motiejus/zig
1
Fork 0
Code Releases Activity

stage1: add support for @mulAdd fused-multiply-add for floats and vectors of floats

Browse Source 
Not all of the softfloat library is being built....

Vector support is very buggy at the moment, but should work when the bugs are fixed.
(as I had the same code working with another vector function, that hasn't been merged yet).
This commit is contained in:
Shawn Landden
2019-06-18 17:28:49 -05:00
parent bbfb53d524
commit fce2d2d18b
8 changed files with 292 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
src/codegen.cpp
View File

@@ -807,31 +807,51 @@ static LLVMValueRef get_int_overflow_fn(CodeGen *g, ZigType *operand_type, AddSu
}
static LLVMValueRef get_float_fn(CodeGen *g, ZigType *type_entry, ZigLLVMFnId fn_id) {
assert(type_entry->id == ZigTypeIdFloat);
assert(type_entry->id == ZigTypeIdFloat ||
type_entry->id == ZigTypeIdVector);
bool is_vector = (type_entry->id == ZigTypeIdVector);
ZigType *float_type = is_vector ? type_entry->data.vector.elem_type : type_entry;
ZigLLVMFnKey key = {};
key.id = fn_id;
key.data.floating.bit_count = (uint32_t)type_entry->data.floating.bit_count;
key.data.floating.bit_count = (uint32_t)float_type->data.floating.bit_count;
key.data.floating.vector_len = is_vector ? (uint32_t)type_entry->data.vector.len : 0;
auto existing_entry = g->llvm_fn_table.maybe_get(key);
if (existing_entry)
return existing_entry->value;
const char *name;
uint32_t num_args;
if (fn_id == ZigLLVMFnIdFloor) {
name = "floor";
num_args = 1;
} else if (fn_id == ZigLLVMFnIdCeil) {
name = "ceil";
num_args = 1;
} else if (fn_id == ZigLLVMFnIdSqrt) {
name = "sqrt";
num_args = 1;
} else if (fn_id == ZigLLVMFnIdFMA) {
name = "fma";
num_args = 3;
} else {
zig_unreachable();
}
char fn_name[64];
sprintf(fn_name, "llvm.%s.f%" ZIG_PRI_usize "", name, type_entry->data.floating.bit_count);
if (is_vector)
sprintf(fn_name, "llvm.%s.v%" PRIu32 "f%" PRIu32, name, key.data.floating.vector_len, key.data.floating.bit_count);
else
sprintf(fn_name, "llvm.%s.f%" PRIu32, name, key.data.floating.bit_count);
LLVMTypeRef float_type_ref = get_llvm_type(g, type_entry);
LLVMTypeRef fn_type = LLVMFunctionType(float_type_ref, &float_type_ref, 1, false);
LLVMTypeRef return_elem_types[3] = {
float_type_ref,
float_type_ref,
float_type_ref,
};
LLVMTypeRef fn_type = LLVMFunctionType(float_type_ref, return_elem_types, num_args, false);
LLVMValueRef fn_val = LLVMAddFunction(g->module, fn_name, fn_type);
assert(LLVMGetIntrinsicID(fn_val));
@@ -5437,6 +5457,21 @@ static LLVMValueRef ir_render_sqrt(CodeGen *g, IrExecutable *executable, IrInstr
return LLVMBuildCall(g->builder, fn_val, &op, 1, "");
}
static LLVMValueRef ir_render_mul_add(CodeGen *g, IrExecutable *executable, IrInstructionMulAdd *instruction) {
LLVMValueRef op1 = ir_llvm_value(g, instruction->op1);
LLVMValueRef op2 = ir_llvm_value(g, instruction->op2);
LLVMValueRef op3 = ir_llvm_value(g, instruction->op3);
assert(instruction->base.value.type->id == ZigTypeIdFloat ||
instruction->base.value.type->id == ZigTypeIdVector);
LLVMValueRef fn_val = get_float_fn(g, instruction->base.value.type, ZigLLVMFnIdFMA);
LLVMValueRef args[3] = {
op1,
op2,
op3,
};
return LLVMBuildCall(g->builder, fn_val, args, 3, "");
}
static LLVMValueRef ir_render_bswap(CodeGen *g, IrExecutable *executable, IrInstructionBswap *instruction) {
LLVMValueRef op = ir_llvm_value(g, instruction->op);
ZigType *int_type = instruction->base.value.type;
@@ -5781,6 +5816,8 @@ static LLVMValueRef ir_render_instruction(CodeGen *g, IrExecutable *executable,
return ir_render_mark_err_ret_trace_ptr(g, executable, (IrInstructionMarkErrRetTracePtr *)instruction);
case IrInstructionIdSqrt:
return ir_render_sqrt(g, executable, (IrInstructionSqrt *)instruction);
case IrInstructionIdMulAdd:
return ir_render_mul_add(g, executable, (IrInstructionMulAdd *)instruction);
case IrInstructionIdArrayToVector:
return ir_render_array_to_vector(g, executable, (IrInstructionArrayToVector *)instruction);
case IrInstructionIdVectorToArray:
@@ -7398,6 +7435,7 @@ static void define_builtin_fns(CodeGen *g) {
create_builtin_fn(g, BuiltinFnIdRem, "rem", 2);
create_builtin_fn(g, BuiltinFnIdMod, "mod", 2);
create_builtin_fn(g, BuiltinFnIdSqrt, "sqrt", 2);
create_builtin_fn(g, BuiltinFnIdMulAdd, "mulAdd", 4);
create_builtin_fn(g, BuiltinFnIdInlineCall, "inlineCall", SIZE_MAX);
create_builtin_fn(g, BuiltinFnIdNoInlineCall, "noInlineCall", SIZE_MAX);
create_builtin_fn(g, BuiltinFnIdNewStackCall, "newStackCall", SIZE_MAX);