numba-cuda 0.21.1__cp313-cp313-win_amd64.whl

numba_cuda/numba/cuda/cgutils.py

@@ -0,0 +1,1294 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: BSD-2-Clause
+
+"""
+Generic helpers for LLVM code generation.
+"""
+
+import collections
+from contextlib import contextmanager, ExitStack
+import functools
+
+from llvmlite import ir
+
+from numba.cuda import types
+from numba.cuda import config, utils, debuginfo
+import numba.cuda.datamodel
+
+
+bool_t = ir.IntType(1)
+int8_t = ir.IntType(8)
+int32_t = ir.IntType(32)
+intp_t = ir.IntType(utils.MACHINE_BITS)
+voidptr_t = int8_t.as_pointer()
+
+true_bit = bool_t(1)
+false_bit = bool_t(0)
+true_byte = int8_t(1)
+false_byte = int8_t(0)
+
+
+def as_bool_bit(builder, value):
+    return builder.icmp_unsigned("!=", value, value.type(0))
+
+
+def make_anonymous_struct(builder, values, struct_type=None):
+    """
+    Create an anonymous struct containing the given LLVM *values*.
+    """
+    if struct_type is None:
+        struct_type = ir.LiteralStructType([v.type for v in values])
+    struct_val = struct_type(ir.Undefined)
+    for i, v in enumerate(values):
+        struct_val = builder.insert_value(struct_val, v, i)
+    return struct_val
+
+
+def make_bytearray(buf):
+    """
+    Make a byte array constant from *buf*.
+    """
+    b = bytearray(buf)
+    n = len(b)
+    return ir.Constant(ir.ArrayType(ir.IntType(8), n), b)
+
+
+_struct_proxy_cache = {}
+
+
+def create_struct_proxy(fe_type, kind="value"):
+    """
+    Returns a specialized StructProxy subclass for the given fe_type.
+    """
+    cache_key = (fe_type, kind)
+    res = _struct_proxy_cache.get(cache_key)
+    if res is None:
+        base = {
+            "value": ValueStructProxy,
+            "data": DataStructProxy,
+        }[kind]
+        clsname = base.__name__ + "_" + str(fe_type)
+        bases = (base,)
+        clsmembers = dict(_fe_type=fe_type)
+        res = type(clsname, bases, clsmembers)
+
+        _struct_proxy_cache[cache_key] = res
+    return res
+
+
+def copy_struct(dst, src, repl={}):
+    """
+    Copy structure from *src* to *dst* with replacement from *repl*.
+    """
+    repl = repl.copy()
+    # copy data from src or use those in repl
+    for k in src._datamodel._fields:
+        v = repl.pop(k, getattr(src, k))
+        setattr(dst, k, v)
+    # use remaining key-values in repl
+    for k, v in repl.items():
+        setattr(dst, k, v)
+    return dst
+
+
+class _StructProxy(object):
+    """
+    Creates a `Structure` like interface that is constructed with information
+    from DataModel instance.  FE type must have a data model that is a
+    subclass of StructModel.
+    """
+
+    # The following class members must be overridden by subclass
+    _fe_type = None
+
+    def __init__(self, context, builder, value=None, ref=None):
+        self._context = context
+        self._datamodel = self._context.data_model_manager[self._fe_type]
+        if not isinstance(self._datamodel, numba.cuda.datamodel.StructModel):
+            raise TypeError(
+                "Not a structure model: {0}".format(self._datamodel)
+            )
+        self._builder = builder
+
+        self._be_type = self._get_be_type(self._datamodel)
+        assert not is_pointer(self._be_type)
+
+        outer_ref, ref = self._make_refs(ref)
+        if ref.type.pointee != self._be_type:
+            raise AssertionError(
+                "bad ref type: expected %s, got %s"
+                % (self._be_type.as_pointer(), ref.type)
+            )
+
+        if value is not None:
+            if value.type != outer_ref.type.pointee:
+                raise AssertionError(
+                    "bad value type: expected %s, got %s"
+                    % (outer_ref.type.pointee, value.type)
+                )
+            self._builder.store(value, outer_ref)
+
+        self._value = ref
+        self._outer_ref = outer_ref
+
+    def _make_refs(self, ref):
+        """
+        Return an (outer ref, value ref) pair.  By default, these are
+        the same pointers, but a derived class may override this.
+        """
+        if ref is None:
+            ref = alloca_once(self._builder, self._be_type, zfill=True)
+        return ref, ref
+
+    def _get_be_type(self, datamodel):
+        raise NotImplementedError
+
+    def _cast_member_to_value(self, index, val):
+        raise NotImplementedError
+
+    def _cast_member_from_value(self, index, val):
+        raise NotImplementedError
+
+    def _get_ptr_by_index(self, index):
+        return gep_inbounds(self._builder, self._value, 0, index)
+
+    def _get_ptr_by_name(self, attrname):
+        index = self._datamodel.get_field_position(attrname)
+        return self._get_ptr_by_index(index)
+
+    def __getattr__(self, field):
+        """
+        Load the LLVM value of the named *field*.
+        """
+        if not field.startswith("_"):
+            return self[self._datamodel.get_field_position(field)]
+        else:
+            raise AttributeError(field)
+
+    def __setattr__(self, field, value):
+        """
+        Store the LLVM *value* into the named *field*.
