1a0c9400aSSong Gao /* SPDX-License-Identifier: GPL-2.0-or-later */ 2a0c9400aSSong Gao /* 31dc33f26SSong Gao * QEMU LoongArch vector helper functions. 4a0c9400aSSong Gao * 5a0c9400aSSong Gao * Copyright (c) 2022-2023 Loongson Technology Corporation Limited 6a0c9400aSSong Gao */ 7c037fbc9SSong Gao 8c037fbc9SSong Gao #include "qemu/osdep.h" 9c037fbc9SSong Gao #include "cpu.h" 10c037fbc9SSong Gao #include "exec/exec-all.h" 11c037fbc9SSong Gao #include "exec/helper-proto.h" 12aca67472SSong Gao #include "fpu/softfloat.h" 13aca67472SSong Gao #include "internals.h" 14d0dfa19aSSong Gao #include "tcg/tcg.h" 15008a3b16SSong Gao #include "vec.h" 1664cf6b99SSong Gao #include "tcg/tcg-gvec-desc.h" 17c037fbc9SSong Gao 18c037fbc9SSong Gao #define DO_ADD(a, b) (a + b) 19c037fbc9SSong Gao #define DO_SUB(a, b) (a - b) 20c037fbc9SSong Gao 21c037fbc9SSong Gao #define DO_ODD_EVEN(NAME, BIT, E1, E2, DO_OP) \ 2204711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 23c037fbc9SSong Gao { \ 24c037fbc9SSong Gao int i; \ 2504711da1SSong Gao VReg *Vd = (VReg *)vd; \ 2604711da1SSong Gao VReg *Vj = (VReg *)vj; \ 2704711da1SSong Gao VReg *Vk = (VReg *)vk; \ 28c037fbc9SSong Gao typedef __typeof(Vd->E1(0)) TD; \ 2964cf6b99SSong Gao int oprsz = simd_oprsz(desc); \ 30c037fbc9SSong Gao \ 3164cf6b99SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 32c037fbc9SSong Gao Vd->E1(i) = DO_OP((TD)Vj->E2(2 * i + 1), (TD)Vk->E2(2 * i)); \ 33c037fbc9SSong Gao } \ 34c037fbc9SSong Gao } 35c037fbc9SSong Gao 36c037fbc9SSong Gao DO_ODD_EVEN(vhaddw_h_b, 16, H, B, DO_ADD) 37c037fbc9SSong Gao DO_ODD_EVEN(vhaddw_w_h, 32, W, H, DO_ADD) 38c037fbc9SSong Gao DO_ODD_EVEN(vhaddw_d_w, 64, D, W, DO_ADD) 39c037fbc9SSong Gao 4004711da1SSong Gao void HELPER(vhaddw_q_d)(void *vd, void *vj, void *vk, uint32_t desc) 41c037fbc9SSong Gao { 4264cf6b99SSong Gao int i; 4304711da1SSong Gao VReg *Vd = (VReg *)vd; 4404711da1SSong Gao VReg *Vj = (VReg *)vj; 4504711da1SSong Gao VReg *Vk = (VReg *)vk; 4664cf6b99SSong Gao int oprsz = simd_oprsz(desc); 47c037fbc9SSong Gao 4864cf6b99SSong Gao for (i = 0; i < oprsz / 16 ; i++) { 4964cf6b99SSong Gao Vd->Q(i) = int128_add(int128_makes64(Vj->D(2 * i + 1)), 5064cf6b99SSong Gao int128_makes64(Vk->D(2 * i))); 5164cf6b99SSong Gao } 52c037fbc9SSong Gao } 53c037fbc9SSong Gao 54c037fbc9SSong Gao DO_ODD_EVEN(vhsubw_h_b, 16, H, B, DO_SUB) 55c037fbc9SSong Gao DO_ODD_EVEN(vhsubw_w_h, 32, W, H, DO_SUB) 56c037fbc9SSong Gao DO_ODD_EVEN(vhsubw_d_w, 64, D, W, DO_SUB) 57c037fbc9SSong Gao 5804711da1SSong Gao void HELPER(vhsubw_q_d)(void *vd, void *vj, void *vk, uint32_t desc) 59c037fbc9SSong Gao { 6064cf6b99SSong Gao int i; 6104711da1SSong Gao VReg *Vd = (VReg *)vd; 6204711da1SSong Gao VReg *Vj = (VReg *)vj; 6304711da1SSong Gao VReg *Vk = (VReg *)vk; 6464cf6b99SSong Gao int oprsz = simd_oprsz(desc); 65c037fbc9SSong Gao 6664cf6b99SSong Gao for (i = 0; i < oprsz / 16; i++) { 6764cf6b99SSong Gao Vd->Q(i) = int128_sub(int128_makes64(Vj->D(2 * i + 1)), 6864cf6b99SSong Gao int128_makes64(Vk->D(2 * i))); 6964cf6b99SSong Gao } 70c037fbc9SSong Gao } 71c037fbc9SSong Gao 72c037fbc9SSong Gao DO_ODD_EVEN(vhaddw_hu_bu, 16, UH, UB, DO_ADD) 73c037fbc9SSong Gao DO_ODD_EVEN(vhaddw_wu_hu, 32, UW, UH, DO_ADD) 74c037fbc9SSong Gao DO_ODD_EVEN(vhaddw_du_wu, 64, UD, UW, DO_ADD) 75c037fbc9SSong Gao 7604711da1SSong Gao void HELPER(vhaddw_qu_du)(void *vd, void *vj, void *vk, uint32_t desc) 77c037fbc9SSong Gao { 7864cf6b99SSong Gao int i; 7904711da1SSong Gao VReg *Vd = (VReg *)vd; 8004711da1SSong Gao VReg *Vj = (VReg *)vj; 8104711da1SSong Gao VReg *Vk = (VReg *)vk; 8264cf6b99SSong Gao int oprsz = simd_oprsz(desc); 83c037fbc9SSong Gao 8464cf6b99SSong Gao for (i = 0; i < oprsz / 16; i ++) { 8564cf6b99SSong Gao Vd->Q(i) = int128_add(int128_make64(Vj->UD(2 * i + 1)), 8664cf6b99SSong Gao int128_make64(Vk->UD(2 * i))); 8764cf6b99SSong Gao } 88c037fbc9SSong Gao } 89c037fbc9SSong Gao 90c037fbc9SSong Gao DO_ODD_EVEN(vhsubw_hu_bu, 16, UH, UB, DO_SUB) 91c037fbc9SSong Gao DO_ODD_EVEN(vhsubw_wu_hu, 32, UW, UH, DO_SUB) 92c037fbc9SSong Gao DO_ODD_EVEN(vhsubw_du_wu, 64, UD, UW, DO_SUB) 93c037fbc9SSong Gao 9404711da1SSong Gao void HELPER(vhsubw_qu_du)(void *vd, void *vj, void *vk, uint32_t desc) 95c037fbc9SSong Gao { 9664cf6b99SSong Gao int i; 9704711da1SSong Gao VReg *Vd = (VReg *)vd; 9804711da1SSong Gao VReg *Vj = (VReg *)vj; 9904711da1SSong Gao VReg *Vk = (VReg *)vk; 10064cf6b99SSong Gao int oprsz = simd_oprsz(desc); 101c037fbc9SSong Gao 10264cf6b99SSong Gao for (i = 0; i < oprsz / 16; i++) { 10364cf6b99SSong Gao Vd->Q(i) = int128_sub(int128_make64(Vj->UD(2 * i + 1)), 10464cf6b99SSong Gao int128_make64(Vk->UD(2 * i))); 10564cf6b99SSong Gao } 106c037fbc9SSong Gao } 1072d5f950cSSong Gao 1082d5f950cSSong Gao #define DO_EVEN(NAME, BIT, E1, E2, DO_OP) \ 10985995f07SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1102d5f950cSSong Gao { \ 1112d5f950cSSong Gao int i; \ 1122d5f950cSSong Gao VReg *Vd = (VReg *)vd; \ 1132d5f950cSSong Gao VReg *Vj = (VReg *)vj; \ 1142d5f950cSSong Gao VReg *Vk = (VReg *)vk; \ 1152d5f950cSSong Gao typedef __typeof(Vd->E1(0)) TD; \ 11685995f07SSong Gao int oprsz = simd_oprsz(desc); \ 11785995f07SSong Gao \ 11885995f07SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 1192d5f950cSSong Gao Vd->E1(i) = DO_OP((TD)Vj->E2(2 * i) ,(TD)Vk->E2(2 * i)); \ 1202d5f950cSSong Gao } \ 1212d5f950cSSong Gao } 1222d5f950cSSong Gao 1232d5f950cSSong Gao #define DO_ODD(NAME, BIT, E1, E2, DO_OP) \ 12485995f07SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1252d5f950cSSong Gao { \ 1262d5f950cSSong Gao int i; \ 1272d5f950cSSong Gao VReg *Vd = (VReg *)vd; \ 1282d5f950cSSong Gao VReg *Vj = (VReg *)vj; \ 1292d5f950cSSong Gao VReg *Vk = (VReg *)vk; \ 1302d5f950cSSong Gao typedef __typeof(Vd->E1(0)) TD; \ 13185995f07SSong Gao int oprsz = simd_oprsz(desc); \ 13285995f07SSong Gao \ 13385995f07SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 1342d5f950cSSong Gao Vd->E1(i) = DO_OP((TD)Vj->E2(2 * i + 1), (TD)Vk->E2(2 * i + 1)); \ 1352d5f950cSSong Gao } \ 1362d5f950cSSong Gao } 1372d5f950cSSong Gao 13885995f07SSong Gao void HELPER(vaddwev_q_d)(void *vd, void *vj, void *vk, uint32_t desc) 1392d5f950cSSong Gao { 14085995f07SSong Gao int i; 1412d5f950cSSong Gao VReg *Vd = (VReg *)vd; 1422d5f950cSSong Gao VReg *Vj = (VReg *)vj; 