X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=rts%2FPrimOps.cmm;h=121102c8e17f1a97dab5e546d89820bd25832166;hb=54075ea5872c151573710d32876314e2005984ba;hp=cd9a5bf2fe755cda285311b24776bee0241f1599;hpb=19f85473301c46020caada24a07a4c8c51ec2f64;p=ghc-hetmet.git

diff --git a/rts/PrimOps.cmm b/rts/PrimOps.cmm
index cd9a5bf..121102c 100644
--- a/rts/PrimOps.cmm
+++ b/rts/PrimOps.cmm
@@ -83,25 +83,65 @@ newByteArrayzh_fast
     RET_P(p);
 }
 
+#define BA_ALIGN 16
+#define BA_MASK  (BA_ALIGN-1)
+
 newPinnedByteArrayzh_fast
 {
-    W_ words, payload_words, n, p;
+    W_ words, bytes, payload_words, p;
 
     MAYBE_GC(NO_PTRS,newPinnedByteArrayzh_fast);
-    n = R1;
-    payload_words = ROUNDUP_BYTES_TO_WDS(n);
+    bytes = R1;
+    /* payload_words is what we will tell the profiler we had to allocate */
+    payload_words = ROUNDUP_BYTES_TO_WDS(bytes);
+    /* When we actually allocate memory, we need to allow space for the
+       header: */
+    bytes = bytes + SIZEOF_StgArrWords;
+    /* And we want to align to BA_ALIGN bytes, so we need to allow space
+       to shift up to BA_ALIGN - 1 bytes: */
+    bytes = bytes + BA_ALIGN - 1;
+    /* Now we convert to a number of words: */
+    words = ROUNDUP_BYTES_TO_WDS(bytes);
+
+    ("ptr" p) = foreign "C" allocatePinned(words) [];
+    TICK_ALLOC_PRIM(SIZEOF_StgArrWords,WDS(payload_words),0);
+
+    /* Now we need to move p forward so that the payload is aligned
+       to BA_ALIGN bytes: */
+    p = p + ((-p - SIZEOF_StgArrWords) & BA_MASK);
+
+    SET_HDR(p, stg_ARR_WORDS_info, W_[CCCS]);
+    StgArrWords_words(p) = payload_words;
+    RET_P(p);
+}
+
+newAlignedPinnedByteArrayzh_fast
+{
+    W_ words, bytes, payload_words, p, alignment;
+
+    MAYBE_GC(NO_PTRS,newAlignedPinnedByteArrayzh_fast);
+    bytes = R1;
+    alignment = R2;
+
+    /* payload_words is what we will tell the profiler we had to allocate */
+    payload_words = ROUNDUP_BYTES_TO_WDS(bytes);
 
-    // We want a 16-byte aligned array.  allocatePinned() gives us
-    // 8-byte aligned memory by default, but we want to align the
-    // *goods* inside the ArrWords object, so we have to check the
-    // size of the ArrWords header and adjust our size accordingly.
-    words = payload_words + ((SIZEOF_StgArrWords + 15) & ~15);
+    /* When we actually allocate memory, we need to allow space for the
+       header: */
+    bytes = bytes + SIZEOF_StgArrWords;
+    /* And we want to align to <alignment> bytes, so we need to allow space
+       to shift up to <alignment - 1> bytes: */
+    bytes = bytes + alignment - 1;
+    /* Now we convert to a number of words: */
+    words = ROUNDUP_BYTES_TO_WDS(bytes);
 
     ("ptr" p) = foreign "C" allocatePinned(words) [];
     TICK_ALLOC_PRIM(SIZEOF_StgArrWords,WDS(payload_words),0);
 
-    // Push the pointer forward so that the goods fall on a 16-byte boundary.
-    p = p + ((p + SIZEOF_StgArrWords) & 15);
+    /* Now we need to move p forward so that the payload is aligned
+       to <alignment> bytes. Note that we are assuming that
+       <alignment> is a power of 2, which is technically not guaranteed */
+    p = p + ((-p - SIZEOF_StgArrWords) & (alignment - 1));
 
     SET_HDR(p, stg_ARR_WORDS_info, W_[CCCS]);
     StgArrWords_words(p) = payload_words;
@@ -1033,7 +1073,8 @@ forkzh_fast
 
   foreign "C" scheduleThread(MyCapability() "ptr", threadid "ptr") [];
 
-  // switch at the earliest opportunity
+  // context switch soon, but not immediately: we don't want every
+  // forkIO to force a context-switch.
   Capability_context_switch(MyCapability()) = 1 :: CInt;
   
   RET_P(threadid);
@@ -1062,7 +1103,8 @@ forkOnzh_fast
 
   foreign "C" scheduleThreadOn(MyCapability() "ptr", cpu, threadid "ptr") [];
 
-  // switch at the earliest opportunity
+  // context switch soon, but not immediately: we don't want every
+  // forkIO to force a context-switch.
   Capability_context_switch(MyCapability()) = 1 :: CInt;
   
   RET_P(threadid);