[PATCH][AArch64] Use conditional negate for abs expansion
Wilco Dijkstra
wdijkstr@arm.com
Tue Mar 3 16:19:00 GMT 2015
Expand abs into a compare and conditional negate. This is the most obvious expansion, enables
merging of the comparison into ALU instructions and is faster on all implementations. Bootstrapped &
regression tested.
int f(int x) { return abs (x + 1); }
Before:
add w0, w0, 1
sxtw x0, w0
eor x1, x0, x0, asr 63
sub x1, x1, x0, asr 63
mov x0, x1
ret
After:
adds w0, w0, 1
csneg w0, w0, w0, pl
ret
ChangeLog:
2015-03-03 Wilco Dijkstra <wdijkstr@arm.com>
* gcc/config/aarch64/aarch64.md (absdi2): optimize abs expansion.
(csneg3<mode>_insn): enable expansion of pattern.
* gcc/testsuite/gcc.target/aarch64/abs_1.c (abs64): update test
for new abs expansion. (abs64_in_dreg): likewise.
---
gcc/config/aarch64/aarch64.md | 33 +++++++-------------------------
gcc/testsuite/gcc.target/aarch64/abs_1.c | 5 ++---
2 files changed, 9 insertions(+), 29 deletions(-)
diff --git a/gcc/config/aarch64/aarch64.md b/gcc/config/aarch64/aarch64.md
index 1f4169e..46b7a63 100644
--- a/gcc/config/aarch64/aarch64.md
+++ b/gcc/config/aarch64/aarch64.md
@@ -2172,35 +2172,16 @@
[(set_attr "type" "alu_ext")]
)
-(define_insn_and_split "absdi2"
- [(set (match_operand:DI 0 "register_operand" "=&r,w")
- (abs:DI (match_operand:DI 1 "register_operand" "r,w")))]
+(define_expand "abs<mode>2"
+ [(match_operand:GPI 0 "register_operand" "")
+ (match_operand:GPI 1 "register_operand" "")]
""
- "@
- #
- abs\\t%d0, %d1"
- "reload_completed
- && GP_REGNUM_P (REGNO (operands[0]))
- && GP_REGNUM_P (REGNO (operands[1]))"
- [(const_int 0)]
{
- emit_insn (gen_rtx_SET (VOIDmode, operands[0],
- gen_rtx_XOR (DImode,
- gen_rtx_ASHIFTRT (DImode,
- operands[1],
- GEN_INT (63)),
- operands[1])));
- emit_insn (gen_rtx_SET (VOIDmode,
- operands[0],
- gen_rtx_MINUS (DImode,
- operands[0],
- gen_rtx_ASHIFTRT (DImode,
- operands[1],
- GEN_INT (63)))));
+ rtx ccreg = aarch64_gen_compare_reg (LT, operands[1], const0_rtx);
+ rtx x = gen_rtx_LT (VOIDmode, ccreg, const0_rtx);
+ emit_insn (gen_csneg3<mode>_insn (operands[0], x, operands[1], operands[1]));
DONE;
}
- [(set_attr "type" "alu_sreg")
- (set_attr "simd" "no,yes")]
)
(define_insn "neg<mode>2"
@@ -2879,7 +2860,7 @@
[(set_attr "type" "csel")]
)
-(define_insn "*csneg3<mode>_insn"
+(define_insn "csneg3<mode>_insn"
[(set (match_operand:GPI 0 "register_operand" "=r")
(if_then_else:GPI
(match_operand 1 "aarch64_comparison_operation" "")
diff --git a/gcc/testsuite/gcc.target/aarch64/abs_1.c b/gcc/testsuite/gcc.target/aarch64/abs_1.c
index 938bc84..11f1095 100644
--- a/gcc/testsuite/gcc.target/aarch64/abs_1.c
+++ b/gcc/testsuite/gcc.target/aarch64/abs_1.c
@@ -7,15 +7,14 @@ extern void abort (void);
long long
abs64 (long long a)
{
- /* { dg-final { scan-assembler "eor\t" } } */
- /* { dg-final { scan-assembler "sub\t" } } */
+ /* { dg-final { scan-assembler "csneg\t" } } */
return llabs (a);
}
long long
abs64_in_dreg (long long a)
{
- /* { dg-final { scan-assembler "abs\td\[0-9\]+, d\[0-9\]+" } } */
+ /* { dg-final { scan-assembler "csneg\t" } } */
register long long x asm ("d8") = a;
register long long y asm ("d9");
asm volatile ("" : : "w" (x));
--
1.9.1
More information about the Gcc-patches
mailing list