From 71e014caecaa54fdd8a0516710d2d9597da41398 Mon Sep 17 00:00:00 2001 From: Shawn Landden Date: Fri, 21 Jun 2019 16:18:59 -0500 Subject: stage1: add @sin @cos @exp @exp2 @ln @log2 @log10 @fabs @floor @ceil @trunc @round and expand @sqrt This revealed that the accuracy of ln is not as good as the current algorithm in musl and glibc, and should be ported again. v2: actually include tests v3: fix reversal of in and out arguments on f128M_sqrt() add test for @sqrt on comptime_float do not include @nearbyInt() until it works on all targets. --- src/ir.cpp | 355 ++++++++++++++++++++++++++++++++++++++++++++++++++----------- 1 file changed, 294 insertions(+), 61 deletions(-) (limited to 'src/ir.cpp') diff --git a/src/ir.cpp b/src/ir.cpp index c2c6cb6154..50d2a06868 100644 --- a/src/ir.cpp +++ b/src/ir.cpp @@ -991,8 +991,8 @@ static constexpr IrInstructionId ir_instruction_id(IrInstructionMarkErrRetTraceP return IrInstructionIdMarkErrRetTracePtr; } -static constexpr IrInstructionId ir_instruction_id(IrInstructionSqrt *) { - return IrInstructionIdSqrt; +static constexpr IrInstructionId ir_instruction_id(IrInstructionFloatOp *) { + return IrInstructionIdFloatOp; } static constexpr IrInstructionId ir_instruction_id(IrInstructionCheckRuntimeScope *) { @@ -2312,6 +2312,59 @@ static IrInstruction *ir_build_overflow_op(IrBuilder *irb, Scope *scope, AstNode return &instruction->base; } + +//TODO Powi, Pow, minnum, maxnum, maximum, minimum, copysign, +// lround, llround, lrint, llrint +// So far this is only non-complicated type functions. +const char *float_op_to_name(BuiltinFnId op, bool llvm_name) { + const bool b = llvm_name; + + switch (op) { + case BuiltinFnIdSqrt: + return "sqrt"; + case BuiltinFnIdSin: + return "sin"; + case BuiltinFnIdCos: + return "cos"; + case BuiltinFnIdExp: + return "exp"; + case BuiltinFnIdExp2: + return "exp2"; + case BuiltinFnIdLn: + return b ? "log" : "ln"; + case BuiltinFnIdLog10: + return "log10"; + case BuiltinFnIdLog2: + return "log2"; + case BuiltinFnIdFabs: + return "fabs"; + case BuiltinFnIdFloor: + return "floor"; + case BuiltinFnIdCeil: + return "ceil"; + case BuiltinFnIdTrunc: + return "trunc"; + case BuiltinFnIdNearbyInt: + return b ? "nearbyint" : "nearbyInt"; + case BuiltinFnIdRound: + return "round"; + default: + zig_unreachable(); + } +} + +static IrInstruction *ir_build_float_op(IrBuilder *irb, Scope *scope, AstNode *source_node, IrInstruction *type, IrInstruction *op1, BuiltinFnId op) { + IrInstructionFloatOp *instruction = ir_build_instruction(irb, scope, source_node); + instruction->type = type; + instruction->op1 = op1; + instruction->op = op; + + if (type != nullptr) ir_ref_instruction(type, irb->current_basic_block); + ir_ref_instruction(op1, irb->current_basic_block); + + return &instruction->base; +} + static IrInstruction *ir_build_mul_add(IrBuilder *irb, Scope *scope, AstNode *source_node, IrInstruction *type_value, IrInstruction *op1, IrInstruction *op2, IrInstruction *op3) { IrInstructionMulAdd *instruction = ir_build_instruction(irb, scope, source_node); @@ -3033,17 +3086,6 @@ static IrInstruction *ir_build_mark_err_ret_trace_ptr(IrBuilder *irb, Scope *sco return &instruction->base; } -static IrInstruction *ir_build_sqrt(IrBuilder *irb, Scope *scope, AstNode *source_node, IrInstruction *type, IrInstruction *op) { - IrInstructionSqrt *instruction = ir_build_instruction(irb, scope, source_node); - instruction->type = type; - instruction->op = op; - - if (type != nullptr) ir_ref_instruction(type, irb->current_basic_block); - ir_ref_instruction(op, irb->current_basic_block); - - return &instruction->base; -} - static IrInstruction *ir_build_has_decl(IrBuilder *irb, Scope *scope, AstNode *source_node, IrInstruction *container, IrInstruction *name) { @@ -4400,6 +4442,19 @@ static IrInstruction *ir_gen_builtin_fn_call(IrBuilder *irb, Scope *scope, AstNo return ir_lval_wrap(irb, scope, bin_op, lval); } case BuiltinFnIdSqrt: + case BuiltinFnIdSin: + case BuiltinFnIdCos: + case BuiltinFnIdExp: + case BuiltinFnIdExp2: + case BuiltinFnIdLn: + case BuiltinFnIdLog2: + case BuiltinFnIdLog10: + case BuiltinFnIdFabs: + case BuiltinFnIdFloor: + case BuiltinFnIdCeil: + case BuiltinFnIdTrunc: + case BuiltinFnIdNearbyInt: + case BuiltinFnIdRound: { AstNode *arg0_node = node->data.fn_call_expr.params.at(0); IrInstruction *arg0_value = ir_gen_node(irb, arg0_node, scope); @@ -4411,7 +4466,7 @@ static IrInstruction *ir_gen_builtin_fn_call(IrBuilder *irb, Scope *scope, AstNo if (arg1_value == irb->codegen->invalid_instruction) return arg1_value; - IrInstruction *ir_sqrt = ir_build_sqrt(irb, scope, node, arg0_value, arg1_value); + IrInstruction *ir_sqrt = ir_build_float_op(irb, scope, node, arg0_value, arg1_value, builtin_fn->id); return ir_lval_wrap(irb, scope, ir_sqrt, lval); } case BuiltinFnIdTruncate: @@ -23214,70 +23269,248 @@ static IrInstruction *ir_analyze_instruction_mark_err_ret_trace_ptr(IrAnalyze *i return result; } -static IrInstruction *ir_analyze_instruction_sqrt(IrAnalyze *ira, IrInstructionSqrt *instruction) { - ZigType *float_type = ir_resolve_type(ira, instruction->type->child); - if (type_is_invalid(float_type)) - return ira->codegen->invalid_instruction; +static void ir_eval_float_op(IrAnalyze *ira, IrInstructionFloatOp *source_instr, ZigType *float_type, + ConstExprValue *op, ConstExprValue *out_val) { + assert(ira && source_instr && float_type && out_val && op); + assert(float_type->id == ZigTypeIdFloat || + float_type->id == ZigTypeIdComptimeFloat); - IrInstruction *op = instruction->op->child; - if (type_is_invalid(op->value.type)) + BuiltinFnId fop = source_instr->op; + unsigned bits; + + if (float_type->id == ZigTypeIdComptimeFloat) { + bits = 128; + } else if (float_type->id == ZigTypeIdFloat) + bits = float_type->data.floating.bit_count; + + switch (bits) { + case 16: { + switch (fop) { + case BuiltinFnIdSqrt: + out_val->data.x_f16 = f16_sqrt(op->data.x_f16); + break; + case BuiltinFnIdSin: + case BuiltinFnIdCos: + case BuiltinFnIdExp: + case BuiltinFnIdExp2: + case BuiltinFnIdLn: + case BuiltinFnIdLog10: + case BuiltinFnIdLog2: + case BuiltinFnIdFabs: + case BuiltinFnIdFloor: + case BuiltinFnIdCeil: + case BuiltinFnIdTrunc: + case BuiltinFnIdNearbyInt: + case BuiltinFnIdRound: + zig_panic("unimplemented f16 builtin"); + default: + zig_unreachable(); + }; + break; + }; + case 32: { + switch (fop) { + case BuiltinFnIdSqrt: + out_val->data.x_f32 = sqrtf(op->data.x_f32); + break; + case BuiltinFnIdSin: + out_val->data.x_f32 = sinf(op->data.x_f32); + break; + case BuiltinFnIdCos: + out_val->data.x_f32 = cosf(op->data.x_f32); + break; + case BuiltinFnIdExp: + out_val->data.x_f32 = expf(op->data.x_f32); + break; + case BuiltinFnIdExp2: + out_val->data.x_f32 = exp2f(op->data.x_f32); + break; + case BuiltinFnIdLn: + out_val->data.x_f32 = logf(op->data.x_f32); + break; + case BuiltinFnIdLog10: + out_val->data.x_f32 = log10f(op->data.x_f32); + break; + case BuiltinFnIdLog2: + out_val->data.x_f32 = log2f(op->data.x_f32); + break; + case BuiltinFnIdFabs: + out_val->data.x_f32 = fabsf(op->data.x_f32); + break; + case BuiltinFnIdFloor: + out_val->data.x_f32 = floorf(op->data.x_f32); + break; + case BuiltinFnIdCeil: + out_val->data.x_f32 = ceilf(op->data.x_f32); + break; + case BuiltinFnIdTrunc: + out_val->data.x_f32 = truncf(op->data.x_f32); + break; + case BuiltinFnIdNearbyInt: + out_val->data.x_f32 = nearbyintf(op->data.x_f32); + break; + case BuiltinFnIdRound: + out_val->data.x_f32 = roundf(op->data.x_f32); + break; + default: + zig_unreachable(); + }; + break; + }; + case 