+        """
+        if field.startswith("_"):
+            return super(_StructProxy, self).__setattr__(field, value)
+        self[self._datamodel.get_field_position(field)] = value
+
+    def __getitem__(self, index):
+        """
+        Load the LLVM value of the field at *index*.
+        """
+        member_val = self._builder.load(self._get_ptr_by_index(index))
+        return self._cast_member_to_value(index, member_val)
+
+    def __setitem__(self, index, value):
+        """
+        Store the LLVM *value* into the field at *index*.
+        """
+        ptr = self._get_ptr_by_index(index)
+        value = self._cast_member_from_value(index, value)
+        if value.type != ptr.type.pointee:
+            if (
+                is_pointer(value.type)
+                and is_pointer(ptr.type.pointee)
+                and value.type.pointee == ptr.type.pointee.pointee
+            ):
+                # Differ by address-space only
+                # Auto coerce it
+                value = self._context.addrspacecast(
+                    self._builder, value, ptr.type.pointee.addrspace
+                )
+            else:
+                raise TypeError(
+                    "Invalid store of {value.type} to "
+                    "{ptr.type.pointee} in "
+                    "{self._datamodel} "
+                    "(trying to write member #{index})".format(
+                        value=value, ptr=ptr, self=self, index=index
+                    )
+                )
+        self._builder.store(value, ptr)
+
+    def __len__(self):
+        """
+        Return the number of fields.
+        """
+        return self._datamodel.field_count
+
+    def _getpointer(self):
+        """
+        Return the LLVM pointer to the underlying structure.
+        """
+        return self._outer_ref
+
+    def _getvalue(self):
+        """
+        Load and return the value of the underlying LLVM structure.
+        """
+        return self._builder.load(self._outer_ref)
+
+    def _setvalue(self, value):
+        """
+        Store the value in this structure.
+        """
+        assert not is_pointer(value.type)
+        assert value.type == self._be_type, (value.type, self._be_type)
+        self._builder.store(value, self._value)
+
+
+class ValueStructProxy(_StructProxy):
+    """
+    Create a StructProxy suitable for accessing regular values
+    (e.g. LLVM values or alloca slots).
+    """
+
+    def _get_be_type(self, datamodel):
+        return datamodel.get_value_type()
+
+    def _cast_member_to_value(self, index, val):
+        return val
+
+    def _cast_member_from_value(self, index, val):
+        return val
+
+
+class DataStructProxy(_StructProxy):
+    """
+    Create a StructProxy suitable for accessing data persisted in memory.
+    """
+
+    def _get_be_type(self, datamodel):
+        return datamodel.get_data_type()
+
+    def _cast_member_to_value(self, index, val):
+        model = self._datamodel.get_model(index)
+        return model.from_data(self._builder, val)
+
+    def _cast_member_from_value(self, index, val):
+        model = self._datamodel.get_model(index)
+        return model.as_data(self._builder, val)
+
+
+class Structure(object):
+    """
+    A high-level object wrapping a alloca'ed LLVM structure, including
+    named fields and attribute access.
+    """
+
+    # XXX Should this warrant several separate constructors?
+    def __init__(self, context, builder, value=None, ref=None, cast_ref=False):
+        self._type = context.get_struct_type(self)
+        self._context = context
+        self._builder = builder
+        if ref is None:
+            self._value = alloca_once(builder, self._type, zfill=True)
+            if value is not None:
+                assert not is_pointer(value.type)
+                assert value.type == self._type, (value.type, self._type)
+                builder.store(value, self._value)
+        else:
+            assert value is None
+            assert is_pointer(ref.type)
+            if self._type != ref.type.pointee:
+                if cast_ref:
+                    ref = builder.bitcast(ref, self._type.as_pointer())
+                else:
+                    raise TypeError(
+                        "mismatching pointer type: got %s, expected %s"
+                        % (ref.type.pointee, self._type)
+                    )
+            self._value = ref
+
+        self._namemap = {}
+        self._fdmap = []
+        self._typemap = []
+        base = int32_t(0)
+        for i, (k, tp) in enumerate(self._fields):
+            self._namemap[k] = i
+            self._fdmap.append((base, int32_t(i)))
+            self._typemap.append(tp)
+
+    def _get_ptr_by_index(self, index):
+        ptr = self._builder.gep(self._value, self._fdmap[index], inbounds=True)
+        return ptr
+
+    def _get_ptr_by_name(self, attrname):
+        return self._get_ptr_by_index(self._namemap[attrname])
+
+    def __getattr__(self, field):
+        """
+        Load the LLVM value of the named *field*.
+        """
+        if not field.startswith("_"):
+            return self[self._namemap[field]]
+        else:
+            raise AttributeError(field)
+
+    def __setattr__(self, field, value):
+        """
+        Store the LLVM *value* into the named *field*.
+        """
+        if field.startswith("_"):
+            return super(Structure, self).__setattr__(field, value)
+        self[self._namemap[field]] = value
+
+    def __getitem__(self, index):
+        """
+        Load the LLVM value of the field at *index*.
+        """
+
+        return self._builder.load(self._get_ptr_by_index(index))
+
+    def __setitem__(self, index, value):
+        """
+        Store the LLVM *value* into the field at *index*.
