+else
     [ "$VERBOSE" = "yes" ] && echo "Unable to determine architecture!"
     exit 2
-}
+fi
 
 $MAKE distclean >/dev/null 2>&1
-
-[ "$VERBOSE" = "yes" ] && echo "    Found '$ARCH' in binary"
-
-echo "$ARCH" > $5

configure

@@ -708,8 +708,10 @@ CFG_PKGCONFIG=auto
 
 # Target architecture
 CFG_ARCH=
+CFG_CPUFEATURES=
 # Host architecture, same as CFG_ARCH when not cross-compiling
 CFG_HOST_ARCH=
+CFG_HOST_CPUFEATURES=
 # Set when the -arch or -host-arch arguments are used
 OPT_OBSOLETE_HOST_ARG=no
 
@@ -3668,9 +3670,9 @@ fi
 
 # Use config.tests/arch/arch.pro to has the compiler tell us what the target architecture is
 OUTFILE=$outpath/arch.result
-"$unixtests/arch.test" "$XQMAKESPEC" $OPT_VERBOSE "$relpath" "$outpath" "$OUTFILE"
+"$unixtests/arch.test" "$XQMAKESPEC" $OPT_VERBOSE "$relpath" "$outpath" "$OUTFILE" "CFG"
 if [ $? -eq 0 ]; then
-    CFG_ARCH=`cat "$OUTFILE"`
+    eval `cat "$OUTFILE"`
 else
     echo
     echo "Could not determine the target architecture!"
@@ -3681,9 +3683,9 @@ rm -f "$OUTFILE" 2>/dev/null
 
 if [ "$QMAKESPEC" != "$XQMAKESPEC" ]; then
     # Do the same test again, using the host compiler
-    SYSROOT_FLAG= "$unixtests/arch.test" "$QMAKESPEC" $OPT_VERBOSE "$relpath" "$outpath" "$OUTFILE"
+    SYSROOT_FLAG= "$unixtests/arch.test" "$QMAKESPEC" $OPT_VERBOSE "$relpath" "$outpath" "$OUTFILE" "CFG_HOST"
     if [ $? -eq 0 ]; then
-        CFG_HOST_ARCH=`cat "$OUTFILE"`
+        eval `cat "$OUTFILE"`
     else
         echo
         echo "Could not determine the host architecture!"
@@ -3694,6 +3696,7 @@ if [ "$QMAKESPEC" != "$XQMAKESPEC" ]; then
 else
     # not cross compiling, host == target
     CFG_HOST_ARCH="$CFG_ARCH"
+    CFG_HOST_CPUFEATURES="$CFG_CPUFEATURES"
 fi
 unset OUTFILE
 
@@ -5719,6 +5722,8 @@ cat >>"$QTCONFIG.tmp" <<EOF
 CONFIG += $QTCONFIG_CONFIG
 QT_ARCH = $CFG_ARCH
 QT_HOST_ARCH = $CFG_HOST_ARCH
+QT_CPU_FEATURES = $CFG_CPUFEATURES
+QT_HOST_CPU_FEATURES = $CFG_HOST_CPUFEATURES
 QT_EDITION = $Edition
 QT_CONFIG += $QT_CONFIG
 
@@ -5882,8 +5887,9 @@ else
     echo "Building for:  $XPLATFORM"
 fi
 
-echo "Architecture:  $CFG_ARCH"
+# the missing space before $CFG_FEATURES is intentional
-echo "Host architecture: $CFG_HOST_ARCH"
+echo "Architecture:  $CFG_ARCH, features:$CFG_CPUFEATURES"
+echo "Host architecture: $CFG_HOST_ARCH, features:$CFG_HOST_CPUFEATURES"
 
 if [ -n "$PLATFORM_NOTES" ]; then
     echo "Platform notes:"

tools/configure/configureapp.cpp

@@ -2531,6 +2531,17 @@ void Configure::generateCachefile()
     }
 }
 
+struct ArchData {
+    const char *qmakespec;
+    const char *key;
+    const char *subarchKey;
+    const char *type;
+    ArchData() {}
+    ArchData(const char *t, const char *qm, const char *k, const char *sak)
+        : qmakespec(qm), key(k), subarchKey(sak), type(t)
+    {}
+};
+
 /*
     Runs qmake on config.tests/arch/arch.pro, which will detect the target arch
     for the compiler we are using
@@ -2551,15 +2562,16 @@ void Configure::detectArch()
         return;
     }
 