64: { + switch (fop) { + case BuiltinFnIdSqrt: + out_val->data.x_f64 = sqrt(op->data.x_f64); + break; + case BuiltinFnIdSin: + out_val->data.x_f64 = sin(op->data.x_f64); + break; + case BuiltinFnIdCos: + out_val->data.x_f64 = cos(op->data.x_f64); + break; + case BuiltinFnIdExp: + out_val->data.x_f64 = exp(op->data.x_f64); + break; + case BuiltinFnIdExp2: + out_val->data.x_f64 = exp2(op->data.x_f64); + break; + case BuiltinFnIdLn: + out_val->data.x_f64 = log(op->data.x_f64); + break; + case BuiltinFnIdLog10: + out_val->data.x_f64 = log10(op->data.x_f64); + break; + case BuiltinFnIdLog2: + out_val->data.x_f64 = log2(op->data.x_f64); + break; + case BuiltinFnIdFabs: + out_val->data.x_f64 = fabs(op->data.x_f64); + break; + case BuiltinFnIdFloor: + out_val->data.x_f64 = floor(op->data.x_f64); + break; + case BuiltinFnIdCeil: + out_val->data.x_f64 = ceil(op->data.x_f64); + break; + case BuiltinFnIdTrunc: + out_val->data.x_f64 = trunc(op->data.x_f64); + break; + case BuiltinFnIdNearbyInt: + out_val->data.x_f64 = nearbyint(op->data.x_f64); + break; + case BuiltinFnIdRound: + out_val->data.x_f64 = round(op->data.x_f64); + break; + default: + zig_unreachable(); + } + break; + }; + case 128: { + float128_t *out, *in; + if (float_type->id == ZigTypeIdComptimeFloat) { + out = &out_val->data.x_bigfloat.value; + in = &op->data.x_bigfloat.value; + } else { + out = &out_val->data.x_f128; + in = &op->data.x_f128; + } + switch (fop) { + case BuiltinFnIdSqrt: + f128M_sqrt(in, out); + break; + case BuiltinFnIdNearbyInt: + case BuiltinFnIdSin: + case BuiltinFnIdCos: + case BuiltinFnIdExp: + case BuiltinFnIdExp2: + case BuiltinFnIdLn: + case BuiltinFnIdLog10: + case BuiltinFnIdLog2: + case BuiltinFnIdFabs: + case BuiltinFnIdFloor: + case BuiltinFnIdCeil: + case BuiltinFnIdTrunc: + case BuiltinFnIdRound: + zig_panic("unimplemented f128 builtin"); + default: + zig_unreachable(); + } + break; + }; + default: + zig_unreachable(); + } +} + +static IrInstruction *ir_analyze_instruction_float_op(IrAnalyze *ira, IrInstructionFloatOp *instruction) { + IrInstruction *type = instruction->type->child; + if (type_is_invalid(type->value.type)) + return ira->codegen->invalid_instruction; + + ZigType *expr_type = ir_resolve_type(ira, type); + if (type_is_invalid(expr_type)) return ira->codegen->invalid_instruction; - bool ok_type = float_type->id == ZigTypeIdComptimeFloat || float_type->id == ZigTypeIdFloat; - if (!ok_type) { - ir_add_error(ira, instruction->type, buf_sprintf("@sqrt does not support type '%s'", buf_ptr(&float_type->name))); + // Only allow float types, and vectors of floats. + ZigType *float_type = (expr_type->id == ZigTypeIdVector) ? expr_type->data.vector.elem_type : expr_type; + if (float_type->id != ZigTypeIdFloat && float_type->id != ZigTypeIdComptimeFloat) { + ir_add_error(ira, instruction->type, buf_sprintf("@%s does not support type '%s'", float_op_to_name(instruction->op, false), buf_ptr(&float_type->name))); return ira->codegen->invalid_instruction; } - IrInstruction *casted_op = ir_implicit_cast(ira, op, float_type); - if (type_is_invalid(casted_op->value.type)) + IrInstruction *op1 = instruction->op1->child; + if (type_is_invalid(op1->value.type)) return ira->codegen->invalid_instruction; - if (instr_is_comptime(casted_op)) { - ConstExprValue *val = ir_resolve_const(ira, casted_op, UndefBad); - if (!val) + IrInstruction *casted_op1 = ir_implicit_cast(ira, op1, float_type); + if (type_is_invalid(casted_op1->value.type)) + return ira->codegen->invalid_instruction; + + if (instr_is_comptime(casted_op1)) { + // Our comptime 16-bit and 128-bit support is quite limited. + if ((float_type->id == ZigTypeIdComptimeFloat || + float_type->data.floating.bit_count == 16 || + float_type->data.floating.bit_count == 128) && + instruction->op != BuiltinFnIdSqrt) { + ir_add_error(ira, instruction->type, buf_sprintf("@%s does not support type '%s'", float_op_to_name(instruction->op, false), buf_ptr(&float_type->name))); return ira->codegen->invalid_instruction; + } - IrInstruction *result = ir_const(ira, &instruction->base, float_type); + ConstExprValue *op1_const = ir_resolve_const(ira, casted_op1, UndefBad); + if (!op1_const) + return ira->codegen->invalid_instruction; + + IrInstruction *result = ir_const(ira, &instruction->base, expr_type); ConstExprValue *out_val = &result->value; - if (float_type->id == ZigTypeIdComptimeFloat) { - bigfloat_sqrt(&out_val->data.x_bigfloat, &val->data.x_bigfloat); - } else if (float_type->id == ZigTypeIdFloat) { - switch (float_type->data.floating.bit_count) { - case 16: - out_val->data.x_f16 = f16_sqrt(val->data.x_f16); - break; - case 32: - out_val->data.x_f32 = sqrtf(val->data.x_f32); - break; - case 64: - out_val->data.x_f64 = sqrt(val->data.x_f64); - break; - case 128: - f128M_sqrt(&val->data.x_f128, &out_val->data.x_f128); - break; - default: - zig_unreachable(); + if (expr_type->id == ZigTypeIdVector) { + expand_undef_array(ira->codegen, op1_const); + out_val->special = ConstValSpecialUndef; + expand_undef_array(ira->codegen, out_val); + size_t len = expr_type->data.vector.len; + for (size_t i = 0; i < len; i += 1) { + ConstExprValue *float_operand_op1 = &op1_const->data.x_array.data.s_none.elements[i]; + ConstExprValue *float_out_val = &out_val->data.x_array.data.s_none.elements[i]; + assert(float_operand_op1->type == float_type); + assert(float_out_val->type == float_type); + ir_eval_float_op(ira, instruction, float_type, + op1_const, float_out_val); + float_out_val->type = float_type; } + out_val->type = expr_type; + out_val->special = ConstValSpecialStatic; } else { - zig_unreachable(); + ir_eval_float_op(ira, instruction, float_type, op1_const, out_val); } - return result; } ir_assert(float_type->id == ZigTypeIdFloat, &instruction->base); - if (float_type->data.floating.bit_count != 16 && - float_type->data.floating.bit_count != 32 && - float_type->data.floating.bit_count != 64) { - ir_add_error(ira, instruction->type, buf_sprintf("compiler TODO: add implementation of sqrt for '%s'", buf_ptr(&float_type->name))); - return ira->codegen->invalid_instruction; - } - IrInstruction *result = ir_build_sqrt(&ira->new_irb, instruction->base.scope, - instruction->base.source_node, nullptr, casted_op); - result->value.type = float_type; + IrInstruction *result = ir_build_float_op(&ira->new_irb, instruction->base.scope, + instruction->base.source_node, nullptr, casted_op1, instruction->op); + result->value.type = expr_type; return result; } @@ -23762,8 +23995,8 @@ static IrInstruction *ir_analyze_instruction_nocast(IrAnalyze *ira, IrInstructio return ir_analyze_instruction_merge_err_ret_traces(ira, (IrInstructionMergeErrRetTraces *)instruction); case IrInstructionIdMarkErrRetTracePtr: return ir_analyze_instruction_mark_err_ret_trace_ptr(ira, (IrInstructionMarkErrRetTracePtr *)instruction); - case IrInstructionIdSqrt: - return ir_analyze_instruction_sqrt(ira, (IrInstructionSqrt *)instruction); + case IrInstructionIdFloatOp: + return ir_analyze_instruction_float_op(ira, (IrInstructionFloatOp *)instruction); case IrInstructionIdMulAdd: return ir_analyze_instruction_mul_add(ira, (IrInstructionMulAdd *)instruction); case IrInstructionIdIntToErr: @@ -24004,7 +24237,7 @@ bool ir_has_side_effects(IrInstruction *instruction) { case IrInstructionIdCoroFree: case IrInstructionIdCoroPromise: case IrInstructionIdPromiseResultType: - case IrInstructionIdSqrt: + case IrInstructionIdFloatOp: case IrInstructionIdMulAdd: case IrInstructionIdAtomicLoad: case IrInstructionIdIntCast: -- cgit v1.2.3