[v8,07/23] tcg: Add generic vector ops for comparisons

Message ID	20180106031346.6650-8-richard.henderson@linaro.org
State	New
Headers	show Delivered-To: patch@linaro.org Received-SPF: pass (google.com: domain of qemu-devel-bounces+patch=linaro.org@nongnu.org designates 2001:4830:134:3::11 as permitted sender) client-ip=2001:4830:134:3::11; From: Richard Henderson <richard.henderson@linaro.org> To: qemu-devel@nongnu.org Date: Fri, 5 Jan 2018 19:13:30 -0800 Message-Id: <20180106031346.6650-8-richard.henderson@linaro.org> In-Reply-To: <20180106031346.6650-1-richard.henderson@linaro.org> References: <20180106031346.6650-1-richard.henderson@linaro.org> Subject: [Qemu-devel] [PATCH v8 07/23] tcg: Add generic vector ops for comparisons Precedence: list Errors-To: qemu-devel-bounces+patch=linaro.org@nongnu.org Sender: "Qemu-devel" <qemu-devel-bounces+patch=linaro.org@nongnu.org>
Series	tcg: generic vector operations \| expand [v8,00/23] tcg: generic vector operations [v8,01/23] tcg: Allow multiple word entries into the constant pool [v8,02/23] tcg: Add types and basic operations for host vectors [v8,03/23] tcg: Standardize integral arguments to expanders [v8,04/23] tcg: Add generic vector expanders [v8,05/23] tcg: Add generic vector ops for interleave [v8,06/23] tcg: Add generic vector ops for constant shifts [v8,07/23] tcg: Add generic vector ops for comparisons [v8,08/23] tcg: Add generic vector ops for multiplication [v8,09/23] tcg: Add generic vector ops for extension [v8,10/23] tcg: Add generic helpers for saturating arithmetic [v8,11/23] tcg: Add generic vector helpers with a scalar immediate operand [v8,12/23] tcg/optimize: Handle vector opcodes during optimize [v8,13/23] target/arm: Align vector registers [v8,14/23] target/arm: Use vector infrastructure for aa64 add/sub/logic [v8,15/23] target/arm: Use vector infrastructure for aa64 mov/not/neg [v8,16/23] target/arm: Use vector infrastructure for aa64 dup/movi [v8,17/23] target/arm: Use vector infrastructure for aa64 zip/uzp/trn/xtn [v8,18/23] target/arm: Use vector infrastructure for aa64 constant shifts [v8,19/23] target/arm: Use vector infrastructure for aa64 compares [v8,20/23] target/arm: Use vector infrastructure for aa64 multiplies [v8,21/23] target/arm: Use vector infrastructure for aa64 widening shifts [v8,22/23] tcg/i386: Add vector operations [v8,23/23] tcg/aarch64: Add vector operations

diff --git a/accel/tcg/tcg-runtime.h b/accel/tcg/tcg-runtime.h index cb05a755b8..28abf30d76 100644 --- a/accel/tcg/tcg-runtime.h +++ b/accel/tcg/tcg-runtime.h @@ -193,3 +193,33 @@ DEF_HELPER_FLAGS_4(gvec_trn8, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) DEF_HELPER_FLAGS_4(gvec_trn16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) DEF_HELPER_FLAGS_4(gvec_trn32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) DEF_HELPER_FLAGS_4(gvec_trn64, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) + +DEF_HELPER_FLAGS_4(gvec_eq8, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_eq16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_eq32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