+        """
+        ptr = self._get_ptr_by_index(index)
+        if ptr.type.pointee != value.type:
+            fmt = "Type mismatch: __setitem__(%d, ...) expected %r but got %r"
+            raise AssertionError(
+                fmt % (index, str(ptr.type.pointee), str(value.type))
+            )
+        self._builder.store(value, ptr)
+
+    def __len__(self):
+        """
+        Return the number of fields.
+        """
+        return len(self._namemap)
+
+    def _getpointer(self):
+        """
+        Return the LLVM pointer to the underlying structure.
+        """
+        return self._value
+
+    def _getvalue(self):
+        """
+        Load and return the value of the underlying LLVM structure.
+        """
+        return self._builder.load(self._value)
+
+    def _setvalue(self, value):
+        """Store the value in this structure"""
+        assert not is_pointer(value.type)
+        assert value.type == self._type, (value.type, self._type)
+        self._builder.store(value, self._value)
+
+    # __iter__ is derived by Python from __len__ and __getitem__
+
+
+def alloca_once(builder, ty, size=None, name="", zfill=False):
+    """Allocate stack memory at the entry block of the current function
+    pointed by ``builder`` with llvm type ``ty``.  The optional ``size`` arg
+    set the number of element to allocate.  The default is 1.  The optional
+    ``name`` arg set the symbol name inside the llvm IR for debugging.
+    If ``zfill`` is set, fill the memory with zeros at the current
+    use-site location.  Note that the memory is always zero-filled after the
+    ``alloca`` at init-site (the entry block).
+    """
+    if isinstance(size, int):
+        size = ir.Constant(intp_t, size)
+    # suspend debug metadata emission else it links up python source lines with
+    # alloca in the entry block as well as their actual location and it makes
+    # the debug info "jump about".
+    with debuginfo.suspend_emission(builder):
+        with builder.goto_entry_block():
+            ptr = builder.alloca(ty, size=size, name=name)
+            # Always zero-fill at init-site.  This is safe.
+            builder.store(ty(None), ptr)
+        # Also zero-fill at the use-site
+        if zfill:
+            builder.store(ptr.type.pointee(None), ptr)
+        return ptr
+
+
+def sizeof(builder, ptr_type):
+    """Compute sizeof using GEP"""
+    null = ptr_type(None)
+    offset = null.gep([int32_t(1)])
+    return builder.ptrtoint(offset, intp_t)
+
+
+def alloca_once_value(builder, value, name="", zfill=False):
+    """
+    Like alloca_once(), but passing a *value* instead of a type.  The
+    type is inferred and the allocated slot is also initialized with the
+    given value.
+    """
+    storage = alloca_once(builder, value.type, zfill=zfill)
+    builder.store(value, storage)
+    return storage
+
+
+def insert_pure_function(module, fnty, name):
+    """
+    Insert a pure function (in the functional programming sense) in the
+    given module.
+    """
+    fn = get_or_insert_function(module, fnty, name)
+    fn.attributes.add("readonly")
+    fn.attributes.add("nounwind")
+    return fn
+
+
+def get_or_insert_function(module, fnty, name):
+    """
+    Get the function named *name* with type *fnty* from *module*, or insert it
+    if it doesn't exist.
+    """
+    fn = module.globals.get(name, None)
+    if fn is None:
+        fn = ir.Function(module, fnty, name)
+    return fn
+
+
+def get_or_insert_named_metadata(module, name):
+    try:
+        return module.get_named_metadata(name)
+    except KeyError:
+        return module.add_named_metadata(name)
+
+
+def add_global_variable(module, ty, name, addrspace=0):
+    unique_name = module.get_unique_name(name)
+    return ir.GlobalVariable(module, ty, unique_name, addrspace)
+
+
+def terminate(builder, bbend):
+    bb = builder.basic_block
+    if bb.terminator is None:
+        builder.branch(bbend)
+
+
+def get_null_value(ltype):
+    return ltype(None)
+
+
+def is_null(builder, val):
+    null = get_null_value(val.type)
+    return builder.icmp_unsigned("==", null, val)
+
+
+def is_not_null(builder, val):
+    null = get_null_value(val.type)
+    return builder.icmp_unsigned("!=", null, val)
+
+
+def if_unlikely(builder, pred):
+    return builder.if_then(pred, likely=False)
+
+
+def if_likely(builder, pred):
+    return builder.if_then(pred, likely=True)
+
+
+def ifnot(builder, pred):
+    return builder.if_then(builder.not_(pred))
+
+
+def increment_index(builder, val):
+    """
+    Increment an index *val*.
+    """
+    one = val.type(1)
+    # We pass the "nsw" flag in the hope that LLVM understands the index
+    # never changes sign.  Unfortunately this doesn't always work
+    # (e.g. ndindex()).
+    return builder.add(val, one, flags=["nsw"])
+
+
+Loop = collections.namedtuple("Loop", ("index", "do_break"))
+
+
+@contextmanager
+def for_range(builder, count, start=None, intp=None):
+    """
+    Generate LLVM IR for a for-loop in [start, count).
+    *start* is equal to 0 by default.
