compiler: intrinsify some math/bits functions
Let the Go frontend recognize some math/bits functions and turn
them into intrinsics.
Corresponding change in the GCC backend:
Index: gcc/go/go-gcc.cc
===================================================================
--- gcc/go/go-gcc.cc (revision 272290)
+++ gcc/go/go-gcc.cc (working copy)
@@ -613,7 +613,7 @@
NULL_TREE),
false, false);
- // Used by runtime/internal/sys.
+ // Used by runtime/internal/sys and math/bits.
this->define_builtin(BUILT_IN_CTZ, "__builtin_ctz", "ctz",
build_function_type_list(integer_type_node,
unsigned_type_node,
@@ -624,6 +624,31 @@
long_long_unsigned_type_node,
NULL_TREE),
true, false);
+ this->define_builtin(BUILT_IN_CLZ, "__builtin_clz", "clz",
+ build_function_type_list(integer_type_node,
+ unsigned_type_node,
+ NULL_TREE),
+ true, false);
+ this->define_builtin(BUILT_IN_CLZLL, "__builtin_clzll", "clzll",
+ build_function_type_list(integer_type_node,
+ long_long_unsigned_type_node,
+ NULL_TREE),
+ true, false);
+ this->define_builtin(BUILT_IN_POPCOUNT, "__builtin_popcount", "popcount",
+ build_function_type_list(integer_type_node,
+ unsigned_type_node,
+ NULL_TREE),
+ true, false);
+ this->define_builtin(BUILT_IN_POPCOUNTLL, "__builtin_popcountll", "popcountll",
+ build_function_type_list(integer_type_node,
+ long_long_unsigned_type_node,
+ NULL_TREE),
+ true, false);
+ this->define_builtin(BUILT_IN_BSWAP16, "__builtin_bswap16", "bswap16",
+ build_function_type_list(uint16_type_node,
+ uint16_type_node,
+ NULL_TREE),
+ true, false);
this->define_builtin(BUILT_IN_BSWAP32, "__builtin_bswap32", "bswap32",
build_function_type_list(uint32_type_node,
uint32_type_node,
Change-Id: If44b18882c885f1c93d4682890f176b89cbc5b1a
Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/183266
Reviewed-by: Ian Lance Taylor <iant@golang.org>
diff --git a/go/expressions.cc b/go/expressions.cc
index b9cf0f3..04aacf8 100644
--- a/go/expressions.cc
+++ b/go/expressions.cc
@@ -11063,6 +11063,25 @@
package = "runtime/internal/atomic";
}
+ if (package == "runtime/internal/sys")
+ {
+ // runtime/internal/sys functions and math/bits functions
+ // are very similar. In order not to duplicate code, we just
+ // redirect to the latter and let the code below to handle them.
+ if (name == "Bswap32")
+ name = "ReverseBytes32";
+ else if (name == "Bswap64")
+ name = "ReverseBytes64";
+ else if (name == "Ctz32")
+ name = "TrailingZeros32";
+ else if (name == "Ctz64")
+ name = "TrailingZeros64";
+ else
+ return NULL;
+
+ package = "math/bits";
+ }
+
if (package == "runtime")
{
// Handle a couple of special runtime functions. In the runtime
@@ -11093,21 +11112,44 @@
return Expression::make_cast(uintptr_type, call, loc);
}
}
- else if (package == "runtime/internal/sys")
+ else if (package == "math/bits")
{
- if (name == "Bswap32"
+ if ((name == "ReverseBytes16" || name == "ReverseBytes32"
+ || name == "ReverseBytes64" || name == "ReverseBytes")
&& this->args_ != NULL && this->args_->size() == 1)
{
+ Runtime::Function code;
+ if (name == "ReverseBytes16")
+ code = Runtime::BUILTIN_BSWAP16;
+ else if (name == "ReverseBytes32")
+ code = Runtime::BUILTIN_BSWAP32;
+ else if (name == "ReverseBytes64")
+ code = Runtime::BUILTIN_BSWAP64;
+ else if (name == "ReverseBytes")
+ code = (int_size == 8 ? Runtime::BUILTIN_BSWAP64 : Runtime::BUILTIN_BSWAP32);
+ else
+ go_unreachable();
Expression* arg = this->args_->front();
- return Runtime::make_call(Runtime::BUILTIN_BSWAP32, loc, 1, arg);
+ Expression* call = Runtime::make_call(code, loc, 1, arg);
+ if (name == "ReverseBytes")
+ return Expression::make_cast(uint_type, call, loc);
+ return call;
}
- else if (name == "Bswap64"
+ else if ((name == "TrailingZeros8" || name == "TrailingZeros16")
&& this->args_ != NULL && this->args_->size() == 1)
{
+ // GCC does not have a ctz8 or ctz16 intrinsic. We do
+ // ctz32(0x100 | arg) or ctz32(0x10000 | arg).
