procyberian/gcc
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

expand: Allow widdening optab when expanding popcount==1 [PR116508]

Browse source 
After adding popcount{qi,hi}2 to the aarch64 backend, I noticed that
the expansion for popcount==1 was no longer trying to do the trick
of handling popcount==1 as `(arg ^ (arg - 1)) > arg - 1`. The problem
is the expansion was using OPTAB_DIRECT, when using OPTAB_WIDEN
will allow modes which are smaller than SImode (in the aarch64 case).

Note QImode's cost still needs some improvements so part of popcnt-eq-1.c
is xfailed. Though there is a check to make sure the costs are compared now.

Built and tested on aarch64-linux-gnu.

	PR middle-end/116508

gcc/ChangeLog:

	* internal-fn.cc (expand_POPCOUNT): Use OPTAB_WIDEN for PLUS and
	XOR/AND expansion.

gcc/testsuite/ChangeLog:

	* gcc.target/aarch64/popcnt-eq-1.c: New test.

Signed-off-by: Andrew Pinski <quic_apinski@quicinc.com>
This commit is contained in:
Andrew Pinski 2024-08-28 15:03:53 -07:00
parent cdd5dd2125
commit 215c7e3084
2 changed files with 47 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
gcc/internal-fn.cc
View file

@ -5332,11 +5332,11 @@ expand_POPCOUNT (internal_fn fn, gcall *stmt)
start_sequence ();
rtx op0 = expand_normal (arg);
rtx argm1 = expand_simple_binop (mode, PLUS, op0, constm1_rtx, NULL_RTX,
1, OPTAB_DIRECT);
1, OPTAB_WIDEN);
if (argm1 == NULL_RTX)
goto fail;
rtx argxorargm1 = expand_simple_binop (mode, nonzero_arg ? AND : XOR, op0,
argm1, NULL_RTX, 1, OPTAB_DIRECT);
argm1, NULL_RTX, 1, OPTAB_WIDEN);
if (argxorargm1 == NULL_RTX)
goto fail;
rtx cmp;

45
gcc/testsuite/gcc.target/aarch64/popcnt-eq-1.c Normal file
View file

@ -0,0 +1,45 @@
/* { dg-do compile } */
/* { dg-options "-O2 -fdump-rtl-expand-details" } */
/* { dg-final { check-function-bodies "**" "" } } */
/* PR middle-end/116508 */
#pragma GCC target "+nocssc"
/*
** h16:
** sub w([0-9]+), w0, #1
** eor w([0-9]+), w0, w\1
** and w([0-9]+), w\1, 65535
** cmp w\3, w\2, uxth
** cset w0, cc
** ret
*/
/* when expanding popcount == 1, should use
`(arg ^ (arg - 1)) > arg - 1` as that has a lower latency
than doing the popcount then comparing against 1.
The popcount/addv can be costly. */
unsigned h16 (const unsigned short a) {
return __builtin_popcountg (a) == 1;
}
/* unsigned char should also do the same trick */
/* Currently xfailed since the cost does not take into account the
moving between gprs and vector regs correctly. */
/*
** h8: { xfail *-*-* }
** sub w([0-9]+), w0, #1
** eor w([0-9]+), w0, w\1
** and w([0-9]+), w\1, 255
** cmp w\3, w\2, uxtb
** cset w0, cc
** ret
*/
unsigned h8 (const unsigned char a) {
return __builtin_popcountg (a) == 1;
}
/* There should be printing out the costs for h8 and h16's popcount == 1 */
/* { dg-final { scan-rtl-dump-times "popcount == 1:" 2 "expand"} } */