+
+    Yields a Loop namedtuple with the following members:
+    - `index` is the loop index's value
+    - `do_break` is a no-argument callable to break out of the loop
+    """
+    if intp is None:
+        intp = count.type
+    if start is None:
+        start = intp(0)
+    stop = count
+
+    bbcond = builder.append_basic_block("for.cond")
+    bbbody = builder.append_basic_block("for.body")
+    bbend = builder.append_basic_block("for.end")
+
+    def do_break():
+        builder.branch(bbend)
+
+    bbstart = builder.basic_block
+    builder.branch(bbcond)
+
+    with builder.goto_block(bbcond):
+        index = builder.phi(intp, name="loop.index")
+        pred = builder.icmp_signed("<", index, stop)
+        builder.cbranch(pred, bbbody, bbend)
+
+    with builder.goto_block(bbbody):
+        yield Loop(index, do_break)
+        # Update bbbody as a new basic block may have been activated
+        bbbody = builder.basic_block
+        incr = increment_index(builder, index)
+        terminate(builder, bbcond)
+
+    index.add_incoming(start, bbstart)
+    index.add_incoming(incr, bbbody)
+
+    builder.position_at_end(bbend)
+
+
+@contextmanager
+def for_range_slice(builder, start, stop, step, intp=None, inc=True):
+    """
+    Generate LLVM IR for a for-loop based on a slice.  Yields a
+    (index, count) tuple where `index` is the slice index's value
+    inside the loop, and `count` the iteration count.
+
+    Parameters
+    -------------
+    builder : object
+        IRBuilder object
+    start : int
+        The beginning value of the slice
+    stop : int
+        The end value of the slice
+    step : int
+        The step value of the slice
+    intp :
+        The data type
+    inc : boolean, optional
+        Signals whether the step is positive (True) or negative (False).
+
+    Returns
+    -----------
+        None
+    """
+    if intp is None:
+        intp = start.type
+
+    bbcond = builder.append_basic_block("for.cond")
+    bbbody = builder.append_basic_block("for.body")
+    bbend = builder.append_basic_block("for.end")
+    bbstart = builder.basic_block
+    builder.branch(bbcond)
+
+    with builder.goto_block(bbcond):
+        index = builder.phi(intp, name="loop.index")
+        count = builder.phi(intp, name="loop.count")
+        if inc:
+            pred = builder.icmp_signed("<", index, stop)
+        else:
+            pred = builder.icmp_signed(">", index, stop)
+        builder.cbranch(pred, bbbody, bbend)
+
+    with builder.goto_block(bbbody):
+        yield index, count
+        bbbody = builder.basic_block
+        incr = builder.add(index, step)
+        next_count = increment_index(builder, count)
+        terminate(builder, bbcond)
+
+    index.add_incoming(start, bbstart)
+    index.add_incoming(incr, bbbody)
+    count.add_incoming(ir.Constant(intp, 0), bbstart)
+    count.add_incoming(next_count, bbbody)
+    builder.position_at_end(bbend)
+
+
+@contextmanager
+def for_range_slice_generic(builder, start, stop, step):
+    """
+    A helper wrapper for for_range_slice().  This is a context manager which
+    yields two for_range_slice()-alike context managers, the first for
+    the positive step case, the second for the negative step case.
+
+    Use:
+        with for_range_slice_generic(...) as (pos_range, neg_range):
+            with pos_range as (idx, count):
+                ...
+            with neg_range as (idx, count):
+                ...
+    """
+    intp = start.type
+    is_pos_step = builder.icmp_signed(">=", step, ir.Constant(intp, 0))
+
+    pos_for_range = for_range_slice(builder, start, stop, step, intp, inc=True)
+    neg_for_range = for_range_slice(builder, start, stop, step, intp, inc=False)
+
+    @contextmanager
+    def cm_cond(cond, inner_cm):
+        with cond:
+            with inner_cm as value:
+                yield value
+
+    with builder.if_else(is_pos_step, likely=True) as (then, otherwise):
+        yield cm_cond(then, pos_for_range), cm_cond(otherwise, neg_for_range)
+
+
+@contextmanager
+def loop_nest(builder, shape, intp, order="C"):
+    """
+    Generate a loop nest walking a N-dimensional array.
+    Yields a tuple of N indices for use in the inner loop body,
+    iterating over the *shape* space.
+
+    If *order* is 'C' (the default), indices are incremented inside-out
+    (i.e. (0,0), (0,1), (0,2), (1,0) etc.).
+    If *order* is 'F', they are incremented outside-in
+    (i.e. (0,0), (1,0), (2,0), (0,1) etc.).
+    This has performance implications when walking an array as it impacts
+    the spatial locality of memory accesses.
+    """
+    assert order in "CF"
+    if not shape:
+        # 0-d array
+        yield ()
+    else:
+        if order == "F":
+            _swap = lambda x: x[::-1]
+        else:
+            _swap = lambda x: x
+        with _loop_nest(builder, _swap(shape), intp) as indices:
+            assert len(indices) == len(shape)
+            yield _swap(indices)
+
+
+@contextmanager
+def _loop_nest(builder, shape, intp):
+    with for_range(builder, shape[0], intp=intp) as loop:
+        if len(shape) > 1:
+            with _loop_nest(builder, shape[1:], intp) as indices:
+                yield (loop.index,) + indices
+        else:
+            yield (loop.index,)
+
+
+def pack_array(builder, values, ty=None):
+    """
+    Pack a sequence of values in a LLVM array.  *ty* should be given
+    if the array may be empty, in which case the type can't be inferred
+    from the values.