1432d5f950cSSong Gao VReg *Vk = (VReg *)vk; 14485995f07SSong Gao int oprsz = simd_oprsz(desc); 1452d5f950cSSong Gao 14685995f07SSong Gao for (i = 0; i < oprsz / 16; i++) { 14785995f07SSong Gao Vd->Q(i) = int128_add(int128_makes64(Vj->D(2 * i)), 14885995f07SSong Gao int128_makes64(Vk->D(2 * i))); 14985995f07SSong Gao } 1502d5f950cSSong Gao } 1512d5f950cSSong Gao 1522d5f950cSSong Gao DO_EVEN(vaddwev_h_b, 16, H, B, DO_ADD) 1532d5f950cSSong Gao DO_EVEN(vaddwev_w_h, 32, W, H, DO_ADD) 1542d5f950cSSong Gao DO_EVEN(vaddwev_d_w, 64, D, W, DO_ADD) 1552d5f950cSSong Gao 15685995f07SSong Gao void HELPER(vaddwod_q_d)(void *vd, void *vj, void *vk, uint32_t desc) 1572d5f950cSSong Gao { 15885995f07SSong Gao int i; 1592d5f950cSSong Gao VReg *Vd = (VReg *)vd; 1602d5f950cSSong Gao VReg *Vj = (VReg *)vj; 1612d5f950cSSong Gao VReg *Vk = (VReg *)vk; 16285995f07SSong Gao int oprsz = simd_oprsz(desc); 1632d5f950cSSong Gao 16485995f07SSong Gao for (i = 0; i < oprsz / 16; i++) { 16585995f07SSong Gao Vd->Q(i) = int128_add(int128_makes64(Vj->D(2 * i +1)), 16685995f07SSong Gao int128_makes64(Vk->D(2 * i +1))); 16785995f07SSong Gao } 1682d5f950cSSong Gao } 1692d5f950cSSong Gao 1702d5f950cSSong Gao DO_ODD(vaddwod_h_b, 16, H, B, DO_ADD) 1712d5f950cSSong Gao DO_ODD(vaddwod_w_h, 32, W, H, DO_ADD) 1722d5f950cSSong Gao DO_ODD(vaddwod_d_w, 64, D, W, DO_ADD) 1732d5f950cSSong Gao 17485995f07SSong Gao void HELPER(vsubwev_q_d)(void *vd, void *vj, void *vk, uint32_t desc) 1752d5f950cSSong Gao { 17685995f07SSong Gao int i; 1772d5f950cSSong Gao VReg *Vd = (VReg *)vd; 1782d5f950cSSong Gao VReg *Vj = (VReg *)vj; 1792d5f950cSSong Gao VReg *Vk = (VReg *)vk; 18085995f07SSong Gao int oprsz = simd_oprsz(desc); 1812d5f950cSSong Gao 18285995f07SSong Gao for (i = 0; i < oprsz / 16; i++) { 18385995f07SSong Gao Vd->Q(i) = int128_sub(int128_makes64(Vj->D(2 * i)), 18485995f07SSong Gao int128_makes64(Vk->D(2 * i))); 18585995f07SSong Gao } 1862d5f950cSSong Gao } 1872d5f950cSSong Gao 1882d5f950cSSong Gao DO_EVEN(vsubwev_h_b, 16, H, B, DO_SUB) 1892d5f950cSSong Gao DO_EVEN(vsubwev_w_h, 32, W, H, DO_SUB) 1902d5f950cSSong Gao DO_EVEN(vsubwev_d_w, 64, D, W, DO_SUB) 1912d5f950cSSong Gao 19285995f07SSong Gao void HELPER(vsubwod_q_d)(void *vd, void *vj, void *vk, uint32_t desc) 1932d5f950cSSong Gao { 19485995f07SSong Gao int i; 1952d5f950cSSong Gao VReg *Vd = (VReg *)vd; 1962d5f950cSSong Gao VReg *Vj = (VReg *)vj; 1972d5f950cSSong Gao VReg *Vk = (VReg *)vk; 19885995f07SSong Gao int oprsz = simd_oprsz(desc); 1992d5f950cSSong Gao 20085995f07SSong Gao for (i = 0; i < oprsz / 16; i++) { 20185995f07SSong Gao Vd->Q(i) = int128_sub(int128_makes64(Vj->D(2 * i + 1)), 20285995f07SSong Gao int128_makes64(Vk->D(2 * i + 1))); 20385995f07SSong Gao } 2042d5f950cSSong Gao } 2052d5f950cSSong Gao 2062d5f950cSSong Gao DO_ODD(vsubwod_h_b, 16, H, B, DO_SUB) 2072d5f950cSSong Gao DO_ODD(vsubwod_w_h, 32, W, H, DO_SUB) 2082d5f950cSSong Gao DO_ODD(vsubwod_d_w, 64, D, W, DO_SUB) 2092d5f950cSSong Gao 21085995f07SSong Gao void HELPER(vaddwev_q_du)(void *vd, void *vj, void *vk, uint32_t desc) 2112d5f950cSSong Gao { 21285995f07SSong Gao int i; 2132d5f950cSSong Gao VReg *Vd = (VReg *)vd; 2142d5f950cSSong Gao VReg *Vj = (VReg *)vj; 2152d5f950cSSong Gao VReg *Vk = (VReg *)vk; 21685995f07SSong Gao int oprsz = simd_oprsz(desc); 2172d5f950cSSong Gao 21885995f07SSong Gao for (i = 0; i < oprsz / 16; i++) { 21985995f07SSong Gao Vd->Q(i) = int128_add(int128_make64(Vj->UD(2 * i)), 22085995f07SSong Gao int128_make64(Vk->UD(2 * i))); 22185995f07SSong Gao } 2222d5f950cSSong Gao } 2232d5f950cSSong Gao 2242d5f950cSSong Gao DO_EVEN(vaddwev_h_bu, 16, UH, UB, DO_ADD) 2252d5f950cSSong Gao DO_EVEN(vaddwev_w_hu, 32, UW, UH, DO_ADD) 2262d5f950cSSong Gao DO_EVEN(vaddwev_d_wu, 64, UD, UW, DO_ADD) 2272d5f950cSSong Gao 22885995f07SSong Gao void HELPER(vaddwod_q_du)(void *vd, void *vj, void *vk, uint32_t desc) 2292d5f950cSSong Gao { 23085995f07SSong Gao int i; 2312d5f950cSSong Gao VReg *Vd = (VReg *)vd; 2322d5f950cSSong Gao VReg *Vj = (VReg *)vj; 2332d5f950cSSong Gao VReg *Vk = (VReg *)vk; 23485995f07SSong Gao int oprsz = simd_oprsz(desc); 2352d5f950cSSong Gao 23685995f07SSong Gao for (i = 0; i < oprsz / 16; i++) { 23785995f07SSong Gao Vd->Q(i) = int128_add(int128_make64(Vj->UD(2 * i + 1)), 23885995f07SSong Gao int128_make64(Vk->UD(2 * i + 1))); 23985995f07SSong Gao } 2402d5f950cSSong Gao } 2412d5f950cSSong Gao 2422d5f950cSSong Gao DO_ODD(vaddwod_h_bu, 16, UH, UB, DO_ADD) 2432d5f950cSSong Gao DO_ODD(vaddwod_w_hu, 32, UW, UH, DO_ADD) 2442d5f950cSSong Gao DO_ODD(vaddwod_d_wu, 64, UD, UW, DO_ADD) 2452d5f950cSSong Gao 24685995f07SSong Gao void HELPER(vsubwev_q_du)(void *vd, void *vj, void *vk, uint32_t desc) 2472d5f950cSSong Gao { 24885995f07SSong Gao int i; 2492d5f950cSSong Gao VReg *Vd = (VReg *)vd; 2502d5f950cSSong Gao VReg *Vj = (VReg *)vj; 2512d5f950cSSong Gao VReg *Vk = (VReg *)vk; 25285995f07SSong Gao int oprsz = simd_oprsz(desc); 2532d5f950cSSong Gao 25485995f07SSong Gao for (i = 0; i < oprsz / 16; i++) { 25585995f07SSong Gao Vd->Q(i) = int128_sub(int128_make64(Vj->UD(2 * i)), 25685995f07SSong Gao int128_make64(Vk->UD(2 * i))); 25785995f07SSong Gao } 2582d5f950cSSong Gao } 2592d5f950cSSong Gao 2602d5f950cSSong Gao DO_EVEN(vsubwev_h_bu, 16, UH, UB, DO_SUB) 2612d5f950cSSong Gao DO_EVEN(vsubwev_w_hu, 32, UW, UH, DO_SUB) 2622d5f950cSSong Gao DO_EVEN(vsubwev_d_wu, 64, UD, UW, DO_SUB) 2632d5f950cSSong Gao 26485995f07SSong Gao void HELPER(vsubwod_q_du)(void *vd, void *vj, void *vk, uint32_t desc) 2652d5f950cSSong Gao { 26685995f07SSong Gao int i; 2672d5f950cSSong Gao VReg *Vd = (VReg *)vd; 2682d5f950cSSong Gao VReg *Vj = (VReg *)vj; 2692d5f950cSSong Gao VReg *Vk = (VReg *)vk; 27085995f07SSong Gao int oprsz = simd_oprsz(desc); 2712d5f950cSSong Gao 27285995f07SSong Gao for (i = 0; i < oprsz / 16; i++) { 27385995f07SSong Gao Vd->Q(i) = int128_sub(int128_make64(Vj->UD(2 * i + 1)), 27485995f07SSong Gao int128_make64(Vk->UD(2 * i + 1))); 27585995f07SSong Gao } 2762d5f950cSSong Gao } 2772d5f950cSSong Gao 2782d5f950cSSong Gao DO_ODD(vsubwod_h_bu, 16, UH, UB, DO_SUB) 2792d5f950cSSong Gao DO_ODD(vsubwod_w_hu, 32, UW, UH, DO_SUB) 2802d5f950cSSong Gao DO_ODD(vsubwod_d_wu, 64, UD, UW, DO_SUB) 2812d5f950cSSong Gao 2822d5f950cSSong Gao #define DO_EVEN_U_S(NAME, BIT, ES1, EU1, ES2, EU2, DO_OP) \ 28385995f07SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 2842d5f950cSSong Gao { \ 2852d5f950cSSong Gao int i; \ 2862d5f950cSSong Gao VReg *Vd = (VReg *)vd; \ 2872d5f950cSSong Gao VReg *Vj = (VReg *)vj; \ 2882d5f950cSSong Gao VReg *Vk = (VReg *)vk; \ 2892d5f950cSSong Gao typedef __typeof(Vd->ES1(0)) TDS; \ 2902d5f950cSSong Gao typedef __typeof(Vd->EU1(0)) TDU; \ 29185995f07SSong Gao int oprsz = simd_oprsz(desc); \ 29285995f07SSong Gao \ 29385995f07SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 2942d5f950cSSong Gao Vd->ES1(i) = DO_OP((TDU)Vj->EU2(2 * i) ,(TDS)Vk->ES2(2 * i)); \ 2952d5f950cSSong Gao } \ 2962d5f950cSSong Gao } 2972d5f950cSSong Gao 2982d5f950cSSong Gao #define DO_ODD_U_S(NAME, BIT, ES1, EU1, ES2, EU2, DO_OP) \ 29985995f07SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 3002d5f950cSSong Gao { \ 3012d5f950cSSong Gao int i; \ 3022d5f950cSSong Gao VReg *Vd = (VReg *)vd; \ 3032d5f950cSSong Gao VReg *Vj = (VReg *)vj; \ 3042d5f950cSSong Gao VReg *Vk = (VReg *)vk; \ 3052d5f950cSSong Gao typedef __typeof(Vd->ES1(0)) TDS; \ 3062d5f950cSSong Gao typedef __typeof(Vd->EU1(0)) TDU; \ 30785995f07SSong Gao int oprsz = simd_oprsz(desc); \ 30885995f07SSong Gao \ 30985995f07SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 3102d5f950cSSong Gao Vd->ES1(i) = DO_OP((TDU)Vj->EU2(2 * i + 1), (TDS)Vk->ES2(2 * i + 1)); \ 3112d5f950cSSong Gao } \ 3122d5f950cSSong Gao } 3132d5f950cSSong Gao 31485995f07SSong Gao void HELPER(vaddwev_q_du_d)(void *vd, void *vj, void *vk, uint32_t desc) 3152d5f950cSSong Gao { 31685995f07SSong Gao int i; 3172d5f950cSSong Gao VReg *Vd = (VReg *)vd; 3182d5f950cSSong Gao VReg *Vj = (VReg *)vj; 3192d5f950cSSong Gao VReg *Vk = (VReg *)vk; 32085995f07SSong Gao int oprsz = simd_oprsz(desc); 3212d5f950cSSong Gao 32285995f07SSong Gao for (i = 0; i < oprsz / 16; i++) { 32385995f07SSong Gao Vd->Q(i) = int128_add(int128_make64(Vj->UD(2 * i)), 32485995f07SSong Gao int128_makes64(Vk->D(2 * i))); 32585995f07SSong Gao } 3262d5f950cSSong Gao } 3272d5f950cSSong Gao 3282d5f950cSSong Gao DO_EVEN_U_S(vaddwev_h_bu_b, 16, H, UH, B, UB, DO_ADD) 3292d5f950cSSong Gao DO_EVEN_U_S(vaddwev_w_hu_h, 32, W, UW, H, UH, DO_ADD) 3302d5f950cSSong Gao DO_EVEN_U_S(vaddwev_d_wu_w, 64, D, UD, W, UW, DO_ADD) 3312d5f950cSSong Gao 33285995f07SSong Gao void HELPER(vaddwod_q_du_d)(void *vd, void *vj, void *vk, uint32_t desc) 3332d5f950cSSong Gao { 33485995f07SSong Gao int i; 3352d5f950cSSong Gao VReg *Vd = (VReg *)vd; 3362d5f950cSSong Gao VReg *Vj = (VReg *)vj; 3372d5f950cSSong Gao VReg *Vk = (VReg *)vk; 33885995f07SSong Gao int oprsz = simd_oprsz(desc); 3392d5f950cSSong Gao 34085995f07SSong Gao for (i = 0; i < oprsz / 16; i++) { 34185995f07SSong Gao Vd->Q(i) = int128_add(int128_make64(Vj->UD(2 * i + 1)), 34285995f07SSong Gao int128_makes64(Vk->D(2 * i + 1))); 34385995f07SSong Gao } 3442d5f950cSSong Gao } 3452d5f950cSSong Gao 3462d5f950cSSong Gao DO_ODD_U_S(vaddwod_h_bu_b, 16, H, UH, B, UB, DO_ADD) 3472d5f950cSSong Gao DO_ODD_U_S(vaddwod_w_hu_h, 32, W, UW, H, UH, DO_ADD) 3482d5f950cSSong Gao DO_ODD_U_S(vaddwod_d_wu_w, 64, D, UD, W, UW, DO_ADD) 34939e9b0a7SSong Gao 35039e9b0a7SSong Gao #define DO_VAVG(a, b) ((a >> 1) + (b >> 1) + (a & b & 1)) 35139e9b0a7SSong Gao #define DO_VAVGR(a, b) ((a >> 1) + (b >> 1) + ((a | b) & 1)) 35239e9b0a7SSong Gao 35339e9b0a7SSong Gao #define DO_3OP(NAME, BIT, E, DO_OP) \ 354ee7250d0SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 35539e9b0a7SSong Gao { \ 35639e9b0a7SSong Gao int i; \ 35739e9b0a7SSong Gao VReg *Vd = (VReg *)vd; \ 35839e9b0a7SSong Gao VReg *Vj = (VReg *)vj; \ 35939e9b0a7SSong Gao VReg *Vk = (VReg *)vk; \ 360ee7250d0SSong Gao int oprsz = simd_oprsz(desc); \ 361ee7250d0SSong Gao \ 362ee7250d0SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 36339e9b0a7SSong Gao Vd->E(i) = DO_OP(Vj->E(i), Vk->E(i)); \ 36439e9b0a7SSong Gao } \ 36539e9b0a7SSong Gao } 36639e9b0a7SSong Gao 36739e9b0a7SSong Gao DO_3OP(vavg_b, 8, B, DO_VAVG) 36839e9b0a7SSong Gao DO_3OP(vavg_h, 16, H, DO_VAVG) 36939e9b0a7SSong Gao DO_3OP(vavg_w, 32, W, DO_VAVG) 37039e9b0a7SSong Gao DO_3OP(vavg_d, 64, D, DO_VAVG) 37139e9b0a7SSong Gao DO_3OP(vavgr_b, 8, B, DO_VAVGR) 37239e9b0a7SSong Gao DO_3OP(vavgr_h, 16, H, DO_VAVGR) 37339e9b0a7SSong Gao DO_3OP(vavgr_w, 32, W, DO_VAVGR) 37439e9b0a7SSong Gao DO_3OP(vavgr_d, 64, D, DO_VAVGR) 37539e9b0a7SSong Gao DO_3OP(vavg_bu, 8, UB, DO_VAVG) 37639e9b0a7SSong Gao DO_3OP(vavg_hu, 16, UH, DO_VAVG) 37739e9b0a7SSong Gao DO_3OP(vavg_wu, 32, UW, DO_VAVG) 37839e9b0a7SSong Gao DO_3OP(vavg_du, 64, UD, DO_VAVG) 37939e9b0a7SSong Gao DO_3OP(vavgr_bu, 8, UB, DO_VAVGR) 38039e9b0a7SSong Gao DO_3OP(vavgr_hu, 16, UH, DO_VAVGR) 38139e9b0a7SSong Gao DO_3OP(vavgr_wu, 32, UW, DO_VAVGR) 38239e9b0a7SSong Gao DO_3OP(vavgr_du, 64, UD, DO_VAVGR) 38349725659SSong Gao 38449725659SSong Gao #define DO_VABSD(a, b) ((a > b) ? (a -b) : (b-a)) 38549725659SSong Gao 38649725659SSong Gao DO_3OP(vabsd_b, 8, B, DO_VABSD) 38749725659SSong Gao DO_3OP(vabsd_h, 16, H, DO_VABSD) 38849725659SSong Gao DO_3OP(vabsd_w, 32, W, DO_VABSD) 38949725659SSong Gao DO_3OP(vabsd_d, 64, D, DO_VABSD) 39049725659SSong Gao DO_3OP(vabsd_bu, 8, UB, DO_VABSD) 39149725659SSong Gao DO_3OP(vabsd_hu, 16, UH, DO_VABSD) 39249725659SSong Gao DO_3OP(vabsd_wu, 32, UW, DO_VABSD) 39349725659SSong Gao DO_3OP(vabsd_du, 64, UD, DO_VABSD) 394af448cb3SSong Gao 395af448cb3SSong Gao #define DO_VABS(a) ((a < 0) ? (-a) : (a)) 396af448cb3SSong Gao 39727f5485dSSong Gao #define DO_VADDA(NAME, BIT, E) \ 39827f5485dSSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 399af448cb3SSong Gao { \ 400af448cb3SSong Gao int i; \ 401af448cb3SSong Gao VReg *Vd = (VReg *)vd; \ 402af448cb3SSong Gao VReg *Vj = (VReg *)vj; \ 403af448cb3SSong Gao VReg *Vk = (VReg *)vk; \ 40427f5485dSSong Gao int oprsz = simd_oprsz(desc); \ 40527f5485dSSong Gao \ 40627f5485dSSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 40727f5485dSSong Gao Vd->E(i) = DO_VABS(Vj->E(i)) + DO_VABS(Vk->E(i)); \ 408af448cb3SSong Gao } \ 409af448cb3SSong Gao } 410af448cb3SSong Gao 41127f5485dSSong Gao DO_VADDA(vadda_b, 8, B) 41227f5485dSSong Gao DO_VADDA(vadda_h, 16, H) 41327f5485dSSong Gao DO_VADDA(vadda_w, 32, W) 41427f5485dSSong Gao DO_VADDA(vadda_d, 64, D) 4159ab29520SSong Gao 4169ab29520SSong Gao #define DO_MIN(a, b) (a < b ? a : b) 4179ab29520SSong Gao #define DO_MAX(a, b) (a > b ? a : b) 4189ab29520SSong Gao 4199ab29520SSong Gao #define VMINMAXI(NAME, BIT, E, DO_OP) \ 420c09360faSSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 4219ab29520SSong Gao { \ 4229ab29520SSong Gao int i; \ 4239ab29520SSong Gao VReg *Vd = (VReg *)vd; \ 4249ab29520SSong Gao VReg *Vj = (VReg *)vj; \ 4259ab29520SSong Gao typedef __typeof(Vd->E(0)) TD; \ 426c09360faSSong Gao int oprsz = simd_oprsz(desc); \ 4279ab29520SSong Gao \ 428c09360faSSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 4299ab29520SSong Gao Vd->E(i) = DO_OP(Vj->E(i), (TD)imm); \ 4309ab29520SSong Gao } \ 4319ab29520SSong Gao } 4329ab29520SSong Gao 4339ab29520SSong Gao VMINMAXI(vmini_b, 8, B, DO_MIN) 4349ab29520SSong Gao VMINMAXI(vmini_h, 16, H, DO_MIN) 4359ab29520SSong Gao VMINMAXI(vmini_w, 32, W, DO_MIN) 4369ab29520SSong Gao VMINMAXI(vmini_d, 64, D, DO_MIN) 4379ab29520SSong Gao VMINMAXI(vmaxi_b, 8, B, DO_MAX) 4389ab29520SSong Gao VMINMAXI(vmaxi_h, 16, H, DO_MAX) 4399ab29520SSong Gao VMINMAXI(vmaxi_w, 32, W, DO_MAX) 4409ab29520SSong Gao VMINMAXI(vmaxi_d, 64, D, DO_MAX) 4419ab29520SSong Gao VMINMAXI(vmini_bu, 8, UB, DO_MIN) 4429ab29520SSong Gao VMINMAXI(vmini_hu, 16, UH, DO_MIN) 4439ab29520SSong Gao VMINMAXI(vmini_wu, 32, UW, DO_MIN) 4449ab29520SSong Gao VMINMAXI(vmini_du, 64, UD, DO_MIN) 4459ab29520SSong Gao VMINMAXI(vmaxi_bu, 8, UB, DO_MAX) 4469ab29520SSong Gao VMINMAXI(vmaxi_hu, 16, UH, DO_MAX) 4479ab29520SSong Gao VMINMAXI(vmaxi_wu, 32, UW, DO_MAX) 4489ab29520SSong Gao VMINMAXI(vmaxi_du, 64, UD, DO_MAX) 449cd1c49adSSong Gao 450cd1c49adSSong Gao #define DO_VMUH(NAME, BIT, E1, E2, DO_OP) \ 451342dc1cfSSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 452cd1c49adSSong Gao { \ 453cd1c49adSSong Gao int i; \ 454cd1c49adSSong Gao VReg *Vd = (VReg *)vd; \ 455cd1c49adSSong Gao VReg *Vj = (VReg *)vj; \ 456cd1c49adSSong Gao VReg *Vk = (VReg *)vk; \ 457cd1c49adSSong Gao typedef __typeof(Vd->E1(0)) T; \ 458342dc1cfSSong Gao int oprsz = simd_oprsz(desc); \ 459cd1c49adSSong Gao \ 460342dc1cfSSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 461cd1c49adSSong Gao Vd->E2(i) = ((T)Vj->E2(i)) * ((T)Vk->E2(i)) >> BIT; \ 462cd1c49adSSong Gao } \ 463cd1c49adSSong Gao } 464cd1c49adSSong Gao 465342dc1cfSSong Gao void HELPER(vmuh_d)(void *vd, void *vj, void *vk, uint32_t desc) 466cd1c49adSSong Gao { 467342dc1cfSSong Gao int i; 468342dc1cfSSong Gao uint64_t l, h; 469cd1c49adSSong Gao VReg *Vd = (VReg *)vd; 470cd1c49adSSong Gao VReg *Vj = (VReg *)vj; 471cd1c49adSSong Gao VReg *Vk = (VReg *)vk; 472342dc1cfSSong Gao int oprsz = simd_oprsz(desc); 473cd1c49adSSong Gao 474342dc1cfSSong Gao for (i = 0; i < oprsz / 8; i++) { 475342dc1cfSSong Gao muls64(&l, &h, Vj->D(i), Vk->D(i)); 476342dc1cfSSong Gao Vd->D(i) = h; 477342dc1cfSSong Gao } 478cd1c49adSSong Gao } 479cd1c49adSSong Gao 480cd1c49adSSong Gao DO_VMUH(vmuh_b, 8, H, B, DO_MUH) 481cd1c49adSSong Gao DO_VMUH(vmuh_h, 16, W, H, DO_MUH) 482cd1c49adSSong Gao DO_VMUH(vmuh_w, 32, D, W, DO_MUH) 483cd1c49adSSong Gao 484342dc1cfSSong Gao void HELPER(vmuh_du)(void *vd, void *vj, void *vk, uint32_t desc) 485cd1c49adSSong Gao { 486342dc1cfSSong Gao int i; 487342dc1cfSSong Gao uint64_t l, h; 488cd1c49adSSong Gao VReg *Vd = (VReg *)vd; 489cd1c49adSSong Gao VReg *Vj = (VReg *)vj; 490cd1c49adSSong Gao VReg *Vk = (VReg *)vk; 491342dc1cfSSong Gao int oprsz = simd_oprsz(desc); 492cd1c49adSSong Gao 493342dc1cfSSong Gao for (i = 0; i < oprsz / 8; i++) { 494342dc1cfSSong Gao mulu64(&l, &h, Vj->D(i), Vk->D(i)); 495342dc1cfSSong Gao Vd->D(i) = h; 496342dc1cfSSong Gao } 497cd1c49adSSong Gao } 498cd1c49adSSong Gao 499cd1c49adSSong Gao DO_VMUH(vmuh_bu, 8, UH, UB, DO_MUH) 500cd1c49adSSong Gao DO_VMUH(vmuh_hu, 16, UW, UH, DO_MUH) 501cd1c49adSSong Gao DO_VMUH(vmuh_wu, 32, UD, UW, DO_MUH) 502cd1c49adSSong Gao 503cd1c49adSSong Gao #define DO_MUL(a, b) (a * b) 504cd1c49adSSong Gao 505cd1c49adSSong Gao DO_EVEN(vmulwev_h_b, 16, H, B, DO_MUL) 506cd1c49adSSong Gao DO_EVEN(vmulwev_w_h, 32, W, H, DO_MUL) 507cd1c49adSSong Gao DO_EVEN(vmulwev_d_w, 64, D, W, DO_MUL) 508cd1c49adSSong Gao 509cd1c49adSSong Gao DO_ODD(vmulwod_h_b, 16, H, B, DO_MUL) 510cd1c49adSSong Gao DO_ODD(vmulwod_w_h, 32, W, H, DO_MUL) 511cd1c49adSSong Gao DO_ODD(vmulwod_d_w, 64, D, W, DO_MUL) 512cd1c49adSSong Gao 513cd1c49adSSong Gao DO_EVEN(vmulwev_h_bu, 16, UH, UB, DO_MUL) 514cd1c49adSSong Gao DO_EVEN(vmulwev_w_hu, 32, UW, UH, DO_MUL) 515cd1c49adSSong Gao DO_EVEN(vmulwev_d_wu, 64, UD, UW, DO_MUL) 516cd1c49adSSong Gao 517cd1c49adSSong Gao DO_ODD(vmulwod_h_bu, 16, UH, UB, DO_MUL) 518cd1c49adSSong Gao DO_ODD(vmulwod_w_hu, 32, UW, UH, DO_MUL) 519cd1c49adSSong Gao DO_ODD(vmulwod_d_wu, 64, UD, UW, DO_MUL) 520cd1c49adSSong Gao 521cd1c49adSSong Gao DO_EVEN_U_S(vmulwev_h_bu_b, 16, H, UH, B, UB, DO_MUL) 522cd1c49adSSong Gao DO_EVEN_U_S(vmulwev_w_hu_h, 32, W, UW, H, UH, DO_MUL) 523cd1c49adSSong Gao DO_EVEN_U_S(vmulwev_d_wu_w, 64, D, UD, W, UW, DO_MUL) 524cd1c49adSSong Gao 525cd1c49adSSong Gao DO_ODD_U_S(vmulwod_h_bu_b, 16, H, UH, B, UB, DO_MUL) 526cd1c49adSSong Gao DO_ODD_U_S(vmulwod_w_hu_h, 32, W, UW, H, UH, DO_MUL) 527cd1c49adSSong Gao DO_ODD_U_S(vmulwod_d_wu_w, 64, D, UD, W, UW, DO_MUL) 528d3aec65bSSong Gao 529d3aec65bSSong Gao #define DO_MADD(a, b, c) (a + b * c) 530d3aec65bSSong Gao #define DO_MSUB(a, b, c) (a - b * c) 531d3aec65bSSong Gao 532d3aec65bSSong Gao #define VMADDSUB(NAME, BIT, E, DO_OP) \ 5333f450c17SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 534d3aec65bSSong Gao { \ 535d3aec65bSSong Gao int i; \ 536d3aec65bSSong Gao VReg *Vd = (VReg *)vd; \ 537d3aec65bSSong Gao VReg *Vj = (VReg *)vj; \ 538d3aec65bSSong Gao VReg *Vk = (VReg *)vk; \ 5393f450c17SSong Gao int oprsz = simd_oprsz(desc); \ 5403f450c17SSong Gao \ 5413f450c17SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 542d3aec65bSSong Gao Vd->E(i) = DO_OP(Vd->E(i), Vj->E(i) ,Vk->E(i)); \ 543d3aec65bSSong Gao } \ 544d3aec65bSSong Gao } 545d3aec65bSSong Gao 546d3aec65bSSong Gao VMADDSUB(vmadd_b, 8, B, DO_MADD) 547d3aec65bSSong Gao VMADDSUB(vmadd_h, 16, H, DO_MADD) 548d3aec65bSSong Gao VMADDSUB(vmadd_w, 32, W, DO_MADD) 549d3aec65bSSong Gao VMADDSUB(vmadd_d, 64, D, DO_MADD) 550d3aec65bSSong Gao VMADDSUB(vmsub_b, 8, B, DO_MSUB) 551d3aec65bSSong Gao VMADDSUB(vmsub_h, 16, H, DO_MSUB) 