Skip to content
Snippets Groups Projects
Commit 8e26ac47 authored by Tamar Christina's avatar Tamar Christina
Browse files

AArch64: Fix codegen regressions around tbz.

We were analyzing code quality after recent changes and have noticed that the
tbz support somehow managed to increase the number of branches overall rather
than decreased them.

While investigating this we figured out that the problem is that when an
existing & <contants> exists in gimple and the instruction is generated because
of the range information gotten from the ANDed constant that we end up with the
situation that you get a NOP AND in the RTL expansion.

This is not a problem as CSE will take care of it normally.   The issue is when
this original AND was done in a location where PRE or FRE "lift" the AND to a
different basic block.  This triggers a problem when the resulting value is not
single use.  Instead of having an AND and tbz, we end up generating an
AND + TST + BR if the mode is HI or QI.

This CSE across BB was a problem before but this change made it worse. Our
branch patterns rely on combine being able to fold AND or zero_extends into the
instructions.

To work around this (since a proper fix is outside of the scope of stage-4) we
are limiting the new tbranch optab to only HI and QI mode values.  This isn't a
problem because these two modes are modes for which we don't have CBZ support,
so they are the problematic cases to begin with.  Additionally booleans are QI.

The second thing we're doing is limiting the only legal bitpos to pos 0. i.e.
only the bottom bit.  This such that we prevent the double ANDs as much as
possible.

Now most other cases, i.e. where we had an explicit & in the source code are
still handled correctly by the anonymous (*tb<optab><ALLI:mode><GPI:mode>1)
pattern that was added along with tbranch support.

This means we don't expand the superflous AND here, and while it doesn't fix the
problem that in the cross BB case we loss tbz, it also doesn't make things worse.

With these tweaks we've now reduced the number of insn uniformly was originally
expected.

gcc/ChangeLog:

	* config/aarch64/aarch64.md (tbranch_<code><mode>3): Restrict to SHORT
	and bottom bit only.

gcc/testsuite/ChangeLog:

	* gcc.target/aarch64/tbz_2.c: New test.
	* gcc.target/aarch64/tbz_3.c: New test.
parent 96abc822
No related branches found
No related tags found
Loading
Loading
0% Loading or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment