xref: /openbmc/qemu/target/arm/tcg/translate-sme.c (revision ad66b5cb)
1 /*
2  * AArch64 SME translation
3  *
4  * Copyright (c) 2022 Linaro, Ltd
5  *
6  * This library is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2.1 of the License, or (at your option) any later version.
10  *
11  * This library is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with this library; if not, see <http://www.gnu.org/licenses/>.
18  */
19 
20 #include "qemu/osdep.h"
21 #include "cpu.h"
22 #include "tcg/tcg-op.h"
23 #include "tcg/tcg-op-gvec.h"
24 #include "tcg/tcg-gvec-desc.h"
25 #include "translate.h"
26 #include "exec/helper-gen.h"
27 #include "translate-a64.h"
28 #include "fpu/softfloat.h"
29 
30 
31 /*
32  * Include the generated decoder.
33  */
34 
35 #include "decode-sme.c.inc"
36 
37 
38 /*
39  * Resolve tile.size[index] to a host pointer, where tile and index
40  * are always decoded together, dependent on the element size.
41  */
42 static TCGv_ptr get_tile_rowcol(DisasContext *s, int esz, int rs,
43                                 int tile_index, bool vertical)
44 {
45     int tile = tile_index >> (4 - esz);
46     int index = esz == MO_128 ? 0 : extract32(tile_index, 0, 4 - esz);
47     int pos, len, offset;
48     TCGv_i32 tmp;
49     TCGv_ptr addr;
50 
51     /* Compute the final index, which is Rs+imm. */
52     tmp = tcg_temp_new_i32();
53     tcg_gen_trunc_tl_i32(tmp, cpu_reg(s, rs));
54     tcg_gen_addi_i32(tmp, tmp, index);
55 
56     /* Prepare a power-of-two modulo via extraction of @len bits. */
57     len = ctz32(streaming_vec_reg_size(s)) - esz;
58 
59     if (vertical) {
60         /*
61          * Compute the byte offset of the index within the tile:
62          *     (index % (svl / size)) * size
63          *   = (index % (svl >> esz)) << esz
64          * Perform the power-of-two modulo via extraction of the low @len bits.
65          * Perform the multiply by shifting left by @pos bits.
66          * Perform these operations simultaneously via deposit into zero.
67          */
68         pos = esz;
69         tcg_gen_deposit_z_i32(tmp, tmp, pos, len);
70 
71         /*
72          * For big-endian, adjust the indexed column byte offset within
73          * the uint64_t host words that make up env->zarray[].
74          */
75         if (HOST_BIG_ENDIAN && esz < MO_64) {
76             tcg_gen_xori_i32(tmp, tmp, 8 - (1 << esz));
77         }
78     } else {
79         /*
80          * Compute the byte offset of the index within the tile:
81          *     (index % (svl / size)) * (size * sizeof(row))
82          *   = (index % (svl >> esz)) << (esz + log2(sizeof(row)))
83          */
84         pos = esz + ctz32(sizeof(ARMVectorReg));
85         tcg_gen_deposit_z_i32(tmp, tmp, pos, len);
86 
87         /* Row slices are always aligned and need no endian adjustment. */
88     }
89 
90     /* The tile byte offset within env->zarray is the row. */
91     offset = tile * sizeof(ARMVectorReg);
92 
93     /* Include the byte offset of zarray to make this relative to env. */
94     offset += offsetof(CPUARMState, zarray);
95     tcg_gen_addi_i32(tmp, tmp, offset);
96 
97     /* Add the byte offset to env to produce the final pointer. */
98     addr = tcg_temp_new_ptr();
99     tcg_gen_ext_i32_ptr(addr, tmp);
100     tcg_gen_add_ptr(addr, addr, cpu_env);
101 
102     return addr;
103 }
104 
105 static bool trans_ZERO(DisasContext *s, arg_ZERO *a)
106 {
107     if (!dc_isar_feature(aa64_sme, s)) {
108         return false;
109     }
110     if (sme_za_enabled_check(s)) {
111         gen_helper_sme_zero(cpu_env, tcg_constant_i32(a->imm),
112                             tcg_constant_i32(streaming_vec_reg_size(s)));
113     }
114     return true;
115 }
116 
117 static bool trans_MOVA(DisasContext *s, arg_MOVA *a)
118 {
119     static gen_helper_gvec_4 * const h_fns[5] = {
120         gen_helper_sve_sel_zpzz_b, gen_helper_sve_sel_zpzz_h,
121         gen_helper_sve_sel_zpzz_s, gen_helper_sve_sel_zpzz_d,
122         gen_helper_sve_sel_zpzz_q
123     };
124     static gen_helper_gvec_3 * const cz_fns[5] = {
125         gen_helper_sme_mova_cz_b, gen_helper_sme_mova_cz_h,
126         gen_helper_sme_mova_cz_s, gen_helper_sme_mova_cz_d,
127         gen_helper_sme_mova_cz_q,
128     };
129     static gen_helper_gvec_3 * const zc_fns[5] = {
130         gen_helper_sme_mova_zc_b, gen_helper_sme_mova_zc_h,
131         gen_helper_sme_mova_zc_s, gen_helper_sme_mova_zc_d,
132         gen_helper_sme_mova_zc_q,
133     };
134 
135     TCGv_ptr t_za, t_zr, t_pg;
136     TCGv_i32 t_desc;
137     int svl;
138 
139     if (!dc_isar_feature(aa64_sme, s)) {
140         return false;
141     }
142     if (!sme_smza_enabled_check(s)) {
143         return true;
144     }
145 
146     t_za = get_tile_rowcol(s, a->esz, a->rs, a->za_imm, a->v);
147     t_zr = vec_full_reg_ptr(s, a->zr);
148     t_pg = pred_full_reg_ptr(s, a->pg);
149 
150     svl = streaming_vec_reg_size(s);
151     t_desc = tcg_constant_i32(simd_desc(svl, svl, 0));
152 
153     if (a->v) {
154         /* Vertical slice -- use sme mova helpers. */
155         if (a->to_vec) {
156             zc_fns[a->esz](t_zr, t_za, t_pg, t_desc);
157         } else {
158             cz_fns[a->esz](t_za, t_zr, t_pg, t_desc);
159         }
160     } else {
161         /* Horizontal slice -- reuse sve sel helpers. */
162         if (a->to_vec) {
163             h_fns[a->esz](t_zr, t_za, t_zr, t_pg, t_desc);
164         } else {
165             h_fns[a->esz](t_za, t_zr, t_za, t_pg, t_desc);
166         }
167     }
168     return true;
169 }
170 
171 static bool trans_LDST1(DisasContext *s, arg_LDST1 *a)
172 {
173     typedef void GenLdSt1(TCGv_env, TCGv_ptr, TCGv_ptr, TCGv, TCGv_i32);
174 
175     /*
176      * Indexed by [esz][be][v][mte][st], which is (except for load/store)
177      * also the order in which the elements appear in the function names,
178      * and so how we must concatenate the pieces.