552d3aec65bSSong Gao VMADDSUB(vmsub_w, 32, W, DO_MSUB) 553d3aec65bSSong Gao VMADDSUB(vmsub_d, 64, D, DO_MSUB) 554d3aec65bSSong Gao 555d3aec65bSSong Gao #define VMADDWEV(NAME, BIT, E1, E2, DO_OP) \ 5563f450c17SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 557d3aec65bSSong Gao { \ 558d3aec65bSSong Gao int i; \ 559d3aec65bSSong Gao VReg *Vd = (VReg *)vd; \ 560d3aec65bSSong Gao VReg *Vj = (VReg *)vj; \ 561d3aec65bSSong Gao VReg *Vk = (VReg *)vk; \ 562d3aec65bSSong Gao typedef __typeof(Vd->E1(0)) TD; \ 5633f450c17SSong Gao int oprsz = simd_oprsz(desc); \ 564d3aec65bSSong Gao \ 5653f450c17SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 566d3aec65bSSong Gao Vd->E1(i) += DO_OP((TD)Vj->E2(2 * i), (TD)Vk->E2(2 * i)); \ 567d3aec65bSSong Gao } \ 568d3aec65bSSong Gao } 569d3aec65bSSong Gao 570d3aec65bSSong Gao VMADDWEV(vmaddwev_h_b, 16, H, B, DO_MUL) 571d3aec65bSSong Gao VMADDWEV(vmaddwev_w_h, 32, W, H, DO_MUL) 572d3aec65bSSong Gao VMADDWEV(vmaddwev_d_w, 64, D, W, DO_MUL) 573d3aec65bSSong Gao VMADDWEV(vmaddwev_h_bu, 16, UH, UB, DO_MUL) 574d3aec65bSSong Gao VMADDWEV(vmaddwev_w_hu, 32, UW, UH, DO_MUL) 575d3aec65bSSong Gao VMADDWEV(vmaddwev_d_wu, 64, UD, UW, DO_MUL) 576d3aec65bSSong Gao 577d3aec65bSSong Gao #define VMADDWOD(NAME, BIT, E1, E2, DO_OP) \ 5783f450c17SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 579d3aec65bSSong Gao { \ 580d3aec65bSSong Gao int i; \ 581d3aec65bSSong Gao VReg *Vd = (VReg *)vd; \ 582d3aec65bSSong Gao VReg *Vj = (VReg *)vj; \ 583d3aec65bSSong Gao VReg *Vk = (VReg *)vk; \ 584d3aec65bSSong Gao typedef __typeof(Vd->E1(0)) TD; \ 5853f450c17SSong Gao int oprsz = simd_oprsz(desc); \ 586d3aec65bSSong Gao \ 5873f450c17SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 588d3aec65bSSong Gao Vd->E1(i) += DO_OP((TD)Vj->E2(2 * i + 1), \ 589d3aec65bSSong Gao (TD)Vk->E2(2 * i + 1)); \ 590d3aec65bSSong Gao } \ 591d3aec65bSSong Gao } 592d3aec65bSSong Gao 593d3aec65bSSong Gao VMADDWOD(vmaddwod_h_b, 16, H, B, DO_MUL) 594d3aec65bSSong Gao VMADDWOD(vmaddwod_w_h, 32, W, H, DO_MUL) 595d3aec65bSSong Gao VMADDWOD(vmaddwod_d_w, 64, D, W, DO_MUL) 596d3aec65bSSong Gao VMADDWOD(vmaddwod_h_bu, 16, UH, UB, DO_MUL) 597d3aec65bSSong Gao VMADDWOD(vmaddwod_w_hu, 32, UW, UH, DO_MUL) 598d3aec65bSSong Gao VMADDWOD(vmaddwod_d_wu, 64, UD, UW, DO_MUL) 599d3aec65bSSong Gao 600d3aec65bSSong Gao #define VMADDWEV_U_S(NAME, BIT, ES1, EU1, ES2, EU2, DO_OP) \ 6013f450c17SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 602d3aec65bSSong Gao { \ 603d3aec65bSSong Gao int i; \ 604d3aec65bSSong Gao VReg *Vd = (VReg *)vd; \ 605d3aec65bSSong Gao VReg *Vj = (VReg *)vj; \ 606d3aec65bSSong Gao VReg *Vk = (VReg *)vk; \ 607d3aec65bSSong Gao typedef __typeof(Vd->ES1(0)) TS1; \ 608d3aec65bSSong Gao typedef __typeof(Vd->EU1(0)) TU1; \ 6093f450c17SSong Gao int oprsz = simd_oprsz(desc); \ 610d3aec65bSSong Gao \ 6113f450c17SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 612d3aec65bSSong Gao Vd->ES1(i) += DO_OP((TU1)Vj->EU2(2 * i), \ 613d3aec65bSSong Gao (TS1)Vk->ES2(2 * i)); \ 614d3aec65bSSong Gao } \ 615d3aec65bSSong Gao } 616d3aec65bSSong Gao 617d3aec65bSSong Gao VMADDWEV_U_S(vmaddwev_h_bu_b, 16, H, UH, B, UB, DO_MUL) 618d3aec65bSSong Gao VMADDWEV_U_S(vmaddwev_w_hu_h, 32, W, UW, H, UH, DO_MUL) 619d3aec65bSSong Gao VMADDWEV_U_S(vmaddwev_d_wu_w, 64, D, UD, W, UW, DO_MUL) 620d3aec65bSSong Gao 621d3aec65bSSong Gao #define VMADDWOD_U_S(NAME, BIT, ES1, EU1, ES2, EU2, DO_OP) \ 6223f450c17SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 623d3aec65bSSong Gao { \ 624d3aec65bSSong Gao int i; \ 625d3aec65bSSong Gao VReg *Vd = (VReg *)vd; \ 626d3aec65bSSong Gao VReg *Vj = (VReg *)vj; \ 627d3aec65bSSong Gao VReg *Vk = (VReg *)vk; \ 628d3aec65bSSong Gao typedef __typeof(Vd->ES1(0)) TS1; \ 629d3aec65bSSong Gao typedef __typeof(Vd->EU1(0)) TU1; \ 6303f450c17SSong Gao int oprsz = simd_oprsz(desc); \ 631d3aec65bSSong Gao \ 6323f450c17SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 633d3aec65bSSong Gao Vd->ES1(i) += DO_OP((TU1)Vj->EU2(2 * i + 1), \ 634d3aec65bSSong Gao (TS1)Vk->ES2(2 * i + 1)); \ 635d3aec65bSSong Gao } \ 636d3aec65bSSong Gao } 637d3aec65bSSong Gao 638d3aec65bSSong Gao VMADDWOD_U_S(vmaddwod_h_bu_b, 16, H, UH, B, UB, DO_MUL) 639d3aec65bSSong Gao VMADDWOD_U_S(vmaddwod_w_hu_h, 32, W, UW, H, UH, DO_MUL) 640d3aec65bSSong Gao VMADDWOD_U_S(vmaddwod_d_wu_w, 64, D, UD, W, UW, DO_MUL) 6414cc4c0f7SSong Gao 6424cc4c0f7SSong Gao #define DO_DIVU(N, M) (unlikely(M == 0) ? 0 : N / M) 6434cc4c0f7SSong Gao #define DO_REMU(N, M) (unlikely(M == 0) ? 0 : N % M) 6444cc4c0f7SSong Gao #define DO_DIV(N, M) (unlikely(M == 0) ? 0 :\ 6454cc4c0f7SSong Gao unlikely((N == -N) && (M == (__typeof(N))(-1))) ? N : N / M) 6464cc4c0f7SSong Gao #define DO_REM(N, M) (unlikely(M == 0) ? 0 :\ 6474cc4c0f7SSong Gao unlikely((N == -N) && (M == (__typeof(N))(-1))) ? 0 : N % M) 6484cc4c0f7SSong Gao 6494cc4c0f7SSong Gao #define VDIV(NAME, BIT, E, DO_OP) \ 65004711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 6514cc4c0f7SSong Gao { \ 6524cc4c0f7SSong Gao int i; \ 65304711da1SSong Gao VReg *Vd = (VReg *)vd; \ 65404711da1SSong Gao VReg *Vj = (VReg *)vj; \ 65504711da1SSong Gao VReg *Vk = (VReg *)vk; \ 656abb693deSSong Gao int oprsz = simd_oprsz(desc); \ 657abb693deSSong Gao \ 658abb693deSSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 6594cc4c0f7SSong Gao Vd->E(i) = DO_OP(Vj->E(i), Vk->E(i)); \ 6604cc4c0f7SSong Gao } \ 6614cc4c0f7SSong Gao } 6624cc4c0f7SSong Gao 6634cc4c0f7SSong Gao VDIV(vdiv_b, 8, B, DO_DIV) 6644cc4c0f7SSong Gao VDIV(vdiv_h, 16, H, DO_DIV) 6654cc4c0f7SSong Gao VDIV(vdiv_w, 32, W, DO_DIV) 6664cc4c0f7SSong Gao VDIV(vdiv_d, 64, D, DO_DIV) 6674cc4c0f7SSong Gao VDIV(vdiv_bu, 8, UB, DO_DIVU) 6684cc4c0f7SSong Gao VDIV(vdiv_hu, 16, UH, DO_DIVU) 6694cc4c0f7SSong Gao VDIV(vdiv_wu, 32, UW, DO_DIVU) 6704cc4c0f7SSong Gao VDIV(vdiv_du, 64, UD, DO_DIVU) 6714cc4c0f7SSong Gao VDIV(vmod_b, 8, B, DO_REM) 6724cc4c0f7SSong Gao VDIV(vmod_h, 16, H, DO_REM) 6734cc4c0f7SSong Gao VDIV(vmod_w, 32, W, DO_REM) 6744cc4c0f7SSong Gao VDIV(vmod_d, 64, D, DO_REM) 6754cc4c0f7SSong Gao VDIV(vmod_bu, 8, UB, DO_REMU) 6764cc4c0f7SSong Gao VDIV(vmod_hu, 16, UH, DO_REMU) 6774cc4c0f7SSong Gao VDIV(vmod_wu, 32, UW, DO_REMU) 6784cc4c0f7SSong Gao VDIV(vmod_du, 64, UD, DO_REMU) 679cbe44190SSong Gao 