On 05/12/15 11:02 PM, Keith Packard wrote:
walter harms <[email protected]> writes:
Just for my curiosity .. is there any benchmark that shows that this
actually improves something
Yeah, I was measuring the new text code during development and saw some
(tiny) improvements -- it's got a ton of branches in an inner loop
loading glyphs pointers to hardware.
Speaking of which, something I've wondered about for a while, but never
had time to test, is if telling the server to optimize for valid requests
would make a difference, such as:
--- a/include/dix.h
+++ b/include/dix.h
@@ -67,25 +67,25 @@ SOFTWARE.
#define ARRAY_SIZE(a) (sizeof((a)) / sizeof((a)[0]))
#define REQUEST_SIZE_MATCH(req)\
- if ((sizeof(req) >> 2) != client->req_len)\
+ if (_X_UNLIKELY((sizeof(req) >> 2) != client->req_len)) \
return(BadLength)
#define REQUEST_AT_LEAST_SIZE(req) \
- if ((sizeof(req) >> 2) > client->req_len )\
+ if (_X_UNLIKELY((sizeof(req) >> 2) > client->req_len )) \
return(BadLength)
#define REQUEST_AT_LEAST_EXTRA_SIZE(req, extra) \
- if (((sizeof(req) + ((uint64_t) extra)) >> 2) > client->req_len ) \
+ if (_X_UNLIKELY(((sizeof(req) + ((uint64_t) extra)) >> 2) > client->req_len
return(BadLength)
#define REQUEST_FIXED_SIZE(req, n)\
- if (((sizeof(req) >> 2) > client->req_len) || \
- (((n) >> 2) >= client->req_len) || \
- ((((uint64_t) sizeof(req) + (n) + 3) >> 2) != (uint64_t) client->req_le
+ if (_X_UNLIKELY(((sizeof(req) >> 2) > client->req_len) || \
+ (((n) >> 2) >= client->req_len) || \
+ ((((uint64_t) sizeof(req) + (n) + 3) >> 2) != (uint64_t) cl
return(BadLength)
#define LEGAL_NEW_RESOURCE(id,client)\
- if (!LegalNewID(id,client)) \
+ if (_X_UNLIKELY(!LegalNewID(id,client))) \
{\
client->errorValue = id;\
return BadIDChoice;\
@@ -94,12 +94,13 @@ SOFTWARE.
#define VALIDATE_DRAWABLE_AND_GC(drawID, pDraw, mode)\
{\
int tmprc = dixLookupDrawable(&(pDraw), drawID, client, M_ANY, mode);\
- if (tmprc != Success)\
+ if (_X_UNLIKELY(tmprc != Success)) \
return tmprc;\
tmprc = dixLookupGC(&(pGC), stuff->gc, client, DixUseAccess);\
- if (tmprc != Success)\
+ if (_X_UNLIKELY(tmprc != Success))\
return tmprc;\
- if ((pGC->depth != pDraw->depth) || (pGC->pScreen != pDraw->pScreen))\
+ if (_X_UNLIKELY((pGC->depth != pDraw->depth) || \
+ (pGC->pScreen != pDraw->pScreen))) \
return BadMatch;\
}\
if (pGC->serialNumber != pDraw->serialNumber)\
Or optimizing for same-endian requests:
#define WriteReplyToClient(pClient, size, pReply) { \
- if ((pClient)->swapped) \
+ if (_X_UNLIKELY((pClient)->swapped)) \
(*ReplySwapVector[((xReq *)(pClient)->requestBuffer)->reqType]) \
(pClient, (int)(size), pReply); \
else WriteToClient(pClient, (int)(size), (pReply)); }
#define WriteSwappedDataToClient(pClient, size, pbuf) \
- if ((pClient)->swapped) \
+ if (_X_UNLIKELY((pClient)->swapped)) \
(*(pClient)->pSwapReplyFunc)(pClient, (int)(size), pbuf); \
else WriteToClient(pClient, (int)(size), (pbuf));
If anyone wants to try it out and report back, please do so!
--
-Alan Coopersmith- [email protected]
Oracle Solaris Engineering - http://blogs.oracle.com/alanc
_______________________________________________
[email protected]: X.Org development
Archives: http://lists.x.org/archives/xorg-devel
Info: http://lists.x.org/mailman/listinfo/xorg-devel
