Reimplement h_generic_calc_GetMSBs8x16 to be more efficient.

h_generic_calc_GetMSBs8x16 concatenates the top bit of each 8-bit lane in a
128-bit value, producing a 16-bit scalar value.  (It is PMOVMSKB, really).
The existing implementation is excessively inefficient and shows up sometimes
in 'perf' profiles of V.  This commit replaces it with a logarithmic (4-stage)
algorithm which is hopefully much faster.

diff --git a/VEX/priv/host_generic_simd128.c b/VEX/priv/host_generic_simd128.c
index 1c0f7cfbafcea544548d9dc578b822e5f13efd0c..f895de46f486f6e5dfa34a03bb5db0f9193891b7 100644 (file)
--- a/VEX/priv/host_generic_simd128.c
+++ b/VEX/priv/host_generic_simd128.c
@@ -383,23 +383,20 @@ void VEX_REGPARM(3)
 UInt /*not-regparm*/
      h_generic_calc_GetMSBs8x16 ( ULong w64hi, ULong w64lo )
 {
-   UInt r = 0;
-   if (w64hi & (1ULL << (64-1))) r |= (1<<15);
-   if (w64hi & (1ULL << (56-1))) r |= (1<<14);
-   if (w64hi & (1ULL << (48-1))) r |= (1<<13);
-   if (w64hi & (1ULL << (40-1))) r |= (1<<12);
-   if (w64hi & (1ULL << (32-1))) r |= (1<<11);
-   if (w64hi & (1ULL << (24-1))) r |= (1<<10);
-   if (w64hi & (1ULL << (16-1))) r |= (1<<9);
-   if (w64hi & (1ULL << ( 8-1))) r |= (1<<8);
-   if (w64lo & (1ULL << (64-1))) r |= (1<<7);
-   if (w64lo & (1ULL << (56-1))) r |= (1<<6);
-   if (w64lo & (1ULL << (48-1))) r |= (1<<5);
-   if (w64lo & (1ULL << (40-1))) r |= (1<<4);
-   if (w64lo & (1ULL << (32-1))) r |= (1<<3);
-   if (w64lo & (1ULL << (24-1))) r |= (1<<2);
-   if (w64lo & (1ULL << (16-1))) r |= (1<<1);
-   if (w64lo & (1ULL << ( 8-1))) r |= (1<<0);
+   /* Some serious bit twiddling going on here.  Mostly we can do it in
+      parallel for the upper and lower 64 bits, assuming the processor offers
+      a suitably high level of ILP. */
+   w64hi &= 0x8080808080808080ULL;
+   w64lo &= 0x8080808080808080ULL;
+   w64hi >>= 7;
+   w64lo >>= 7;
+   w64hi |= (w64hi >> 7);
+   w64lo |= (w64lo >> 7);
+   w64hi |= (w64hi >> 14);
+   w64lo |= (w64lo >> 14);
+   w64hi |= (w64hi >> 28);
+   w64lo |= (w64lo >> 28);
+   UInt r = ((w64hi & 0xFF) << 8) | (w64lo & 0xFF);
    return r;
 }