Optimize framing a little

- rely on the fact that data-to-come holds a reference to
data-being-written, so there's no need to add a ref for every frame
written
- provide an 'inlined' version of grpc_slice_malloc (via a #define) that
gives the compiler more information about small allocations to enable
better optimization

4 files changed, 107 insertions, 32 deletions

diff --git a/src/core/lib/slice/b64.c b/src/core/lib/slice/b64.c
index 2007cc4810..d9091646e0 100644
--- a/src/core/lib/slice/b64.c
+++ b/src/core/lib/slice/b64.c
@@ -202,7 +202,7 @@ static int decode_group(const unsigned char *codes, size_t num_codes,
 
 grpc_slice grpc_base64_decode_with_len(grpc_exec_ctx *exec_ctx, const char *b64,
                                        size_t b64_len, int url_safe) {
-  grpc_slice result = grpc_slice_malloc(b64_len);
+  grpc_slice result = GRPC_SLICE_MALLOC(b64_len);
   unsigned char *current = GRPC_SLICE_START_PTR(result);
   size_t result_size = 0;
   unsigned char codes[4];
diff --git a/src/core/lib/slice/percent_encoding.c b/src/core/lib/slice/percent_encoding.c
index c76c58d371..a77c69763f 100644
--- a/src/core/lib/slice/percent_encoding.c
+++ b/src/core/lib/slice/percent_encoding.c
@@ -71,7 +71,7 @@ grpc_slice grpc_percent_encode_slice(grpc_slice slice,
     return grpc_slice_ref_internal(slice);
   }
   // second pass: actually encode
-  grpc_slice out = grpc_slice_malloc(output_length);
+  grpc_slice out = GRPC_SLICE_MALLOC(output_length);
   uint8_t *q = GRPC_SLICE_START_PTR(out);
   for (p = slice_start; p < slice_end; p++) {
     if (is_unreserved_character(*p, unreserved_bytes)) {
@@ -125,7 +125,7 @@ bool grpc_strict_percent_decode_slice(grpc_slice slice_in,
     return true;
   }
   p = GRPC_SLICE_START_PTR(slice_in);
-  *slice_out = grpc_slice_malloc(out_length);
+  *slice_out = GRPC_SLICE_MALLOC(out_length);
   uint8_t *q = GRPC_SLICE_START_PTR(*slice_out);
   while (p != in_end) {
     if (*p == '%') {
@@ -163,7 +163,7 @@ grpc_slice grpc_permissive_percent_decode_slice(grpc_slice slice_in) {
     return grpc_slice_ref_internal(slice_in);
   }
   p = GRPC_SLICE_START_PTR(slice_in);
-  grpc_slice out = grpc_slice_malloc(out_length);
+  grpc_slice out = GRPC_SLICE_MALLOC(out_length);
   uint8_t *q = GRPC_SLICE_START_PTR(out);
   while (p != in_end) {
     if (*p == '%') {
diff --git a/src/core/lib/slice/slice.c b/src/core/lib/slice/slice.c
index 1cddf062cd..8896f1dfd3 100644
--- a/src/core/lib/slice/slice.c
+++ b/src/core/lib/slice/slice.c
@@ -197,7 +197,7 @@ grpc_slice grpc_slice_new_with_len(void *p, size_t len,
 }
 
 grpc_slice grpc_slice_from_copied_buffer(const char *source, size_t length) {
-  grpc_slice slice = grpc_slice_malloc(length);
+  grpc_slice slice = GRPC_SLICE_MALLOC(length);
   memcpy(GRPC_SLICE_START_PTR(slice), source, length);
   return slice;
 }
@@ -227,35 +227,42 @@ static const grpc_slice_refcount_vtable malloc_vtable = {
     malloc_ref, malloc_unref, grpc_slice_default_eq_impl,
     grpc_slice_default_hash_impl};
 