_eq64, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) + +DEF_HELPER_FLAGS_4(gvec_ne8, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_ne16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_ne32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_ne64, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) + +DEF_HELPER_FLAGS_4(gvec_lt8, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_lt16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_lt32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_lt64, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) + +DEF_HELPER_FLAGS_4(gvec_le8, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_le16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_le32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_le64, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) + +DEF_HELPER_FLAGS_4(gvec_ltu8, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_ltu16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_ltu32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_ltu64, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) + +DEF_HELPER_FLAGS_4(gvec_leu8, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_leu16, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_leu32, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_4(gvec_leu64, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, i32) diff --git a/tcg/tcg-op-gvec.h b/tcg/tcg-op-gvec.h index 7ca78853f2..c71aded066 100644 --- a/tcg/tcg-op-gvec.h +++ b/tcg/tcg-op-gvec.h @@ -220,6 +220,10 @@ void tcg_gen_gvec_trne(unsigned vece, uint32_t dofs, uint32_t aofs, void tcg_gen_gvec_trno(unsigned vece, uint32_t dofs, uint32_t aofs, uint32_t bofs, uint32_t oprsz, uint32_t maxsz); +void tcg_gen_gvec_cmp(TCGCond cond, unsigned vece, uint32_t dofs, + uint32_t aofs, uint32_t bofs, + uint32_t oprsz, uint32_t maxsz); + /* * 64-bit vector operations. Use these when the register has been allocated * with tcg_global_mem_new_i64, and so we cannot also address it via pointer. diff --git a/tcg/tcg-op.h b/tcg/tcg-op.h index c002523bea..777db6ad59 100644 --- a/tcg/tcg-op.h +++ b/tcg/tcg-op.h @@ -939,6 +939,9 @@ void tcg_gen_uzpo_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b); void tcg_gen_trne_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b); void tcg_gen_trno_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b); +void tcg_gen_cmp_vec(TCGCond cond, unsigned vece, TCGv_vec r, + TCGv_vec a, TCGv_vec b); + void tcg_gen_ld_vec(TCGv_vec r, TCGv_ptr base, TCGArg offset); void tcg_gen_st_vec(TCGv_vec r, TCGv_ptr base, TCGArg offset); void tcg_gen_stl_vec(TCGv_vec r, TCGv_ptr base, TCGArg offset, TCGType t); diff --git a/tcg/tcg-opc.h b/tcg/tcg-opc.h index a085fc077b..d3fa014507 100644 --- a/tcg/tcg-opc.h +++ b/tcg/tcg-opc.h @@ -248,6 +248,8 @@ DEF(uzpo_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_uzp_vec)) DEF(trne_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_trn_vec)) DEF(trno_vec, 1, 2, 0, IMPLVEC | IMPL(TCG_TARGET_HAS_trn_vec)) +DEF(cmp_vec, 1, 2, 1, IMPLVEC) + DEF(last_generic, 0, 0, 0, TCG_OPF_NOT_PRESENT) #if