From 65a55fbbd949868f7df3c65d39eb8d4455515132 Mon Sep 17 00:00:00 2001 From: Tamar Christina Date: Wed, 28 Jun 2017 11:09:01 +0100 Subject: [PATCH] [AArch64] Add dot product support for AArch64 to binutils gas/ * config/tc-aarch64.c (aarch64_reg_parse_32_64): Accept 4B. (aarch64_features): Added dotprod. * doc/c-aarch64.texi: Added dotprod. * testsuite/gas/aarch64/dotproduct.d: New. * testsuite/gas/aarch64/dotproduct.s: New. opcodes/ * aarch64-asm.c (aarch64_ins_reglane): Added 4B dotprod. * aarch64-dis.c (aarch64_ext_reglane): Likewise. * aarch64-tbl.h (QL_V3DOT, QL_V2DOT): New. (aarch64_feature_dotprod, DOT_INSN): New. (udot, sdot): New. * aarch64-dis-2.c: Regenerated. include/ * opcode/aarch64.h: (AARCH64_FEATURE_DOTPROD): New. (aarch64_insn_class): Added dotprod. --- gas/ChangeLog | 8 + gas/config/tc-aarch64.c | 10 +- gas/doc/c-aarch64.texi | 2 + gas/testsuite/gas/aarch64/dotproduct.d | 548 +++++++++++++++++++++++++ gas/testsuite/gas/aarch64/dotproduct.s | 540 ++++++++++++++++++++++++ include/ChangeLog | 5 + include/opcode/aarch64.h | 2 + opcodes/ChangeLog | 9 + opcodes/aarch64-asm.c | 14 + opcodes/aarch64-dis-2.c | 382 +++++++++-------- opcodes/aarch64-dis.c | 15 + opcodes/aarch64-tbl.h | 24 ++ 12 files changed, 1377 insertions(+), 182 deletions(-) create mode 100644 gas/testsuite/gas/aarch64/dotproduct.d create mode 100644 gas/testsuite/gas/aarch64/dotproduct.s diff --git a/gas/ChangeLog b/gas/ChangeLog index b149e5e2c9e..92d1f86f0d6 100644 --- a/gas/ChangeLog +++ b/gas/ChangeLog @@ -1,3 +1,11 @@ +2017-06-28 Tamar Christina + + * config/tc-aarch64.c (aarch64_reg_parse_32_64): Accept 4B. + (aarch64_features): Added dotprod. + * doc/c-aarch64.texi: Added dotprod. + * testsuite/gas/aarch64/dotproduct.d: New. + * testsuite/gas/aarch64/dotproduct.s: New. + 2017-06-28 Jiong Wang * config/tc-arm.c (fpu_neon_ext_dotprod): New variable. diff --git a/gas/config/tc-aarch64.c b/gas/config/tc-aarch64.c index fae29b4f47e..f095ab99284 100644 --- a/gas/config/tc-aarch64.c +++ b/gas/config/tc-aarch64.c @@ -798,7 +798,7 @@ aarch64_reg_parse_32_64 (char **ccp, aarch64_opnd_qualifier_t *qualifier) succeeds; otherwise return FALSE. Accept only one occurrence of: - 8b 16b 2h 4h 8h 2s 4s 1d 2d + 4b 8b 16b 2h 4h 8h 2s 4s 1d 2d b h s d q */ static bfd_boolean parse_vector_type_for_operand (aarch64_reg_type reg_type, @@ -859,8 +859,10 @@ elt_size: first_error (_("missing element size")); return FALSE; } - if (width != 0 && width * element_size != 64 && width * element_size != 128 - && !(width == 2 && element_size == 16)) + if (width != 0 && width * element_size != 64 + && width * element_size != 128 + && !(width == 2 && element_size == 16) + && !(width == 4 && element_size == 8)) { first_error_fmt (_ ("invalid element size %d and vector size combination %c"), @@ -8485,6 +8487,8 @@ static const struct aarch64_option_cpu_value_table aarch64_features[] = { | AARCH64_FEATURE_SIMD, 0)}, {"rcpc", AARCH64_FEATURE (AARCH64_FEATURE_RCPC, 0), AARCH64_ARCH_NONE}, + {"dotprod", AARCH64_FEATURE (AARCH64_FEATURE_DOTPROD, 0), + AARCH64_ARCH_NONE}, {NULL, AARCH64_ARCH_NONE, AARCH64_ARCH_NONE}, }; diff --git a/gas/doc/c-aarch64.texi b/gas/doc/c-aarch64.texi index 24775553cda..677420594e1 100644 --- a/gas/doc/c-aarch64.texi +++ b/gas/doc/c-aarch64.texi @@ -165,6 +165,8 @@ automatically cause those extensions to be disabled. @item @code{sve} @tab ARMv8.2-A @tab No @tab Enable the Scalable Vector Extensions. This implies @code{fp16}, @code{simd} and @code{compnum}. +@item @code{dotprod} @tab ARMv8.2-A @tab No + @tab Enable the Dot Product extension. This implies @code{simd}. @end multitable @node AArch64 Syntax diff --git a/gas/testsuite/gas/aarch64/dotproduct.d b/gas/testsuite/gas/aarch64/dotproduct.d new file mode 100644 index 00000000000..58669ff3476 --- /dev/null +++ b/gas/testsuite/gas/aarch64/dotproduct.d @@ -0,0 +1,548 @@ +#as: -march=armv8.2-a+dotprod +#objdump: -dr + +.*: file format .* + +Disassembly of section \.text: + +0+ <.*>: +[^:]+:\s+2e809400 udot v0.2s, v0.8b, v0.8b +[^:]+:\s+2e8b9400 udot v0.2s, v0.8b, v11.8b +[^:]+:\s+2e969400 udot v0.2s, v0.8b, v22.8b +[^:]+:\s+2e809560 udot v0.2s, v11.8b, v0.8b +[^:]+:\s+2e8b9560 udot v0.2s, v11.8b, v11.8b +[^:]+:\s+2e969560 udot v0.2s, v11.8b, v22.8b +[^:]+:\s+2e8096c0 udot v0.2s, v22.8b, v0.8b +[^:]+:\s+2e8b96c0 udot v0.2s, v22.8b, v11.8b +[^:]+:\s+2e9696c0 udot v0.2s, v22.8b, v22.8b +[^:]+:\s+2e80940b udot v11.2s, v0.8b, v0.8b +[^:]+:\s+2e8b940b udot v11.2s, v0.8b, v11.8b +[^:]+:\s+2e96940b udot v11.2s, v0.8b, v22.8b +[^:]+:\s+2e80956b udot v11.2s, v11.8b, v0.8b +[^:]+:\s+2e8b956b udot v11.2s, v11.8b, v11.8b +[^:]+:\s+2e96956b udot v11.2s, v11.8b, v22.8b +[^:]+:\s+2e8096cb udot v11.2s, v22.8b, v0.8b +[^:]+:\s+2e8b96cb udot v11.2s, v22.8b, v11.8b +[^:]+:\s+2e9696cb udot v11.2s, v22.8b, v22.8b +[^:]+:\s+2e809416 udot v22.2s, v0.8b, v0.8b +[^:]+:\s+2e8b9416 udot v22.2s, v0.8b, v11.8b +[^:]+:\s+2e969416 udot v22.2s, v0.8b, v22.8b +[^:]+:\s+2e809576 udot v22.2s, v11.8b, v0.8b +[^:]+:\s+2e8b9576 udot v22.2s, v11.8b, v11.8b +[^:]+:\s+2e969576 udot v22.2s, v11.8b, v22.8b +[^:]+:\s+2e8096d6 udot v22.2s, v22.8b, v0.8b +[^:]+:\s+2e8b96d6 udot v22.2s, v22.8b, v11.8b +[^:]+:\s+2e9696d6 udot v22.2s, v22.8b, v22.8b +[^:]+:\s+0e809400 sdot v0.2s, v0.8b, v0.8b +[^:]+:\s+0e8b9400 sdot v0.2s, v0.8b, v11.8b +[^:]+:\s+0e969400 sdot v0.2s, v0.8b, v22.8b +[^:]+:\s+0e809560 sdot v0.2s, v11.8b, v0.8b +[^:]+:\s+0e8b9560 sdot v0.2s, v11.8b, v11.8b +[^:]+:\s+0e969560 sdot v0.2s, v11.8b, v22.8b +[^:]+:\s+0e8096c0 sdot v0.2s, v22.8b, v0.8b +[^:]+:\s+0e8b96c0 sdot v0.2s, v22.8b, v11.8b +[^:]+:\s+0e9696c0 sdot v0.2s, v22.8b, v22.8b +[^:]+:\s+0e80940b sdot v11.2s, v0.8b, v0.8b +[^:]+:\s+0e8b940b sdot v11.2s, v0.8b, v11.8b +[^:]+:\s+0e96940b sdot v11.2s, v0.8b, v22.8b +[^:]+:\s+0e80956b sdot v11.2s, v11.8b, v0.8b +[^:]+:\s+0e8b956b sdot v11.2s, v11.8b, v11.8b +[^:]+:\s+0e96956b sdot v11.2s, v11.8b, v22.8b +[^:]+:\s+0e8096cb sdot v11.2s, v22.8b, v0.8b +[^:]+:\s+0e8b96cb sdot v11.2s, v22.8b, v11.8b +[^:]+:\s+0e9696cb sdot v11.2s, v22.8b, v22.8b +[^:]+:\s+0e809416 sdot v22.2s, v0.8b, v0.8b +[^:]+:\s+0e8b9416 sdot v22.2s, v0.8b, v11.8b +[^:]+:\s+0e969416 sdot v22.2s, v0.8b, v22.8b +[^:]+:\s+0e809576 sdot v22.2s, v11.8b, v0.8b +[^:]+:\s+0e8b9576 sdot v22.2s, v11.8b, v11.8b +[^:]+:\s+0e969576 sdot v22.2s, v11.8b, v22.8b +[^:]+:\s+0e8096d6 sdot v22.2s, v22.8b, v0.8b +[^:]+:\s+0e8b96d6 sdot v22.2s, v22.8b, v11.8b +[^:]+:\s+0e9696d6 sdot v22.2s, v22.8b, v22.8b +[^:]+:\s+6e809400 udot v0.4s, v0.16b, v0.16b +[^:]+:\s+6e8b9400 udot v0.4s, v0.16b, v11.16b +[^:]+:\s+6e969400 udot v0.4s, v0.16b, v22.16b +[^:]+:\s+6e809560 udot v0.4s, v11.16b, v0.16b +[^:]+:\s+6e8b9560 udot v0.4s, v11.16b, v11.16b +[^:]+:\s+6e969560 udot v0.4s, v11.16b, v22.16b +[^:]+:\s+6e8096c0 udot v0.4s, v22.16b, v0.16b +[^:]+:\s+6e8b96c0 udot v0.4s, v22.16b, v11.16b +[^:]+:\s+6e9696c0 udot v0.4s, v22.16b, v22.16b +[^:]+:\s+6e80940b udot v11.4s, v0.16b, v0.16b +[^:]+:\s+6e8b940b udot v11.4s, v0.16b, v11.16b +[^:]+:\s+6e96940b udot v11.4s, v0.16b, v22.16b +[^:]+:\s+6e80956b udot v11.4s, v11.16b, v0.16b +[^:]+:\s+6e8b956b udot v11.4s, v11.16b, v11.16b +[^:]+:\s+6e96956b udot v11.4s, v11.16b, v22.16b +[^:]+:\s+6e8096cb udot v11.4s, v22.16b, v0.16b +[^:]+:\s+6e8b96cb udot v11.4s, v22.16b, v11.16b +[^:]+:\s+6e9696cb udot v11.4s, v22.16b, v22.16b +[^:]+:\s+6e809416 udot v22.4s, v0.16b, v0.16b +[^:]+:\s+6e8b9416 udot v22.4s, v0.16b, v11.16b +[^:]+:\s+6e969416 udot v22.4s, v0.16b, v22.16b +[^:]+:\s+6e809576 udot v22.4s, v11.16b, v0.16b +[^:]+:\s+6e8b9576 udot v22.4s, v11.16b, v11.16b +[^:]+:\s+6e969576 udot v22.4s, v11.16b, v22.16b +[^:]+:\s+6e8096d6 udot v22.4s, v22.16b, v0.16b +[^:]+:\s+6e8b96d6 udot v22.4s, v22.16b, v11.16b +[^:]+:\s+6e9696d6 udot v22.4s, v22.16b, v22.16b +[^:]+:\s+4e809400 sdot v0.4s, v0.16b, v0.16b +[^:]+:\s+4e8b9400 sdot v0.4s, v0.16b, v11.16b +[^:]+:\s+4e969400 sdot v0.4s, v0.16b, v22.16b +[^:]+:\s+4e809560 sdot v0.4s, v11.16b, v0.16b +[^:]+:\s+4e8b9560 sdot v0.4s, v11.16b, v11.16b +[^:]+:\s+4e969560 sdot v0.4s, v11.16b, v22.16b +[^:]+:\s+4e8096c0 sdot v0.4s, v22.16b, v0.16b +[^:]+:\s+4e8b96c0 sdot v0.4s, v22.16b, v11.16b +[^:]+:\s+4e9696c0 sdot v0.4s, v22.16b, v22.16b +[^:]+:\s+4e80940b sdot v11.4s, v0.16b, v0.16b +[^:]+:\s+4e8b940b sdot v11.4s, v0.16b, v11.16b +[^:]+:\s+4e96940b sdot v11.4s, v0.16b, v22.16b +[^:]+:\s+4e80956b sdot v11.4s, v11.16b, v0.16b +[^:]+:\s+4e8b956b sdot v11.4s, v11.16b, v11.16b +[^:]+:\s+4e96956b sdot v11.4s, v11.16b, v22.16b +[^:]+:\s+4e8096cb sdot v11.4s, v22.16b, v0.16b +[^:]+:\s+4e8b96cb sdot v11.4s, v22.16b, v11.16b +[^:]+:\s+4e9696cb sdot v11.4s, v22.16b, v22.16b +[^:]+:\s+4e809416 sdot v22.4s, v0.16b, v0.16b +[^:]+:\s+4e8b9416 sdot v22.4s, v0.16b, v11.16b +[^:]+:\s+4e969416 sdot v22.4s, v0.16b, v22.16b +[^:]+:\s+4e809576 sdot v22.4s, v11.16b, v0.16b +[^:]+:\s+4e8b9576 sdot v22.4s, v11.16b, v11.16b +[^:]+:\s+4e969576 sdot v22.4s, v11.16b, v22.16b +[^:]+:\s+4e8096d6 sdot v22.4s, v22.16b, v0.16b +[^:]+:\s+4e8b96d6 sdot v22.4s, v22.16b, v11.16b +[^:]+:\s+4e9696d6 sdot v22.4s, v22.16b, v22.16b +[^:]+:\s+2f80e000 udot v0.2s, v0.8b, v0.b\[0\] +[^:]+:\s+2f8be000 udot v0.2s, v0.8b, v11.b\[0\] +[^:]+:\s+2f96e000 udot v0.2s, v0.8b, v22.b\[0\] +[^:]+:\s+2fa0e000 udot v0.2s, v0.8b, v0.b\[1\] +[^:]+:\s+2fabe000 udot v0.2s, v0.8b, v11.b\[1\] +[^:]+:\s+2fb6e000 udot v0.2s, v0.8b, v22.b\[1\] +[^:]+:\s+2f80e800 udot v0.2s, v0.8b, v0.b\[2\] +[^:]+:\s+2f8be800 udot v0.2s, v0.8b, v11.b\[2\] +[^:]+:\s+2f96e800 udot v0.2s, v0.8b, v22.b\[2\] +[^:]+:\s+2fa0e800 udot v0.2s, v0.8b, v0.b\[3\] +[^:]+:\s+2fabe800 udot v0.2s, v0.8b, v11.b\[3\] +[^:]+:\s+2fb6e800 udot v0.2s, v0.8b, v22.b\[3\] +[^:]+:\s+2f80e160 udot v0.2s, v11.8b, v0.b\[0\] +[^:]+:\s+2f8be160 udot v0.2s, v11.8b, v11.b\[0\] +[^:]+:\s+2f96e160 udot v0.2s, v11.8b, v22.b\[0\] +[^:]+:\s+2fa0e160 udot v0.2s, v11.8b, v0.b\[1\] +[^:]+:\s+2fabe160 udot v0.2s, v11.8b, v11.b\[1\] +[^:]+:\s+2fb6e160 udot v0.2s, v11.8b, v22.b\[1\] +[^:]+:\s+2f80e960 udot v0.2s, v11.8b, v0.b\[2\] +[^:]+:\s+2f8be960 udot v0.2s, v11.8b, v11.b\[2\] +[^:]+:\s+2f96e960 udot v0.2s, v11.8b, v22.b\[2\] +[^:]+:\s+2fa0e960 udot v0.2s, v11.8b, v0.b\[3\] +[^:]+:\s+2fabe960 udot v0.2s, v11.8b, v11.b\[3\] +[^:]+:\s+2fb6e960 udot v0.2s, v11.8b, v22.b\[3\] +[^:]+:\s+2f80e2c0 udot v0.2s, v22.8b, v0.b\[0\] +[^:]+:\s+2f8be2c0 udot v0.2s, v22.8b, v11.b\[0\] +[^:]+:\s+2f96e2c0 udot v0.2s, v22.8b, v22.b\[0\] +[^:]+:\s+2fa0e2c0 udot v0.2s, v22.8b, v0.b\[1\] +[^:]+:\s+2fabe2c0 udot v0.2s, v22.8b, v11.b\[1\] +[^:]+:\s+2fb6e2c0 udot v0.2s, v22.8b, v22.b\[1\] +[^:]+:\s+2f80eac0 udot v0.2s, v22.8b, v0.b\[2\] +[^:]+:\s+2f8beac0 udot v0.2s, v22.8b, v11.b\[2\] +[^:]+:\s+2f96eac0 udot v0.2s, v22.8b, v22.b\[2\] +[^:]+:\s+2fa0eac0 udot v0.2s, v22.8b, v0.b\[3\] +[^:]+:\s+2fabeac0 udot v0.2s, v22.8b, v11.b\[3\] +[^:]+:\s+2fb6eac0 udot v0.2s, v22.8b, v22.b\[3\] +[^:]+:\s+2f80e00b udot v11.2s, v0.8b, v0.b\[0\] +[^:]+:\s+2f8be00b udot v11.2s, v0.8b, v11.b\[0\] +[^:]+:\s+2f96e00b udot v11.2s, v0.8b, v22.b\[0\] +[^:]+:\s+2fa0e00b udot v11.2s, v0.8b, v0.b\[1\] +[^:]+:\s+2fabe00b udot v11.2s, v0.8b, v11.b\[1\] +[^:]+:\s+2fb6e00b udot v11.2s, v0.8b, v22.b\[1\] +[^:]+:\s+2f80e80b udot v11.2s, v0.8b, v0.b\[2\] +[^:]+:\s+2f8be80b udot v11.2s, v0.8b, v11.b\[2\] +[^:]+:\s+2f96e80b udot v11.2s, v0.8b, v22.b\[2\] +[^:]+:\s+2fa0e80b udot v11.2s, v0.8b, v0.b\[3\] +[^:]+:\s+2fabe80b udot v11.2s, v0.8b, v11.b\[3\] +[^:]+:\s+2fb6e80b udot v11.2s, v0.8b, v22.b\[3\] +[^:]+:\s+2f80e16b udot v11.2s, v11.8b, v0.b\[0\] +[^:]+:\s+2f8be16b udot v11.2s, v11.8b, v11.b\[0\] +[^:]+:\s+2f96e16b udot v11.2s, v11.8b, v22.b\[0\] +[^:]+:\s+2fa0e16b udot v11.2s, v11.8b, v0.b\[1\] +[^:]+:\s+2fabe16b udot v11.2s, v11.8b, v11.b\[1\] +[^:]+:\s+2fb6e16b udot v11.2s, v11.8b, v22.b\[1\] +[^:]+:\s+2f80e96b udot v11.2s, v11.8b, v0.b\[2\] +[^:]+:\s+2f8be96b udot v11.2s, v11.8b, v11.b\[2\] +[^:]+:\s+2f96e96b udot v11.2s, v11.8b, v22.b\[2\] +[^:]+:\s+2fa0e96b udot v11.2s, v11.8b, v0.b\[3\] +[^:]+:\s+2fabe96b udot v11.2s, v11.8b, v11.b\[3\] +[^:]+:\s+2fb6e96b udot v11.2s, v11.8b, v22.b\[3\] +[^:]+:\s+2f80e2cb udot v11.2s, v22.8b, v0.b\[0\] +[^:]+:\s+2f8be2cb udot v11.2s, v22.8b, v11.b\[0\] +[^:]+:\s+2f96e2cb udot v11.2s, v22.8b, v22.b\[0\] +[^:]+:\s+2fa0e2cb udot v11.2s, v22.8b, v0.b\[1\] +[^:]+:\s+2fabe2cb udot v11.2s, v22.8b, v11.b\[1\] +[^:]+:\s+2fb6e2cb udot v11.2s, v22.8b, v22.b\[1\] +[^:]+:\s+2f80eacb udot v11.2s, v22.8b, v0.b\[2\] +[^:]+:\s+2f8beacb udot v11.2s, v22.8b, v11.b\[2\] +[^:]+:\s+2f96eacb udot v11.2s, v22.8b, v22.b\[2\] +[^:]+:\s+2fa0eacb udot v11.2s, v22.8b, v0.b\[3\] +[^:]+:\s+2fabeacb udot v11.2s, v22.8b, v11.b\[3\] +[^:]+:\s+2fb6eacb udot v11.2s, v22.8b, v22.b\[3\] +[^:]+:\s+2f80e016 udot v22.2s, v0.8b, v0.b\[0\] +[^:]+:\s+2f8be016 udot v22.2s, v0.8b, v11.b\[0\] +[^:]+:\s+2f96e016 udot v22.2s, v0.8b, v22.b\[0\] +[^:]+:\s+2fa0e016 udot v22.2s, v0.8b, v0.b\[1\] +[^:]+:\s+2fabe016 udot v22.2s, v0.8b, v11.b\[1\] +[^:]+:\s+2fb6e016 udot v22.2s, v0.8b, v22.b\[1\] +[^:]+:\s+2f80e816 udot v22.2s, v0.8b, v0.b\[2\] +[^:]+:\s+2f8be816 udot v22.2s, v0.8b, v11.b\[2\] +[^:]+:\s+2f96e816 udot v22.2s, v0.8b, v22.b\[2\] +[^:]+:\s+2fa0e816 udot v22.2s, v0.8b, v0.b\[3\] +[^:]+:\s+2fabe816 udot v22.2s, v0.8b, v11.b\[3\] +[^:]+:\s+2fb6e816 udot v22.2s, v0.8b, v22.b\[3\] +[^:]+:\s+2f80e176 udot v22.2s, v11.8b, v0.b\[0\] +[^:]+:\s+2f8be176 udot v22.2s, v11.8b, v11.b\[0\] +[^:]+:\s+2f96e176 udot v22.2s, v11.8b, v22.b\[0\] +[^:]+:\s+2fa0e176 udot v22.2s, v11.8b, v0.b\[1\] +[^:]+:\s+2fabe176 udot v22.2s, v11.8b, v11.b\[1\] +[^:]+:\s+2fb6e176 udot v22.2s, v11.8b, v22.b\[1\] +[^:]+:\s+2f80e976 udot v22.2s, v11.8b, v0.b\[2\] +[^:]+:\s+2f8be976 udot v22.2s, v11.8b, v11.b\[2\] +[^:]+:\s+2f96e976 udot v22.2s, v11.8b, v22.b\[2\] +[^:]+:\s+2fa0e976 udot v22.2s, v11.8b, v0.b\[3\] +[^:]+:\s+2fabe976 udot v22.2s, v11.8b, v11.b\[3\] +[^:]+:\s+2fb6e976 udot v22.2s, v11.8b, v22.b\[3\] +[^:]+:\s+2f80e2d6 udot v22.2s, v22.8b, v0.b\[0\] +[^:]+:\s+2f8be2d6 udot v22.2s, v22.8b, v11.b\[0\] +[^:]+:\s+2f96e2d6 udot v22.2s, v22.8b, v22.b\[0\] +[^:]+:\s+2fa0e2d6 udot v22.2s, v22.8b, v0.b\[1\] +[^:]+:\s+2fabe2d6 udot v22.2s, v22.8b, v11.b\[1\] +[^:]+:\s+2fb6e2d6 udot v22.2s, v22.8b, v22.b\[1\] +[^:]+:\s+2f80ead6 udot v22.2s, v22.8b, v0.b\[2\] +[^:]+:\s+2f8bead6 udot v22.2s, v22.8b, v11.b\[2\] +[^:]+:\s+2f96ead6 udot v22.2s, v22.8b, v22.b\[2\] +[^:]+:\s+2fa0ead6 udot v22.2s, v22.8b, v0.b\[3\] +[^:]+:\s+2fabead6 udot v22.2s, v22.8b, v11.b\[3\] +[^:]+:\s+2fb6ead6 udot v22.2s, v22.8b, v22.b\[3\] +[^:]+:\s+0f80e000 sdot v0.2s, v0.8b, v0.b\[0\] +[^:]+:\s+0f8be000 sdot v0.2s, v0.8b, v11.b\[0\] +[^:]+:\s+0f96e000 sdot v0.2s, v0.8b, v22.b\[0\] +[^:]+:\s+0fa0e000 sdot v0.2s, v0.8b, v0.b\[1\] +[^:]+:\s+0fabe000 sdot v0.2s, v0.8b, v11.b\[1\] +[^:]+:\s+0fb6e000 sdot v0.2s, v0.8b, v22.b\[1\] +[^:]+:\s+0f80e800 sdot v0.2s, v0.8b, v0.b\[2\] +[^:]+:\s+0f8be800 sdot v0.2s, v0.8b, v11.b\[2\] +[^:]+:\s+0f96e800 sdot v0.2s, v0.8b, v22.b\[2\] +[^:]+:\s+0fa0e800 sdot v0.2s, v0.8b, v0.b\[3\] +[^:]+:\s+0fabe800 sdot v0.2s, v0.8b, v11.b\[3\] +[^:]+:\s+0fb6e800 sdot v0.2s, v0.8b, v22.b\[3\] +[^:]+:\s+0f80e160 sdot v0.2s, v11.8b, v0.b\[0\] +[^:]+:\s+0f8be160 sdot v0.2s, v11.8b, v11.b\[0\] +[^:]+:\s+0f96e160 sdot v0.2s, v11.8b, v22.b\[0\] +[^:]+:\s+0fa0e160 sdot v0.2s, v11.8b, v0.b\[1\] +[^:]+:\s+0fabe160 sdot v0.2s, v11.8b, v11.b\[1\] +[^:]+:\s+0fb6e160 sdot v0.2s, v11.8b, v22.b\[1\] +[^:]+:\s+0f80e960 sdot v0.2s, v11.8b, v0.b\[2\] +[^:]+:\s+0f8be960 sdot v0.2s, v11.8b, v11.b\[2\] +[^:]+:\s+0f96e960 sdot v0.2s, v11.8b, v22.b\[2\] +[^:]+:\s+0fa0e960 sdot v0.2s, v11.8b, v0.b\[3\] +[^:]+:\s+0fabe960 sdot v0.2s, v11.8b, v11.b\[3\] +[^:]+:\s+0fb6e960 sdot v0.2s, v11.8b, v22.b\[3\] +[^:]+:\s+0f80e2c0 sdot v0.2s, v22.8b, v0.b\[0\] +[^:]+:\s+0f8be2c0 sdot v0.2s, v22.8b, v11.b\[0\] +[^:]+:\s+0f96e2c0 sdot v0.2s, v22.8b, v22.b\[0\] +[^:]+:\s+0fa0e2c0 sdot v0.2s, v22.8b, v0.b\[1\] +[^:]+:\s+0fabe2c0 