+grpc_slice grpc_slice_malloc_large(size_t length) {
+  grpc_slice slice;
+
+  /* Memory layout used by the slice created here:
+
+     +-----------+----------------------------------------------------------+
+     | refcount  | bytes                                                    |
+     +-----------+----------------------------------------------------------+
+
+     refcount is a malloc_refcount
+     bytes is an array of bytes of the requested length
+     Both parts are placed in the same allocation returned from gpr_malloc */
+  malloc_refcount *rc = gpr_malloc(sizeof(malloc_refcount) + length);
+
+  /* Initial refcount on rc is 1 - and it's up to the caller to release
+     this reference. */
+  gpr_ref_init(&rc->refs, 1);
+
+  rc->base.vtable = &malloc_vtable;
+  rc->base.sub_refcount = &rc->base;
+
+  /* Build up the slice to be returned. */
+  /* The slices refcount points back to the allocated block. */
+  slice.refcount = &rc->base;
+  /* The data bytes are placed immediately after the refcount struct */
+  slice.data.refcounted.bytes = (uint8_t *)(rc + 1);
+  /* And the length of the block is set to the requested length */
+  slice.data.refcounted.length = length;
+  return slice;
+}
+
 grpc_slice grpc_slice_malloc(size_t length) {
   grpc_slice slice;
 
   if (length > sizeof(slice.data.inlined.bytes)) {
-    /* Memory layout used by the slice created here:
-
-       +-----------+----------------------------------------------------------+
-       | refcount  | bytes                                                    |
-       +-----------+----------------------------------------------------------+
-
-       refcount is a malloc_refcount
-       bytes is an array of bytes of the requested length
-       Both parts are placed in the same allocation returned from gpr_malloc */
-    malloc_refcount *rc = gpr_malloc(sizeof(malloc_refcount) + length);
-
-    /* Initial refcount on rc is 1 - and it's up to the caller to release
-       this reference. */
-    gpr_ref_init(&rc->refs, 1);
-
-    rc->base.vtable = &malloc_vtable;
-    rc->base.sub_refcount = &rc->base;
-
-    /* Build up the slice to be returned. */
-    /* The slices refcount points back to the allocated block. */
-    slice.refcount = &rc->base;
-    /* The data bytes are placed immediately after the refcount struct */
-    slice.data.refcounted.bytes = (uint8_t *)(rc + 1);
-    /* And the length of the block is set to the requested length */
-    slice.data.refcounted.length = length;
+    return grpc_slice_malloc_large(length);
   } else {
     /* small slice: just inline the data */
     slice.refcount = NULL;
@@ -341,6 +348,40 @@ grpc_slice grpc_slice_split_tail(grpc_slice *source, size_t split) {
   return tail;
 }
 
+grpc_slice grpc_slice_split_tail_no_ref(grpc_slice *source, size_t split) {
+  grpc_slice tail;
+
+  if (source->refcount == NULL) {
+    /* inlined data, copy it out */
+    GPR_ASSERT(source->data.inlined.length >= split);
+    tail.refcount = NULL;
+    tail.data.inlined.length = (uint8_t)(source->data.inlined.length - split);
+    memcpy(tail.data.inlined.bytes, source->data.inlined.bytes + split,
+           tail.data.inlined.length);
+    source->data.inlined.length = (uint8_t)split;
+  } else {
+    size_t tail_length = source->data.refcounted.length - split;
+    GPR_ASSERT(source->data.refcounted.length >= split);
+    if (tail_length < sizeof(tail.data.inlined.bytes)) {
+      /* Copy out the bytes - it'll be cheaper than refcounting */
+      tail.refcount = NULL;
+      tail.data.inlined.length = (uint8_t)tail_length;
+      memcpy(tail.data.inlined.bytes, source->data.refcounted.bytes + split,
+             tail_length);
+    } else {
+      /* Build the result */
+      tail.refcount = &noop_refcount;
+      /* Point into the source array */
+      tail.data.refcounted.bytes = source->data.refcounted.bytes + split;
+      tail.data.refcounted.length = tail_length;
+    }
+    source->refcount = source->refcount->sub_refcount;
+    source->data.refcounted.length = split;
+  }
+
+  return tail;
+}
+
 grpc_slice grpc_slice_split_head(grpc_slice *source, size_t split) {
   grpc_slice head;
 
