diff options
| author | Kyrylo Tkachov <kyrylo.tkachov@arm.com> | 2021-01-07 14:02:02 +0000 |
|---|---|---|
| committer | Kyrylo Tkachov <kyrylo.tkachov@arm.com> | 2021-01-08 10:29:25 +0000 |
| commit | c9d25aa7489fd478098e0ef098438e797d597d3b (patch) | |
| tree | 1528daf7cbf593bba5f6952c23e0fad2511495aa | |
| parent | 21c1a30fc73105af50c5e717cb99dc3becabf8fa (diff) | |
| download | gcc-c9d25aa7489fd478098e0ef098438e797d597d3b.zip gcc-c9d25aa7489fd478098e0ef098438e797d597d3b.tar.gz gcc-c9d25aa7489fd478098e0ef098438e797d597d3b.tar.bz2 | |
aarch64: Fix RTL patterns for UABA/SABA
Sometime ago we changed the RTL representation of the (SU)ABD
instructions in RTL to a (MINUS (MAX) (MIN)) rather than a (MINUS (ABS) (ABS))
as it is more correctly models the semantics.
We should do the same for the accumulation forms of these instructions:
UABA/SABA.
This patch does that and allows the new pattern to generate the unsigned
UABA form as well.
The new form also allows it to more easily be re-used to implement the
relevant arm_neon.h intrinsics in the future.
The testcase takes an -fno-tree-reassoc to work around a side-effect of
PR98581.
gcc/
* config/aarch64/aarch64-simd.md (aba<mode>_3): Rename to...
(aarch64_<su>aba<mode>): ... This. Handle uaba as well.
Change RTL pattern to match.
gcc/testsuite/
* gcc.target/aarch64/usaba_1.c: New test.
| -rw-r--r-- | gcc/config/aarch64/aarch64-simd.md | 18 | ||||
| -rw-r--r-- | gcc/testsuite/gcc.target/aarch64/usaba_1.c | 29 |
2 files changed, 40 insertions, 7 deletions
diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md index 85770c8..d23398e 100644 --- a/gcc/config/aarch64/aarch64-simd.md +++ b/gcc/config/aarch64/aarch64-simd.md @@ -859,14 +859,18 @@ } ) -(define_insn "aba<mode>_3" +(define_insn "aarch64_<su>aba<mode>" [(set (match_operand:VDQ_BHSI 0 "register_operand" "=w") - (plus:VDQ_BHSI (abs:VDQ_BHSI (minus:VDQ_BHSI - (match_operand:VDQ_BHSI 1 "register_operand" "w") - (match_operand:VDQ_BHSI 2 "register_operand" "w"))) - (match_operand:VDQ_BHSI 3 "register_operand" "0")))] - "TARGET_SIMD" - "saba\t%0.<Vtype>, %1.<Vtype>, %2.<Vtype>" + (plus:VDQ_BHSI (minus:VDQ_BHSI + (USMAX:VDQ_BHSI + (match_operand:VDQ_BHSI 2 "register_operand" "w") + (match_operand:VDQ_BHSI 3 "register_operand" "w")) + (<max_opp>:VDQ_BHSI + (match_dup 2) + (match_dup 3))) + (match_operand:VDQ_BHSI 1 "register_operand" "0")))] + "TARGET_SIMD" + "<su>aba\t%0.<Vtype>, %2.<Vtype>, %3.<Vtype>" [(set_attr "type" "neon_arith_acc<q>")] ) diff --git a/gcc/testsuite/gcc.target/aarch64/usaba_1.c b/gcc/testsuite/gcc.target/aarch64/usaba_1.c new file mode 100644 index 0000000..58b5beb --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/usaba_1.c @@ -0,0 +1,29 @@ +/* { dg-do compile } */ +/* { dg-options "-O3 -fno-tree-reassoc" } */ + +#pragma GCC target "+nosve" + +#define MIN(a, b) ((a) < (b) ? (a) : (b)) +#define MAX(a, b) ((a) > (b) ? (a) : (b)) + +#define FUNC(T, N, S) \ +void saba_##S (T * __restrict__ a, T * __restrict__ b, T * __restrict__ c) \ +{ \ + int i; \ + for (i = 0; i < N; i++) \ + c[i] += (MAX (a[i], b[i]) - MIN (a[i], b[i])); \ +} + +FUNC (signed char, 16, qi) +/* { dg-final { scan-assembler-times {saba\tv[0-9]+\.16b, v[0-9]+\.16b, v[0-9]+\.16b} 1 } } */ +FUNC (short, 8, hi) +/* { dg-final { scan-assembler-times {saba\tv[0-9]+\.8h, v[0-9]+\.8h, v[0-9]+\.8h} 1 } } */ +FUNC (int, 4, si) +/* { dg-final { scan-assembler-times {saba\tv[0-9]+\.4s, v[0-9]+\.4s, v[0-9]+\.4s} 1 } } */ +FUNC (unsigned char, 16, uqi) +/* { dg-final { scan-assembler-times {uaba\tv[0-9]+\.16b, v[0-9]+\.16b, v[0-9]+.16b} 1 } } */ +FUNC (unsigned short, 8, uhi) +/* { dg-final { scan-assembler-times {uaba\tv[0-9]+\.8h, v[0-9]+\.8h, v[0-9]+\.8h} 1 } } */ +FUNC (unsigned int, 4, usi) +/* { dg-final { scan-assembler-times {uaba\tv[0-9]+\.4s, v[0-9]+\.4s, v[0-9]+\.4s} 1 } } */ + |