sdot v0.2s, v22.8b, v11.b\[1\] +[^:]+:\s+0fb6e2c0 sdot v0.2s, v22.8b, v22.b\[1\] +[^:]+:\s+0f80eac0 sdot v0.2s, v22.8b, v0.b\[2\] +[^:]+:\s+0f8beac0 sdot v0.2s, v22.8b, v11.b\[2\] +[^:]+:\s+0f96eac0 sdot v0.2s, v22.8b, v22.b\[2\] +[^:]+:\s+0fa0eac0 sdot v0.2s, v22.8b, v0.b\[3\] +[^:]+:\s+0fabeac0 sdot v0.2s, v22.8b, v11.b\[3\] +[^:]+:\s+0fb6eac0 sdot v0.2s, v22.8b, v22.b\[3\] +[^:]+:\s+0f80e00b sdot v11.2s, v0.8b, v0.b\[0\] +[^:]+:\s+0f8be00b sdot v11.2s, v0.8b, v11.b\[0\] +[^:]+:\s+0f96e00b sdot v11.2s, v0.8b, v22.b\[0\] +[^:]+:\s+0fa0e00b sdot v11.2s, v0.8b, v0.b\[1\] +[^:]+:\s+0fabe00b sdot v11.2s, v0.8b, v11.b\[1\] +[^:]+:\s+0fb6e00b sdot v11.2s, v0.8b, v22.b\[1\] +[^:]+:\s+0f80e80b sdot v11.2s, v0.8b, v0.b\[2\] +[^:]+:\s+0f8be80b sdot v11.2s, v0.8b, v11.b\[2\] +[^:]+:\s+0f96e80b sdot v11.2s, v0.8b, v22.b\[2\] +[^:]+:\s+0fa0e80b sdot v11.2s, v0.8b, v0.b\[3\] +[^:]+:\s+0fabe80b sdot v11.2s, v0.8b, v11.b\[3\] +[^:]+:\s+0fb6e80b sdot v11.2s, v0.8b, v22.b\[3\] +[^:]+:\s+0f80e16b sdot v11.2s, v11.8b, v0.b\[0\] +[^:]+:\s+0f8be16b sdot v11.2s, v11.8b, v11.b\[0\] +[^:]+:\s+0f96e16b sdot v11.2s, v11.8b, v22.b\[0\] +[^:]+:\s+0fa0e16b sdot v11.2s, v11.8b, v0.b\[1\] +[^:]+:\s+0fabe16b sdot v11.2s, v11.8b, v11.b\[1\] +[^:]+:\s+0fb6e16b sdot v11.2s, v11.8b, v22.b\[1\] +[^:]+:\s+0f80e96b sdot v11.2s, v11.8b, v0.b\[2\] +[^:]+:\s+0f8be96b sdot v11.2s, v11.8b, v11.b\[2\] +[^:]+:\s+0f96e96b sdot v11.2s, v11.8b, v22.b\[2\] +[^:]+:\s+0fa0e96b sdot v11.2s, v11.8b, v0.b\[3\] +[^:]+:\s+0fabe96b sdot v11.2s, v11.8b, v11.b\[3\] +[^:]+:\s+0fb6e96b sdot v11.2s, v11.8b, v22.b\[3\] +[^:]+:\s+0f80e2cb sdot v11.2s, v22.8b, v0.b\[0\] +[^:]+:\s+0f8be2cb sdot v11.2s, v22.8b, v11.b\[0\] +[^:]+:\s+0f96e2cb sdot v11.2s, v22.8b, v22.b\[0\] +[^:]+:\s+0fa0e2cb sdot v11.2s, v22.8b, v0.b\[1\] +[^:]+:\s+0fabe2cb sdot v11.2s, v22.8b, v11.b\[1\] +[^:]+:\s+0fb6e2cb sdot v11.2s, v22.8b, v22.b\[1\] +[^:]+:\s+0f80eacb sdot v11.2s, v22.8b, v0.b\[2\] +[^:]+:\s+0f8beacb sdot v11.2s, v22.8b, v11.b\[2\] +[^:]+:\s+0f96eacb sdot v11.2s, v22.8b, v22.b\[2\] +[^:]+:\s+0fa0eacb sdot v11.2s, v22.8b, v0.b\[3\] +[^:]+:\s+0fabeacb sdot v11.2s, v22.8b, v11.b\[3\] +[^:]+:\s+0fb6eacb sdot v11.2s, v22.8b, v22.b\[3\] +[^:]+:\s+0f80e016 sdot v22.2s, v0.8b, v0.b\[0\] +[^:]+:\s+0f8be016 sdot v22.2s, v0.8b, v11.b\[0\] +[^:]+:\s+0f96e016 sdot v22.2s, v0.8b, v22.b\[0\] +[^:]+:\s+0fa0e016 sdot v22.2s, v0.8b, v0.b\[1\] +[^:]+:\s+0fabe016 sdot v22.2s, v0.8b, v11.b\[1\] +[^:]+:\s+0fb6e016 sdot v22.2s, v0.8b, v22.b\[1\] +[^:]+:\s+0f80e816 sdot v22.2s, v0.8b, v0.b\[2\] +[^:]+:\s+0f8be816 sdot v22.2s, v0.8b, v11.b\[2\] +[^:]+:\s+0f96e816 sdot v22.2s, v0.8b, v22.b\[2\] +[^:]+:\s+0fa0e816 sdot v22.2s, v0.8b, v0.b\[3\] +[^:]+:\s+0fabe816 sdot v22.2s, v0.8b, v11.b\[3\] +[^:]+:\s+0fb6e816 sdot v22.2s, v0.8b, v22.b\[3\] +[^:]+:\s+0f80e176 sdot v22.2s, v11.8b, v0.b\[0\] +[^:]+:\s+0f8be176 sdot v22.2s, v11.8b, v11.b\[0\] +[^:]+:\s+0f96e176 sdot v22.2s, v11.8b, v22.b\[0\] +[^:]+:\s+0fa0e176 sdot v22.2s, v11.8b, v0.b\[1\] +[^:]+:\s+0fabe176 sdot v22.2s, v11.8b, v11.b\[1\] +[^:]+:\s+0fb6e176 sdot v22.2s, v11.8b, v22.b\[1\] +[^:]+:\s+0f80e976 sdot v22.2s, v11.8b, v0.b\[2\] +[^:]+:\s+0f8be976 sdot v22.2s, v11.8b, v11.b\[2\] +[^:]+:\s+0f96e976 sdot v22.2s, v11.8b, v22.b\[2\] +[^:]+:\s+0fa0e976 sdot v22.2s, v11.8b, v0.b\[3\] +[^:]+:\s+0fabe976 sdot v22.2s, v11.8b, v11.b\[3\] +[^:]+:\s+0fb6e976 sdot v22.2s, v11.8b, v22.b\[3\] +[^:]+:\s+0f80e2d6 sdot v22.2s, v22.8b, v0.b\[0\] +[^:]+:\s+0f8be2d6 sdot v22.2s, v22.8b, v11.b\[0\] +[^:]+:\s+0f96e2d6 sdot v22.2s, v22.8b, v22.b\[0\] +[^:]+:\s+0fa0e2d6 sdot v22.2s, v22.8b, v0.b\[1\] +[^:]+:\s+0fabe2d6 sdot v22.2s, v22.8b, v11.b\[1\] +[^:]+:\s+0fb6e2d6 sdot v22.2s, v22.8b, v22.b\[1\] +[^:]+:\s+0f80ead6 sdot v22.2s, v22.8b, v0.b\[2\] +[^:]+:\s+0f8bead6 sdot v22.2s, v22.8b, v11.b\[2\] +[^:]+:\s+0f96ead6 sdot v22.2s, v22.8b, v22.b\[2\] +[^:]+:\s+0fa0ead6 sdot v22.2s, v22.8b, v0.b\[3\] +[^:]+:\s+0fabead6 sdot v22.2s, v22.8b, v11.b\[3\] +[^:]+:\s+0fb6ead6 sdot v22.2s, v22.8b, v22.b\[3\] +[^:]+:\s+6f80e000 udot v0.4s, v0.16b, v0.b\[0\] +[^:]+:\s+6f8be000 udot v0.4s, v0.16b, v11.b\[0\] +[^:]+:\s+6f96e000 udot v0.4s, v0.16b, v22.b\[0\] +[^:]+:\s+6fa0e000 udot v0.4s, v0.16b, v0.b\[1\] +[^:]+:\s+6fabe000 udot v0.4s, v0.16b, v11.b\[1\] +[^:]+:\s+6fb6e000 udot v0.4s, v0.16b, v22.b\[1\] +[^:]+:\s+6f80e800 udot v0.4s, v0.16b, v0.b\[2\] +[^:]+:\s+6f8be800 udot v0.4s, v0.16b, v11.b\[2\] +[^:]+:\s+6f96e800 udot v0.4s, v0.16b, v22.b\[2\] +[^:]+:\s+6fa0e800 udot v0.4s, v0.16b, v0.b\[3\] +[^:]+:\s+6fabe800 udot v0.4s, v0.16b, v11.b\[3\] +[^:]+:\s+6fb6e800 udot v0.4s, v0.16b, v22.b\[3\] +[^:]+:\s+6f80e160 udot v0.4s, v11.16b, v0.b\[0\] +[^:]+:\s+6f8be160 udot v0.4s, v11.16b, v11.b\[0\] +[^:]+:\s+6f96e160 udot v0.4s, v11.16b, v22.b\[0\] +[^:]+:\s+6fa0e160 udot v0.4s, v11.16b, v0.b\[1\] +[^:]+:\s+6fabe160 udot v0.4s, v11.16b, v11.b\[1\] +[^:]+:\s+6fb6e160 udot v0.4s, v11.16b, v22.b\[1\] +[^:]+:\s+6f80e960 udot v0.4s, v11.16b, v0.b\[2\] +[^:]+:\s+6f8be960 udot v0.4s, v11.16b, v11.b\[2\] +[^:]+:\s+6f96e960 udot v0.4s, v11.16b, v22.b\[2\] +[^:]+:\s+6fa0e960 udot v0.4s, v11.16b, v0.b\[3\] +[^:]+:\s+6fabe960 udot v0.4s, v11.16b, v11.b\[3\] +[^:]+:\s+6fb6e960 udot v0.4s, v11.16b, v22.b\[3\] +[^:]+:\s+6f80e2c0 udot v0.4s, v22.16b, v0.b\[0\] +[^:]+:\s+6f8be2c0 udot v0.4s, v22.16b, v11.b\[0\] +[^:]+:\s+6f96e2c0 udot v0.4s, v22.16b, v22.b\[0\] +[^:]+:\s+6fa0e2c0 udot v0.4s, v22.16b, v0.b\[1\] +[^:]+:\s+6fabe2c0 udot v0.4s, v22.16b, v11.b\[1\] +[^:]+:\s+6fb6e2c0 udot v0.4s, v22.16b, v22.b\[1\] +[^:]+:\s+6f80eac0 udot v0.4s, v22.16b, v0.b\[2\] +[^:]+:\s+6f8beac0 udot v0.4s, v22.16b, v11.b\[2\] +[^:]+:\s+6f96eac0 udot v0.4s, v22.16b, v22.b\[2\] +[^:]+:\s+6fa0eac0 udot v0.4s, v22.16b, v0.b\[3\] +[^:]+:\s+6fabeac0 udot v0.4s, v22.16b, v11.b\[3\] +[^:]+:\s+6fb6eac0 udot v0.4s, v22.16b, v22.b\[3\] +[^:]+:\s+6f80e00b udot v11.4s, v0.16b, v0.b\[0\] +[^:]+:\s+6f8be00b udot v11.4s, v0.16b, v11.b\[0\] +[^:]+:\s+6f96e00b udot v11.4s, v0.16b, v22.b\[0\] +[^:]+:\s+6fa0e00b udot v11.4s, v0.16b, v0.b\[1\] +[^:]+:\s+6fabe00b udot v11.4s, v0.16b, v11.b\[1\] +[^:]+:\s+6fb6e00b udot v11.4s, v0.16b, v22.b\[1\] +[^:]+:\s+6f80e80b udot v11.4s, v0.16b, v0.b\[2\] +[^:]+:\s+6f8be80b udot v11.4s, v0.16b, v11.b\[2\] +[^:]+:\s+6f96e80b udot v11.4s, v0.16b, v22.b\[2\] +[^:]+:\s+6fa0e80b udot v11.4s, v0.16b, v0.b\[3\] +[^:]+:\s+6fabe80b udot v11.4s, v0.16b, v11.b\[3\] +[^:]+:\s+6fb6e80b udot v11.4s, v0.16b, v22.b\[3\] +[^:]+:\s+6f80e16b udot v11.4s, v11.16b, v0.b\[0\] +[^:]+:\s+6f8be16b udot v11.4s, v11.16b, v11.b\[0\] +[^:]+:\s+6f96e16b udot v11.4s, v11.16b, v22.b\[0\] +[^:]+:\s+6fa0e16b udot v11.4s, v11.16b, v0.b\[1\] +[^:]+:\s+6fabe16b udot v11.4s, v11.16b, v11.b\[1\] +[^:]+:\s+6fb6e16b udot v11.4s, v11.16b, v22.b\[1\] +[^:]+:\s+6f80e96b udot v11.4s, v11.16b, v0.b\[2\] +[^:]+:\s+6f8be96b udot v11.4s, v11.16b, v11.b\[2\] +[^:]+:\s+6f96e96b udot v11.4s, v11.16b, v22.b\[2\] +[^:]+:\s+6fa0e96b udot v11.4s, v11.16b, v0.b\[3\] +[^:]+:\s+6fabe96b udot v11.4s, v11.16b, v11.b\[3\] +[^:]+:\s+6fb6e96b udot v11.4s, v11.16b, v22.b\[3\] +[^:]+:\s+6f80e2cb udot v11.4s, v22.16b, v0.b\[0\] +[^:]+:\s+6f8be2cb udot v11.4s, v22.16b, v11.b\[0\] +[^:]+:\s+6f96e2cb udot v11.4s, v22.16b, v22.b\[0\] +[^:]+:\s+6fa0e2cb udot v11.4s, v22.16b, v0.b\[1\] +[^:]+:\s+6fabe2cb udot v11.4s, v22.16b, v11.b\[1\] +[^:]+:\s+6fb6e2cb udot v11.4s, v22.16b, v22.b\[1\] +[^:]+:\s+6f80eacb udot v11.4s, v22.16b, v0.b\[2\] +[^:]+:\s+6f8beacb udot v11.4s, v22.16b, v11.b\[2\] +[^:]+:\s+6f96eacb udot v11.4s, v22.16b, v22.b\[2\] +[^:]+:\s+6fa0eacb udot v11.4s, v22.16b, v0.b\[3\] +[^:]+:\s+6fabeacb udot v11.4s, v22.16b, v11.b\[3\] +[^:]+:\s+6fb6eacb udot v11.4s, v22.16b, v22.b\[3\] +[^:]+:\s+6f80e016 udot v22.4s, v0.16b, v0.b\[0\] +[^:]+:\s+6f8be016 udot v22.4s, v0.16b, v11.b\[0\] +[^:]+:\s+6f96e016 udot v22.4s, v0.16b, v22.b\[0\] +[^:]+:\s+6fa0e016 udot v22.4s, v0.16b, v0.b\[1\] +[^:]+:\s+6fabe016 udot