Drop the "?" from "copy_history32?".

This makes the caller, not the implementation, responsible for detecting
short reads and writes. This (the lack of "?") is consistent with the
slice.copy_from method.

name                           old speed     new speed     delta
puffs_flate_decode_1k/clang    117MB/s ± 6%  120MB/s ± 0%   +2.31%  (p=0.016 n=5+5)
puffs_flate_decode_10k/clang   153MB/s ± 1%  155MB/s ± 0%   +1.16%  (p=0.008 n=5+5)
puffs_flate_decode_100k/clang  168MB/s ± 1%  177MB/s ± 0%   +5.41%  (p=0.008 n=5+5)
puffs_flate_decode_1k/gcc      127MB/s ± 1%  143MB/s ± 0%  +12.15%  (p=0.008 n=5+5)
puffs_flate_decode_10k/gcc     145MB/s ± 0%  180MB/s ± 1%  +24.00%  (p=0.008 n=5+5)
puffs_flate_decode_100k/gcc    163MB/s ± 0%  212MB/s ± 1%  +29.90%  (p=0.008 n=5+5)
This commit is contained in:
Nigel Tao committed 2017-10-06 17:02:09 +11:00
1 parent a1271222a4
commit 2e883fcc23
11 files changed
+155 -116

No files matched your search

+35
View File
@@ -132,4 +132,39 @@ static inline uint64_t puffs_base_slice_u8_copy_from(puffs_base_slice_u8 dst,
return n;
}
static inline uint32_t puffs_base_writer1_copy_history32(uint8_t** ptr_ptr,
uint8_t* start,
uint8_t* end,
uint32_t distance,
uint32_t length) {
uint8_t* ptr = *ptr_ptr;
size_t d = ptr - start;
if ((d == 0) || (d < (size_t)(distance))) {
return 0;
}
start = ptr - distance;
size_t l = end - ptr;
if ((size_t)(length) > l) {
length = l;
} else {
l = length;
}
// TODO: is manual unrolling actually helpful?
for (; l >= 8; l -= 8) {
*ptr++ = *start++;
*ptr++ = *start++;
*ptr++ = *start++;
*ptr++ = *start++;
*ptr++ = *start++;
*ptr++ = *start++;
*ptr++ = *start++;
*ptr++ = *start++;
}
for (; l; l--) {
*ptr++ = *start++;
}
*ptr_ptr = ptr;
return length;
}
#endif // PUFFS_BASE_IMPL_H
+2 -1
View File
@@ -16,7 +16,8 @@ const baseImpl = "" +
"fine PUFFS_UNLIKELY(expr) (expr)\n#endif\n\n// ---------------- Static Inline Functions\n//\n// The helpers below are functions, instead of macros, because their arguments\n// can be an expression that we shouldn't evaluate more than once.\n//\n// They are in base-impl.h and hence copy/pasted into every generated C file,\n// instead of being in some \"base.c\" file, since a design goal is that users of\n// the generated C code can often just #include a single .c file, such as\n// \"gif.c\", without having to additionally include or otherwise build and link\n// a \"base.c\" file.\n//\n// They are static, so that linking multiple puffs .o files won't complain about\n// duplicate function definitions.\n//\n// They are explicitly marked inline, even if modern compilers don't use the\n// inline attribute to guide optimizations such as inlining, to avoid the\n// -Wunused-function warning, and we like to compile with -Wall -Werror.\n\nstatic inline uint16_t puffs_base_load_u16be(uint8_t* p) {\n return ((uint16_t)(p[0]) << 8) | ((uint16_t)(p[1" +
"]) << 0);\n}\n\nstatic inline uint16_t puffs_base_load_u16le(uint8_t* p) {\n return ((uint16_t)(p[0]) << 0) | ((uint16_t)(p[1]) << 8);\n}\n\nstatic inline uint32_t puffs_base_load_u32be(uint8_t* p) {\n return ((uint32_t)(p[0]) << 24) | ((uint32_t)(p[1]) << 16) |\n ((uint32_t)(p[2]) << 8) | ((uint32_t)(p[3]) << 0);\n}\n\nstatic inline uint32_t puffs_base_load_u32le(uint8_t* p) {\n return ((uint32_t)(p[0]) << 0) | ((uint32_t)(p[1]) << 8) |\n ((uint32_t)(p[2]) << 16) | ((uint32_t)(p[3]) << 24);\n}\n\nstatic inline puffs_base_slice_u8 puffs_base_slice_u8_subslice_i(\n puffs_base_slice_u8 s,\n uint64_t i) {\n if ((i <= SIZE_MAX) && (i <= s.len)) {\n return ((puffs_base_slice_u8){\n .ptr = s.ptr + i,\n .len = s.len - i,\n });\n }\n return ((puffs_base_slice_u8){});\n}\n\nstatic inline puffs_base_slice_u8 puffs_base_slice_u8_subslice_j(\n puffs_base_slice_u8 s,\n uint64_t j) {\n if ((j <= SIZE_MAX) && (j <= s.len)) {\n return ((puffs_base_slice_u8){.ptr = s.ptr, .len = j});\n }\n return ((pu" +