@@ -457,7 +498,7 @@ int grpc_slice_slice(grpc_slice haystack, grpc_slice needle) {
 }
 
 grpc_slice grpc_slice_dup(grpc_slice a) {
-  grpc_slice copy = grpc_slice_malloc(GRPC_SLICE_LENGTH(a));
+  grpc_slice copy = GRPC_SLICE_MALLOC(GRPC_SLICE_LENGTH(a));
   memcpy(GRPC_SLICE_START_PTR(copy), GRPC_SLICE_START_PTR(a),
          GRPC_SLICE_LENGTH(a));
   return copy;
diff --git a/src/core/lib/slice/slice_buffer.c b/src/core/lib/slice/slice_buffer.c
index c96b9c3b28..a13941e5ac 100644
--- a/src/core/lib/slice/slice_buffer.c
+++ b/src/core/lib/slice/slice_buffer.c
@@ -255,14 +255,47 @@ void grpc_slice_buffer_move_into(grpc_slice_buffer *src,
 
 void grpc_slice_buffer_move_first(grpc_slice_buffer *src, size_t n,
                                   grpc_slice_buffer *dst) {
+  GPR_ASSERT(src->length >= n);
+  if (src->length == n) {
+    grpc_slice_buffer_move_into(src, dst);
+    return;
+  }
+
   size_t output_len = dst->length + n;
   size_t new_input_len = src->length - n;
+
+  while (src->count > 0) {
+    grpc_slice slice = grpc_slice_buffer_take_first(src);
+    size_t slice_len = GRPC_SLICE_LENGTH(slice);
+    if (n > slice_len) {
+      grpc_slice_buffer_add(dst, slice);
+      n -= slice_len;
+    } else if (n == slice_len) {
+      grpc_slice_buffer_add(dst, slice);
+      break;
+    } else { /* n < slice_len */
+      grpc_slice_buffer_undo_take_first(src, grpc_slice_split_tail(&slice, n));
+      GPR_ASSERT(GRPC_SLICE_LENGTH(slice) == n);
+      grpc_slice_buffer_add(dst, slice);
+      break;
+    }
+  }
+  GPR_ASSERT(dst->length == output_len);
+  GPR_ASSERT(src->length == new_input_len);
+  GPR_ASSERT(src->count > 0);
+}
+
+void grpc_slice_buffer_move_first_no_ref(grpc_slice_buffer *src, size_t n,
+                                         grpc_slice_buffer *dst) {
   GPR_ASSERT(src->length >= n);
   if (src->length == n) {
     grpc_slice_buffer_move_into(src, dst);
     return;
   }
 
+  size_t output_len = dst->length + n;
+  size_t new_input_len = src->length - n;
+
   while (src->count > 0) {
     grpc_slice slice = grpc_slice_buffer_take_first(src);
     size_t slice_len = GRPC_SLICE_LENGTH(slice);
@@ -273,7 +306,8 @@ void grpc_slice_buffer_move_first(grpc_slice_buffer *src, size_t n,
       grpc_slice_buffer_add(dst, slice);
       break;
     } else { /* n < slice_len */
-      grpc_slice_buffer_undo_take_first(src, grpc_slice_split_tail(&slice, n));
+      grpc_slice_buffer_undo_take_first(
+          src, grpc_slice_split_tail_no_ref(&slice, n));
       GPR_ASSERT(GRPC_SLICE_LENGTH(slice) == n);
       grpc_slice_buffer_add(dst, slice);
       break;