v22.4s, v0.16b, v11.b\[1\] +[^:]+:\s+6fb6e016 udot v22.4s, v0.16b, v22.b\[1\] +[^:]+:\s+6f80e816 udot v22.4s, v0.16b, v0.b\[2\] +[^:]+:\s+6f8be816 udot v22.4s, v0.16b, v11.b\[2\] +[^:]+:\s+6f96e816 udot v22.4s, v0.16b, v22.b\[2\] +[^:]+:\s+6fa0e816 udot v22.4s, v0.16b, v0.b\[3\] +[^:]+:\s+6fabe816 udot v22.4s, v0.16b, v11.b\[3\] +[^:]+:\s+6fb6e816 udot v22.4s, v0.16b, v22.b\[3\] +[^:]+:\s+6f80e176 udot v22.4s, v11.16b, v0.b\[0\] +[^:]+:\s+6f8be176 udot v22.4s, v11.16b, v11.b\[0\] +[^:]+:\s+6f96e176 udot v22.4s, v11.16b, v22.b\[0\] +[^:]+:\s+6fa0e176 udot v22.4s, v11.16b, v0.b\[1\] +[^:]+:\s+6fabe176 udot v22.4s, v11.16b, v11.b\[1\] +[^:]+:\s+6fb6e176 udot v22.4s, v11.16b, v22.b\[1\] +[^:]+:\s+6f80e976 udot v22.4s, v11.16b, v0.b\[2\] +[^:]+:\s+6f8be976 udot v22.4s, v11.16b, v11.b\[2\] +[^:]+:\s+6f96e976 udot v22.4s, v11.16b, v22.b\[2\] +[^:]+:\s+6fa0e976 udot v22.4s, v11.16b, v0.b\[3\] +[^:]+:\s+6fabe976 udot v22.4s, v11.16b, v11.b\[3\] +[^:]+:\s+6fb6e976 udot v22.4s, v11.16b, v22.b\[3\] +[^:]+:\s+6f80e2d6 udot v22.4s, v22.16b, v0.b\[0\] +[^:]+:\s+6f8be2d6 udot v22.4s, v22.16b, v11.b\[0\] +[^:]+:\s+6f96e2d6 udot v22.4s, v22.16b, v22.b\[0\] +[^:]+:\s+6fa0e2d6 udot v22.4s, v22.16b, v0.b\[1\] +[^:]+:\s+6fabe2d6 udot v22.4s, v22.16b, v11.b\[1\] +[^:]+:\s+6fb6e2d6 udot v22.4s, v22.16b, v22.b\[1\] +[^:]+:\s+6f80ead6 udot v22.4s, v22.16b, v0.b\[2\] +[^:]+:\s+6f8bead6 udot v22.4s, v22.16b, v11.b\[2\] +[^:]+:\s+6f96ead6 udot v22.4s, v22.16b, v22.b\[2\] +[^:]+:\s+6fa0ead6 udot v22.4s, v22.16b, v0.b\[3\] +[^:]+:\s+6fabead6 udot v22.4s, v22.16b, v11.b\[3\] +[^:]+:\s+6fb6ead6 udot v22.4s, v22.16b, v22.b\[3\] +[^:]+:\s+4f80e000 sdot v0.4s, v0.16b, v0.b\[0\] +[^:]+:\s+4f8be000 sdot v0.4s, v0.16b, v11.b\[0\] +[^:]+:\s+4f96e000 sdot v0.4s, v0.16b, v22.b\[0\] +[^:]+:\s+4fa0e000 sdot v0.4s, v0.16b, v0.b\[1\] +[^:]+:\s+4fabe000 sdot v0.4s, v0.16b, v11.b\[1\] +[^:]+:\s+4fb6e000 sdot v0.4s, v0.16b, v22.b\[1\] +[^:]+:\s+4f80e800 sdot v0.4s, v0.16b, v0.b\[2\] +[^:]+:\s+4f8be800 sdot v0.4s, v0.16b, v11.b\[2\] +[^:]+:\s+4f96e800 sdot v0.4s, v0.16b, v22.b\[2\] +[^:]+:\s+4fa0e800 sdot v0.4s, v0.16b, v0.b\[3\] +[^:]+:\s+4fabe800 sdot v0.4s, v0.16b, v11.b\[3\] +[^:]+:\s+4fb6e800 sdot v0.4s, v0.16b, v22.b\[3\] +[^:]+:\s+4f80e160 sdot v0.4s, v11.16b, v0.b\[0\] +[^:]+:\s+4f8be160 sdot v0.4s, v11.16b, v11.b\[0\] +[^:]+:\s+4f96e160 sdot v0.4s, v11.16b, v22.b\[0\] +[^:]+:\s+4fa0e160 sdot v0.4s, v11.16b, v0.b\[1\] +[^:]+:\s+4fabe160 sdot v0.4s, v11.16b, v11.b\[1\] +[^:]+:\s+4fb6e160 sdot v0.4s, v11.16b, v22.b\[1\] +[^:]+:\s+4f80e960 sdot v0.4s, v11.16b, v0.b\[2\] +[^:]+:\s+4f8be960 sdot v0.4s, v11.16b, v11.b\[2\] +[^:]+:\s+4f96e960 sdot v0.4s, v11.16b, v22.b\[2\] +[^:]+:\s+4fa0e960 sdot v0.4s, v11.16b, v0.b\[3\] +[^:]+:\s+4fabe960 sdot v0.4s, v11.16b, v11.b\[3\] +[^:]+:\s+4fb6e960 sdot v0.4s, v11.16b, v22.b\[3\] +[^:]+:\s+4f80e2c0 sdot v0.4s, v22.16b, v0.b\[0\] +[^:]+:\s+4f8be2c0 sdot v0.4s, v22.16b, v11.b\[0\] +[^:]+:\s+4f96e2c0 sdot v0.4s, v22.16b, v22.b\[0\] +[^:]+:\s+4fa0e2c0 sdot v0.4s, v22.16b, v0.b\[1\] +[^:]+:\s+4fabe2c0 sdot v0.4s, v22.16b, v11.b\[1\] +[^:]+:\s+4fb6e2c0 sdot v0.4s, v22.16b, v22.b\[1\] +[^:]+:\s+4f80eac0 sdot v0.4s, v22.16b, v0.b\[2\] +[^:]+:\s+4f8beac0 sdot v0.4s, v22.16b, v11.b\[2\] +[^:]+:\s+4f96eac0 sdot v0.4s, v22.16b, v22.b\[2\] +[^:]+:\s+4fa0eac0 sdot v0.4s, v22.16b, v0.b\[3\] +[^:]+:\s+4fabeac0 sdot v0.4s, v22.16b, v11.b\[3\] +[^:]+:\s+4fb6eac0 sdot v0.4s, v22.16b, v22.b\[3\] +[^:]+:\s+4f80e00b sdot v11.4s, v0.16b, v0.b\[0\] +[^:]+:\s+4f8be00b sdot v11.4s, v0.16b, v11.b\[0\] +[^:]+:\s+4f96e00b sdot v11.4s, v0.16b, v22.b\[0\] +[^:]+:\s+4fa0e00b sdot v11.4s, v0.16b, v0.b\[1\] +[^:]+:\s+4fabe00b sdot v11.4s, v0.16b, v11.b\[1\] +[^:]+:\s+4fb6e00b sdot v11.4s, v0.16b, v22.b\[1\] +[^:]+:\s+4f80e80b sdot v11.4s, v0.16b, v0.b\[2\] +[^:]+:\s+4f8be80b sdot v11.4s, v0.16b, v11.b\[2\] +[^:]+:\s+4f96e80b sdot v11.4s, v0.16b, v22.b\[2\] +[^:]+:\s+4fa0e80b sdot v11.4s, v0.16b, v0.b\[3\] +[^:]+:\s+4fabe80b sdot v11.4s, v0.16b, v11.b\[3\] +[^:]+:\s+4fb6e80b sdot v11.4s, v0.16b, v22.b\[3\] +[^:]+:\s+4f80e16b sdot v11.4s, v11.16b, v0.b\[0\] +[^:]+:\s+4f8be16b sdot v11.4s, v11.16b, v11.b\[0\] +[^:]+:\s+4f96e16b sdot v11.4s, v11.16b, v22.b\[0\] +[^:]+:\s+4fa0e16b sdot v11.4s, v11.16b, v0.b\[1\] +[^:]+:\s+4fabe16b sdot v11.4s, v11.16b, v11.b\[1\] +[^:]+:\s+4fb6e16b sdot v11.4s, v11.16b, v22.b\[1\] +[^:]+:\s+4f80e96b sdot v11.4s, v11.16b, v0.b\[2\] +[^:]+:\s+4f8be96b sdot v11.4s, v11.16b, v11.b\[2\] +[^:]+:\s+4f96e96b sdot v11.4s, v11.16b, v22.b\[2\] +[^:]+:\s+4fa0e96b sdot v11.4s, v11.16b, v0.b\[3\] +[^:]+:\s+4fabe96b sdot v11.4s, v11.16b, v11.b\[3\] +[^:]+:\s+4fb6e96b sdot v11.4s, v11.16b, v22.b\[3\] +[^:]+:\s+4f80e2cb sdot v11.4s, v22.16b, v0.b\[0\] +[^:]+:\s+4f8be2cb sdot v11.4s, v22.16b, v11.b\[0\] +[^:]+:\s+4f96e2cb sdot v11.4s, v22.16b, v22.b\[0\] +[^:]+:\s+4fa0e2cb sdot v11.4s, v22.16b, v0.b\[1\] +[^:]+:\s+4fabe2cb sdot v11.4s, v22.16b, v11.b\[1\] +[^:]+:\s+4fb6e2cb sdot v11.4s, v22.16b, v22.b\[1\] +[^:]+:\s+4f80eacb sdot v11.4s, v22.16b, v0.b\[2\] +[^:]+:\s+4f8beacb sdot v11.4s, v22.16b, v11.b\[2\] +[^:]+:\s+4f96eacb sdot v11.4s, v22.16b, v22.b\[2\] +[^:]+:\s+4fa0eacb sdot v11.4s, v22.16b, v0.b\[3\] +[^:]+:\s+4fabeacb sdot v11.4s, v22.16b, v11.b\[3\] +[^:]+:\s+4fb6eacb sdot v11.4s, v22.16b, v22.b\[3\] +[^:]+:\s+4f80e016 sdot v22.4s, v0.16b, v0.b\[0\] +[^:]+:\s+4f8be016 sdot v22.4s, v0.16b, v11.b\[0\] +[^:]+:\s+4f96e016 sdot v22.4s, v0.16b, v22.b\[0\] +[^:]+:\s+4fa0e016 sdot v22.4s, v0.16b, v0.b\[1\] +[^:]+:\s+4fabe016 sdot v22.4s, v0.16b, v11.b\[1\] +[^:]+:\s+4fb6e016 sdot v22.4s, v0.16b, v22.b\[1\] +[^:]+:\s+4f80e816 sdot v22.4s, v0.16b, v0.b\[2\] +[^:]+:\s+4f8be816 sdot v22.4s, v0.16b, v11.b\[2\] +[^:]+:\s+4f96e816 sdot v22.4s, v0.16b, v22.b\[2\] +[^:]+:\s+4fa0e816 sdot v22.4s, v0.16b, v0.b\[3\] +[^:]+:\s+4fabe816 sdot v22.4s, v0.16b, v11.b\[3\] +[^:]+:\s+4fb6e816 sdot v22.4s, v0.16b, v22.b\[3\] +[^:]+:\s+4f80e176 sdot v22.4s, v11.16b, v0.b\[0\] +[^:]+:\s+4f8be176 sdot v22.4s, v11.16b, v11.b\[0\] +[^:]+:\s+4f96e176 sdot v22.4s, v11.16b, v22.b\[0\] +[^:]+:\s+4fa0e176 sdot v22.4s, v11.16b, v0.b\[1\] +[^:]+:\s+4fabe176 sdot v22.4s, v11.16b, v11.b\[1\] +[^:]+:\s+4fb6e176 sdot v22.4s, v11.16b, v22.b\[1\] +[^:]+:\s+4f80e976 sdot v22.4s, v11.16b, v0.b\[2\] +[^:]+:\s+4f8be976 sdot v22.4s, v11.16b, v11.b\[2\] +[^:]+:\s+4f96e976 sdot v22.4s, v11.16b, v22.b\[2\] +[^:]+:\s+4fa0e976 sdot v22.4s, v11.16b, v0.b\[3\] +[^:]+:\s+4fabe976 sdot v22.4s, v11.16b, v11.b\[3\] +[^:]+:\s+4fb6e976 sdot v22.4s, v11.16b, v22.b\[3\] +[^:]+:\s+4f80e2d6 sdot v22.4s, v22.16b, v0.b\[0\] +[^:]+:\s+4f8be2d6 sdot v22.4s, v22.16b, v11.b\[0\] +[^:]+:\s+4f96e2d6 sdot v22.4s, v22.16b, v22.b\[0\] +[^:]+:\s+4fa0e2d6 sdot v22.4s, v22.16b, v0.b\[1\] +[^:]+:\s+4fabe2d6 sdot v22.4s, v22.16b, v11.b\[1\] +[^:]+:\s+4fb6e2d6 sdot v22.4s, v22.16b, v22.b\[1\] +[^:]+:\s+4f80ead6 sdot v22.4s, v22.16b, v0.b\[2\] +[^:]+:\s+4f8bead6 sdot v22.4s, v22.16b, v11.b\[2\] +[^:]+:\s+4f96ead6 sdot v22.4s, v22.16b, v22.b\[2\] +[^:]+:\s+4fa0ead6 sdot v22.4s, v22.16b, v0.b\[3\] +[^:]+:\s+4fabead6 sdot v22.4s, v22.16b, v11.b\[3\] +[^:]+:\s+4fb6ead6 sdot v22.4s, v22.16b, v22.b\[3\] diff --git a/gas/testsuite/gas/aarch64/dotproduct.s b/gas/testsuite/gas/aarch64/dotproduct.s new file mode 100644 index 00000000000..b3844e1dbbf --- /dev/null +++ b/gas/testsuite/gas/aarch64/dotproduct.s @@ -0,0 +1,540 @@ +UDOT V0.2S, V0.8B, V0.8B +UDOT V0.2S, V0.8B, V11.8B +UDOT V0.2S, V0.8B, V22.8B +UDOT V0.2S, V11.8B, V0.8B +UDOT V0.2S, V11.8B, V11.8B +UDOT V0.2S, V11.8B, V22.8B +UDOT V0.2S, V22.8B, V0.8B +UDOT V0.2S, V22.8B, V11.8B +UDOT V0.2S, V22.8B, V22.8B +UDOT V11.2S, V0.8B, V0.8B +UDOT V11.2S, V0.8B, V11.8B +UDOT V11.2S, V0.8B, V22.8B +UDOT V11.2S, V11.8B, V0.8B +UDOT V11.2S, V11.8B, V11.8B +UDOT V11.2S, V11.8B, V22.8B +UDOT V11.2S, V22.8B, V0.8B +UDOT V11.2S, V22.8B, V11.8B +UDOT V11.2S, V22.8B, V22.8B +UDOT V22.2S, V0.8B, V0.8B +UDOT V22.2S, V0.8B, V11.8B +UDOT