# HG changeset patch # User LYH # Date 1383408847 -3600 # Node ID 1867774aaa2e775eb068d456a2e1f710766fab12 # Parent e2770dfa0bc437fe5e6776a57465c45b15fb5ab0 jit compiler: Add support for int64 type * libinterp/corefcn/pt-jit.cc: Add basic operations, casts und overloads diff -r e2770dfa0bc4 -r 1867774aaa2e libinterp/corefcn/jit-typeinfo.cc --- a/libinterp/corefcn/jit-typeinfo.cc Fri Sep 27 02:58:43 2013 +0800 +++ b/libinterp/corefcn/jit-typeinfo.cc Sat Nov 02 17:14:07 2013 +0100 @@ -70,6 +70,7 @@ #include "ov-scalar.h" #include "ov-float.h" #include "ov-int8.h" +#include "ov-int64.h" #include "ov-uint16.h" #include "pager.h" @@ -223,6 +224,20 @@ return new octave_int8_scalar (value); } +extern "C" int64_t +octave_jit_cast_int64_any (octave_base_value *obv) +{ + int64_t ret = obv->int64_scalar_value (); + obv->release (); + return ret; +} + +extern "C" octave_base_value * +octave_jit_cast_any_int64 (int64_t value) +{ + return new octave_int64_scalar (value); +} + extern "C" uint16_t octave_jit_cast_uint16_any (octave_base_value *obv) { @@ -597,6 +612,89 @@ /************************************************************ * + * int64 related external helper function + * + ************************************************************/ + +extern "C" int64_t +octave_jit_add_int64_int64 (int64_t lhs, int64_t rhs) +{ + uint64_t ulhs = lhs; + uint64_t urhs = rhs; + uint64_t res = ulhs + urhs; + + /* Calculate overflowed result. (Don't change the sign bit of ux) */ + ulhs = (ulhs >> 63) + LONG_MAX; + + /* Force compiler to use cmovns instruction */ + if ((int64_t) ((ulhs ^ urhs) | ~(urhs ^ res)) >= 0) + { + res = ulhs; + } + + return res; +} + +extern "C" int64_t +octave_jit_sub_int64_int64 (int64_t lhs, int64_t rhs) +{ + uint64_t ulhs = lhs; + uint64_t urhs = rhs; + uint64_t res = ulhs - urhs; + + ulhs = (ulhs >> 63) + LONG_MAX; + + /* Force compiler to use cmovns instruction */ + if ((int64_t)((ulhs ^ urhs) & (ulhs ^ res)) < 0) + { + res = ulhs; + } + + return res; +} + +extern "C" int64_t +octave_jit_mul_int64_int64 (int64_t lhs, int64_t rhs) +{ +#ifdef OCTAVE_INT_USE_LONG_DOUBLE + long double p = static_cast (lhs) * static_cast (rhs); + if (p > static_cast (std::numeric_limits::max ())) + { + return std::numeric_limits::max (); + } + else if (p < static_cast (std::numeric_limits::min ())) + { + return std::numeric_limits::min (); + } + else + return static_cast (p); +#else + __int128_t res = (__int128_t) lhs * (__int128_t) rhs; + uint64_t res2 = ((uint64_t) (lhs ^ rhs) >> 63) + LONG_MAX; + + int64_t hi = (res >> 64); + int64_t lo = res; + + if (hi != (lo >> 63)) res = res2; + + return res; +#endif +} + +extern "C" int64_t +octave_jit_incr_int64 (int64_t val) +{ + return octave_jit_add_int64_int64 (val, 1); +} + +extern "C" int64_t +octave_jit_decr_int64 (int64_t val) +{ + return octave_jit_sub_int64_int64 (val, 1); +} + +/************************************************************ + * * uint16 related external helper function * ************************************************************/ @@ -1268,6 +1366,7 @@ llvm::Type *scalar_t = llvm::Type::getDoubleTy (context); llvm::Type *single_t = llvm::Type::getFloatTy (context); llvm::Type *int8__t = llvm::Type::getIntNTy (context, 8); + llvm::Type *int64__t = llvm::Type::getIntNTy (context, 64); llvm::Type *uint16__t = llvm::Type::getIntNTy (context, 16); llvm::Type *bool_t = llvm::Type::getInt1Ty (context); llvm::Type *string_t = llvm::Type::getInt8Ty (context); @@ -1418,6 +1517,7 @@ grab_fn.add_overload (create_identity (scalar)); grab_fn.add_overload (create_identity (single)); grab_fn.add_overload (create_identity (intN (8))); + grab_fn.add_overload (create_identity (intN (64))); grab_fn.add_overload (create_identity (uintN (16))); grab_fn.add_overload (create_identity (scalar_ptr)); grab_fn.add_overload (create_identity (any_ptr)); @@ -1440,6 +1540,7 @@ destroy_fn.add_overload (create_identity(scalar)); destroy_fn.add_overload (create_identity(single)); destroy_fn.add_overload (create_identity(intN (8))); + destroy_fn.add_overload (create_identity(intN (64))); destroy_fn.add_overload (create_identity(uintN (16))); destroy_fn.add_overload (create_identity(boolean)); destroy_fn.add_overload (create_identity(index)); @@ -1750,6 +1851,62 @@ /************************************************************ * + * int64 related operations + * + ************************************************************/ + + // FIXME: overflow occurs at minus + minus, minus - plus + // now for binary int64 operations + fn = create_external (JIT_FN (octave_jit_add_int64_int64), intN (64), intN (64), + intN (64)); + binary_ops[octave_value::op_add].add_overload (fn); + fn = create_external (JIT_FN (octave_jit_sub_int64_int64), intN (64), intN (64), + intN (64)); + binary_ops[octave_value::op_sub].add_overload (fn); + fn = create_external (JIT_FN (octave_jit_mul_int64_int64), intN (64), intN (64), + intN (64)); + binary_ops[octave_value::op_mul].add_overload (fn); + binary_ops[octave_value::op_el_mul].add_overload (fn); + + add_binary_icmp (intN (64), octave_value::op_lt, llvm::CmpInst::ICMP_SLT); + add_binary_icmp (intN (64), octave_value::op_le, llvm::CmpInst::ICMP_SLE); + add_binary_icmp (intN (64), octave_value::op_eq, llvm::CmpInst::ICMP_EQ); + add_binary_icmp (intN (64), octave_value::op_ge, llvm::CmpInst::ICMP_SGE); + add_binary_icmp (intN (64), octave_value::op_gt, llvm::CmpInst::ICMP_SGT); + add_binary_icmp (intN (64), octave_value::op_ne, llvm::CmpInst::ICMP_NE); + + // FIXME: saturation divide definition? interpreter convert int to double, calculate and round. + // divide is annoying because it might error + // FIXME: Implement div + + // FIXME: Implement pow + + // now for unary int8 operations + // FIXME: Impelment not + fn = create_external (JIT_FN (octave_jit_incr_int64), intN (64), intN (64)); + unary_ops[octave_value::op_incr].add_overload (fn); + + fn = create_external (JIT_FN (octave_jit_decr_int64), intN (64), intN (64)); + unary_ops[octave_value::op_decr].add_overload (fn); + + fn = create_internal ("octave_jit_uminus", intN (64), intN (64)); + body = fn.new_block (); + builder.SetInsertPoint (body); + { + llvm::Value *mone = llvm::ConstantInt::get (int64__t, -1); + llvm::Value *val = fn.argument (builder, 0); + val = builder.CreateMul (val, mone); + fn.do_return (builder, val); + } + unary_ops[octave_value::op_uminus].add_overload (fn); + + fn = create_identity (intN (64)); + unary_ops[octave_value::op_uplus].add_overload (fn); + unary_ops[octave_value::op_transpose].add_overload (fn); + unary_ops[octave_value::op_hermitian].add_overload (fn); + + /************************************************************ + * * uint16 related operations * ************************************************************/ @@ -2210,6 +2367,7 @@ casts[scalar->type_id ()].stash_name ("(scalar)"); casts[single->type_id ()].stash_name ("(single)"); casts[intN (8)->type_id ()].stash_name ("(int8)"); + casts[intN (64)->type_id ()].stash_name ("(int64)"); casts[uintN (16)->type_id ()].stash_name ("(uint16)"); casts[complex->type_id ()].stash_name ("(complex)"); casts[matrix->type_id ()].stash_name ("(matrix)"); @@ -2255,6 +2413,14 @@ fn = create_external (JIT_FN (octave_jit_cast_int8_any), intN (8), any); casts[intN (8)->type_id ()].add_overload (fn); + // cast any <- int64 + fn = create_external (JIT_FN (octave_jit_cast_any_int64), any, intN (64)); + casts[any->type_id ()].add_overload (fn); + + // cast int64 <- any + fn = create_external (JIT_FN (octave_jit_cast_int64_any), intN (64), any); + casts[intN (64)->type_id ()].add_overload (fn); + // cast any <- uint16 fn = create_external (JIT_FN (octave_jit_cast_any_uint16), any, uintN (16)); casts[any->type_id ()].add_overload (fn); @@ -2304,6 +2470,10 @@ fn = create_identity (intN (8)); casts[intN (8)->type_id ()].add_overload (fn); + // cast int64 <- int64 + fn = create_identity (intN (64)); + casts[intN (64)->type_id ()].add_overload (fn); + // cast uint16 <- uint16 fn = create_identity (uintN (16)); casts[uintN (16)->type_id ()].add_overload (fn); @@ -2771,6 +2941,11 @@ return intN (8); } + if (ov.is_int64_type()) + { + return intN (64); + } + if (ov.is_uint16_type()) { return uintN (16);