diff options
author | Kyrylo Tkachov <kyrylo.tkachov@arm.com> | 2023-05-15 09:55:44 +0100 |
---|---|---|
committer | Kyrylo Tkachov <kyrylo.tkachov@arm.com> | 2023-05-15 09:55:44 +0100 |
commit | e90791e5a02b021d22ffb4c36673b9af623e2063 (patch) | |
tree | b5f51ed43487cd5687023fd889b7c6de8de4634f | |
parent | 676d33f95eede2dfa629c9b0174c15cc55c4a45a (diff) | |
download | gcc-e90791e5a02b021d22ffb4c36673b9af623e2063.zip gcc-e90791e5a02b021d22ffb4c36673b9af623e2063.tar.gz gcc-e90791e5a02b021d22ffb4c36673b9af623e2063.tar.bz2 |
aarch64: PR target/99195 annotate vector compare patterns for vec-concat-zero
This instalment of the series goes through the vector comparison patterns in the backend.
One wart are the int64x1_t comparisons that this patch doesn't touch.
Those are a bit trickier because they have define_insn_and_split mechanisms for falling back to
GP reg comparisons after reload and I don't think a simple annotation will catch those cases correctly.
Those will need more custom thinking.
As said, this patch doesn't touch those and is a decent straightforward improvement on its own.
Bootstrapped and tested on aarch64-none-linux-gnu and aarch64_be-none-elf.
gcc/ChangeLog:
PR target/99195
* config/aarch64/aarch64-simd.md (aarch64_cm<optab><mode>): Rename to...
(aarch64_cm<optab><mode><vczle><vczbe>): ... This.
(aarch64_cmtst<mode>): Rename to...
(aarch64_cmtst<mode><vczle><vczbe>): ... This.
(*aarch64_cmtst_same_<mode>): Rename to...
(*aarch64_cmtst_same_<mode><vczle><vczbe>): ... This.
(*aarch64_cmtstdi): Rename to...
(*aarch64_cmtstdi<vczle><vczbe>): ... This.
(aarch64_fac<optab><mode>): Rename to...
(aarch64_fac<optab><mode><vczle><vczbe>): ... This.
gcc/testsuite/ChangeLog:
PR target/99195
* gcc.target/aarch64/simd/pr99195_7.c: New test.
-rw-r--r-- | gcc/config/aarch64/aarch64-simd.md | 14 | ||||
-rw-r--r-- | gcc/testsuite/gcc.target/aarch64/simd/pr99195_7.c | 96 |
2 files changed, 103 insertions, 7 deletions
diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md index 26320f2..61c815b 100644 --- a/gcc/config/aarch64/aarch64-simd.md +++ b/gcc/config/aarch64/aarch64-simd.md @@ -6615,7 +6615,7 @@ ;; Note, we have constraints for Dz and Z as different expanders ;; have different ideas of what should be passed to this pattern. -(define_insn "aarch64_cm<optab><mode>" +(define_insn "aarch64_cm<optab><mode><vczle><vczbe>" [(set (match_operand:<V_INT_EQUIV> 0 "register_operand" "=w,w") (neg:<V_INT_EQUIV> (COMPARISONS:<V_INT_EQUIV> @@ -6680,7 +6680,7 @@ ;; cm(hs|hi) -(define_insn "aarch64_cm<optab><mode>" +(define_insn "aarch64_cm<optab><mode><vczle><vczbe>" [(set (match_operand:<V_INT_EQUIV> 0 "register_operand" "=w") (neg:<V_INT_EQUIV> (UCOMPARISONS:<V_INT_EQUIV> @@ -6747,7 +6747,7 @@ ;; which is rewritten by simplify_rtx as ;; plus (eq (and x y) 0) -1. -(define_insn "aarch64_cmtst<mode>" +(define_insn "aarch64_cmtst<mode><vczle><vczbe>" [(set (match_operand:<V_INT_EQUIV> 0 "register_operand" "=w") (plus:<V_INT_EQUIV> (eq:<V_INT_EQUIV> @@ -6766,7 +6766,7 @@ ;; not (neq (eq x 0)) in which case you rewrite it to ;; a comparison against itself -(define_insn "*aarch64_cmtst_same_<mode>" +(define_insn "*aarch64_cmtst_same_<mode><vczle><vczbe>" [(set (match_operand:<V_INT_EQUIV> 0 "register_operand" "=w") (plus:<V_INT_EQUIV> (eq:<V_INT_EQUIV> @@ -6817,7 +6817,7 @@ [(set_attr "type" "neon_tst,multiple")] ) -(define_insn "*aarch64_cmtstdi" +(define_insn "*aarch64_cmtstdi<vczle><vczbe>" [(set (match_operand:DI 0 "register_operand" "=w") (neg:DI (ne:DI @@ -6832,7 +6832,7 @@ ;; fcm(eq|ge|gt|le|lt) -(define_insn "aarch64_cm<optab><mode>" +(define_insn "aarch64_cm<optab><mode><vczle><vczbe>" [(set (match_operand:<V_INT_EQUIV> 0 "register_operand" "=w,w") (neg:<V_INT_EQUIV> (COMPARISONS:<V_INT_EQUIV> @@ -6850,7 +6850,7 @@ ;; Note we can also handle what would be fac(le|lt) by ;; generating fac(ge|gt). -(define_insn "aarch64_fac<optab><mode>" +(define_insn "aarch64_fac<optab><mode><vczle><vczbe>" [(set (match_operand:<V_INT_EQUIV> 0 "register_operand" "=w") (neg:<V_INT_EQUIV> (FAC_COMPARISONS:<V_INT_EQUIV> diff --git a/gcc/testsuite/gcc.target/aarch64/simd/pr99195_7.c b/gcc/testsuite/gcc.target/aarch64/simd/pr99195_7.c new file mode 100644 index 0000000..86bd729 --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/simd/pr99195_7.c @@ -0,0 +1,96 @@ +/* PR target/99195. */ +/* Check that we take advantage of 64-bit Advanced SIMD operations clearing + the top half of the vector register and no explicit zeroing instructions + are emitted. */ +/* { dg-do compile } */ +/* { dg-options "-O" } */ + +#include <arm_neon.h> + +#define MYOP(OT,IT,IMT,OP,IS,OS) \ +OT \ +foo_##OP##_##OS##_##IT##_##IS (IT a, IT b) \ +{ \ + IMT zeros = vcreate_##OS (0); \ + return vcombine_##OS (v##OP##_##IS (a, b), zeros); \ +} + +#define FUNC(OT,IT,IMT,IS,OS) \ +MYOP (OT, IT, IMT, ceq, IS, OS) \ +MYOP (OT, IT, IMT, clt, IS, OS) \ +MYOP (OT, IT, IMT, cge, IS, OS) \ +MYOP (OT, IT, IMT, cle, IS, OS) \ +MYOP (OT, IT, IMT, cgt, IS, OS) \ +MYOP (OT, IT, IMT, tst, IS, OS) + +#define MYFUNC(PFX, T, S, N, DN) \ +FUNC (uint##S##x##DN##_t, T##S##x##N##_t, uint##S##x##N##_t, PFX##S, u##S) + +MYFUNC (s, int, 8, 8, 16) +MYFUNC (s, int, 16, 4, 8) +MYFUNC (s, int, 32, 2, 4) +MYFUNC (u, uint, 8, 8, 16) +MYFUNC (u, uint, 16, 4, 8) +MYFUNC (u, uint, 32, 2, 4) + +#undef FUNC +#define FUNC(OT,IT,IMT,IS,OS) \ +MYOP (OT, IT, IMT, ceq, IS, OS) \ +MYOP (OT, IT, IMT, clt, IS, OS) \ +MYOP (OT, IT, IMT, cge, IS, OS) \ +MYOP (OT, IT, IMT, cle, IS, OS) \ +MYOP (OT, IT, IMT, cgt, IS, OS) + + +#pragma GCC push_options +#pragma GCC target ("arch=armv8.2-a+fp16") +MYFUNC (f, float, 16, 4, 8) +#pragma GCC pop_options +MYFUNC (f, float, 32, 2, 4) +MYFUNC (f, float, 64, 1, 2) + +#undef FUNC +#define FUNC(OT,IT,IMT,IS,OS) \ +MYOP (OT, IT, IMT, cale, IS, OS) \ +MYOP (OT, IT, IMT, cagt, IS, OS) \ +MYOP (OT, IT, IMT, calt, IS, OS) \ +MYOP (OT, IT, IMT, cage, IS, OS) \ + +#pragma GCC push_options +#pragma GCC target ("arch=armv8.2-a+fp16") +MYFUNC (f, float, 16, 4, 8) +#pragma GCC pop_options +MYFUNC (f, float, 32, 2, 4) +MYFUNC (f, float, 64, 1, 2) + +#undef MYOP +#define MYOP(OT,IT,IMT,OP,IS,OS) \ +OT \ +foo_##OP##_##OS##_##IT##_z (IT a) \ +{ \ + IMT zeros = vcreate_##OS (0); \ + return vcombine_##OS (v##OP##_##IS (a), zeros); \ +} + +#undef FUNC +#define FUNC(OT,IT,IMT,IS,OS) \ +MYOP (OT, IT, IMT, cltz, IS, OS) \ +MYOP (OT, IT, IMT, ceqz, IS, OS) \ +MYOP (OT, IT, IMT, cgez, IS, OS) \ +MYOP (OT, IT, IMT, cgtz, IS, OS) \ +MYOP (OT, IT, IMT, clez, IS, OS) \ + +MYFUNC (s, int, 8, 8, 16) +MYFUNC (s, int, 16, 4, 8) +MYFUNC (s, int, 32, 2, 4) + +#pragma GCC push_options +#pragma GCC target ("arch=armv8.2-a+fp16") +MYFUNC (f, float, 16, 4, 8) +#pragma GCC pop_options +MYFUNC (f, float, 32, 2, 4) +MYFUNC (f, float, 64, 1, 2) + +/* { dg-final { scan-assembler-not {\tfmov\t} } } */ +/* { dg-final { scan-assembler-not {\tmov\t} } } */ + |