V22.2S, V0.8B, V22.8B +UDOT V22.2S, V11.8B, V0.8B +UDOT V22.2S, V11.8B, V11.8B +UDOT V22.2S, V11.8B, V22.8B +UDOT V22.2S, V22.8B, V0.8B +UDOT V22.2S, V22.8B, V11.8B +UDOT V22.2S, V22.8B, V22.8B +SDOT V0.2S, V0.8B, V0.8B +SDOT V0.2S, V0.8B, V11.8B +SDOT V0.2S, V0.8B, V22.8B +SDOT V0.2S, V11.8B, V0.8B +SDOT V0.2S, V11.8B, V11.8B +SDOT V0.2S, V11.8B, V22.8B +SDOT V0.2S, V22.8B, V0.8B +SDOT V0.2S, V22.8B, V11.8B +SDOT V0.2S, V22.8B, V22.8B +SDOT V11.2S, V0.8B, V0.8B +SDOT V11.2S, V0.8B, V11.8B +SDOT V11.2S, V0.8B, V22.8B +SDOT V11.2S, V11.8B, V0.8B +SDOT V11.2S, V11.8B, V11.8B +SDOT V11.2S, V11.8B, V22.8B +SDOT V11.2S, V22.8B, V0.8B +SDOT V11.2S, V22.8B, V11.8B +SDOT V11.2S, V22.8B, V22.8B +SDOT V22.2S, V0.8B, V0.8B +SDOT V22.2S, V0.8B, V11.8B +SDOT V22.2S, V0.8B, V22.8B +SDOT V22.2S, V11.8B, V0.8B +SDOT V22.2S, V11.8B, V11.8B +SDOT V22.2S, V11.8B, V22.8B +SDOT V22.2S, V22.8B, V0.8B +SDOT V22.2S, V22.8B, V11.8B +SDOT V22.2S, V22.8B, V22.8B +UDOT V0.4S, V0.16B, V0.16B +UDOT V0.4S, V0.16B, V11.16B +UDOT V0.4S, V0.16B, V22.16B +UDOT V0.4S, V11.16B, V0.16B +UDOT V0.4S, V11.16B, V11.16B +UDOT V0.4S, V11.16B, V22.16B +UDOT V0.4S, V22.16B, V0.16B +UDOT V0.4S, V22.16B, V11.16B +UDOT V0.4S, V22.16B, V22.16B +UDOT V11.4S, V0.16B, V0.16B +UDOT V11.4S, V0.16B, V11.16B +UDOT V11.4S, V0.16B, V22.16B +UDOT V11.4S, V11.16B, V0.16B +UDOT V11.4S, V11.16B, V11.16B +UDOT V11.4S, V11.16B, V22.16B +UDOT V11.4S, V22.16B, V0.16B +UDOT V11.4S, V22.16B, V11.16B +UDOT V11.4S, V22.16B, V22.16B +UDOT V22.4S, V0.16B, V0.16B +UDOT V22.4S, V0.16B, V11.16B +UDOT V22.4S, V0.16B, V22.16B +UDOT V22.4S, V11.16B, V0.16B +UDOT V22.4S, V11.16B, V11.16B +UDOT V22.4S, V11.16B, V22.16B +UDOT V22.4S, V22.16B, V0.16B +UDOT V22.4S, V22.16B, V11.16B +UDOT V22.4S, V22.16B, V22.16B +SDOT V0.4S, V0.16B, V0.16B +SDOT V0.4S, V0.16B, V11.16B +SDOT V0.4S, V0.16B, V22.16B +SDOT V0.4S, V11.16B, V0.16B +SDOT V0.4S, V11.16B, V11.16B +SDOT V0.4S, V11.16B, V22.16B +SDOT V0.4S, V22.16B, V0.16B +SDOT V0.4S, V22.16B, V11.16B +SDOT V0.4S, V22.16B, V22.16B +SDOT V11.4S, V0.16B, V0.16B +SDOT V11.4S, V0.16B, V11.16B +SDOT V11.4S, V0.16B, V22.16B +SDOT V11.4S, V11.16B, V0.16B +SDOT V11.4S, V11.16B, V11.16B +SDOT V11.4S, V11.16B, V22.16B +SDOT V11.4S, V22.16B, V0.16B +SDOT V11.4S, V22.16B, V11.16B +SDOT V11.4S, V22.16B, V22.16B +SDOT V22.4S, V0.16B, V0.16B +SDOT V22.4S, V0.16B, V11.16B +SDOT V22.4S, V0.16B, V22.16B +SDOT V22.4S, V11.16B, V0.16B +SDOT V22.4S, V11.16B, V11.16B +SDOT V22.4S, V11.16B, V22.16B +SDOT V22.4S, V22.16B, V0.16B +SDOT V22.4S, V22.16B, V11.16B +SDOT V22.4S, V22.16B, V22.16B +UDOT V0.2S, V0.8B, V0.4B[0] +UDOT V0.2S, V0.8B, V11.4B[0] +UDOT V0.2S, V0.8B, V22.4B[0] +UDOT V0.2S, V0.8B, V0.4B[1] +UDOT V0.2S, V0.8B, V11.4B[1] +UDOT V0.2S, V0.8B, V22.4B[1] +UDOT V0.2S, V0.8B, V0.4B[2] +UDOT V0.2S, V0.8B, V11.4B[2] +UDOT V0.2S, V0.8B, V22.4B[2] +UDOT V0.2S, V0.8B, V0.4B[3] +UDOT V0.2S, V0.8B, V11.4B[3] +UDOT V0.2S, V0.8B, V22.4B[3] +UDOT V0.2S, V11.8B, V0.4B[0] +UDOT V0.2S, V11.8B, V11.4B[0] +UDOT V0.2S, V11.8B, V22.4B[0] +UDOT V0.2S, V11.8B, V0.4B[1] +UDOT V0.2S, V11.8B, V11.4B[1] +UDOT V0.2S, V11.8B, V22.4B[1] +UDOT V0.2S, V11.8B, V0.4B[2] +UDOT V0.2S, V11.8B, V11.4B[2] +UDOT V0.2S, V11.8B, V22.4B[2] +UDOT V0.2S, V11.8B, V0.4B[3] +UDOT V0.2S, V11.8B, V11.4B[3] +UDOT V0.2S, V11.8B, V22.4B[3] +UDOT V0.2S, V22.8B, V0.4B[0] +UDOT V0.2S, V22.8B, V11.4B[0] +UDOT V0.2S, V22.8B, V22.4B[0] +UDOT V0.2S, V22.8B, V0.4B[1] +UDOT V0.2S, V22.8B, V11.4B[1] +UDOT V0.2S, V22.8B, V22.4B[1] +UDOT V0.2S, V22.8B, V0.4B[2] +UDOT V0.2S, V22.8B, V11.4B[2] +UDOT V0.2S, V22.8B, V22.4B[2] +UDOT V0.2S, V22.8B, V0.4B[3] +UDOT V0.2S, V22.8B, V11.4B[3] +UDOT V0.2S, V22.8B, V22.4B[3] +UDOT V11.2S, V0.8B, V0.4B[0] +UDOT V11.2S, V0.8B, V11.4B[0] +UDOT V11.2S, V0.8B, V22.4B[0] +UDOT V11.2S, V0.8B, V0.4B[1] +UDOT V11.2S, V0.8B, V11.4B[1] +UDOT V11.2S, V0.8B, V22.4B[1] +UDOT V11.2S, V0.8B, V0.4B[2] +UDOT V11.2S, V0.8B, V11.4B[2] +UDOT V11.2S, V0.8B, V22.4B[2] +UDOT V11.2S, V0.8B, V0.4B[3] +UDOT V11.2S, V0.8B, V11.4B[3] +UDOT V11.2S, V0.8B, V22.4B[3] +UDOT V11.2S, V11.8B, V0.4B[0] +UDOT V11.2S, V11.8B, V11.4B[0] +UDOT V11.2S, V11.8B, V22.4B[0] +UDOT V11.2S, V11.8B, V0.4B[1] +UDOT V11.2S, V11.8B, V11.4B[1] +UDOT V11.2S, V11.8B, V22.4B[1] +UDOT V11.2S, V11.8B, V0.4B[2] +UDOT V11.2S, V11.8B, V11.4B[2] +UDOT V11.2S, V11.8B, V22.4B[2] +UDOT V11.2S, V11.8B, V0.4B[3] +UDOT V11.2S, V11.8B, V11.4B[3] +UDOT V11.2S, V11.8B, V22.4B[3] +UDOT V11.2S, V22.8B, V0.4B[0] +UDOT V11.2S, V22.8B, V11.4B[0] +UDOT V11.2S, V22.8B, V22.4B[0] +UDOT V11.2S, V22.8B, V0.4B[1] +UDOT V11.2S, V22.8B, V11.4B[1] +UDOT V11.2S, V22.8B, V22.4B[1] +UDOT V11.2S, V22.8B, V0.4B[2] +UDOT V11.2S, V22.8B, V11.4B[2] +UDOT V11.2S, V22.8B, V22.4B[2] +UDOT V11.2S, V22.8B, V0.4B[3] +UDOT V11.2S, V22.8B, V11.4B[3] +UDOT V11.2S, V22.8B, V22.4B[3] +UDOT V22.2S, V0.8B, V0.4B[0] +UDOT V22.2S, V0.8B, V11.4B[0] +UDOT V22.2S, V0.8B, V22.4B[0] +UDOT V22.2S, V0.8B, V0.4B[1] +UDOT V22.2S, V0.8B, V11.4B[1] +UDOT V22.2S, V0.8B, V22.4B[1] +UDOT V22.2S, V0.8B, V0.4B[2] +UDOT V22.2S, V0.8B, V11.4B[2] +UDOT V22.2S, V0.8B, V22.4B[2] +UDOT V22.2S, V0.8B, V0.4B[3] +UDOT V22.2S, V0.8B, V11.4B[3] +UDOT V22.2S, V0.8B, V22.4B[3] +UDOT V22.2S, V11.8B, V0.4B[0] +UDOT V22.2S, V11.8B, V11.4B[0] +UDOT V22.2S, V11.8B, V22.4B[0] +UDOT V22.2S, V11.8B, V0.4B[1] +UDOT V22.2S, V11.8B, V11.4B[1] +UDOT V22.2S, V11.8B, V22.4B[1] +UDOT V22.2S, V11.8B, V0.4B[2] +UDOT V22.2S, V11.8B, V11.4B[2] +UDOT V22.2S, V11.8B, V22.4B[2] +UDOT V22.2S, V11.8B, V0.4B[3] +UDOT V22.2S, V11.8B, V11.4B[3] +UDOT V22.2S, V11.8B, V22.4B[3] +UDOT V22.2S, V22.8B, V0.4B[0] +UDOT V22.2S, V22.8B, V11.4B[0] +UDOT V22.2S, V22.8B, V22.4B[0] +UDOT V22.2S, V22.8B, V0.4B[1] +UDOT V22.2S, V22.8B, V11.4B[1] +UDOT V22.2S, V22.8B, V22.4B[1] +UDOT V22.2S, V22.8B, V0.4B[2] +UDOT V22.2S, V22.8B, V11.4B[2] +UDOT V22.2S, V22.8B, V22.4B[2] +UDOT V22.2S, V22.8B, V0.4B[3] +UDOT V22.2S, V22.8B, V11.4B[3] +UDOT V22.2S, V22.8B, V22.4B[3] +SDOT V0.2S, V0.8B, V0.4B[0] +SDOT V0.2S, V0.8B, V11.4B[0] +SDOT V0.2S, V0.8B, V22.4B[0] +SDOT V0.2S, V0.8B, V0.4B[1] +SDOT V0.2S, V0.8B, V11.4B[1] +SDOT V0.2S, V0.8B, V22.4B[1] +SDOT V0.2S, V0.8B, V0.4B[2] +SDOT V0.2S, V0.8B, V11.4B[2] +SDOT V0.2S, V0.8B, V22.4B[2] +SDOT V0.2S, V0.8B, V0.4B[3] +SDOT V0.2S, V0.8B, V11.4B[3] +SDOT V0.2S, V0.8B, V22.4B[3] +SDOT V0.2S, V11.8B, V0.4B[0] +SDOT V0.2S, V11.8B, V11.4B[0] +SDOT V0.2S, V11.8B, V22.4B[0] +SDOT V0.2S, V11.8B, V0.4B[1] +SDOT V0.2S, V11.8B, V11.4B[1] +SDOT V0.2S, V11.8B, V22.4B[1] +SDOT V0.2S, V11.8B, V0.4B[2] +SDOT V0.2S, V11.8B, V11.4B[2] +SDOT V0.2S, V11.8B, V22.4B[2] +SDOT V0.2S, V11.8B, V0.4B[3] +SDOT V0.2S, V11.8B, V11.4B[3] +SDOT V0.2S, V11.8B, V22.4B[3] +SDOT V0.2S, V22.8B, V0.4B[0] +SDOT V0.2S, V22.8B, V11.4B[0] +SDOT V0.2S, V22.8B, V22.4B[0] +SDOT V0.2S, V22.8B, V0.4B[1] +SDOT V0.2S, V22.8B, V11.4B[1] +SDOT V0.2S, V22.8B, V22.4B[1] +SDOT V0.2S, V22.8B, V0.4B[2] +SDOT V0.2S, V22.8B, V11.4B[2] +SDOT V0.2S, V22.8B, V22.4B[2] +SDOT V0.2S, V22.8B, V0.4B[3] +SDOT V0.2S, V22.8B, V11.4B[3] +SDOT V0.2S, V22.8B, V22.4B[3] +SDOT V11.2S, V0.8B, V0.4B[0] +SDOT V11.2S, V0.8B, V11.4B[0] +SDOT V11.2S, V0.8B, V22.4B[0] +SDOT V11.2S, V0.8B, V0.4B[1] +SDOT V11.2S, V0.8B, V11.4B[1] +SDOT V11.2S, V0.8B, V22.4B[1] +SDOT V11.2S, V0.8B, V0.4B[2] +SDOT V11.2S, V0.8B, V11.4B[2] +SDOT V11.2S, V0.8B, V22.4B[2] +SDOT V11.2S, V0.8B, V0.4B[3] +SDOT V11.2S, V0.8B, V11.4B[3] +SDOT V11.2S, V0.8B, V22.4B[3] +SDOT V11.2S, V11.8B, V0.4B[0] +SDOT V11.2S, V11.8B, V11.4B[0] +SDOT V11.2S, V11.8B, V22.4B[0] +SDOT V11.2S, V11.8B, V0.4B[1] +SDOT V11.2S, V11.8B, V11.4B[1] +SDOT V11.2S, V11.8B, V22.4B[1] +SDOT V11.2S, V11.8B, V0.4B[2] +SDOT V11.2S, V11.8B, V11.4B[2] +SDOT V11.2S, V11.8B, V22.4B[2] +SDOT V11.2S, V11.8B, V0.4B[3] +SDOT V11.2S, V11.8B, V11.4B[3] +SDOT V11.2S, V11.8B, V22.4B[3] +SDOT V11.2S, V22.8B, V0.4B[0] +SDOT V11.2S, V22.8B, V11.4B[0] +SDOT V11.2S, V22.8B, V22.4B[0] +SDOT V11.2S, V22.8B, V0.4B[1] +SDOT V11.2S, V22.8B, V11.4B[1] +SDOT