TCG_TARGET_MAYBE_vec diff --git a/tcg/tcg.h b/tcg/tcg.h index bdfe058c45..ceef8742b5 100644 --- a/tcg/tcg.h +++ b/tcg/tcg.h @@ -184,6 +184,7 @@ typedef uint64_t TCGRegSet; #define TCG_TARGET_HAS_zip_vec 0 #define TCG_TARGET_HAS_uzp_vec 0 #define TCG_TARGET_HAS_trn_vec 0 +#define TCG_TARGET_HAS_cmp_vec 0 #else #define TCG_TARGET_MAYBE_vec 1 #endif diff --git a/accel/tcg/tcg-runtime-gvec.c b/accel/tcg/tcg-runtime-gvec.c index fba62f1192..e0cde3216f 100644 --- a/accel/tcg/tcg-runtime-gvec.c +++ b/accel/tcg/tcg-runtime-gvec.c @@ -520,3 +520,27 @@ DO_TRN(gvec_trn8, uint8_t) DO_TRN(gvec_trn16, uint16_t) DO_TRN(gvec_trn32, uint32_t) DO_TRN(gvec_trn64, uint64_t) + +#define DO_CMP1(NAME, TYPE, OP) \ +void HELPER(NAME)(void *d, void *a, void *b, uint32_t desc) \ +{ \ + intptr_t oprsz = simd_oprsz(desc); \ + intptr_t i; \ + for (i = 0; i < oprsz; i += sizeof(vec64)) { \ + *(TYPE *)(d + i) = *(TYPE *)(a + i) OP *(TYPE *)(b + i); \ + } \ + clear_high(d, oprsz, desc); \ +} + +#define DO_CMP2(SZ) \ + DO_CMP1(gvec_eq##SZ, vec##SZ, ==) \ + DO_CMP1(gvec_ne##SZ, vec##SZ, !=) \ + DO_CMP1(gvec_lt##SZ, svec##SZ, <) \ + DO_CMP1(gvec_le##SZ, svec##SZ, <=) \ + DO_CMP1(gvec_ltu##SZ, vec##SZ, <) \ + DO_CMP1(gvec_leu##SZ, vec##SZ, <=) + +DO_CMP2(8) +DO_CMP2(16) +DO_CMP2(32) +DO_CMP2(64) diff --git a/tcg/tcg-op-gvec.c b/tcg/tcg-op-gvec.c index 0ba61dd049..664dcae041 100644 --- a/tcg/tcg-op-gvec.c +++ b/tcg/tcg-op-gvec.c @@ -2020,3 +2020,205 @@ void tcg_gen_gvec_trno(unsigned vece, uint32_t dofs, uint32_t aofs, tcg_debug_assert(vece <= MO_64); tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]); } + +/* Expand OPSZ bytes worth of three-operand operations using i32 elements. */ +static void expand_cmp_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs, + uint32_t opsz, TCGCond cond) +{ + TCGv_i32 t0 = tcg_temp_new_i32(); + TCGv_i32 t1 = tcg_temp_new_i32(); + uint32_t i; + + for (i = 0; i < opsz; i += 4) { + tcg_gen_ld_i32(t0, cpu_env, aofs + i); + tcg_gen_ld_i32(t1, cpu_env, bofs + i); + tcg_gen_setcond_i32(cond, t0, t0, t1); + tcg_gen_neg_i32(t0, t0); + tcg_gen_st_i32(t0, cpu_env, dofs + i); + } + tcg_temp_free_i32(t1); + tcg_temp_free_i32(t0); +} + +static void expand_cmp_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs, + uint32_t opsz, TCGCond cond) +{ + TCGv_i64 t0 = tcg_temp_new_i64(); + TCGv_i64 t1 = tcg_temp_new_i64(); + uint32_t i; + + for (i = 0; i < opsz; i += 8) { + tcg_gen_ld_i64(t0, cpu_env, aofs + i); + tcg_gen_ld_i64(t1, cpu_env, bofs + i); + tcg_gen_setcond_i64(cond, t0, t0, t1); + tcg_gen_neg_i64(t0, t0); + tcg_gen_st_i64(t0, cpu_env, dofs + i); + } + tcg_temp_free_i64(t1); + tcg_temp_free_i64(t0); +} + +static void expand_cmp_vec(unsigned vece, uint32_t dofs, uint32_t aofs, + uint32_t bofs, uint32_t opsz, uint32_t tysz, + TCGType type, TCGCond cond) +{ + TCGv_vec t0 = tcg_temp_new_vec(type); + TCGv_vec t1 = tcg_temp_new_vec(type); + uint32_t i; + + for (i = 0; i < opsz; i += tysz) { + tcg_gen_ld_vec(t0, cpu_env, aofs + i); + tcg_gen_ld_vec(t1, cpu_env, bofs + i); + tcg_gen_cmp_vec(cond, vece, t0, t0, t1); + tcg_gen_st_vec(t0, cpu_env, dofs + i); + } + tcg_temp_free_vec(t1); + tcg_temp_free_vec(t0); +} + +void tcg_gen_gvec_cmp(TCGCond cond, unsigned vece, uint32_t dofs, + uint32_t aofs, uint32_t bofs, + uint32_t oprsz, uint32_t maxsz) +{ + static gen_helper_gvec_3 * const eq_fn[4] = { + gen_helper_gvec_eq8, gen_helper_gvec_eq16, + gen_helper_gvec_eq32, gen_helper_gvec_eq64 + }; + static gen_helper_gvec_3 * const ne_fn[4] = { + gen_helper_gvec_ne8, gen_helper_gvec_ne16, + gen_helper_gvec_ne32, gen_helper_gvec_ne64 + }; + static gen_helper_gvec_3 * const lt_fn[4] = { + gen_helper_gvec_lt8, gen_helper_gvec_lt16, + gen_helper_gvec_lt32, gen_helper_gvec_lt64 + }; + static gen_helper_gvec_3 * const le_fn[4] = { + gen_helper_gvec_le8, gen_helper_gvec_le16, + gen_helper_gvec_le32, gen_helper_gvec_le64 + }; + static gen_helper_gvec_3 * const ltu_fn[4] = { + gen_helper_gvec_ltu8, gen_helper_gvec_ltu16, + gen_helper_gvec_ltu32, gen_helper_gvec_ltu64 + }; + static gen_helper_gvec_3 * const leu_fn[4] = { + gen_helper_gvec_leu8, gen_helper_gvec_leu16, + gen_helper_gvec_leu32, gen_helper_gvec_leu64 + }; + gen_helper_gvec_3 *fn; + uint32_t tmp; + + check_size_align(oprsz, maxsz, dofs | aofs | bofs); + check_overlap_3(dofs, aofs, bofs, maxsz); + + if (cond == TCG_COND_NEVER || cond == TCG_COND_ALWAYS) { + tcg_gen_gvec_dup32i(dofs, oprsz, maxsz, -(cond == TCG_COND_ALWAYS)); + return; + } + + /* Quick check for sizes we won't support inline. */ + if (oprsz > MAX_UNROLL * 32 || maxsz > MAX_UNROLL * 32) { + goto do_ool; + } + + /* Recall that ARM SVE allows vector sizes that are not a power of 2. + Expand with successively smaller host vector sizes. The intent is + that e.g. oprsz == 80 would be expanded with 2x32 + 1x16. */ + /* ??? For maxsz > oprsz, the host may be able to use an op-sized + operation, zeroing the balance of the register. We can then + use a cl-sized store to implement the clearing without an extra + store operation. This is true for aarch64 and x86_64 hosts. */ + + if (TCG_TARGET_HAS_v256 && check_size_impl(oprsz, 32) + && tcg_can_emit_vec_op(INDEX_op_cmp_vec, TCG_TYPE_V256, vece)) { + uint32_t done = QEMU_ALIGN_DOWN(oprsz, 32); + expand_cmp_vec(vece, dofs, aofs, bofs, done, 32, TCG_TYPE_V256, cond); + dofs += done; + aofs += done; + bofs += done; + oprsz -= done; + maxsz -= done; + } + + if (TCG_TARGET_HAS_v128 && check_size_impl(oprsz, 16) + && tcg_can_emit_vec_op(INDEX_op_cmp_vec, TCG_TYPE_V128, vece)) { + uint32_t done = QEMU_ALIGN_DOWN(oprsz, 16); + expand_cmp_vec(vece, dofs, aofs, bofs, done, 16, TCG_TYPE_V128, cond); + dofs += done; + aofs += done; + bofs += done; + oprsz -= done; + maxsz -= done; + } + + if (check_size_impl(oprsz, 8)) { + uint32_t done = QEMU_ALIGN_DOWN(oprsz, 8); + if (TCG_TARGET_HAS_v64 + && (TCG_TARGET_REG_BITS == 32 || vece != MO_64) + && tcg_can_emit_vec_op(INDEX_op_cmp_vec, TCG_TYPE_V64, vece)) { + expand_cmp_vec(vece, dofs, aofs, bofs, done, 8, TCG_TYPE_V64, cond); + } else if (vece == MO_64) { + expand_cmp_i64(dofs, aofs, bofs, done, cond); + } else { + done = 0; + } + dofs += done; + aofs += done; + bofs += done; + oprsz -= done; + maxsz -= done; + } + + if (vece == MO_32 && check_size_impl(oprsz, 4)) { + uint32_t done = QEMU_ALIGN_DOWN(oprsz, 4); + expand_cmp_i32(dofs, aofs, bofs, done, cond); + dofs += done; + aofs += done; + bofs += done; + oprsz -= done; + maxsz -= done; + } + + if (oprsz == 0) { + if (maxsz != 0) { + expand_clr(dofs, maxsz); + } + return; + } + + do_ool: + switch (cond) { + case TCG_COND_EQ: + fn = eq_fn[vece]; + break; + case TCG_COND_NE: + fn = ne_fn[vece]; + break; + case TCG_COND_GT: + tmp = aofs, aofs = bofs, bofs = tmp; + /* fallthru */ + case TCG_COND_LT: + fn = lt_fn[vece]; + break; + case TCG_COND_GE: + tmp = aofs, aofs = bofs, bofs = tmp; + /* fallthru */ + case TCG_COND_LE: + fn = le_fn[vece]; + break; + case TCG_COND_GTU: + tmp = aofs, aofs = bofs, bofs = tmp; + /* fallthru */ + case TCG_COND_LTU: + fn = ltu_fn[vece]; + break; + case TCG_COND_GEU: + tmp = aofs, aofs = bofs, bofs = tmp; + /* fallthru */ + case TCG_COND_LEU: + fn = leu_fn[vece]; + break; + default: + g_assert_not_reached(); + } + tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz, maxsz, 0, fn); +} diff --git a/tcg/tcg-op-vec.c b/tcg/tcg-op-vec.c index b54d3de19f..b3f0ec324a 100644 --- a/tcg/tcg-op-vec.c +++ b/tcg/tcg-op-vec.c @@ -480,3 +480,26 @@ void tcg_gen_trno_vec(unsigned vece, TCGv_vec r, TCGv_vec a, TCGv_vec b) { do_interleave(INDEX_op_trno_vec, vece, r, a, b); } + +void tcg_gen_cmp_vec(TCGCond cond, unsigned vece, + TCGv_vec r, TCGv_vec a, TCGv_vec b) +{ + TCGTemp *rt = tcgv_vec_temp(r); + TCGTemp *at = tcgv_vec_temp(a); + TCGTemp *bt = tcgv_vec_temp(b); + TCGArg ri = temp_arg(rt); + TCGArg ai = temp_arg(at); + TCGArg bi = temp_arg(bt); + TCGType type = rt->base_type; + int can; + + tcg_debug_assert(at->base_type == type); + tcg_debug_assert(bt->base_type == type); + can = tcg_can_emit_vec_op(INDEX_op_cmp_vec, type, vece); + if (can > 0) { + vec_gen_4(INDEX_op_cmp_vec, type, vece, ri, ai, bi, cond); + } else { + tcg_debug_assert(can < 0); + tcg_expand_vec_op(INDEX_op_cmp_vec, type, vece, ri, ai, bi, cond); + } +} diff --git a/tcg/tcg.c b/tcg/tcg.c index f82d6e80b0..a85547a6d2 100644 --- a/tcg/tcg.c +++ b/tcg/tcg.c @@ -1392,6 +1392,7 @@ bool tcg_op_supported(TCGOpcode op) case INDEX_op_and_vec: case INDEX_op_or_vec: case INDEX_op_xor_vec: + case INDEX_op_cmp_vec: return have_vec; case INDEX_op_dup2_vec: return have_vec && TCG_TARGET_REG_BITS == 32; @@ -1791,6 +1792,7 @@ void tcg_dump_ops(TCGContext *s) case INDEX_op_brcond_i64: case INDEX_op_setcond_i64: case INDEX_op_movcond_i64: + case INDEX_op_cmp_vec: if (op->args[k] < ARRAY_SIZE(cond_name) && cond_name[op->args[k]]) { col += qemu_log(",%s", cond_name[op->args[k++]]); diff --git a/tcg/README b/tcg/README index 75db47922d..18b6bbd8f1 100644 --- a/tcg/README +++ b/tcg/README @@ -630,6 +630,10 @@ E.g. VECL=1 -> 64 << 1 -> v128, and VECE=2 -> 1 << 2 -> i32. v0[2i + 1] = v2[2i + part]; } +* cmp_vec v0, v1, v2, cond + + Compare vectors by element, storing -1 for true and 0 for false. + ********* Note 1: Some shortcuts are defined when the last operand is known to be

[v8,07/23] tcg: Add generic vector ops for comparisons

Commit Message

Patch