Vectorise copy_address_range_perms for common cases. This gives about

40% speedup on artificial programs which just do realloc() and nothing
else, and about a 3-4% speedup on starting kpresenter-1.5.0 and
loading a 16-slide presentation.

git-svn-id: svn://svn.valgrind.org/valgrind/trunk@5880

diff --git a/docs/internals/performance.txt b/docs/internals/performance.txt
index 8a12bfcd3ee36bffc340edcf4accaf2283f56bb3..2dbcfc6ae2a2db27286f907af96e64e1e9925f0f 100644 (file)
--- a/docs/internals/performance.txt
+++ b/docs/internals/performance.txt
@@ -29,6 +29,10 @@ Post 3.1.0:
 - Nick changed ExeContext gathering to not record/save extra zeroes at the
   end.  Saved 7% on perf/heap with --num-callers=50, and about 1% on
   perf/tinycc.
+- Julian vectorised copy_address_range_perms for common cases, which
+  gives about 40% speedup on artificial programs which just do
+  realloc() and nothing else, and about a 3-4% speedup on starting
+  kpresenter-1.5.0 and loading a 16-slide presentation.
 
 COMPVBITS branch:
 - Nick converted to compress V bits, initial version saved 0--5% on most

diff --git a/memcheck/mc_main.c b/memcheck/mc_main.c
index 0ee073b25ad10ef511d133b3d94c4b6c1f7bd471..b28e80eb27fdad5d9bc19a5eacf8d8fc4527f397 100644 (file)
--- a/memcheck/mc_main.c
+++ b/memcheck/mc_main.c
@@ -589,6 +589,28 @@ UChar get_vabits2 ( Addr a )
    return extract_vabits2_from_vabits8(a, vabits8);
 }
 
+// *** WARNING! ***
+// Any time this function is called, if it is possible that any of the
+// 4 2-bit fields in vabits8 are equal to VA_BITS2_PARTDEFINED, then the 
+// corresponding entry(s) in the sec-V-bits table must also be set!
+static INLINE
+UChar get_vabits8_for_aligned_word32 ( Addr a )
+{
+   SecMap* sm       = get_secmap_for_reading(a);
+   UWord   sm_off   = SM_OFF(a);
+   UChar   vabits8  = sm->vabits8[sm_off];
+   return vabits8;
+}
+
+static INLINE
+void set_vabits8_for_aligned_word32 ( Addr a, UChar vabits8 )
+{
+   SecMap* sm       = get_secmap_for_writing(a);
+   UWord   sm_off   = SM_OFF(a);
+   sm->vabits8[sm_off] = vabits8;
+}
+
+
 // Forward declarations
 static UWord get_sec_vbits8(Addr a);
 static void  set_sec_vbits8(Addr a, UWord vbits8);
@@ -1227,35 +1249,81 @@ static void make_mem_defined_if_addressable ( Addr a, SizeT len )
 void MC_(copy_address_range_state) ( Addr src, Addr dst, SizeT len )
 {
    SizeT i, j;
-   UChar vabits2;
+   UChar vabits2, vabits8;
+   Bool  aligned, nooverlap;
 
    DEBUG("MC_(copy_address_range_state)\n");
    PROF_EVENT(50, "MC_(copy_address_range_state)");
 
-   if (len == 0)
+   if (len == 0 || src == dst)
       return;
 
-   if (src < dst) {
-      for (i = 0, j = len-1; i < len; i++, j--) {
-         PROF_EVENT(51, "MC_(copy_address_range_state)(loop)");
-         vabits2 = get_vabits2( src+j );
-         set_vabits2( dst+j, vabits2 );
-         if (VA_BITS2_PARTDEFINED == vabits2) {
-            set_sec_vbits8( dst+j, get_sec_vbits8( src+j ) );
+   aligned   = VG_IS_4_ALIGNED(src) && VG_IS_4_ALIGNED(dst);
+   nooverlap = src+len <= dst || dst+len <= src;
+
+   if (nooverlap && aligned) {
+
+      /* Vectorised fast case, when no overlap and suitably aligned */
+      /* vector loop */
+      i = 0;
+      while (len >= 4) {
+         vabits8 = get_vabits8_for_aligned_word32( src+i );
+         set_vabits8_for_aligned_word32( dst+i, vabits8 );
+         if (EXPECTED_TAKEN(VA_BITS8_DEFINED == vabits8 
+                            || VA_BITS8_UNDEFINED == vabits8 
+                            || VA_BITS8_NOACCESS == vabits8)) {
+            /* do nothing */
+         } else {
+            /* have to copy secondary map info */
+            if (VA_BITS2_PARTDEFINED == get_vabits2( src+i+0 ))
+               set_sec_vbits8( dst+i+0, get_sec_vbits8( src+i+0 ) );
+            if (VA_BITS2_PARTDEFINED == get_vabits2( src+i+1 ))
+               set_sec_vbits8( dst+i+1, get_sec_vbits8( src+i+1 ) );
+            if (VA_BITS2_PARTDEFINED == get_vabits2( src+i+2 ))
+               set_sec_vbits8( dst+i+2, get_sec_vbits8( src+i+2 ) );
+            if (VA_BITS2_PARTDEFINED == get_vabits2( src+i+3 ))
+               set_sec_vbits8( dst+i+3, get_sec_vbits8( src+i+3 ) );
          }
+         i += 4;
+         len -= 4;
       }
-   }
-
-   if (src > dst) {
-      for (i = 0; i < len; i++) {
-         PROF_EVENT(52, "MC_(copy_address_range_state)(loop)");
+      /* fixup loop */
+      while (len >= 1) {
          vabits2 = get_vabits2( src+i );
          set_vabits2( dst+i, vabits2 );
          if (VA_BITS2_PARTDEFINED == vabits2) {
             set_sec_vbits8( dst+i, get_sec_vbits8( src+i ) );
          }
+         i++;
+         len--;
+      }
+
+   } else {
+
+      /* We have to do things the slow way */
+      if (src < dst) {
+         for (i = 0, j = len-1; i < len; i++, j--) {
+            PROF_EVENT(51, "MC_(copy_address_range_state)(loop)");
+            vabits2 = get_vabits2( src+j );
+            set_vabits2( dst+j, vabits2 );
+            if (VA_BITS2_PARTDEFINED == vabits2) {
+               set_sec_vbits8( dst+j, get_sec_vbits8( src+j ) );
+            }
+         }
+      }
+
+      if (src > dst) {
+         for (i = 0; i < len; i++) {
+            PROF_EVENT(52, "MC_(copy_address_range_state)(loop)");
+            vabits2 = get_vabits2( src+i );
+            set_vabits2( dst+i, vabits2 );
+            if (VA_BITS2_PARTDEFINED == vabits2) {
+               set_sec_vbits8( dst+i, get_sec_vbits8( src+i ) );
+            }
+         }
       }
    }
+
 }
 
 
@@ -4422,6 +4490,3 @@ VG_DETERMINE_INTERFACE_VERSION(mc_pre_clo_init)
 /*--------------------------------------------------------------------*/
 /*--- end                                                          ---*/
 /*--------------------------------------------------------------------*/
-
-
-