V11.2S, V22.8B, V22.4B[1] +SDOT V11.2S, V22.8B, V0.4B[2] +SDOT V11.2S, V22.8B, V11.4B[2] +SDOT V11.2S, V22.8B, V22.4B[2] +SDOT V11.2S, V22.8B, V0.4B[3] +SDOT V11.2S, V22.8B, V11.4B[3] +SDOT V11.2S, V22.8B, V22.4B[3] +SDOT V22.2S, V0.8B, V0.4B[0] +SDOT V22.2S, V0.8B, V11.4B[0] +SDOT V22.2S, V0.8B, V22.4B[0] +SDOT V22.2S, V0.8B, V0.4B[1] +SDOT V22.2S, V0.8B, V11.4B[1] +SDOT V22.2S, V0.8B, V22.4B[1] +SDOT V22.2S, V0.8B, V0.4B[2] +SDOT V22.2S, V0.8B, V11.4B[2] +SDOT V22.2S, V0.8B, V22.4B[2] +SDOT V22.2S, V0.8B, V0.4B[3] +SDOT V22.2S, V0.8B, V11.4B[3] +SDOT V22.2S, V0.8B, V22.4B[3] +SDOT V22.2S, V11.8B, V0.4B[0] +SDOT V22.2S, V11.8B, V11.4B[0] +SDOT V22.2S, V11.8B, V22.4B[0] +SDOT V22.2S, V11.8B, V0.4B[1] +SDOT V22.2S, V11.8B, V11.4B[1] +SDOT V22.2S, V11.8B, V22.4B[1] +SDOT V22.2S, V11.8B, V0.4B[2] +SDOT V22.2S, V11.8B, V11.4B[2] +SDOT V22.2S, V11.8B, V22.4B[2] +SDOT V22.2S, V11.8B, V0.4B[3] +SDOT V22.2S, V11.8B, V11.4B[3] +SDOT V22.2S, V11.8B, V22.4B[3] +SDOT V22.2S, V22.8B, V0.4B[0] +SDOT V22.2S, V22.8B, V11.4B[0] +SDOT V22.2S, V22.8B, V22.4B[0] +SDOT V22.2S, V22.8B, V0.4B[1] +SDOT V22.2S, V22.8B, V11.4B[1] +SDOT V22.2S, V22.8B, V22.4B[1] +SDOT V22.2S, V22.8B, V0.4B[2] +SDOT V22.2S, V22.8B, V11.4B[2] +SDOT V22.2S, V22.8B, V22.4B[2] +SDOT V22.2S, V22.8B, V0.4B[3] +SDOT V22.2S, V22.8B, V11.4B[3] +SDOT V22.2S, V22.8B, V22.4B[3] +UDOT V0.4S, V0.16B, V0.4B[0] +UDOT V0.4S, V0.16B, V11.4B[0] +UDOT V0.4S, V0.16B, V22.4B[0] +UDOT V0.4S, V0.16B, V0.4B[1] +UDOT V0.4S, V0.16B, V11.4B[1] +UDOT V0.4S, V0.16B, V22.4B[1] +UDOT V0.4S, V0.16B, V0.4B[2] +UDOT V0.4S, V0.16B, V11.4B[2] +UDOT V0.4S, V0.16B, V22.4B[2] +UDOT V0.4S, V0.16B, V0.4B[3] +UDOT V0.4S, V0.16B, V11.4B[3] +UDOT V0.4S, V0.16B, V22.4B[3] +UDOT V0.4S, V11.16B, V0.4B[0] +UDOT V0.4S, V11.16B, V11.4B[0] +UDOT V0.4S, V11.16B, V22.4B[0] +UDOT V0.4S, V11.16B, V0.4B[1] +UDOT V0.4S, V11.16B, V11.4B[1] +UDOT V0.4S, V11.16B, V22.4B[1] +UDOT V0.4S, V11.16B, V0.4B[2] +UDOT V0.4S, V11.16B, V11.4B[2] +UDOT V0.4S, V11.16B, V22.4B[2] +UDOT V0.4S, V11.16B, V0.4B[3] +UDOT V0.4S, V11.16B, V11.4B[3] +UDOT V0.4S, V11.16B, V22.4B[3] +UDOT V0.4S, V22.16B, V0.4B[0] +UDOT V0.4S, V22.16B, V11.4B[0] +UDOT V0.4S, V22.16B, V22.4B[0] +UDOT V0.4S, V22.16B, V0.4B[1] +UDOT V0.4S, V22.16B, V11.4B[1] +UDOT V0.4S, V22.16B, V22.4B[1] +UDOT V0.4S, V22.16B, V0.4B[2] +UDOT V0.4S, V22.16B, V11.4B[2] +UDOT V0.4S, V22.16B, V22.4B[2] +UDOT V0.4S, V22.16B, V0.4B[3] +UDOT V0.4S, V22.16B, V11.4B[3] +UDOT V0.4S, V22.16B, V22.4B[3] +UDOT V11.4S, V0.16B, V0.4B[0] +UDOT V11.4S, V0.16B, V11.4B[0] +UDOT V11.4S, V0.16B, V22.4B[0] +UDOT V11.4S, V0.16B, V0.4B[1] +UDOT V11.4S, V0.16B, V11.4B[1] +UDOT V11.4S, V0.16B, V22.4B[1] +UDOT V11.4S, V0.16B, V0.4B[2] +UDOT V11.4S, V0.16B, V11.4B[2] +UDOT V11.4S, V0.16B, V22.4B[2] +UDOT V11.4S, V0.16B, V0.4B[3] +UDOT V11.4S, V0.16B, V11.4B[3] +UDOT V11.4S, V0.16B, V22.4B[3] +UDOT V11.4S, V11.16B, V0.4B[0] +UDOT V11.4S, V11.16B, V11.4B[0] +UDOT V11.4S, V11.16B, V22.4B[0] +UDOT V11.4S, V11.16B, V0.4B[1] +UDOT V11.4S, V11.16B, V11.4B[1] +UDOT V11.4S, V11.16B, V22.4B[1] +UDOT V11.4S, V11.16B, V0.4B[2] +UDOT V11.4S, V11.16B, V11.4B[2] +UDOT V11.4S, V11.16B, V22.4B[2] +UDOT V11.4S, V11.16B, V0.4B[3] +UDOT V11.4S, V11.16B, V11.4B[3] +UDOT V11.4S, V11.16B, V22.4B[3] +UDOT V11.4S, V22.16B, V0.4B[0] +UDOT V11.4S, V22.16B, V11.4B[0] +UDOT V11.4S, V22.16B, V22.4B[0] +UDOT V11.4S, V22.16B, V0.4B[1] +UDOT V11.4S, V22.16B, V11.4B[1] +UDOT V11.4S, V22.16B, V22.4B[1] +UDOT V11.4S, V22.16B, V0.4B[2] +UDOT V11.4S, V22.16B, V11.4B[2] +UDOT V11.4S, V22.16B, V22.4B[2] +UDOT V11.4S, V22.16B, V0.4B[3] +UDOT V11.4S, V22.16B, V11.4B[3] +UDOT V11.4S, V22.16B, V22.4B[3] +UDOT V22.4S, V0.16B, V0.4B[0] +UDOT V22.4S, V0.16B, V11.4B[0] +UDOT V22.4S, V0.16B, V22.4B[0] +UDOT V22.4S, V0.16B, V0.4B[1] +UDOT V22.4S, V0.16B, V11.4B[1] +UDOT V22.4S, V0.16B, V22.4B[1] +UDOT V22.4S, V0.16B, V0.4B[2] +UDOT V22.4S, V0.16B, V11.4B[2] +UDOT V22.4S, V0.16B, V22.4B[2] +UDOT V22.4S, V0.16B, V0.4B[3] +UDOT V22.4S, V0.16B, V11.4B[3] +UDOT V22.4S, V0.16B, V22.4B[3] +UDOT V22.4S, V11.16B, V0.4B[0] +UDOT V22.4S, V11.16B, V11.4B[0] +UDOT V22.4S, V11.16B, V22.4B[0] +UDOT V22.4S, V11.16B, V0.4B[1] +UDOT V22.4S, V11.16B, V11.4B[1] +UDOT V22.4S, V11.16B, V22.4B[1] +UDOT V22.4S, V11.16B, V0.4B[2] +UDOT V22.4S, V11.16B, V11.4B[2] +UDOT V22.4S, V11.16B, V22.4B[2] +UDOT V22.4S, V11.16B, V0.4B[3] +UDOT V22.4S, V11.16B, V11.4B[3] +UDOT V22.4S, V11.16B, V22.4B[3] +UDOT V22.4S, V22.16B, V0.4B[0] +UDOT V22.4S, V22.16B, V11.4B[0] +UDOT V22.4S, V22.16B, V22.4B[0] +UDOT V22.4S, V22.16B, V0.4B[1] +UDOT V22.4S, V22.16B, V11.4B[1] +UDOT V22.4S, V22.16B, V22.4B[1] +UDOT V22.4S, V22.16B, V0.4B[2] +UDOT V22.4S, V22.16B, V11.4B[2] +UDOT V22.4S, V22.16B, V22.4B[2] +UDOT V22.4S, V22.16B, V0.4B[3] +UDOT V22.4S, V22.16B, V11.4B[3] +UDOT V22.4S, V22.16B, V22.4B[3] +SDOT V0.4S, V0.16B, V0.4B[0] +SDOT V0.4S, V0.16B, V11.4B[0] +SDOT V0.4S, V0.16B, V22.4B[0] +SDOT V0.4S, V0.16B, V0.4B[1] +SDOT V0.4S, V0.16B, V11.4B[1] +SDOT V0.4S, V0.16B, V22.4B[1] +SDOT V0.4S, V0.16B, V0.4B[2] +SDOT V0.4S, V0.16B, V11.4B[2] +SDOT V0.4S, V0.16B, V22.4B[2] +SDOT V0.4S, V0.16B, V0.4B[3] +SDOT V0.4S, V0.16B, V11.4B[3] +SDOT V0.4S, V0.16B, V22.4B[3] +SDOT V0.4S, V11.16B, V0.4B[0] +SDOT V0.4S, V11.16B, V11.4B[0] +SDOT V0.4S, V11.16B, V22.4B[0] +SDOT V0.4S, V11.16B, V0.4B[1] +SDOT V0.4S, V11.16B, V11.4B[1] +SDOT V0.4S, V11.16B, V22.4B[1] +SDOT V0.4S, V11.16B, V0.4B[2] +SDOT V0.4S, V11.16B, V11.4B[2] +SDOT V0.4S, V11.16B, V22.4B[2] +SDOT V0.4S, V11.16B, V0.4B[3] +SDOT V0.4S, V11.16B, V11.4B[3] +SDOT V0.4S, V11.16B, V22.4B[3] +SDOT V0.4S, V22.16B, V0.4B[0] +SDOT V0.4S, V22.16B, V11.4B[0] +SDOT V0.4S, V22.16B, V22.4B[0] +SDOT V0.4S, V22.16B, V0.4B[1] +SDOT V0.4S, V22.16B, V11.4B[1] +SDOT V0.4S, V22.16B, V22.4B[1] +SDOT V0.4S, V22.16B, V0.4B[2] +SDOT V0.4S, V22.16B, V11.4B[2] +SDOT V0.4S, V22.16B, V22.4B[2] +SDOT V0.4S, V22.16B, V0.4B[3] +SDOT V0.4S, V22.16B, V11.4B[3] +SDOT V0.4S, V22.16B, V22.4B[3] +SDOT V11.4S, V0.16B, V0.4B[0] +SDOT V11.4S, V0.16B, V11.4B[0] +SDOT V11.4S, V0.16B, V22.4B[0] +SDOT V11.4S, V0.16B, V0.4B[1] +SDOT V11.4S, V0.16B, V11.4B[1] +SDOT V11.4S, V0.16B, V22.4B[1] +SDOT V11.4S, V0.16B, V0.4B[2] +SDOT V11.4S, V0.16B, V11.4B[2] +SDOT V11.4S, V0.16B, V22.4B[2] +SDOT V11.4S, V0.16B, V0.4B[3] +SDOT V11.4S, V0.16B, V11.4B[3] +SDOT V11.4S, V0.16B, V22.4B[3] +SDOT V11.4S, V11.16B, V0.4B[0] +SDOT V11.4S, V11.16B, V11.4B[0] +SDOT V11.4S, V11.16B, V22.4B[0] +SDOT V11.4S, V11.16B, V0.4B[1] +SDOT V11.4S, V11.16B, V11.4B[1] +SDOT V11.4S, V11.16B, V22.4B[1] +SDOT V11.4S, V11.16B, V0.4B[2] +SDOT V11.4S, V11.16B, V11.4B[2] +SDOT V11.4S, V11.16B, V22.4B[2] +SDOT V11.4S, V11.16B, V0.4B[3] +SDOT V11.4S, V11.16B, V11.4B[3] +SDOT V11.4S, V11.16B, V22.4B[3] +SDOT V11.4S, V22.16B, V0.4B[0] +SDOT V11.4S, V22.16B, V11.4B[0] +SDOT V11.4S, V22.16B, V22.4B[0] +SDOT V11.4S, V22.16B, V0.4B[1] +SDOT V11.4S, V22.16B, V11.4B[1] +SDOT V11.4S, V22.16B, V22.4B[1] +SDOT V11.4S, V22.16B, V0.4B[2] +SDOT V11.4S, V22.16B, V11.4B[2] +SDOT V11.4S, V22.16B, V22.4B[2] +SDOT V11.4S, V22.16B, V0.4B[3] +SDOT V11.4S, V22.16B, V11.4B[3] +SDOT V11.4S, V22.16B, V22.4B[3] +SDOT V22.4S, V0.16B, V0.4B[0] +SDOT V22.4S, V0.16B, V11.4B[0] +SDOT V22.4S, V0.16B, V22.4B[0] +SDOT V22.4S, V0.16B, V0.4B[1] +SDOT V22.4S, V0.16B, V11.4B[1] +SDOT V22.4S, V0.16B, V22.4B[1] +SDOT V22.4S, V0.16B, V0.4B[2] +SDOT V22.4S, V0.16B, V11.4B[2] +SDOT V22.4S, V0.16B, V22.4B[2] +SDOT V22.4S, V0.16B, V0.4B[3] +SDOT V22.4S, V0.16B, V11.4B[3] +SDOT V22.4S, V0.16B, V22.4B[3] +SDOT V22.4S, V11.16B, V0.4B[0] +SDOT V22.4S, V11.16B, V11.4B[0] +SDOT V22.4S, V11.16B, V22.4B[0] +SDOT V22.4S, V11.16B, V0.4B[1] +SDOT V22.4S, V11.16B, V11.4B[1] +SDOT V22.4S, V11.16B, V22.4B[1] +SDOT V22.4S, V11.16B, V0.4B[2] +SDOT V22.4S, V11.16B, V11.4B[2] +SDOT V22.4S, V11.16B, V22.4B[2] +SDOT V22.4S, V11.16B, V0.4B[3] +SDOT V22.4S, V11.16B, V11.4B[3] +SDOT V22.4S, V11.16B, V22.4B[3] +SDOT V22.4S, V22.16B, V0.4B[0] +SDOT