Expression* arg = this->args_->front();
- return Runtime::make_call(Runtime::BUILTIN_BSWAP64, loc, 1, arg);
+ arg = Expression::make_cast(uint32_type, arg, loc);
+ unsigned long mask = (name == "TrailingZeros8" ? 0x100 : 0x10000);
+ Expression* c = Expression::make_integer_ul(mask, uint32_type, loc);
+ arg = Expression::make_binary(OPERATOR_OR, arg, c, loc);
+ Expression* call = Runtime::make_call(Runtime::BUILTIN_CTZ, loc, 1, arg);
+ return Expression::make_cast(int_type, call, loc);
}
- else if (name == "Ctz32"
+ else if ((name == "TrailingZeros32"
+ || (name == "TrailingZeros" && int_size == 4))
&& this->args_ != NULL && this->args_->size() == 1)
{
Expression* arg = this->args_->front();
@@ -11125,7 +11167,8 @@
call = Expression::make_cast(int_type, call, loc);
return Expression::make_conditional(cmp, c32, call, loc);
}
- else if (name == "Ctz64"
+ else if ((name == "TrailingZeros64"
+ || (name == "TrailingZeros" && int_size == 8))
&& this->args_ != NULL && this->args_->size() == 1)
{
Expression* arg = this->args_->front();
@@ -11143,6 +11186,99 @@
call = Expression::make_cast(int_type, call, loc);
return Expression::make_conditional(cmp, c64, call, loc);
}
+ else if ((name == "LeadingZeros8" || name == "LeadingZeros16"
+ || name == "Len8" || name == "Len16")
+ && this->args_ != NULL && this->args_->size() == 1)
+ {
+ // GCC does not have a clz8 ir clz16 intrinsic. We do
+ // clz32(arg<<24 | 0xffffff) or clz32(arg<<16 | 0xffff).
+ Expression* arg = this->args_->front();
+ arg = Expression::make_cast(uint32_type, arg, loc);
+ unsigned long shift =
+ ((name == "LeadingZeros8" || name == "Len8") ? 24 : 16);
+ Expression* c = Expression::make_integer_ul(shift, uint32_type, loc);
+ arg = Expression::make_binary(OPERATOR_LSHIFT, arg, c, loc);
+ unsigned long mask =
+ ((name == "LeadingZeros8" || name == "Len8") ? 0xffffff : 0xffff);
+ c = Expression::make_integer_ul(mask, uint32_type, loc);
+ arg = Expression::make_binary(OPERATOR_OR, arg, c, loc);
+ Expression* call = Runtime::make_call(Runtime::BUILTIN_CLZ, loc, 1, arg);
+ call = Expression::make_cast(int_type, call, loc);
+ // len = width - clz
+ if (name == "Len8")
+ {
+ c = Expression::make_integer_ul(8, int_type, loc);
+ return Expression::make_binary(OPERATOR_MINUS, c, call, loc);
+ }
+ else if (name == "Len16")
+ {
+ c = Expression::make_integer_ul(16, int_type, loc);
+ return Expression::make_binary(OPERATOR_MINUS, c, call, loc);
+ }
+ return call;
+ }
+ else if ((name == "LeadingZeros32" || name == "Len32"
+ || ((name == "LeadingZeros" || name == "Len") && int_size == 4))
+ && this->args_ != NULL && this->args_->size() == 1)
+ {
+ Expression* arg = this->args_->front();
+ if (!arg->is_variable())
+ {
+ Temporary_statement* ts = Statement::make_temporary(uint32_type, arg, loc);
+ inserter->insert(ts);
+ arg = Expression::make_temporary_reference(ts, loc);
+ }
+ // arg == 0 ? 32 : __builtin_clz(arg)
+ Expression* zero = Expression::make_integer_ul(0, uint32_type, loc);
+ Expression* cmp = Expression::make_binary(OPERATOR_EQEQ, arg, zero, loc);
+ Expression* c32 = Expression::make_integer_ul(32, int_type, loc);
+ Expression* call = Runtime::make_call(Runtime::BUILTIN_CLZ, loc, 1, arg->copy());
+ call = Expression::make_cast(int_type, call, loc);
+ Expression* cond = Expression::make_conditional(cmp, c32, call, loc);
+ // len = 32 - clz
+ if (name == "Len32" || name == "Len")