"ffs_base_slice_u8){});\n}\n\nstatic inline puffs_base_slice_u8\npuffs_base_slice_u8_subslice_ij(puffs_base_slice_u8 s, uint64_t i, uint64_t j) {\n if ((i <= j) && (j <= SIZE_MAX) && (j <= s.len)) {\n return ((puffs_base_slice_u8){\n .ptr = s.ptr + i,\n .len = j - i,\n });\n }\n return ((puffs_base_slice_u8){});\n}\n\n// puffs_base_slice_u8_prefix returns up to the first n bytes of s.\nstatic inline puffs_base_slice_u8 puffs_base_slice_u8_prefix(\n puffs_base_slice_u8 s,\n uint64_t n) {\n if ((uint64_t)(s.len) > n) {\n s.len = n;\n }\n return s;\n}\n\n// puffs_base_slice_u8_suffix returns up to the last n bytes of s.\nstatic inline puffs_base_slice_u8 puffs_base_slice_u8_suffix(\n puffs_base_slice_u8 s,\n uint64_t n) {\n if ((uint64_t)(s.len) > n) {\n s.ptr += (uint64_t)(s.len) - n;\n s.len = n;\n }\n return s;\n}\n\n// puffs_base_slice_u8_copy_from calls memmove(dst.ptr, src.ptr, n) where n is\n// the minimum of dst.len and src.len.\n//\n// Passing a puffs_base_slice_u8 with all fields NULL or zer" +
"o (a valid, empty\n// slice) is valid and results in a no-op.\nstatic inline uint64_t puffs_base_slice_u8_copy_from(puffs_base_slice_u8 dst,\n puffs_base_slice_u8 src) {\n size_t n = dst.len < src.len ? dst.len : src.len;\n if (n > 0) {\n memmove(dst.ptr, src.ptr, n);\n }\n return n;\n}\n\n#endif // PUFFS_BASE_IMPL_H\n" +
"o (a valid, empty\n// slice) is valid and results in a no-op.\nstatic inline uint64_t puffs_base_slice_u8_copy_from(puffs_base_slice_u8 dst,\n puffs_base_slice_u8 src) {\n size_t n = dst.len < src.len ? dst.len : src.len;\n if (n > 0) {\n memmove(dst.ptr, src.ptr, n);\n }\n return n;\n}\n\nstatic inline uint32_t puffs_base_writer1_copy_history32(uint8_t** ptr_ptr,\n uint8_t* start,\n uint8_t* end,\n uint32_t distance,\n uint32_t length) {\n uint8_t* ptr = *ptr_ptr;\n size_t d = ptr - start;\n if ((d == 0) || (d < (size_t)(distance))) {\n return 0;\n }\n start = ptr - distance;\n size_t l = end - ptr;\n if ((size_t)(length) > l) {\n length = l;\n } else {\n l = length;\n }\n // TODO: is manual unrolling actually helpful?\n for (; l >= 8; l -" +
"= 8) {\n *ptr++ = *start++;\n *ptr++ = *start++;\n *ptr++ = *start++;\n *ptr++ = *start++;\n *ptr++ = *start++;\n *ptr++ = *start++;\n *ptr++ = *start++;\n *ptr++ = *start++;\n }\n for (; l; l--) {\n *ptr++ = *start++;\n }\n *ptr_ptr = ptr;\n return length;\n}\n\n#endif // PUFFS_BASE_IMPL_H\n" +
""
type template_args_short_read struct {
+12
View File
@@ -148,6 +148,18 @@ func (g *gen) writeExprOther(b *buffer, n *a.Expr, rp replacementPolicy, pp pare
b.writes(")")
return nil
}
if isInDst(g.tm, n, t.KeyCopyHistory32, 2) {
b.printf("puffs_base_writer1_copy_history32(&%swptr_dst, %swstart_dst, %swend_dst",
bPrefix, bPrefix, bPrefix)
for _, o := range n.Args() {
b.writeb(',')
if err := g.writeExpr(b, o.Arg().Value(), rp, parenthesesOptional, depth); err != nil {
return err
}
}
b.writeb(')')
return nil
}
if isInDst(g.tm, n, t.KeySlice, 0) {
b.printf("((puffs_base_slice_u8){ .ptr = %swstart_dst, .len = %swptr_dst - %swstart_dst, })",
bPrefix, bPrefix, bPrefix)
+1 -67
View File
@@ -417,72 +417,6 @@ func (g *gen) writeCallSuspendibles(b *buffer, n *a.Expr, depth uint32) error {
b.writes("}\n")
} else if isInDst(g.tm, n, t.KeyCopyHistory32, 2) {
if g.currFunk.tempW > maxTemp-2 {
return fmt.Errorf("too many temporary variables required")
}
temp0 := g.currFunk.tempW + 0
temp1 := g.currFunk.tempW + 1
temp2 := g.currFunk.tempW + 2
g.currFunk.tempW += 3
g.currFunk.tempR += 3
b.writes("{\n")
g.currFunk.usesScratch = true
// TODO: don't hard-code [0], and allow recursive coroutines.