179      */
180 
181 #define FN_LS(F)     { gen_helper_sme_ld1##F, gen_helper_sme_st1##F }
182 #define FN_MTE(F)    { FN_LS(F), FN_LS(F##_mte) }
183 #define FN_HV(F)     { FN_MTE(F##_h), FN_MTE(F##_v) }
184 #define FN_END(L, B) { FN_HV(L), FN_HV(B) }
185 
186     static GenLdSt1 * const fns[5][2][2][2][2] = {
187         FN_END(b, b),
188         FN_END(h_le, h_be),
189         FN_END(s_le, s_be),
190         FN_END(d_le, d_be),
191         FN_END(q_le, q_be),
192     };
193 
194 #undef FN_LS
195 #undef FN_MTE
196 #undef FN_HV
197 #undef FN_END
198 
199     TCGv_ptr t_za, t_pg;
200     TCGv_i64 addr;
201     int svl, desc = 0;
202     bool be = s->be_data == MO_BE;
203     bool mte = s->mte_active[0];
204 
205     if (!dc_isar_feature(aa64_sme, s)) {
206         return false;
207     }
208     if (!sme_smza_enabled_check(s)) {
209         return true;
210     }
211 
212     t_za = get_tile_rowcol(s, a->esz, a->rs, a->za_imm, a->v);
213     t_pg = pred_full_reg_ptr(s, a->pg);
214     addr = tcg_temp_new_i64();
215 
216     tcg_gen_shli_i64(addr, cpu_reg(s, a->rm), a->esz);
217     tcg_gen_add_i64(addr, addr, cpu_reg_sp(s, a->rn));
218 
219     if (mte) {
220         desc = FIELD_DP32(desc, MTEDESC, MIDX, get_mem_index(s));
221         desc = FIELD_DP32(desc, MTEDESC, TBI, s->tbid);
222         desc = FIELD_DP32(desc, MTEDESC, TCMA, s->tcma);
223         desc = FIELD_DP32(desc, MTEDESC, WRITE, a->st);
224         desc = FIELD_DP32(desc, MTEDESC, SIZEM1, (1 << a->esz) - 1);
225         desc <<= SVE_MTEDESC_SHIFT;
226     } else {
227         addr = clean_data_tbi(s, addr);
228     }
229     svl = streaming_vec_reg_size(s);
230     desc = simd_desc(svl, svl, desc);
231 
232     fns[a->esz][be][a->v][mte][a->st](cpu_env, t_za, t_pg, addr,
233                                       tcg_constant_i32(desc));
234     return true;
235 }
236 
237 typedef void GenLdStR(DisasContext *, TCGv_ptr, int, int, int, int);
238 
239 static bool do_ldst_r(DisasContext *s, arg_ldstr *a, GenLdStR *fn)
240 {
241     int svl = streaming_vec_reg_size(s);
242     int imm = a->imm;
243     TCGv_ptr base;
244 
245     if (!sme_za_enabled_check(s)) {
246         return true;
247     }
248 
249     /* ZA[n] equates to ZA0H.B[n]. */
250     base = get_tile_rowcol(s, MO_8, a->rv, imm, false);
251 
252     fn(s, base, 0, svl, a->rn, imm * svl);
253     return true;
254 }
255 
256 TRANS_FEAT(LDR, aa64_sme, do_ldst_r, a, gen_sve_ldr)
257 TRANS_FEAT(STR, aa64_sme, do_ldst_r, a, gen_sve_str)
258 
259 static bool do_adda(DisasContext *s, arg_adda *a, MemOp esz,
260                     gen_helper_gvec_4 *fn)
261 {
262     int svl = streaming_vec_reg_size(s);
263     uint32_t desc = simd_desc(svl, svl, 0);
264     TCGv_ptr za, zn, pn, pm;
265 
266     if (!sme_smza_enabled_check(s)) {
267         return true;
268     }
269 
270     /* Sum XZR+zad to find ZAd. */
271     za = get_tile_rowcol(s, esz, 31, a->zad, false);
272     zn = vec_full_reg_ptr(s, a->zn);
273     pn = pred_full_reg_ptr(s, a->pn);