+ return Expression::make_binary(OPERATOR_MINUS, c32->copy(), cond, loc);
+ return cond;
+ }
+ else if ((name == "LeadingZeros64" || name == "Len64"
+ || ((name == "LeadingZeros" || name == "Len") && int_size == 8))
+ && this->args_ != NULL && this->args_->size() == 1)
+ {
+ Expression* arg = this->args_->front();
+ if (!arg->is_variable())
+ {
+ Temporary_statement* ts = Statement::make_temporary(uint64_type, arg, loc);
+ inserter->insert(ts);
+ arg = Expression::make_temporary_reference(ts, loc);
+ }
+ // arg == 0 ? 64 : __builtin_clzll(arg)
+ Expression* zero = Expression::make_integer_ul(0, uint64_type, loc);
+ Expression* cmp = Expression::make_binary(OPERATOR_EQEQ, arg, zero, loc);
+ Expression* c64 = Expression::make_integer_ul(64, int_type, loc);
+ Expression* call = Runtime::make_call(Runtime::BUILTIN_CLZLL, loc, 1, arg->copy());
+ call = Expression::make_cast(int_type, call, loc);
+ Expression* cond = Expression::make_conditional(cmp, c64, call, loc);
+ // len = 64 - clz
+ if (name == "Len64" || name == "Len")
+ return Expression::make_binary(OPERATOR_MINUS, c64->copy(), cond, loc);
+ return cond;
+ }
+ else if ((name == "OnesCount8" || name == "OnesCount16"
+ || name == "OnesCount32" || name == "OnesCount64"
+ || name == "OnesCount")
+ && this->args_ != NULL && this->args_->size() == 1)
+ {
+ Runtime::Function code;
+ if (name == "OnesCount64")
+ code = Runtime::BUILTIN_POPCOUNTLL;
+ else if (name == "OnesCount")
+ code = (int_size == 8 ? Runtime::BUILTIN_POPCOUNTLL : Runtime::BUILTIN_POPCOUNT);
+ else
+ code = Runtime::BUILTIN_POPCOUNT;
+ Expression* arg = this->args_->front();
+ Expression* call = Runtime::make_call(code, loc, 1, arg);
+ return Expression::make_cast(int_type, call, loc);
+ }
}
else if (package == "runtime/internal/atomic")
{
diff --git a/go/runtime.cc b/go/runtime.cc
index 28aca44..e35658b 100644
--- a/go/runtime.cc
+++ b/go/runtime.cc
@@ -32,6 +32,8 @@
RFT_INT,
// Go type uint8, C type uint8_t.
RFT_UINT8,
+ // Go type uint16, C type uint16_t.
+ RFT_UINT16,
// Go type int32, C type int32_t.
RFT_INT32,
// Go type uint32, C type uint32_t.
@@ -115,6 +117,10 @@
t = Type::lookup_integer_type("uint8");
break;
+ case RFT_UINT16:
+ t = Type::lookup_integer_type("uint16");
+ break;
+
case RFT_INT32:
t = Type::lookup_integer_type("int32");
break;
@@ -257,6 +263,7 @@
case RFT_BOOLPTR:
case RFT_INT:
case RFT_UINT8:
+ case RFT_UINT16:
case RFT_INT32:
case RFT_UINT32:
case RFT_INT64:
diff --git a/go/runtime.def b/go/runtime.def
index a966cd4..27d5e47 100644
--- a/go/runtime.def
+++ b/go/runtime.def
@@ -428,6 +428,8 @@
R1(POINTER))
// Swap bytes.
+DEF_GO_RUNTIME(BUILTIN_BSWAP16, "__builtin_bswap16", P1(UINT16),
+ R1(UINT16))
DEF_GO_RUNTIME(BUILTIN_BSWAP32, "__builtin_bswap32", P1(UINT32),
R1(UINT32))
DEF_GO_RUNTIME(BUILTIN_BSWAP64, "__builtin_bswap64", P1(UINT64),
@@ -437,6 +439,14 @@
DEF_GO_RUNTIME(BUILTIN_CTZ, "__builtin_ctz", P1(UINT32), R1(INT32))
DEF_GO_RUNTIME(BUILTIN_CTZLL, "__builtin_ctzll", P1(UINT64), R1(INT32))
+// Count leading zeros.
+DEF_GO_RUNTIME(BUILTIN_CLZ, "__builtin_clz", P1(UINT32), R1(INT32))
+DEF_GO_RUNTIME(BUILTIN_CLZLL, "__builtin_clzll", P1(UINT64), R1(INT32))
+
+// Count one bits.
+DEF_GO_RUNTIME(BUILTIN_POPCOUNT, "__builtin_popcount", P1(UINT32), R1(INT32))
+DEF_GO_RUNTIME(BUILTIN_POPCOUNTLL, "__builtin_popcountll", P1(UINT64), R1(INT32))
+
// Atomics.
DEF_GO_RUNTIME(ATOMIC_LOAD_4, "__atomic_load_4", P2(POINTER, INT32),
R1(UINT32))