Windows: Update the build script and README-Windows.txt.

The 32-bit build is now for i686 or newer because the
prebuilt MinGW-w64 toolchains include i686 code in the
executables even if one uses -march=i486.

The build script builds 32-bit SSE2 enabled version too.
Run-time detection of SSE2 support would be nice (on any OS)
but it's not implemented in XZ Utils yet.

diff --git a/windows/README-Windows.txt b/windows/README-Windows.txt
index b6a85b83eedbac3fdb37071a97186d6ac242f942..f9a00248d7cb641e181cc0f81bce0eb31b14f962 100644 (file)
--- a/windows/README-Windows.txt
+++ b/windows/README-Windows.txt
@@ -21,14 +21,18 @@ Package contents
 ----------------
 
     All executables and libraries in this package require msvcrt.dll.
-    It's included in all recent Windows versions. On Windows 95 it
-    might be missing, but once you get it somewhere, XZ Utils should
-    run even on Windows 95.
+    It's included in all recent Windows versions. (On Windows 95 it
+    might be missing, but once you get it somewhere, the i686 binaries
+    should run even on Windows 95 if the processor is new enough.)
 
-    There are two different versions of the executable and library files.
-    There is one directory for each type of binaries:
+    There is a SSE2 optimization in the compression code but this
+    version of XZ Utils doesn't include run-time processor detection.
+    This is why there is a separate i686-SSE2 version.
 
-        bin_i486        32-bit x86 (i486 and up), Windows 95 and later
+    There is one directory for each type of executable and library files:
+
+        bin_i686        32-bit x86 (i686 and newer), Windows 95 and later
+        bin_i686-sse2   32-bit x86 (i686 with SSE2), Windows 98 and later
         bin_x86-64      64-bit x86-64, Windows Vista and later
 
     Each of the above directories have the following files:
@@ -90,15 +94,13 @@ Microsoft Visual C++
 
         lib /def:liblzma.def /out:liblzma.lib /machine:x64
 
-    Linking against static liblzma might work too, but usually you
-    should use liblzma.dll if possible. (Or, if having a decompressor
-    is enough, consider using XZ Embedded or LZMA SDK which can be
-    compiled with MSVC.)
+    If you need to link statically against liblzma, you should build
+    liblzma with MSVC 2013 update 2 or later. Alternatively, if having
+    a decompressor is enough, consider using XZ Embedded or LZMA SDK.
 
-    To try linking against static liblzma, rename liblzma.a to e.g.
-    liblzma_static.lib and tell MSVC to link against it. You also need
-    to tell lzma.h to not use __declspec(dllimport) by defining the
-    macro LZMA_API_STATIC. You can do it either in the C/C++ code
+    When you plan to link against static liblzma, you need to tell
+    lzma.h to not use __declspec(dllimport) by defining the macro
+    LZMA_API_STATIC. You can do it either in the C/C++ code
 
         #define LZMA_API_STATIC
         #include <lzma.h>

diff --git a/windows/build.bash b/windows/build.bash
index 85e1f4fcc27f3d8ccf18e97080b72a3b6b1cb09c..3d8fb559b842b18f927a7f6a851bc4defa690b5a 100644 (file)
--- a/windows/build.bash
+++ b/windows/build.bash
@@ -75,6 +75,8 @@ buildit()
        # threading. So I don't include a size-optimized liblzma for now.
        ./configure \
                --prefix= \
+               --enable-silent-rules \
+               --disable-dependency-tracking \
                --disable-nls \
                --disable-scripts \
                --disable-threads \
@@ -89,9 +91,12 @@ buildit()
 
        make distclean
 
-       # Build the normal speed-optimized binaries.
+       # Build the normal speed-optimized binaries. The type of threading
+       # (win95 vs. vista) will be autodetect from the target architecture.
        ./configure \
                --prefix= \
+               --enable-silent-rules \
+               --disable-dependency-tracking \
                --disable-nls \
                --disable-scripts \
                --build="$BUILD" \
@@ -125,19 +130,19 @@ txtcp()
        done
 }
 
-# FIXME: Make sure that we don't get i686 or i586 code from the runtime.
-# Or if we do, update the strings here to match the generated code.
-# i686 has cmov which can help like maybe 1 % in performance but things
-# like SSE don't help, so i486 isn't horrible for performance.
-#
-# FIXME: Using i486 in the configure triplet may be wrong.
 if [ -d "$MINGW_W32_DIR" ]; then
        # 32-bit x86, Win95 or later, using MinGW-w32
        PATH=$MINGW_W32_DIR/bin:$MINGW_W32_DIR/i686-w64-mingw32/bin:$PATH \
                        buildit \
                        pkg/bin_i486 \
-                       i486-w64-mingw32 \
-                       '-march=i486 -mtune=generic'
+                       i686-w64-mingw32 \
+                       '-march=i686 -mtune=generic'
+       # 32-bit x86 with SSE2, Win98 or later, using MinGW-w32
+       PATH=$MINGW_W32_DIR/bin:$MINGW_W32_DIR/i686-w64-mingw32/bin:$PATH \
+                       buildit \
+                       pkg/bin_i686-sse2 \
+                       i686-w64-mingw32 \
+                       '-march=i686 -msse2 -mfpmath=sse -mtune=generic'
 elif [ -d "$MINGW_DIR" ]; then
        # 32-bit x86, Win95 or later, using MinGW
        PATH=$MINGW_DIR/bin:$PATH \