Hi everybody, *** Motivation ***
Currently gcc is capable of replacing hand-crafted implementation of byteswap
by a suitable instruction thanks to the bswap optimization pass. The patch
proposed here aims at extending this pass to also optimize load in a specific
endianness, independent of the host endianness.
*** Methodology ***
The patch adds support for dealing with a memory source (array or structure)
and detect whether the result of a bitwise operation happens to be equivalent
to a big endian or little endian load and replace it by a load or a load and a
byteswap according to the host endianness. The original code used the concept
of symbolic number: a number where the value of each byte indicates its
position (in terms of weight) before the bitwise manipulation. After performing
the bit manipulation on that symbolic number, the result tells how the byte
were shuffled (see variable cmp in function find_bswap). Detecting an operation
resulting in a number in the host endianness is thus pretty straightforward:
look if the symbolic number has *not* changed.
As to supporting read from array and structure, there is some logic to
recognize the base of the array/structure and the offset of entries/fields
accessed to check if the range of memory accessed would fit in an integer. Each
entries is initially treated independently and when they are ORed together the
values in the symbolic number are updated according to the host endianness: the
entry of higher address would see its values incremented on a little endian
machine.
Note that as it stands the patch does not work for arrays indexed with variable
(such a tab[a] || (tab[a+1] << 8)) because fold_const does not fold (a + 1) -
a. If such cases were folded, the number of cases detected would automatically
be increased due to the use of fold_build2 to compare two offsets.
This patch also adds a few testcases to check both (i) that the optimization
works as expected and (ii) that the result are correct. It also define new
effective targets (bswap16, bswap32 and bswap64) to centralize the information
about what target supports byte swap instructions for the testsuite and modify
existing tests to use these new effective targets.
The patch is quite big but could be split if necessary. A big part of the code
added is for handling memory source and it would be difficult to split it but
variable renaming and introduction of bwapXX effective target could be made
separately to reduce the noise. The patch is too big so is only in attachment
of this email.
The ChangeLog are as follows:
*** gcc/ChangeLog ***
2014-03-19 Thomas Preud'homme <thomas.preudho...@arm.com>
PR tree-optimization/54733
* tree-ssa-math-opts.c (find_bswap_1): Renamed to ...
(find_bswap_or_nop_1): This. Also add support for memory source.
(find_bswap): Renamed to ...
(find_bswap_or_nop): This. Also add support for memory source and
detection of noop bitwise operations.
(execute_optimize_bswap): Likewise.
*** gcc/testsuite/ChangeLog ***
2014-03-19 Thomas Preud'homme <thomas.preudho...@arm.com>
PR tree-optimization/54733
* lib/target-supports.exp: New effective targets for architectures
capable of performing byte swap.
* gcc.dg/optimize-bswapdi-1.c: Convert to new bswap target.
* gcc.dg/optimize-bswapdi-2.c: Likewise.
* gcc.dg/optimize-bswapsi-1.c: Likewise.
* gcc.dg/optimize-bswapdi-3.c: New test to check extension of bswap
optimization to support memory sources.
* gcc.dg/optimize-bswaphi-1.c: Likewise.
* gcc.dg/optimize-bswapsi-2.c: Likewise.
* gcc.c-torture/execute/bswap-2.c: Likewise.
Is this ok for stage 1?
Best regards,
Thomas
gcc32rm-84.3.diff
Description: Binary data