+    """
+    n = len(values)
+    if ty is None:
+        ty = values[0].type
+    ary = ir.ArrayType(ty, n)(ir.Undefined)
+    for i, v in enumerate(values):
+        ary = builder.insert_value(ary, v, i)
+    return ary
+
+
+def pack_struct(builder, values):
+    """
+    Pack a sequence of values into a LLVM struct.
+    """
+    structty = ir.LiteralStructType([v.type for v in values])
+    st = structty(ir.Undefined)
+    for i, v in enumerate(values):
+        st = builder.insert_value(st, v, i)
+    return st
+
+
+def unpack_tuple(builder, tup, count=None):
+    """
+    Unpack an array or structure of values, return a Python tuple.
+    """
+    if count is None:
+        # Assuming *tup* is an aggregate
+        count = len(tup.type.elements)
+    vals = [builder.extract_value(tup, i) for i in range(count)]
+    return vals
+
+
+def get_item_pointer(
+    context, builder, aryty, ary, inds, wraparound=False, boundscheck=False
+):
+    # Set boundscheck=True for any pointer access that should be
+    # boundschecked. do_boundscheck() will handle enabling or disabling the
+    # actual boundschecking based on the user config.
+    shapes = unpack_tuple(builder, ary.shape, count=aryty.ndim)
+    strides = unpack_tuple(builder, ary.strides, count=aryty.ndim)
+    return get_item_pointer2(
+        context,
+        builder,
+        data=ary.data,
+        shape=shapes,
+        strides=strides,
+        layout=aryty.layout,
+        inds=inds,
+        wraparound=wraparound,
+        boundscheck=boundscheck,
+    )
+
+
+def do_boundscheck(context, builder, ind, dimlen, axis=None):
+    def _dbg():
+        # Remove this when we figure out how to include this information
+        # in the error message.
+        if axis is not None:
+            if isinstance(axis, int):
+                printf(
+                    builder,
+                    "debug: IndexError: index %d is out of bounds "
+                    "for axis {} with size %d\n".format(axis),
+                    ind,
+                    dimlen,
+                )
+            else:
+                printf(
+                    builder,
+                    "debug: IndexError: index %d is out of bounds "
+                    "for axis %d with size %d\n",
+                    ind,
+                    axis,
+                    dimlen,
+                )
+        else:
+            printf(
+                builder,
+                "debug: IndexError: index %d is out of bounds for size %d\n",
+                ind,
+                dimlen,
+            )
+
+    msg = "index is out of bounds"
+    out_of_bounds_upper = builder.icmp_signed(">=", ind, dimlen)
+    with if_unlikely(builder, out_of_bounds_upper):
+        if config.FULL_TRACEBACKS:
+            _dbg()
+        context.call_conv.return_user_exc(builder, IndexError, (msg,))
+    out_of_bounds_lower = builder.icmp_signed("<", ind, ind.type(0))
+    with if_unlikely(builder, out_of_bounds_lower):
+        if config.FULL_TRACEBACKS:
+            _dbg()
+        context.call_conv.return_user_exc(builder, IndexError, (msg,))
+
+
+def get_item_pointer2(
+    context,
+    builder,
+    data,
+    shape,
+    strides,
+    layout,
+    inds,
+    wraparound=False,
+    boundscheck=False,
+):
+    # Set boundscheck=True for any pointer access that should be
+    # boundschecked. do_boundscheck() will handle enabling or disabling the
+    # actual boundschecking based on the user config.
+    if wraparound:
+        # Wraparound
+        indices = []
+        for ind, dimlen in zip(inds, shape):
+            negative = builder.icmp_signed("<", ind, ind.type(0))
+            wrapped = builder.add(dimlen, ind)
+            selected = builder.select(negative, wrapped, ind)
+            indices.append(selected)
+    else:
+        indices = inds
+    if boundscheck:
+        for axis, (ind, dimlen) in enumerate(zip(indices, shape)):
+            do_boundscheck(context, builder, ind, dimlen, axis)
+
+    if not indices:
+        # Indexing with empty tuple
+        return builder.gep(data, [int32_t(0)])
+    intp = indices[0].type
+    # Indexing code
+    if layout in "CF":
+        steps = []
+        # Compute steps for each dimension
+        if layout == "C":
+            # C contiguous
+            for i in range(len(shape)):
+                last = intp(1)
+                for j in shape[i + 1 :]:
+                    last = builder.mul(last, j)
+                steps.append(last)
+        elif layout == "F":
+            # F contiguous
+            for i in range(len(shape)):
+                last = intp(1)
+                for j in shape[:i]:
+                    last = builder.mul(last, j)
+                steps.append(last)
+        else:
+            raise Exception("unreachable")
+
+        # Compute index
+        loc = intp(0)
+        for i, s in zip(indices, steps):
+            tmp = builder.mul(i, s)
+            loc = builder.add(loc, tmp)
+        ptr = builder.gep(data, [loc])
+        return ptr
+    else:
+        # Any layout
+        dimoffs = [builder.mul(s, i) for s, i in zip(strides, indices)]
+        offset = functools.reduce(builder.add, dimoffs)
+        return pointer_add(builder, data, offset)
+
+
+def _scalar_pred_against_zero(builder, value, fpred, icond):
+    nullval = value.type(0)
+    if isinstance(value.type, (ir.FloatType, ir.DoubleType)):
+        isnull = fpred(value, nullval)
+    elif isinstance(value.type, ir.IntType):
+        isnull = builder.icmp_signed(icond, value, nullval)
+    else:
+        raise TypeError("unexpected value type %s" % (value.type,))
+    return isnull
+
+
+def is_scalar_zero(builder, value):
+    """
+    Return a predicate representing whether *value* is equal to zero.