V22.4S, V22.16B, V11.4B[0] +SDOT V22.4S, V22.16B, V22.4B[0] +SDOT V22.4S, V22.16B, V0.4B[1] +SDOT V22.4S, V22.16B, V11.4B[1] +SDOT V22.4S, V22.16B, V22.4B[1] +SDOT V22.4S, V22.16B, V0.4B[2] +SDOT V22.4S, V22.16B, V11.4B[2] +SDOT V22.4S, V22.16B, V22.4B[2] +SDOT V22.4S, V22.16B, V0.4B[3] +SDOT V22.4S, V22.16B, V11.4B[3] +SDOT V22.4S, V22.16B, V22.4B[3] diff --git a/include/ChangeLog b/include/ChangeLog index ce1e93d3a94..cd5a51fc5a9 100644 --- a/include/ChangeLog +++ b/include/ChangeLog @@ -1,3 +1,8 @@ +2017-06-28 Tamar Christina + + * opcode/aarch64.h: (AARCH64_FEATURE_DOTPROD): New. + (aarch64_insn_class): Added dotprod. + 2017-06-28 Jiong Wang * opcode/arm.h (FPU_NEON_EXT_DOTPROD): New macro. diff --git a/include/opcode/aarch64.h b/include/opcode/aarch64.h index 37e2486b3c4..c5788e2d09d 100644 --- a/include/opcode/aarch64.h +++ b/include/opcode/aarch64.h @@ -55,6 +55,7 @@ typedef uint32_t aarch64_insn; #define AARCH64_FEATURE_SVE 0x10000000 /* SVE instructions. */ #define AARCH64_FEATURE_RCPC 0x20000000 /* RCPC instructions. */ #define AARCH64_FEATURE_COMPNUM 0x40000000 /* Complex # instructions. */ +#define AARCH64_FEATURE_DOTPROD 0x080000000 /* Dot Product instructions. */ /* Architectures are the sum of the base and extensions. */ #define AARCH64_ARCH_V8 AARCH64_FEATURE (AARCH64_FEATURE_V8, \ @@ -508,6 +509,7 @@ enum aarch64_insn_class sve_size_hsd, sve_size_sd, testbranch, + dotproduct, }; /* Opcode enumerators. */ diff --git a/opcodes/ChangeLog b/opcodes/ChangeLog index 8cc85205d00..db23da32065 100644 --- a/opcodes/ChangeLog +++ b/opcodes/ChangeLog @@ -1,3 +1,12 @@ +2017-06-28 Tamar Christina + + * aarch64-asm.c (aarch64_ins_reglane): Added 4B dotprod. + * aarch64-dis.c (aarch64_ext_reglane): Likewise. + * aarch64-tbl.h (QL_V3DOT, QL_V2DOT): New. + (aarch64_feature_dotprod, DOT_INSN): New. + (udot, sdot): New. + * aarch64-dis-2.c: Regenerated. + 2017-06-28 Jiong Wang * arm-dis.c (coprocessor_opcodes): New entries for vsdot and vudot. diff --git a/opcodes/aarch64-asm.c b/opcodes/aarch64-asm.c index 6d2c75a068e..345d59902e0 100644 --- a/opcodes/aarch64-asm.c +++ b/opcodes/aarch64-asm.c @@ -121,6 +121,20 @@ aarch64_ins_reglane (const aarch64_operand *self, const aarch64_opnd_info *info, insert_field (FLD_imm5, code, value, 0); } } + else if (inst->opcode->iclass == dotproduct) + { + unsigned reglane_index = info->reglane.index; + switch (info->qualifier) + { + case AARCH64_OPND_QLF_S_B: + /* L:H */ + assert (reglane_index < 4); + insert_fields (code, reglane_index, 0, 2, FLD_L, FLD_H); + break; + default: + assert (0); + } + } else { /* index for e.g. SQDMLAL , , .[] diff --git a/opcodes/aarch64-dis-2.c b/opcodes/aarch64-dis-2.c index fe71ebc5189..bbe066929e7 100644 --- a/opcodes/aarch64-dis-2.c +++ b/opcodes/aarch64-dis-2.c @@ -12771,11 +12771,33 @@ aarch64_opcode_lookup_1 (uint32_t word) { if (((word >> 11) & 0x1) == 0) { - /* 33222222222211111111110000000000 - 10987654321098765432109876543210 - xxxxxxxxxx10x001xxxxx0xx01110xxx - sqrdmlah. */ - return 364; + if (((word >> 12) & 0x1) == 0) + { + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxx100001xxxxx0xx01110xxx + sqrdmlah. */ + return 364; + } + else + { + if (((word >> 29) & 0x1) == 0) + { + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxx101001xxxxx0xx011100xx + sdot. */ + return 1987; + } + else + { + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxx101001xxxxx0xx011101xx + udot. */ + return 1986; + } + } } else { @@ -17931,13 +17953,13 @@ aarch64_opcode_lookup_1 (uint32_t word) { if (((word >> 28) & 0x1) == 0) { - if (((word >> 10) & 0x1) == 0) + if (((word >> 14) & 0x1) == 0) { - if (((word >> 12) & 0x1) == 0) + if (((word >> 10) & 0x1) == 0) { - if (((word >> 13) & 0x1) == 0) + if (((word >> 12) & 0x1) == 0) { - if (((word >> 14) & 0x1) == 0) + if (((word >> 13) & 0x1) == 0) { /* 33222222222211111111110000000000 10987654321098765432109876543210 @@ -17947,60 +17969,49 @@ aarch64_opcode_lookup_1 (uint32_t word) } else { - /* 33222222222211111111110000000000 - 10987654321098765432109876543210 - xxxxxxxxxx0x0011xxxxxxxx11110xxx - sqdmulh. */ - return 109; - } - } - else - { - if (((word >> 29) & 0x1) == 0) - { - if (((word >> 30) & 0x1) == 0) - { - /* 33222222222211111111110000000000 - 10987654321098765432109876543210 - xxxxxxxxxx0x01x1xxxxxxxx1111000x - smull. */ - return 105; - } - else - { - /* 33222222222211111111110000000000 - 10987654321098765432109876543210 - xxxxxxxxxx0x01x1xxxxxxxx1111001x - smull2. */ - return 106; - } - } - else - { - if (((word >> 30) & 0x1) == 0) + if (((word >> 29) & 0x1) == 0) { - /* 33222222222211111111110000000000 - 10987654321098765432109876543210 - xxxxxxxxxx0x01x1xxxxxxxx1111010x - umull. */ - return 123; + if (((word >> 30) & 0x1) == 0) + { + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxx0x0101xxxxxxxx1111000x + smull. */ + return 105; + } + else + { + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxx0x0101xxxxxxxx1111001x + smull2. */ + return 106; + } } else { - /* 33222222222211111111110000000000 - 10987654321098765432109876543210 - xxxxxxxxxx0x01x1xxxxxxxx1111011x - umull2. */ - return 124; + if (((word >> 30) & 0x1) == 0) + { + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxx0x0101xxxxxxxx1111010x + umull. */ + return 123; + } + else + { + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxx0x0101xxxxxxxx1111011x + umull2. */ + return 124; + } } } } - } - else - { - if (((word >> 13) & 0x1) == 0) + else { - if (((word >> 14) & 0x1) == 0) + if (((word >> 13) & 0x1) == 0) { if (((word >> 23) & 0x1) == 0) { @@ -18042,28 +18053,6 @@ aarch64_opcode_lookup_1 (uint32_t word) } } else - { - if (((word >> 29) & 0x1) == 0) - { - /* 33222222222211111111110000000000 - 10987654321098765432109876543210 - xxxxxxxxxx0x1011xxxxxxxx111100xx - sqrdmulh. */ - return 110; - } - else - { - /* 33222222222211111111110000000000 - 10987654321098765432109876543210 - xxxxxxxxxx0x1011xxxxxxxx111101xx - sqrdmlah. */ - return 127; - } - } - } - else - { - if (((word >> 14) & 0x1) == 0) { if (((word >> 30) & 0x1) == 0) { @@ -18082,22 +18071,11 @@ aarch64_opcode_lookup_1 (uint32_t word) return 108; } } - else - { - /* 33222222222211111111110000000000 - 10987654321098765432109876543210 - xxxxxxxxxx0x1111xxxxxxxx11110xxx - sqrdmlsh. */ - return 128; - } } } - } - else - { - if (((word >> 11) & 0x1) == 0) + else { - if (((word >> 14) & 0x1) == 0) + if (((word >> 11) & 0x1) == 0) { if (((word >> 12) & 0x1) == 0) { @@ -18140,173 +18118,217 @@ aarch64_opcode_lookup_1 (uint32_t word) } else { - if (((word >> 13) & 0x1) == 0) + if (((word >> 12) & 0x1) == 0) { if (((word >> 29) & 0x1) == 0) { - /* 33222222222211111111110000000000 - 10987654321098765432109876543210 - xxxxxxxxxx10x011xxxxxxxx111100xx - movi. */ - return 135; + if (((word >> 30) & 0x1) == 0) + { + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxx110x01xxxxxxxx1111000x + rshrn. */ + return 376; + } + else + { + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxx110x01xxxxxxxx1111001x + rshrn2. */ + return 377; + } } else { - /* 33222222222211111111110000000000 - 10987654321098765432109876543210 - xxxxxxxxxx10x011xxxxxxxx111101xx - mvni. */ - return 143; + if (((word >> 30) & 0x1) == 0) + { + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxx110x01xxxxxxxx1111010x + sqrshrun. */ + return 400; + } + else + { + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxx110x01xxxxxxxx1111011x + sqrshrun2. */ + return 401; + } } } else { - if (((word >> 12) & 0x1) == 0) + if (((word >> 29) & 0x1) == 0) { - if (((word >> 29) & 0x1) == 0) + if (((word >> 30) & 0x1) == 0) { /* 33222222222211111111110000000000 10987654321098765432109876543210 - xxxxxxxxxx100111xxxxxxxx111100xx - movi. */ - return 136; + xxxxxxxxxx111x01xxxxxxxx1111000x + sqrshrn. */ + return 380; } else { /* 33222222222211111111110000000000 10987654321098765432109876543210 - xxxxxxxxxx100111xxxxxxxx111101xx - movi. */ - return 