#406: working vmx_memchr() in NSPR, passes JS tests, runs browser

configure.in

@@ -9337,6 +9337,9 @@ ac_configure_args="$ac_configure_args --prefix=$dist"
 if test "$MOZ_MEMORY"; then
    ac_configure_args="$ac_configure_args --enable-jemalloc"
 fi
+if test "$TENFOURFOX_VMX"; then
+   ac_configure_args="$ac_configure_args --enable-tenfourfox-vmx"
+fi
 if test -n "$ZLIB_IN_MOZGLUE"; then
    MOZ_ZLIB_LIBS=
 fi
@@ -9355,6 +9358,7 @@ export DIST
 export MOZ_LINKER
 export ZLIB_IN_MOZGLUE
 export MOZ_MEMORY
+export TENFOURFOX_VMX
 
 if ! test -e js; then
     mkdir js

js/src/jsstr.cpp

@@ -51,6 +51,9 @@
 #include "vm/StringObject-inl.h"
 #include "vm/TypeInference-inl.h"
 
+#include "mozilla-config.h"
+#include "plvmx.h"
+
 using namespace js;
 using namespace js::gc;
 using namespace js::unicode;
@@ -1155,10 +1158,12 @@ FirstCharMatcherUnrolled(const TextChar* text, uint32_t n, const PatChar pat)
 static const char*
 FirstCharMatcher8bit(const char* text, uint32_t n, const char pat)
 {
-#if  defined(__clang__)
+#ifndef TENFOURFOX_VMX
+#warning using non-VMX memchr
     return FirstCharMatcherUnrolled<char, char>(text, n, pat);
 #else
-    return reinterpret_cast<const char*>(memchr(text, pat, n));
+#warning using VMX memchr
+    return reinterpret_cast<const char*>(vmx_memchr(text, pat, n));
 #endif
 }
 

nsprpub/config/autoconf.mk.in

@@ -54,6 +54,7 @@ USE_N32		= @USE_N32@
 USE_X32		= @USE_X32@
 USE_64		= @USE_64@
 ENABLE_STRIP	= @ENABLE_STRIP@
+TENFOURFOX_VMX  = @TENFOURFOX_VMX@
 
 USE_PTHREADS	= @USE_PTHREADS@
 USE_BTHREADS	= @USE_BTHREADS@

nsprpub/configure

@@ -667,6 +667,7 @@ DLL_SUFFIX
 LIB_SUFFIX
 OBJ_SUFFIX
 CPU_ARCH
+TENFOURFOX_VMX
 PR_MD_ARCH_DIR
 PR_MD_ASFILES
 PR_MD_CSRCS
@@ -801,6 +802,7 @@ enable_x32
 enable_64bit
 enable_mdupdate
 enable_cplus
+enable_tenfourfox_vmx
 with_arm_kuser
 with_macos_sdk
 enable_macos_target
@@ -1465,6 +1467,7 @@ Optional Features:
   --enable-macos-target=VER
                           Set the minimum MacOS version needed at runtime
                           10.2 for ppc, 10.4 for x86
+  --enable-tenfourfox-vmx Enable VMX SIMD features for PowerPC
   --disable-os2-high-mem  Disable high-memory support on OS/2
 
   --enable-strip          Enable stripping of shared libs and programs
@@ -2495,6 +2498,7 @@ _HAVE_PTHREADS=
 USE_PTHREADS=
 USE_USER_PTHREADS=
 USE_NSPR_THREADS=
+TENFOURFOX_VMX=
 USE_N32=
 USE_X32=
 USE_64=
@@ -2904,6 +2908,16 @@ if test "${enable_n32+set}" = set; then :
     fi
 fi
 
+# Check whether --enable-tenfourfox-vmx was given.
+if test "${enable_tenfourfox_vmx+set}" = set; then :
+  enableval=$enable_tenfourfox_vmx;  if test "$enableval" = "yes"; then
+	TENFOURFOX_VMX=1
+      else if test "$enableval" = "no"; then
+	TENFOURFOX_VMX=
+      fi
+    fi
+fi
+
 
 # Check whether --enable-x32 was given.
 if test "${enable_x32+set}" = set; then :
@@ -6524,6 +6538,10 @@ fi
 
     $as_echo "#define HAVE_SOCKLEN_T 1" >>confdefs.h
 
+    if test -n "$TENFOURFOX_VMX"; then
+      $as_echo "#define TENFOURFOX_VMX 1" >>confdefs.h
+    fi
+
     AS='$(CC) -x assembler-with-cpp'
     CFLAGS="$CFLAGS -Wall -fno-common"
     case "${target_cpu}" in

nsprpub/configure.in

@@ -313,6 +313,15 @@ AC_ARG_ENABLE(optimize,
           MOZ_OPTIMIZE=
       fi ])
 