+    """
+    return _scalar_pred_against_zero(
+        builder, value, functools.partial(builder.fcmp_ordered, "=="), "=="
+    )
+
+
+def is_not_scalar_zero(builder, value):
+    """
+    Return a predicate representing whether a *value* is not equal to zero.
+    (not exactly "not is_scalar_zero" because of nans)
+    """
+    return _scalar_pred_against_zero(
+        builder, value, functools.partial(builder.fcmp_unordered, "!="), "!="
+    )
+
+
+def is_scalar_zero_or_nan(builder, value):
+    """
+    Return a predicate representing whether *value* is equal to either zero
+    or NaN.
+    """
+    return _scalar_pred_against_zero(
+        builder, value, functools.partial(builder.fcmp_unordered, "=="), "=="
+    )
+
+
+is_true = is_not_scalar_zero
+is_false = is_scalar_zero
+
+
+def is_scalar_neg(builder, value):
+    """
+    Is *value* negative?  Assumes *value* is signed.
+    """
+    return _scalar_pred_against_zero(
+        builder, value, functools.partial(builder.fcmp_ordered, "<"), "<"
+    )
+
+
+@contextmanager
+def early_exit_if(builder, stack: ExitStack, cond):
+    """
+    The Python code::
+
+        with contextlib.ExitStack() as stack:
+            with early_exit_if(builder, stack, cond):
+                cleanup()
+            body()
+
+    emits the code::
+
+        if (cond) {
+            <cleanup>
+        }
+        else {
+            <body>
+        }
+
+    This can be useful for generating code with lots of early exits, without
+    having to increase the indentation each time.
+    """
+    then, otherwise = stack.enter_context(builder.if_else(cond, likely=False))
+    with then:
+        yield
+    stack.enter_context(otherwise)
+
+
+def early_exit_if_null(builder, stack, obj):
+    """
+    A convenience wrapper for :func:`early_exit_if`, for the common case where
+    the CPython API indicates an error by returning ``NULL``.
+    """
+    return early_exit_if(builder, stack, is_null(builder, obj))
+
+
+def guard_null(context, builder, value, exc_tuple):
+    """
+    Guard against *value* being null or zero.
+    *exc_tuple* should be a (exception type, arguments...) tuple.
+    """
+    with builder.if_then(is_scalar_zero(builder, value), likely=False):
+        exc = exc_tuple[0]
+        exc_args = exc_tuple[1:] or None
+        context.call_conv.return_user_exc(builder, exc, exc_args)
+
+
+def guard_memory_error(context, builder, pointer, msg=None):
+    """
+    Guard against *pointer* being NULL (and raise a MemoryError).
+    """
+    assert isinstance(pointer.type, ir.PointerType), pointer.type
+    exc_args = (msg,) if msg else ()
+    with builder.if_then(is_null(builder, pointer), likely=False):
+        context.call_conv.return_user_exc(builder, MemoryError, exc_args)
+
+
+@contextmanager
+def if_zero(builder, value, likely=False):
+    """
+    Execute the given block if the scalar value is zero.
+    """
+    with builder.if_then(is_scalar_zero(builder, value), likely=likely):
+        yield
+
+
+guard_zero = guard_null
+
+
+def is_pointer(ltyp):
+    """
+    Whether the LLVM type *typ* is a struct type.
+    """
+    return isinstance(ltyp, ir.PointerType)
+
+
+def get_record_member(builder, record, offset, typ):
+    pval = gep_inbounds(builder, record, 0, offset)
+    assert not is_pointer(pval.type.pointee)
+    return builder.bitcast(pval, typ.as_pointer())
+
+
+def is_neg_int(builder, val):
+    return builder.icmp_signed("<", val, val.type(0))
+
+
+def gep_inbounds(builder, ptr, *inds, **kws):
+    """
+    Same as *gep*, but add the `inbounds` keyword.
+    """
+    return gep(builder, ptr, *inds, inbounds=True, **kws)
+
+
+def gep(builder, ptr, *inds, **kws):
+    """
+    Emit a getelementptr instruction for the given pointer and indices.
+    The indices can be LLVM values or Python int constants.