144; + xxxxxxxxxx111x01xxxxxxxx1111001x + sqrshrn2. */ + return 381; } } else { - if (((word >> 29) & 0x1) == 0) + if (((word >> 30) & 0x1) == 0) { /* 33222222222211111111110000000000 10987654321098765432109876543210 - xxxxxxxxxx101111xxxxxxxx111100xx - fmov. */ - return 137; + xxxxxxxxxx111x01xxxxxxxx1111010x + uqrshrn. */ + return 404; } else { /* 33222222222211111111110000000000 10987654321098765432109876543210 - xxxxxxxxxx101111xxxxxxxx111101xx - fmov. */ - return 146; + xxxxxxxxxx111x01xxxxxxxx1111011x + uqrshrn2. */ + return 405; } } } } } - else + } + else + { + if (((word >> 13) & 0x1) == 0) { - if (((word >> 12) & 0x1) == 0) + if (((word >> 10) & 0x1) == 0) { - if (((word >> 29) & 0x1) == 0) + if (((word >> 12) & 0x1) == 0) { - if (((word >> 30) & 0x1) == 0) - { - /* 33222222222211111111110000000000 - 10987654321098765432109876543210 - xxxxxxxxxx110xx1xxxxxxxx1111000x - rshrn. */ - return 376; - } - else - { - /* 33222222222211111111110000000000 - 10987654321098765432109876543210 - xxxxxxxxxx110xx1xxxxxxxx1111001x - rshrn2. */ - return 377; - } + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxx0x0011xxxxxxxx11110xxx + sqdmulh. */ + return 109; } else { - if (((word >> 30) & 0x1) == 0) + if (((word >> 29) & 0x1) == 0) { /* 33222222222211111111110000000000 10987654321098765432109876543210 - xxxxxxxxxx110xx1xxxxxxxx1111010x - sqrshrun. */ - return 400; + xxxxxxxxxx0x1011xxxxxxxx111100xx + sqrdmulh. */ + return 110; } else { /* 33222222222211111111110000000000 10987654321098765432109876543210 - xxxxxxxxxx110xx1xxxxxxxx1111011x - sqrshrun2. */ - return 401; + xxxxxxxxxx0x1011xxxxxxxx111101xx + sqrdmlah. */ + return 127; } } } else { - if (((word >> 13) & 0x1) == 0) + if (((word >> 29) & 0x1) == 0) { - if (((word >> 29) & 0x1) == 0) + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxx1xx011xxxxxxxx111100xx + movi. */ + return 135; + } + else + { + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxx1xx011xxxxxxxx111101xx + mvni. */ + return 143; + } + } + } + else + { + if (((word >> 12) & 0x1) == 0) + { + if (((word >> 29) & 0x1) == 0) + { + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxxxx0111xxxxxxxx111100xx + movi. */ + return 136; + } + else + { + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxxxx0111xxxxxxxx111101xx + movi. */ + return 144; + } + } + else + { + if (((word >> 10) & 0x1) == 0) + { + /* 33222222222211111111110000000000 + 10987654321098765432109876543210 + xxxxxxxxxx0x1111xxxxxxxx11110xxx + sqrdmlsh. */ + return 128; + } + else + { + if (((word >> 11) & 0x1) == 0) { - if (((word >> 30) & 0x1) == 0) + if (((word >> 29) & 0x1) == 0) { /* 33222222222211111111110000000000 10987654321098765432109876543210 - xxxxxxxxxx1110x1xxxxxxxx1111000x - sqrshrn. */ - return 380; + xxxxxxxxxx101111xxxxxxxx111100xx + fmov. */ + return 137; } else { /* 33222222222211111111110000000000 10987654321098765432109876543210 - xxxxxxxxxx1110x1xxxxxxxx1111001x - sqrshrn2. */ - return 381; + xxxxxxxxxx101111xxxxxxxx111101xx + fmov. */ + return 146; } } else { - if (((word >> 30) & 0x1) == 0) + if (((word >> 29) & 0x1) == 0) { /* 33222222222211111111110000000000 10987654321098765432109876543210 - xxxxxxxxxx1110x1xxxxxxxx1111010x - uqrshrn. */ - return 404; + xxxxxxxxxx111111xxxxxxxx111100xx + fmov. */ + return 138; } else { /* 33222222222211111111110000000000 10987654321098765432109876543210 - xxxxxxxxxx1110x1xxxxxxxx1111011x - uqrshrn2. */ - return 405; + xxxxxxxxxx111111xxxxxxxx111101xx + fcvtzu. */ + return 412; } } } - else - { - if (((word >> 29) & 0x1) == 0) - { - /* 33222222222211111111110000000000 - 10987654321098765432109876543210 - xxxxxxxxxx1111x1xxxxxxxx111100xx - fmov. */ - return 138; - } - else - { - /* 33222222222211111111110000000000 - 10987654321098765432109876543210 - xxxxxxxxxx1111x1xxxxxxxx111101xx - fcvtzu. */ - return 412; - } - } } } } @@ -18743,11 +18765,13 @@ aarch64_find_next_opcode (const aarch64_opcode *opcode) case 403: return NULL; /* uqshrn2 --> NULL. */ case 136: value = 386; break; /* movi --> scvtf. */ case 386: value = 387; break; /* scvtf --> scvtf. */ - case 387: return NULL; /* scvtf --> NULL. */ + case 387: value = 1989; break; /* scvtf --> sdot. */ + case 1989: return NULL; /* sdot --> NULL. */ case 144: value = 145; break; /* movi --> movi. */ case 145: value = 410; break; /* movi --> ucvtf. */ case 410: value = 411; break; /* ucvtf --> ucvtf. */ - case 411: return NULL; /* ucvtf --> NULL. */ + case 411: value = 1988; break; /* ucvtf --> udot. */ + case 1988: return NULL; /* udot --> NULL. */ case 138: value = 388; break; /* fmov --> fcvtzs. */ case 388: value = 389; break; /* fcvtzs --> fcvtzs. */ case 389: return NULL; /* fcvtzs --> NULL. */ diff --git a/opcodes/aarch64-dis.c b/opcodes/aarch64-dis.c index a860dbcb003..4fd94e8f7a9 100644 --- a/opcodes/aarch64-dis.c +++ b/opcodes/aarch64-dis.c @@ -325,6 +325,21 @@ aarch64_ext_reglane (const aarch64_operand *self, aarch64_opnd_info *info, info->reglane.index = (unsigned) (value >> 1); } } + else if (inst->opcode->iclass == dotproduct) + { + /* Need information in other operand(s) to help decoding. */ + info->qualifier = get_expected_qualifier (inst, info->idx); + switch (info->qualifier) + { + case AARCH64_OPND_QLF_S_B: + /* L:H */ + info->reglane.index = extract_fields (code, 0, 2, FLD_H, FLD_L); + info->reglane.regno &= 0x1f; + break; + default: + return 0; + } + } else { /* Index only for e.g. SQDMLAL , , .[] diff --git a/opcodes/aarch64-tbl.h b/opcodes/aarch64-tbl.h index a7510ff5ead..61c2582d011 100644 --- a/opcodes/aarch64-tbl.h +++ b/opcodes/aarch64-tbl.h @@ -1982,6 +1982,19 @@ { \ QLF3(X,X,NIL), \ } +/* e.g. UDOT .2S, .8B, .8B. */ +#define QL_V3DOT \ +{ \ + QLF3(V_2S, V_8B, V_8B), \ + QLF3(V_4S, V_16B, V_16B),\ +} + +/* e.g. UDOT .2S, .8B, .4B[]. */ +#define QL_V2DOT \ +{ \ + QLF3(V_2S, V_8B, S_B),\ + QLF3(V_4S, V_16B, S_B),\ +} /* Opcode table. */ @@ -2021,6 +2034,8 @@ static const aarch64_feature_set aarch64_feature_compnum = AARCH64_FEATURE (AARCH64_FEATURE_COMPNUM, 0); static const aarch64_feature_set aarch64_feature_rcpc = AARCH64_FEATURE (AARCH64_FEATURE_RCPC, 0); +static const aarch64_feature_set aarch64_feature_dotprod = + AARCH64_FEATURE (AARCH64_FEATURE_V8_2 | AARCH64_FEATURE_DOTPROD, 0); #define CORE &aarch64_feature_v8 #define FP &aarch64_feature_fp @@ -2040,6 +2055,7 @@ static const aarch64_feature_set aarch64_feature_rcpc = #define FP_V8_3 &aarch64_feature_fp_v8_3 #define COMPNUM &aarch64_feature_compnum #define RCPC &aarch64_feature_rcpc +#define DOTPROD &aarch64_feature_dotprod #define CORE_INSN(NAME,OPCODE,MASK,CLASS,OP,OPS,QUALS,FLAGS) \ { NAME, OPCODE, MASK, CLASS, OP, CORE, OPS, QUALS, FLAGS, 0, NULL } @@ -2072,6 +2088,8 @@ static const aarch64_feature_set aarch64_feature_rcpc = { NAME, OPCODE, MASK, CLASS, OP, COMPNUM, OPS, QUALS, FLAGS, 0, NULL } #define RCPC_INSN(NAME,OPCODE,MASK,CLASS,OPS,QUALS,FLAGS) \ { NAME, OPCODE, MASK, CLASS, 0, RCPC, OPS, QUALS, FLAGS, 0, NULL } +#define DOT_INSN(NAME,OPCODE,MASK,CLASS,OPS,QUALS,FLAGS) \ + { NAME, OPCODE, MASK, CLASS, 0, DOTPROD, OPS, QUALS, FLAGS, 0, NULL } struct aarch64_opcode aarch64_opcode_table[] = { @@ -4137,6 +4155,12 @@ struct aarch64_opcode aarch64_opcode_table[] = _SVE_INSN ("fmov", 0x05104000, 0xff30ffe0, sve_size_hsd, 0, OP3 (SVE_Zd, SVE_Pg4_16, FPIMM0), OP_SVE_VM_HSD, F_ALIAS | F_PSEUDO, 0), _SVE_INSN ("orn", 0x05000000, 0xfffc0000, sve_limm, 0, OP3 (SVE_Zd, SVE_Zd, SVE_INV_LIMM), OP_SVE_VVU_BHSD, F_ALIAS | F_PSEUDO, 1), + /* SIMD Dot Product (optional in v8.2-A). */ + DOT_INSN ("udot", 0x2e009400, 0xbf20fc00, dotproduct, OP3 (Vd, Vn, Vm), QL_V3DOT, F_SIZEQ), + DOT_INSN ("sdot", 0xe009400, 0xbf20fc00, dotproduct, OP3 (Vd, Vn, Vm), QL_V3DOT, F_SIZEQ), + DOT_INSN ("udot", 0x2f00e000, 0xbf00f000, dotproduct, OP3 (Vd, Vn, Em), QL_V2DOT, F_SIZEQ), + DOT_INSN ("sdot", 0xf00e000, 0xbf00f000, dotproduct, OP3 (Vd, Vn, Em), QL_V2DOT, F_SIZEQ), + {0, 0, 0, 0, 0, 0, {}, {}, 0, 0, NULL}, }; -- 2.30.2