[tl;dr: This is a rewrite of value-range-storage.* such that global
ranges and the internal ranger cache can use the same efficient
storage mechanism. It is optimized such that when wide_ints are
dropped into irange, the copying back and forth from storage will be
very fast, while being able to hold any number of sub-ranges
dynamically allocated at run-time. This replaces the global storage
mechanism which was limited to 6-subranges.]
Previously we had a vrange allocator for use in the ranger cache. It
worked with trees and could be used in place (fast), but it was not
memory efficient. With the upcoming switch to wide_ints for irange,
we can't afford to allocate ranges that can be used in place, because
an irange will be significantly larger, as it will hold full
wide_ints. We need a trailing_wide_int mechanism similar to what we
use for global ranges, but fast enough to use in the ranger's cache.
The global ranges had another allocation mechanism that was
trailing_wide_int based. It was memory efficient but slow given the
constant conversions from trees to wide_ints.
This patch gets us the best of both worlds by providing a storage
mechanism with a custom trailing wide int interface, while at the same
time being fast enough to use in the ranger cache.
We use a custom trailing wide_int mechanism but more flexible than
trailing_wide_int, since the latter has compile-time fixed-sized
wide_ints. The original TWI structure has the current length of each
wide_int in a static portion preceeding the variable length:
template <int N>
struct GTY((user)) trailing_wide_ints
{
...
...
/* The current length of each number.
that will, in turn, turn off TBAA on gimple, trees and RTL. */
struct {unsigned char len;} m_len[N];
/* The variable-length part of the structure, which always contains
at least one HWI. Element I starts at index I * M_MAX_LEN. */
HOST_WIDE_INT m_val[1];
};
We need both m_len[] and m_val[] to be variable-length at run-time.
In the previous incarnation of the storage mechanism the limitation of
m_len[] being static meant that we were limited to whatever [N] could
use up the unused bits in the TWI control world. In practice this
meant we were limited to 6 sub-ranges. This worked fine for global
ranges, but is a no go for our internal cache, where we must represent
things exactly (ranges for switches, etc).
The new implementation removes this restriction by making both m_len[]
and m_val[] variable length. Also, rolling our own allows future
optimization be using some of the leftover bits in the control world.
Also, in preparation for the wide_int conversion, vrange_storage is
now optimized to blast the bits directly into the ultimate irange
instead of going through the irange API. So ultimately copying back
and forth between the ranger cache and the storage mechanism is just a
matter of copying a few bits for the control word, and copying an
array of HOST_WIDE_INTs. These changes were heavily profiled, and
yielded a good chunk of the overall speedup for the wide_int
conversion.
Finally, vrange_storage is now a first class structure with GTY
markers and all, thus alleviating the void * hack in struct
tree_ssa_name and friends. This removes a few warts in the API and
looks cleaner overall.
git://gcc.gnu.org / gcc.git / commit