+    """
+    name = kws.pop("name", "")
+    inbounds = kws.pop("inbounds", False)
+    assert not kws
+    idx = []
+    for i in inds:
+        if isinstance(i, int):
+            # NOTE: llvm only accepts int32 inside structs, not int64
+            ind = int32_t(i)
+        else:
+            ind = i
+        idx.append(ind)
+    return builder.gep(ptr, idx, name=name, inbounds=inbounds)
+
+
+def pointer_add(builder, ptr, offset, return_type=None):
+    """
+    Add an integral *offset* to pointer *ptr*, and return a pointer
+    of *return_type* (or, if omitted, the same type as *ptr*).
+
+    Note the computation is done in bytes, and ignores the width of
+    the pointed item type.
+    """
+    intptr = builder.ptrtoint(ptr, intp_t)
+    if isinstance(offset, int):
+        offset = intp_t(offset)
+    intptr = builder.add(intptr, offset)
+    return builder.inttoptr(intptr, return_type or ptr.type)
+
+
+def memset(builder, ptr, size, value):
+    """
+    Fill *size* bytes starting from *ptr* with *value*.
+    """
+    fn = builder.module.declare_intrinsic("llvm.memset", (voidptr_t, size.type))
+    ptr = builder.bitcast(ptr, voidptr_t)
+    if isinstance(value, int):
+        value = int8_t(value)
+    builder.call(fn, [ptr, value, size, bool_t(0)])
+
+
+def memset_padding(builder, ptr):
+    """
+    Fill padding bytes of the pointee with zeros.
+    """
+    # Load existing value
+    val = builder.load(ptr)
+    # Fill pointee with zeros
+    memset(builder, ptr, sizeof(builder, ptr.type), 0)
+    # Store value back
+    builder.store(val, ptr)
+
+
+def global_constant(builder_or_module, name, value, linkage="internal"):
+    """
+    Get or create a (LLVM module-)global constant with *name* or *value*.
+    """
+    if isinstance(builder_or_module, ir.Module):
+        module = builder_or_module
+    else:
+        module = builder_or_module.module
+    data = add_global_variable(module, value.type, name)
+    data.linkage = linkage
+    data.global_constant = True
+    data.initializer = value
+    return data
+
+
+def divmod_by_constant(builder, val, divisor):
+    """
+    Compute the (quotient, remainder) of *val* divided by the constant
+    positive *divisor*.  The semantics reflects those of Python integer
+    floor division, rather than C's / LLVM's signed division and modulo.
+    The difference lies with a negative *val*.
+    """
+    assert divisor > 0
+    divisor = val.type(divisor)
+    one = val.type(1)
+
+    quot = alloca_once(builder, val.type)
+
+    with builder.if_else(is_neg_int(builder, val)) as (if_neg, if_pos):
+        with if_pos:
+            # quot = val / divisor
+            quot_val = builder.sdiv(val, divisor)
+            builder.store(quot_val, quot)
+        with if_neg:
+            # quot = -1 + (val + 1) / divisor
+            val_plus_one = builder.add(val, one)
+            quot_val = builder.sdiv(val_plus_one, divisor)
+            builder.store(builder.sub(quot_val, one), quot)
+
+    # rem = val - quot * divisor
+    # (should be slightly faster than a separate modulo operation)
+    quot_val = builder.load(quot)
+    rem_val = builder.sub(val, builder.mul(quot_val, divisor))
+    return quot_val, rem_val
+
+
+def cbranch_or_continue(builder, cond, bbtrue):
+    """
+    Branch conditionally or continue.
+
+    Note: a new block is created and builder is moved to the end of the new
+          block.
+    """
+    bbcont = builder.append_basic_block(".continue")
+    builder.cbranch(cond, bbtrue, bbcont)
+    builder.position_at_end(bbcont)
+    return bbcont
+
+
+def memcpy(builder, dst, src, count):
+    """
+    Emit a memcpy to the builder.
+
+    Copies each element of dst to src. Unlike the C equivalent, each element
+    can be any LLVM type.
+
+    Assumes
+    -------
+    * dst.type == src.type
+    * count is positive
+    """
+    # Note this does seem to be optimized as a raw memcpy() by LLVM
+    # whenever possible...
+    assert dst.type == src.type
+    with for_range(builder, count, intp=count.type) as loop:
+        out_ptr = builder.gep(dst, [loop.index])
+        in_ptr = builder.gep(src, [loop.index])
+        builder.store(builder.load(in_ptr), out_ptr)
+
+
+def _raw_memcpy(builder, func_name, dst, src, count, itemsize, align):
+    size_t = count.type
+    if isinstance(itemsize, int):
+        itemsize = ir.Constant(size_t, itemsize)
+
+    memcpy = builder.module.declare_intrinsic(
+        func_name, [voidptr_t, voidptr_t, size_t]
+    )
+    is_volatile = false_bit
+    builder.call(
+        memcpy,
+        [
+            builder.bitcast(dst, voidptr_t),
+            builder.bitcast(src, voidptr_t),
+            builder.mul(count, itemsize),
+            is_volatile,
+        ],
+    )
+
+
+def raw_memcpy(builder, dst, src, count, itemsize, align=1):
+    """
+    Emit a raw memcpy() call for `count` items of size `itemsize`
+    from `src` to `dest`.
+    """
+    return _raw_memcpy(builder, "llvm.memcpy", dst, src, count, itemsize, align)
+
+
+def raw_memmove(builder, dst, src, count, itemsize, align=1):
+    """
+    Emit a raw memmove() call for `count` items of size `itemsize`
+    from `src` to `dest`.