scratchName := fmt.Sprintf("self->private_impl.%s%s[0].scratch",
cPrefix, g.currFunk.astFunc.Name().String(g.tm))
b.printf("%s = ((uint64_t)(", scratchName)
x0 := n.Args()[0].Arg().Value()
if err := g.writeExpr(b, x0, replaceCallSuspendibles, parenthesesMandatory, depth); err != nil {
return err
}
b.writes(")<<32) | (uint64_t)(")
x1 := n.Args()[1].Arg().Value()
if err := g.writeExpr(b, x1, replaceCallSuspendibles, parenthesesMandatory, depth); err != nil {
return err
}
b.writes(");\n")
if err := g.writeCoroSuspPoint(b); err != nil {
return err
}
const wName = "dst"
b.printf("size_t %s%d = (size_t)(%s >> 32);", tPrefix, temp0, scratchName)
// TODO: it's not a BAD_ARGUMENT if we can copy from the sliding window.
b.printf("if (PUFFS_UNLIKELY((%s%d == 0) || (%s%d > (%swptr_%s - %swstart_%s)))) { "+
"status = %sERROR_BAD_ARGUMENT; goto exit; }\n",
tPrefix, temp0, tPrefix, temp0, bPrefix, wName, bPrefix, wName, g.PKGPREFIX)
b.printf("uint8_t* %s%d = %swptr_%s - %s%d;\n", tPrefix, temp1, bPrefix, wName, tPrefix, temp0)
b.printf("uint32_t %s%d = (uint32_t)(%s);", tPrefix, temp2, scratchName)
b.printf("if (PUFFS_LIKELY((size_t)(%s%d) <= (%swend_%s - %swptr_%s))) {",
tPrefix, temp2, bPrefix, wName, bPrefix, wName)
b.printf("for (; %s%d >= 8; %s%d -= 8) {", tPrefix, temp2, tPrefix, temp2)
for i := 0; i < 8; i++ {
b.printf("*%swptr_%s++ = *%s%d++;\n", bPrefix, wName, tPrefix, temp1)
}
b.writes("}\n")
b.printf("for (; %s%d; %s%d--) { *%swptr_%s++ = *%s%d++; }\n",
tPrefix, temp2, tPrefix, temp2, bPrefix, wName, tPrefix, temp1)
b.writes("} else {\n")
b.printf("%s%d = (uint32_t)(%swend_%s - %swptr_%s);\n",
tPrefix, temp2, bPrefix, wName, bPrefix, wName)
b.printf("%s -= (uint64_t)(%s%d);\n", scratchName, tPrefix, temp2)
b.printf("for (; %s%d; %s%d--) { *%swptr_%s++ = *%s%d++; }\n",
tPrefix, temp2, tPrefix, temp2, bPrefix, wName, tPrefix, temp1)
// TODO: SHORT_WRITE vs LIMITED_WRITE?
b.printf("status = %sSUSPENSION_SHORT_WRITE; goto suspend;\n", g.PKGPREFIX)
b.writes("}\n")
b.writes("}\n")
} else if isThisMethod(g.tm, n, "decode_header", 1) {
b.printf("status = %s%s_decode_header(self, %ssrc);\n",
g.pkgPrefix, g.currFunk.astFunc.Receiver().String(g.tm), aPrefix)
@@ -705,7 +639,7 @@ func isInSrc(tm *t.Map, n *a.Expr, methodName t.Key, nArgs int) bool {
}
func isInDst(tm *t.Map, n *a.Expr, methodName t.Key, nArgs int) bool {
callSuspendible := methodName != t.KeySlice
callSuspendible := methodName != t.KeyCopyHistory32 && methodName != t.KeySlice
// TODO: check that n.Args() is "(x:bar)".
if n.ID0().Key() != t.KeyOpenParen || n.CallSuspendible() != callSuspendible || len(n.Args()) != nArgs {
return false