+AC_ARG_ENABLE(tenfourfox-vmx,
+[ --enable-tenfourfox-vmx Turn on AltiVec-VMX specific options for TenFourFox ],
+[ if test "$enableval" != "no"; then
+    AC_DEFINE(TENFOURFOX_VMX)
+    TENFOURFOX_VMX=1
+  else
+    TENFOURFOX_VMX=
+  fi ])
+
 AC_ARG_ENABLE(debug,
     [  --enable-debug[=DBG]    Enable debugging (using compiler flags DBG)],
     [ if test "$enableval" != "no"; then
@@ -3049,6 +3058,7 @@ AC_SUBST(PR_MD_CSRCS)
 AC_SUBST(PR_MD_ASFILES)
 AC_SUBST(PR_MD_ARCH_DIR)
 AC_SUBST(CPU_ARCH)
+AC_SUBST(TENFOURFOX_VMX)
 
 AC_SUBST(OBJ_SUFFIX)
 AC_SUBST(LIB_SUFFIX)

nsprpub/lib/libc/src/Makefile.in

@@ -39,6 +39,11 @@ LIBRARY_VERSION	= $(MOD_MAJOR_VERSION)
 
 RELEASE_LIBS = $(TARGETS)
 
+ifeq ($(TENFOURFOX_VMX),1)
+CSRCS += plvmx.c
+CFLAGS += -faltivec
+endif
+
 ifeq ($(OS_ARCH),WINNT)
 RES=$(OBJDIR)/plc.res
 RESNAME=plc.rc

nsprpub/lib/libc/src/plvmx.c

@@ -0,0 +1,64 @@
+#include <stdio.h>
+#include "plvmx.h"
+
+void *vmx_memchr(const void *b, int c, size_t length) {
+
+
+/* From:
+https://github.com/ridiculousfish/HexFiend/blob/4d5bcee5715f5f288649f7471d1da5bd06376f46/framework/sources/HFFastMemchr.m
+   with some optimizations */
+
+    const unsigned char *haystack = (const unsigned char *)b;
+    unsigned char needle = (unsigned char)c;
+
+    unsigned prefixLength = (unsigned)((16 - ((unsigned long)haystack) % 16) % 16);
+    unsigned suffixLength = (unsigned)(((unsigned long)(haystack + length)) % 16);
+    // It is possible for altivecLength to be < 0 for short strings.
+    int altivecLength = length - prefixLength - suffixLength;
+
+    if (altivecLength < 16) {
+        while (length--) {
+            if (*haystack == needle) return (void *)haystack;
+            haystack++;
+        }
+        return NULL;
+    }
+
+    size_t numVectors = altivecLength >> 4;
+    while (prefixLength--) {
+        if (*haystack == needle) return (void *)haystack;
+        haystack++;
+    }
+    
+    unsigned int mashedByte = (needle << 24 ) | (needle << 16) | (needle << 8) | needle;
+    const vector unsigned char searchVector = (vector unsigned int){mashedByte, mashedByte, mashedByte, mashedByte};
+    while (numVectors--) {
+        if (vec_any_eq(*(const vector unsigned char*)haystack, searchVector)) goto foundResult;
+        haystack += 16;
+    }
+    
+    while (suffixLength--) {
+        if (*haystack == needle) return (void *)haystack;
+        haystack++;
+    }
+    
+    return NULL;
+    
+foundResult:
+        ;
+    /* some byte has the result - look in groups of 4 to find which it is */
+    unsigned numWords = 4;
+    unsigned int val;
+    while (numWords--) {
+        val = *(unsigned int*)haystack;
+        if (((val >> 24) & 0xFF) == needle) return (void *)haystack;
+        if (((val >> 16) & 0xFF) == needle) return (void *)(1 + haystack);
+        if (((val >> 8) & 0xFF)  == needle) return (void *)(2 + haystack);
+        if ((val & 0xFF)         == needle) return (void *)(3 + haystack);
+        haystack += 4;
+    }
+    
+    /* should never get here */
+    fprintf(stderr, "failed vmx_memchr()\n");
+    return NULL;
+}