-    QList<QPair<QString, QString> > qmakespecs;
+    QVector<ArchData> qmakespecs;
     if (dictionary.contains("XQMAKESPEC"))
-        qmakespecs << qMakePair(QString("XQMAKESPEC"), QString("QT_ARCH"));
+        qmakespecs << ArchData("target", "XQMAKESPEC", "QT_ARCH", "QT_CPU_FEATURES");
-    qmakespecs << qMakePair(QString("QMAKESPEC"), QString("QT_HOST_ARCH"));
+    qmakespecs << ArchData("host", "QMAKESPEC", "QT_HOST_ARCH", "QT_HOST_CPU_FEATURES");
 
     for (int i = 0; i < qmakespecs.count(); ++i) {
-        const QPair<QString, QString> &pair = qmakespecs.at(i);
+        const ArchData &data = qmakespecs.at(i);
-        QString qmakespec = dictionary.value(pair.first);
+        QString qmakespec = dictionary.value(data.qmakespec);
-        QString key = pair.second;
+        QString key = data.key;
+        QString subarchKey = data.subarchKey;
 
         // run qmake
         QString command =
@@ -2584,10 +2596,11 @@ void Configure::detectArch()
         QByteArray exeContents = exe.readAll();
         exe.close();
 
-        static const char magic[] = "==Qt=magic=Qt== Architecture:";
+        static const char archMagic[] = "==Qt=magic=Qt== Architecture:";
-        int magicPos = exeContents.indexOf(magic);
+        int magicPos = exeContents.indexOf(archMagic);
         if (magicPos == -1) {
-            cout << "Internal error, could not find the architecture of the executable" << endl;
+            cout << "Internal error, could not find the architecture of the "
+                 << data.type << " executable" << endl;
             dictionary["DONE"] = "error";
             return;
         }
@@ -2595,10 +2608,24 @@ void Configure::detectArch()
 
         // the conversion from QByteArray will stop at the ending NUL anyway
         QString arch = QString::fromLatin1(exeContents.constData() + magicPos
-                                           + sizeof(magic) - 1);
+                                           + sizeof(archMagic) - 1);
         dictionary[key] = arch;
 
+        static const char subarchMagic[] = "==Qt=magic=Qt== Sub-architecture:";
+        magicPos = exeContents.indexOf(subarchMagic);
+        if (magicPos == -1) {
+            cout << "Internal error, could not find the sub-architecture of the "
+                 << data.type << " executable" << endl;
+            dictionary["DONE"] = "error";
+            return;
+        }
+
+        QString subarch = QString::fromLatin1(exeContents.constData() + magicPos
+                                              + sizeof(subarchMagic) - 1);
+        dictionary[subarchKey] = subarch;
+
         //cout << "Detected arch '" << qPrintable(arch) << "'\n";
+        //cout << "Detected sub-arch '" << qPrintable(subarch) << "'\n";
 
         // clean up
         Environment::execute(command + " distclean");
@@ -2606,8 +2633,10 @@ void Configure::detectArch()
 
     if (!dictionary.contains("QT_HOST_ARCH"))
         dictionary["QT_HOST_ARCH"] = "unknown";
-    if (!dictionary.contains("QT_ARCH"))
+    if (!dictionary.contains("QT_ARCH")) {
         dictionary["QT_ARCH"] = dictionary["QT_HOST_ARCH"];
+        dictionary["QT_CPU_FEATURES"] = dictionary["QT_HOST_CPU_FEATURES"];
+    }
 
     QDir::setCurrent(oldpwd);
 }
@@ -2650,6 +2679,8 @@ void Configure::generateQConfigPri()
         configStream << endl;
         configStream << "QT_ARCH = " << dictionary["QT_ARCH"] << endl;
         configStream << "QT_HOST_ARCH = " << dictionary["QT_HOST_ARCH"] << endl;
+        configStream << "QT_CPU_FEATURES = " << dictionary["QT_CPU_FEATURES"] << endl;
+        configStream << "QT_HOST_CPU_FEATURES = " << dictionary["QT_HOST_CPU_FEATURES"] << endl;
         if (dictionary["QT_EDITION"].contains("OPENSOURCE"))
             configStream << "QT_EDITION = " << QLatin1String("OpenSource") << endl;
         else
@@ -3037,8 +3068,10 @@ void Configure::displayConfig()
         sout << "QMAKESPEC..................." << dictionary[ "XQMAKESPEC" ] << " (" << dictionary["QMAKESPEC_FROM"] << ")" << endl;
     else
         sout << "QMAKESPEC..................." << dictionary[ "QMAKESPEC" ] << " (" << dictionary["QMAKESPEC_FROM"] << ")" << endl;
-    sout << "Architecture................" << dictionary["QT_ARCH"] << endl;
+    sout << "Architecture................" << dictionary["QT_ARCH"]
-    sout << "Host Architecture..........." << dictionary["QT_HOST_ARCH"] << endl;
+         << ", features:" << dictionary["QT_CPU_FEATURES"] << endl;
+    sout << "Host Architecture..........." << dictionary["QT_HOST_ARCH"]
+         << ", features:" << dictionary["QT_HOST_CPU_FEATURES"]  << endl;
     sout << "Maketool...................." << dictionary[ "MAKE" ] << endl;
     sout << "Debug symbols..............." << (dictionary[ "BUILD" ] == "debug" ? "yes" : "no") << endl;
     sout << "Link Time Code Generation..." << dictionary[ "LTCG" ] << endl;