274     pm = pred_full_reg_ptr(s, a->pm);
275 
276     fn(za, zn, pn, pm, tcg_constant_i32(desc));
277     return true;
278 }
279 
280 TRANS_FEAT(ADDHA_s, aa64_sme, do_adda, a, MO_32, gen_helper_sme_addha_s)
281 TRANS_FEAT(ADDVA_s, aa64_sme, do_adda, a, MO_32, gen_helper_sme_addva_s)
282 TRANS_FEAT(ADDHA_d, aa64_sme_i16i64, do_adda, a, MO_64, gen_helper_sme_addha_d)
283 TRANS_FEAT(ADDVA_d, aa64_sme_i16i64, do_adda, a, MO_64, gen_helper_sme_addva_d)
284 
285 static bool do_outprod(DisasContext *s, arg_op *a, MemOp esz,
286                        gen_helper_gvec_5 *fn)
287 {
288     int svl = streaming_vec_reg_size(s);
289     uint32_t desc = simd_desc(svl, svl, a->sub);
290     TCGv_ptr za, zn, zm, pn, pm;
291 
292     if (!sme_smza_enabled_check(s)) {
293         return true;
294     }
295 
296     /* Sum XZR+zad to find ZAd. */
297     za = get_tile_rowcol(s, esz, 31, a->zad, false);
298     zn = vec_full_reg_ptr(s, a->zn);
299     zm = vec_full_reg_ptr(s, a->zm);
300     pn = pred_full_reg_ptr(s, a->pn);
301     pm = pred_full_reg_ptr(s, a->pm);
302 
303     fn(za, zn, zm, pn, pm, tcg_constant_i32(desc));
304     return true;
305 }
306 
307 static bool do_outprod_fpst(DisasContext *s, arg_op *a, MemOp esz,
308                             gen_helper_gvec_5_ptr *fn)
309 {
310     int svl = streaming_vec_reg_size(s);
311     uint32_t desc = simd_desc(svl, svl, a->sub);
312     TCGv_ptr za, zn, zm, pn, pm, fpst;
313 
314     if (!sme_smza_enabled_check(s)) {
315         return true;
316     }
317 
318     /* Sum XZR+zad to find ZAd. */
319     za = get_tile_rowcol(s, esz, 31, a->zad, false);
320     zn = vec_full_reg_ptr(s, a->zn);
321     zm = vec_full_reg_ptr(s, a->zm);
322     pn = pred_full_reg_ptr(s, a->pn);
323     pm = pred_full_reg_ptr(s, a->pm);
324     fpst = fpstatus_ptr(FPST_FPCR);
325 
326     fn(za, zn, zm, pn, pm, fpst, tcg_constant_i32(desc));
327     return true;
328 }
329 
330 TRANS_FEAT(FMOPA_h, aa64_sme, do_outprod_fpst, a, MO_32, gen_helper_sme_fmopa_h)
331 TRANS_FEAT(FMOPA_s, aa64_sme, do_outprod_fpst, a, MO_32, gen_helper_sme_fmopa_s)
332 TRANS_FEAT(FMOPA_d, aa64_sme_f64f64, do_outprod_fpst, a, MO_64, gen_helper_sme_fmopa_d)
333 
334 /* TODO: FEAT_EBF16 */
335 TRANS_FEAT(BFMOPA, aa64_sme, do_outprod, a, MO_32, gen_helper_sme_bfmopa)
336 
337 TRANS_FEAT(SMOPA_s, aa64_sme, do_outprod, a, MO_32, gen_helper_sme_smopa_s)
338 TRANS_FEAT(UMOPA_s, aa64_sme, do_outprod, a, MO_32, gen_helper_sme_umopa_s)
339 TRANS_FEAT(SUMOPA_s, aa64_sme, do_outprod, a, MO_32, gen_helper_sme_sumopa_s)
340 TRANS_FEAT(USMOPA_s, aa64_sme, do_outprod, a, MO_32, gen_helper_sme_usmopa_s)
341 
342 TRANS_FEAT(SMOPA_d, aa64_sme_i16i64, do_outprod, a, MO_64, gen_helper_sme_smopa_d)
343 TRANS_FEAT(UMOPA_d, aa64_sme_i16i64, do_outprod, a, MO_64, gen_helper_sme_umopa_d)
344 TRANS_FEAT(SUMOPA_d, aa64_sme_i16i64, do_outprod, a, MO_64, gen_helper_sme_sumopa_d)
345 TRANS_FEAT(USMOPA_d, aa64_sme_i16i64, do_outprod, a, MO_64, gen_helper_sme_usmopa_d)
346