+    """
+    return _raw_memcpy(
+        builder, "llvm.memmove", dst, src, count, itemsize, align
+    )
+
+
+def muladd_with_overflow(builder, a, b, c):
+    """
+    Compute (a * b + c) and return a (result, overflow bit) pair.
+    The operands must be signed integers.
+    """
+    p = builder.smul_with_overflow(a, b)
+    prod = builder.extract_value(p, 0)
+    prod_ovf = builder.extract_value(p, 1)
+    s = builder.sadd_with_overflow(prod, c)
+    res = builder.extract_value(s, 0)
+    ovf = builder.or_(prod_ovf, builder.extract_value(s, 1))
+    return res, ovf
+
+
+def printf(builder, format, *args):
+    """
+    Calls printf().
+    Argument `format` is expected to be a Python string.
+    Values to be printed are listed in `args`.
+
+    Note: There is no checking to ensure there is correct number of values
+    in `args` and there type matches the declaration in the format string.
+    """
+    assert isinstance(format, str)
+    mod = builder.module
+    # Make global constant for format string
+    cstring = voidptr_t
+    fmt_bytes = make_bytearray((format + "\00").encode("ascii"))
+    global_fmt = global_constant(mod, "printf_format", fmt_bytes)
+    fnty = ir.FunctionType(int32_t, [cstring], var_arg=True)
+    # Insert printf()
+    try:
+        fn = mod.get_global("printf")
+    except KeyError:
+        fn = ir.Function(mod, fnty, name="printf")
+    # Call
+    ptr_fmt = builder.bitcast(global_fmt, cstring)
+    return builder.call(fn, [ptr_fmt] + list(args))
+
+
+def snprintf(builder, buffer, bufsz, format, *args):
+    """Calls libc snprintf(buffer, bufsz, format, ...args)"""
+    assert isinstance(format, str)
+    mod = builder.module
+    # Make global constant for format string
+    cstring = voidptr_t
+    fmt_bytes = make_bytearray((format + "\00").encode("ascii"))
+    global_fmt = global_constant(mod, "snprintf_format", fmt_bytes)
+    fnty = ir.FunctionType(
+        int32_t,
+        [cstring, intp_t, cstring],
+        var_arg=True,
+    )
+    # Actual symbol name of snprintf is different on win32.
+    symbol = "snprintf"
+    if config.IS_WIN32:
+        symbol = "_" + symbol
+    # Insert snprintf()
+    try:
+        fn = mod.get_global(symbol)
+    except KeyError:
+        fn = ir.Function(mod, fnty, name=symbol)
+    # Call
+    ptr_fmt = builder.bitcast(global_fmt, cstring)
+    return builder.call(fn, [buffer, bufsz, ptr_fmt] + list(args))
+
+
+def snprintf_stackbuffer(builder, bufsz, format, *args):
+    """Similar to `snprintf()` but the buffer is stack allocated to size
+    *bufsz*.
+
+    Returns the buffer pointer as i8*.
+    """
+    assert isinstance(bufsz, int)
+    spacety = ir.ArrayType(ir.IntType(8), bufsz)
+    space = alloca_once(builder, spacety, zfill=True)
+    buffer = builder.bitcast(space, voidptr_t)
+    snprintf(builder, buffer, intp_t(bufsz), format, *args)
+    return buffer
+
+
+def normalize_ir_text(text):
+    """
+    Normalize the given string to latin1 compatible encoding that is
+    suitable for use in LLVM IR.
+    """
+    # Just re-encoding to latin1 is enough
+    return text.encode("utf8").decode("latin1")
+
+
+def hexdump(builder, ptr, nbytes):
+    """Debug print the memory region in *ptr* to *ptr + nbytes*
+    as hex.
+    """
+    bytes_per_line = 16
+    nbytes = builder.zext(nbytes, intp_t)
+    printf(builder, "hexdump p=%p n=%zu", ptr, nbytes)
+    byte_t = ir.IntType(8)
+    ptr = builder.bitcast(ptr, byte_t.as_pointer())
+    # Loop to print the bytes in *ptr* as hex
+    with for_range(builder, nbytes) as idx:
+        div_by = builder.urem(idx.index, intp_t(bytes_per_line))
+        do_new_line = builder.icmp_unsigned("==", div_by, intp_t(0))
+        with builder.if_then(do_new_line):
+            printf(builder, "\n")
+
+        offset = builder.gep(ptr, [idx.index])
+        val = builder.load(offset)
+        printf(builder, " %02x", val)
+    printf(builder, "\n")
+
+
+def is_nonelike(ty):
+    """returns if 'ty' is none"""
+    return (
+        ty is None
+        or isinstance(ty, types.NoneType)
+        or isinstance(ty, types.Omitted)
+    )
+
+
+def is_empty_tuple(ty):
+    """returns if 'ty' is an empty tuple"""
+    return isinstance(ty, types.Tuple) and len(ty.types) == 0
+
+
+def create_constant_array(ty, val):
+    """
+    Create an LLVM-constant of a fixed-length array from Python values.
+
+    The type provided is the type of the elements.
+    """
+    return ir.Constant(ir.ArrayType(ty, len(val)), val)