680cbe44190SSong Gao #define VSAT_S(NAME, BIT, E) \ 681e5c7f031SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t max, uint32_t desc) \ 682cbe44190SSong Gao { \ 683cbe44190SSong Gao int i; \ 684cbe44190SSong Gao VReg *Vd = (VReg *)vd; \ 685cbe44190SSong Gao VReg *Vj = (VReg *)vj; \ 686cbe44190SSong Gao typedef __typeof(Vd->E(0)) TD; \ 687e5c7f031SSong Gao int oprsz = simd_oprsz(desc); \ 688cbe44190SSong Gao \ 689e5c7f031SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 690cbe44190SSong Gao Vd->E(i) = Vj->E(i) > (TD)max ? (TD)max : \ 691cbe44190SSong Gao Vj->E(i) < (TD)~max ? (TD)~max: Vj->E(i); \ 692cbe44190SSong Gao } \ 693cbe44190SSong Gao } 694cbe44190SSong Gao 695cbe44190SSong Gao VSAT_S(vsat_b, 8, B) 696cbe44190SSong Gao VSAT_S(vsat_h, 16, H) 697cbe44190SSong Gao VSAT_S(vsat_w, 32, W) 698cbe44190SSong Gao VSAT_S(vsat_d, 64, D) 699cbe44190SSong Gao 700cbe44190SSong Gao #define VSAT_U(NAME, BIT, E) \ 701e5c7f031SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t max, uint32_t desc) \ 702cbe44190SSong Gao { \ 703cbe44190SSong Gao int i; \ 704cbe44190SSong Gao VReg *Vd = (VReg *)vd; \ 705cbe44190SSong Gao VReg *Vj = (VReg *)vj; \ 706cbe44190SSong Gao typedef __typeof(Vd->E(0)) TD; \ 707e5c7f031SSong Gao int oprsz = simd_oprsz(desc); \ 708cbe44190SSong Gao \ 709e5c7f031SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 710cbe44190SSong Gao Vd->E(i) = Vj->E(i) > (TD)max ? (TD)max : Vj->E(i); \ 711cbe44190SSong Gao } \ 712cbe44190SSong Gao } 713cbe44190SSong Gao 714cbe44190SSong Gao VSAT_U(vsat_bu, 8, UB) 715cbe44190SSong Gao VSAT_U(vsat_hu, 16, UH) 716cbe44190SSong Gao VSAT_U(vsat_wu, 32, UW) 717cbe44190SSong Gao VSAT_U(vsat_du, 64, UD) 7183734ad93SSong Gao 7193734ad93SSong Gao #define VEXTH(NAME, BIT, E1, E2) \ 720ff27e335SSong Gao void HELPER(NAME)(void *vd, void *vj, uint32_t desc) \ 7213734ad93SSong Gao { \ 722f0db0bebSSong Gao int i, j, ofs; \ 723ff27e335SSong Gao VReg *Vd = (VReg *)vd; \ 724ff27e335SSong Gao VReg *Vj = (VReg *)vj; \ 725f0db0bebSSong Gao int oprsz = simd_oprsz(desc); \ 7263734ad93SSong Gao \ 727f0db0bebSSong Gao ofs = LSX_LEN / BIT; \ 728f0db0bebSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 729f0db0bebSSong Gao for (j = 0; j < ofs; j++) { \ 730f0db0bebSSong Gao Vd->E1(j + i * ofs) = Vj->E2(j + ofs + ofs * 2 * i); \ 731f0db0bebSSong Gao } \ 7323734ad93SSong Gao } \ 7333734ad93SSong Gao } 7343734ad93SSong Gao 735ff27e335SSong Gao void HELPER(vexth_q_d)(void *vd, void *vj, uint32_t desc) 7363734ad93SSong Gao { 737f0db0bebSSong Gao int i; 738ff27e335SSong Gao VReg *Vd = (VReg *)vd; 739ff27e335SSong Gao VReg *Vj = (VReg *)vj; 740f0db0bebSSong Gao int oprsz = simd_oprsz(desc); 7413734ad93SSong Gao 742f0db0bebSSong Gao for (i = 0; i < oprsz / 16; i++) { 743f0db0bebSSong Gao Vd->Q(i) = int128_makes64(Vj->D(2 * i + 1)); 744f0db0bebSSong Gao } 7453734ad93SSong Gao } 7463734ad93SSong Gao 747ff27e335SSong Gao void HELPER(vexth_qu_du)(void *vd, void *vj, uint32_t desc) 7483734ad93SSong Gao { 749f0db0bebSSong Gao int i; 750ff27e335SSong Gao VReg *Vd = (VReg *)vd; 751ff27e335SSong Gao VReg *Vj = (VReg *)vj; 752f0db0bebSSong Gao int oprsz = simd_oprsz(desc); 7533734ad93SSong Gao 754f0db0bebSSong Gao for (i = 0; i < oprsz / 16; i++) { 755f0db0bebSSong Gao Vd->Q(i) = int128_make64(Vj->UD(2 * i + 1)); 756f0db0bebSSong Gao } 7573734ad93SSong Gao } 7583734ad93SSong Gao 7593734ad93SSong Gao VEXTH(vexth_h_b, 16, H, B) 7603734ad93SSong Gao VEXTH(vexth_w_h, 32, W, H) 7613734ad93SSong Gao VEXTH(vexth_d_w, 64, D, W) 7623734ad93SSong Gao VEXTH(vexth_hu_bu, 16, UH, UB) 7633734ad93SSong Gao VEXTH(vexth_wu_hu, 32, UW, UH) 7643734ad93SSong Gao VEXTH(vexth_du_wu, 64, UD, UW) 765f0e395dfSSong Gao 766790acb2aSSong Gao #define VEXT2XV(NAME, BIT, E1, E2) \ 767790acb2aSSong Gao void HELPER(NAME)(void *vd, void *vj, uint32_t desc) \ 768790acb2aSSong Gao { \ 769790acb2aSSong Gao int i; \ 770790acb2aSSong Gao VReg temp = {}; \ 771790acb2aSSong Gao VReg *Vd = (VReg *)vd; \ 772790acb2aSSong Gao VReg *Vj = (VReg *)vj; \ 773790acb2aSSong Gao int oprsz = simd_oprsz(desc); \ 774790acb2aSSong Gao \ 775790acb2aSSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 776790acb2aSSong Gao temp.E1(i) = Vj->E2(i); \ 777790acb2aSSong Gao } \ 778790acb2aSSong Gao *Vd = temp; \ 779790acb2aSSong Gao } 780790acb2aSSong Gao 781790acb2aSSong Gao VEXT2XV(vext2xv_h_b, 16, H, B) 782790acb2aSSong Gao VEXT2XV(vext2xv_w_b, 32, W, B) 783790acb2aSSong Gao VEXT2XV(vext2xv_d_b, 64, D, B) 784790acb2aSSong Gao VEXT2XV(vext2xv_w_h, 32, W, H) 785790acb2aSSong Gao VEXT2XV(vext2xv_d_h, 64, D, H) 786790acb2aSSong Gao VEXT2XV(vext2xv_d_w, 64, D, W) 787790acb2aSSong Gao VEXT2XV(vext2xv_hu_bu, 16, UH, UB) 788790acb2aSSong Gao VEXT2XV(vext2xv_wu_bu, 32, UW, UB) 789790acb2aSSong Gao VEXT2XV(vext2xv_du_bu, 64, UD, UB) 790790acb2aSSong Gao VEXT2XV(vext2xv_wu_hu, 32, UW, UH) 791790acb2aSSong Gao VEXT2XV(vext2xv_du_hu, 64, UD, UH) 792790acb2aSSong Gao VEXT2XV(vext2xv_du_wu, 64, UD, UW) 793790acb2aSSong Gao 794f0e395dfSSong Gao #define DO_SIGNCOV(a, b) (a == 0 ? 0 : a < 0 ? -b : b) 795f0e395dfSSong Gao 796f0e395dfSSong Gao DO_3OP(vsigncov_b, 8, B, DO_SIGNCOV) 797f0e395dfSSong Gao DO_3OP(vsigncov_h, 16, H, DO_SIGNCOV) 798f0e395dfSSong Gao DO_3OP(vsigncov_w, 32, W, DO_SIGNCOV) 799f0e395dfSSong Gao DO_3OP(vsigncov_d, 64, D, DO_SIGNCOV) 800789f4a4cSSong Gao 801789f4a4cSSong Gao static uint64_t do_vmskltz_b(int64_t val) 802789f4a4cSSong Gao { 803789f4a4cSSong Gao uint64_t m = 0x8080808080808080ULL; 804789f4a4cSSong Gao uint64_t c = val & m; 805789f4a4cSSong Gao c |= c << 7; 806789f4a4cSSong Gao c |= c << 14; 807789f4a4cSSong Gao c |= c << 28; 808789f4a4cSSong Gao return c >> 56; 809789f4a4cSSong Gao } 810789f4a4cSSong Gao 811ff27e335SSong Gao void HELPER(vmskltz_b)(void *vd, void *vj, uint32_t desc) 812789f4a4cSSong Gao { 81397074674SSong Gao int i; 814789f4a4cSSong Gao uint16_t temp = 0; 815ff27e335SSong Gao VReg *Vd = (VReg *)vd; 816ff27e335SSong Gao VReg *Vj = (VReg *)vj; 81797074674SSong Gao int oprsz = simd_oprsz(desc); 818789f4a4cSSong Gao 81997074674SSong Gao for (i = 0; i < oprsz / 16; i++) { 82097074674SSong Gao temp = 0; 82197074674SSong Gao temp = do_vmskltz_b(Vj->D(2 * i)); 82297074674SSong Gao temp |= (do_vmskltz_b(Vj->D(2 * i + 1)) << 8); 82397074674SSong Gao Vd->D(2 * i) = temp; 82497074674SSong Gao Vd->D(2 * i + 1) = 0; 82597074674SSong Gao } 826789f4a4cSSong Gao } 827789f4a4cSSong Gao 828789f4a4cSSong Gao static uint64_t do_vmskltz_h(int64_t val) 829789f4a4cSSong Gao { 830789f4a4cSSong Gao uint64_t m = 0x8000800080008000ULL; 831789f4a4cSSong Gao uint64_t c = val & m; 832789f4a4cSSong Gao c |= c << 15; 833789f4a4cSSong Gao c |= c << 30; 834789f4a4cSSong Gao return c >> 60; 835789f4a4cSSong Gao } 836789f4a4cSSong Gao 837ff27e335SSong Gao void HELPER(vmskltz_h)(void *vd, void *vj, uint32_t desc) 838789f4a4cSSong Gao { 83997074674SSong Gao int i; 840789f4a4cSSong Gao uint16_t temp = 0; 841ff27e335SSong Gao VReg *Vd = (VReg *)vd; 842ff27e335SSong Gao VReg *Vj = (VReg *)vj; 84397074674SSong Gao int oprsz = simd_oprsz(desc); 844789f4a4cSSong Gao 84597074674SSong Gao for (i = 0; i < oprsz / 16; i++) { 84697074674SSong Gao temp = 0; 84797074674SSong Gao temp = do_vmskltz_h(Vj->D(2 * i)); 84897074674SSong Gao temp |= (do_vmskltz_h(Vj->D(2 * i + 1)) << 4); 84997074674SSong Gao Vd->D(2 * i) = temp; 85097074674SSong Gao Vd->D(2 * i + 1) = 0; 85197074674SSong Gao } 852789f4a4cSSong Gao } 853789f4a4cSSong Gao 854789f4a4cSSong Gao static uint64_t do_vmskltz_w(int64_t val) 855789f4a4cSSong Gao { 856789f4a4cSSong Gao uint64_t m = 0x8000000080000000ULL; 857789f4a4cSSong Gao uint64_t c = val & m; 858789f4a4cSSong Gao c |= c << 31; 859789f4a4cSSong Gao return c >> 62; 860789f4a4cSSong Gao } 861789f4a4cSSong Gao 862ff27e335SSong Gao void HELPER(vmskltz_w)(void *vd, void *vj, uint32_t desc) 863789f4a4cSSong Gao { 86497074674SSong Gao int i; 865789f4a4cSSong Gao uint16_t temp = 0; 866ff27e335SSong Gao VReg *Vd = (VReg *)vd; 867ff27e335SSong Gao VReg *Vj = (VReg *)vj; 86897074674SSong Gao int oprsz = simd_oprsz(desc); 869789f4a4cSSong Gao 87097074674SSong Gao for (i = 0; i < oprsz / 16; i++) { 87197074674SSong Gao temp = 0; 87297074674SSong Gao temp = do_vmskltz_w(Vj->D(2 * i)); 87397074674SSong Gao temp |= (do_vmskltz_w(Vj->D(2 * i + 1)) << 2); 87497074674SSong Gao Vd->D(2 * i) = temp; 87597074674SSong Gao Vd->D(2 * i + 1) = 0; 87697074674SSong Gao } 877789f4a4cSSong Gao } 878789f4a4cSSong Gao 879789f4a4cSSong Gao static uint64_t do_vmskltz_d(int64_t val) 880789f4a4cSSong Gao { 881789f4a4cSSong Gao return (uint64_t)val >> 63; 882789f4a4cSSong Gao } 883ff27e335SSong Gao void HELPER(vmskltz_d)(void *vd, void *vj, uint32_t desc) 884789f4a4cSSong Gao { 88597074674SSong Gao int i; 886789f4a4cSSong Gao uint16_t temp = 0; 887ff27e335SSong Gao VReg *Vd = (VReg *)vd; 888ff27e335SSong Gao VReg *Vj = (VReg *)vj; 88997074674SSong Gao int oprsz = simd_oprsz(desc); 890789f4a4cSSong Gao 89197074674SSong Gao for (i = 0; i < oprsz / 16; i++) { 89297074674SSong Gao temp = 0; 89397074674SSong Gao temp = do_vmskltz_d(Vj->D(2 * i)); 89497074674SSong Gao temp |= (do_vmskltz_d(Vj->D(2 * i + 1)) << 1); 89597074674SSong Gao Vd->D(2 * i) = temp; 89697074674SSong Gao Vd->D(2 * i + 1) = 0; 89797074674SSong Gao } 898789f4a4cSSong Gao } 899789f4a4cSSong Gao 900ff27e335SSong Gao void HELPER(vmskgez_b)(void *vd, void *vj, uint32_t desc) 901789f4a4cSSong Gao { 90297074674SSong Gao int i; 903789f4a4cSSong Gao uint16_t temp = 0; 904ff27e335SSong Gao VReg *Vd = (VReg *)vd; 905ff27e335SSong Gao VReg *Vj = (VReg *)vj; 90697074674SSong Gao int oprsz = simd_oprsz(desc); 907789f4a4cSSong Gao 90897074674SSong Gao for (i = 0; i < oprsz / 16; i++) { 90997074674SSong Gao temp = 0; 91097074674SSong Gao temp = do_vmskltz_b(Vj->D(2 * i)); 91197074674SSong Gao temp |= (do_vmskltz_b(Vj->D(2 * i + 1)) << 8); 91297074674SSong Gao Vd->D(2 * i) = (uint16_t)(~temp); 91397074674SSong Gao Vd->D(2 * i + 1) = 0; 91497074674SSong Gao } 915789f4a4cSSong Gao } 916789f4a4cSSong Gao 917789f4a4cSSong Gao static uint64_t do_vmskez_b(uint64_t a) 918789f4a4cSSong Gao { 919789f4a4cSSong Gao uint64_t m = 0x7f7f7f7f7f7f7f7fULL; 920789f4a4cSSong Gao uint64_t c = ~(((a & m) + m) | a | m); 921789f4a4cSSong Gao c |= c << 7; 922789f4a4cSSong Gao c |= c << 14; 923789f4a4cSSong Gao c |= c << 28; 924789f4a4cSSong Gao return c >> 56; 925789f4a4cSSong Gao } 926789f4a4cSSong Gao 927ff27e335SSong Gao void HELPER(vmsknz_b)(void *vd, void *vj, uint32_t desc) 928789f4a4cSSong Gao { 92997074674SSong Gao int i; 930789f4a4cSSong Gao uint16_t temp = 0; 931ff27e335SSong Gao VReg *Vd = (VReg *)vd; 932ff27e335SSong Gao VReg *Vj = (VReg *)vj; 93397074674SSong Gao int oprsz = simd_oprsz(desc); 934789f4a4cSSong Gao 93597074674SSong Gao for (i = 0; i < oprsz / 16; i++) { 93697074674SSong Gao temp = 0; 93797074674SSong Gao temp = do_vmskez_b(Vj->D(2 * i)); 93897074674SSong Gao temp |= (do_vmskez_b(Vj->D(2 * i + 1)) << 8); 93997074674SSong Gao Vd->D(2 * i) = (uint16_t)(~temp); 94097074674SSong Gao Vd->D(2 * i + 1) = 0; 94197074674SSong Gao } 942789f4a4cSSong Gao } 943f205a539SSong Gao 9444472a45aSSong Gao void HELPER(vnori_b)(void *vd, void *vj, uint64_t imm, uint32_t desc) 945f205a539SSong Gao { 946f205a539SSong Gao int i; 947f205a539SSong Gao VReg *Vd = (VReg *)vd; 948f205a539SSong Gao VReg *Vj = (VReg *)vj; 949f205a539SSong Gao 9504472a45aSSong Gao for (i = 0; i < simd_oprsz(desc); i++) { 951f205a539SSong Gao Vd->B(i) = ~(Vj->B(i) | (uint8_t)imm); 952f205a539SSong Gao } 953f205a539SSong Gao } 9549b21a7a5SSong Gao 9559b21a7a5SSong Gao #define VSLLWIL(NAME, BIT, E1, E2) \ 956329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 9579b21a7a5SSong Gao { \ 9586567eac7SSong Gao int i, j, ofs; \ 9596567eac7SSong Gao VReg temp = {}; \ 960329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 961329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 9626567eac7SSong Gao int oprsz = simd_oprsz(desc); \ 9639b21a7a5SSong Gao typedef __typeof(temp.E1(0)) TD; \ 9649b21a7a5SSong Gao \ 9656567eac7SSong Gao ofs = LSX_LEN / BIT; \ 9666567eac7SSong Gao for (i = 0; i < oprsz / 16; i++) { \ 9676567eac7SSong Gao for (j = 0; j < ofs; j++) { \ 9686567eac7SSong Gao temp.E1(j + ofs * i) = (TD)Vj->E2(j + ofs * 2 * i) << (imm % BIT); \ 9696567eac7SSong Gao } \ 9709b21a7a5SSong Gao } \ 9719b21a7a5SSong Gao *Vd = temp; \ 9729b21a7a5SSong Gao } 9739b21a7a5SSong Gao 9746567eac7SSong Gao 975ff27e335SSong Gao void HELPER(vextl_q_d)(void *vd, void *vj, uint32_t desc) 9769b21a7a5SSong Gao { 9776567eac7SSong Gao int i; 978ff27e335SSong Gao VReg *Vd = (VReg *)vd; 979ff27e335SSong Gao VReg *Vj = (VReg *)vj; 9806567eac7SSong Gao int oprsz = simd_oprsz(desc); 9819b21a7a5SSong Gao 9826567eac7SSong Gao for (i = 0; i < oprsz / 16; i++) { 9836567eac7SSong Gao Vd->Q(i) = int128_makes64(Vj->D(2 * i)); 9846567eac7SSong Gao } 9859b21a7a5SSong Gao } 9869b21a7a5SSong Gao 987ff27e335SSong Gao void HELPER(vextl_qu_du)(void *vd, void *vj, uint32_t desc) 9889b21a7a5SSong Gao { 9896567eac7SSong Gao int i; 990ff27e335SSong Gao VReg *Vd = (VReg *)vd; 991ff27e335SSong Gao VReg *Vj = (VReg *)vj; 9926567eac7SSong Gao int oprsz = simd_oprsz(desc); 9939b21a7a5SSong Gao 9946567eac7SSong Gao for (i = 0; i < oprsz / 16; i++) { 9956567eac7SSong Gao Vd->Q(i) = int128_make64(Vj->UD(2 * i)); 9966567eac7SSong Gao } 9979b21a7a5SSong Gao } 9989b21a7a5SSong Gao 9999b21a7a5SSong Gao VSLLWIL(vsllwil_h_b, 16, H, B) 10009b21a7a5SSong Gao VSLLWIL(vsllwil_w_h, 32, W, H) 10019b21a7a5SSong Gao VSLLWIL(vsllwil_d_w, 64, D, W) 10029b21a7a5SSong Gao VSLLWIL(vsllwil_hu_bu, 16, UH, UB) 10039b21a7a5SSong Gao VSLLWIL(vsllwil_wu_hu, 32, UW, UH) 10049b21a7a5SSong Gao VSLLWIL(vsllwil_du_wu, 64, UD, UW) 1005ecb93716SSong Gao 1006ecb93716SSong Gao #define do_vsrlr(E, T) \ 1007ecb93716SSong Gao static T do_vsrlr_ ##E(T s1, int sh) \ 1008ecb93716SSong Gao { \ 1009ecb93716SSong Gao if (sh == 0) { \ 1010ecb93716SSong Gao return s1; \ 1011ecb93716SSong Gao } else { \ 1012ecb93716SSong Gao return (s1 >> sh) + ((s1 >> (sh - 1)) & 0x1); \ 1013ecb93716SSong Gao } \ 1014ecb93716SSong Gao } 1015ecb93716SSong Gao 1016ecb93716SSong Gao do_vsrlr(B, uint8_t) 1017ecb93716SSong Gao do_vsrlr(H, uint16_t) 1018ecb93716SSong Gao do_vsrlr(W, uint32_t) 1019ecb93716SSong Gao do_vsrlr(D, uint64_t) 1020ecb93716SSong Gao 1021ecb93716SSong Gao #define VSRLR(NAME, BIT, T, E) \ 102204711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1023ecb93716SSong Gao { \ 1024ecb93716SSong Gao int i; \ 102504711da1SSong Gao VReg *Vd = (VReg *)vd; \ 102604711da1SSong Gao VReg *Vj = (VReg *)vj; \ 102704711da1SSong Gao VReg *Vk = (VReg *)vk; \ 10288c272fe8SSong Gao int oprsz = simd_oprsz(desc); \ 1029ecb93716SSong Gao \ 10308c272fe8SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 1031ecb93716SSong Gao Vd->E(i) = do_vsrlr_ ## E(Vj->E(i), ((T)Vk->E(i))%BIT); \ 1032ecb93716SSong Gao } \ 1033ecb93716SSong Gao } 1034ecb93716SSong Gao 1035ecb93716SSong Gao VSRLR(vsrlr_b, 8, uint8_t, B) 1036ecb93716SSong Gao VSRLR(vsrlr_h, 16, uint16_t, H) 1037ecb93716SSong Gao VSRLR(vsrlr_w, 32, uint32_t, W) 1038ecb93716SSong Gao VSRLR(vsrlr_d, 64, uint64_t, D) 1039ecb93716SSong Gao 1040ecb93716SSong Gao #define VSRLRI(NAME, BIT, E) \ 1041329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 1042ecb93716SSong Gao { \ 1043ecb93716SSong Gao int i; \ 1044329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1045329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 10468c272fe8SSong Gao int oprsz = simd_oprsz(desc); \ 1047ecb93716SSong Gao \ 10488c272fe8SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 1049ecb93716SSong Gao Vd->E(i) = do_vsrlr_ ## E(Vj->E(i), imm); \ 1050ecb93716SSong Gao } \ 1051ecb93716SSong Gao } 1052ecb93716SSong Gao 1053ecb93716SSong Gao VSRLRI(vsrlri_b, 8, B) 1054ecb93716SSong Gao VSRLRI(vsrlri_h, 16, H) 1055ecb93716SSong Gao VSRLRI(vsrlri_w, 32, W) 1056ecb93716SSong Gao VSRLRI(vsrlri_d, 64, D) 1057ecb93716SSong Gao 1058ecb93716SSong Gao #define do_vsrar(E, T) \ 1059ecb93716SSong Gao static T do_vsrar_ ##E(T s1, int sh) \ 1060ecb93716SSong Gao { \ 1061ecb93716SSong Gao if (sh == 0) { \ 1062ecb93716SSong Gao return s1; \ 1063ecb93716SSong Gao } else { \ 1064ecb93716SSong Gao return (s1 >> sh) + ((s1 >> (sh - 1)) & 0x1); \ 1065ecb93716SSong Gao } \ 1066ecb93716SSong Gao } 1067ecb93716SSong Gao 1068ecb93716SSong Gao do_vsrar(B, int8_t) 1069ecb93716SSong Gao do_vsrar(H, int16_t) 1070ecb93716SSong Gao do_vsrar(W, int32_t) 1071ecb93716SSong Gao do_vsrar(D, int64_t) 1072ecb93716SSong Gao 1073ecb93716SSong Gao #define VSRAR(NAME, BIT, T, E) \ 107404711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1075ecb93716SSong Gao { \ 1076ecb93716SSong Gao int i; \ 107704711da1SSong Gao VReg *Vd = (VReg *)vd; \ 107804711da1SSong Gao VReg *Vj = (VReg *)vj; \ 107904711da1SSong Gao VReg *Vk = (VReg *)vk; \ 10808c272fe8SSong Gao int oprsz = simd_oprsz(desc); \ 1081ecb93716SSong Gao \ 10828c272fe8SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 1083ecb93716SSong Gao Vd->E(i) = do_vsrar_ ## E(Vj->E(i), ((T)Vk->E(i))%BIT); \ 1084ecb93716SSong Gao } \ 1085ecb93716SSong Gao } 1086ecb93716SSong Gao 1087ecb93716SSong Gao VSRAR(vsrar_b, 8, uint8_t, B) 1088ecb93716SSong Gao VSRAR(vsrar_h, 16, uint16_t, H) 1089ecb93716SSong Gao VSRAR(vsrar_w, 32, uint32_t, W) 1090ecb93716SSong Gao VSRAR(vsrar_d, 64, uint64_t, D) 1091ecb93716SSong Gao 1092ecb93716SSong Gao #define VSRARI(NAME, BIT, E) \ 1093329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 1094ecb93716SSong Gao { \ 1095ecb93716SSong Gao int i; \ 1096329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1097329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 10988c272fe8SSong Gao int oprsz = simd_oprsz(desc); \ 1099ecb93716SSong Gao \ 11008c272fe8SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) { \ 1101ecb93716SSong Gao Vd->E(i) = do_vsrar_ ## E(Vj->E(i), imm); \ 1102ecb93716SSong Gao } \ 1103ecb93716SSong Gao } 1104ecb93716SSong Gao 1105ecb93716SSong Gao VSRARI(vsrari_b, 8, B) 1106ecb93716SSong Gao VSRARI(vsrari_h, 16, H) 1107ecb93716SSong Gao VSRARI(vsrari_w, 32, W) 1108ecb93716SSong Gao VSRARI(vsrari_d, 64, D) 1109d79fb8ddSSong Gao 1110d79fb8ddSSong Gao #define R_SHIFT(a, b) (a >> b) 1111d79fb8ddSSong Gao 111240c7674eSSong Gao #define VSRLN(NAME, BIT, E1, E2) \ 111304711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1114d79fb8ddSSong Gao { \ 111540c7674eSSong Gao int i, j, ofs; \ 111604711da1SSong Gao VReg *Vd = (VReg *)vd; \ 111704711da1SSong Gao VReg *Vj = (VReg *)vj; \ 111804711da1SSong Gao VReg *Vk = (VReg *)vk; \ 111940c7674eSSong Gao int oprsz = simd_oprsz(desc); \ 1120d79fb8ddSSong Gao \ 112140c7674eSSong Gao ofs = LSX_LEN / BIT; \ 112240c7674eSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 112340c7674eSSong Gao for (j = 0; j < ofs; j++) { \ 112440c7674eSSong Gao Vd->E1(j + ofs * 2 * i) = R_SHIFT(Vj->E2(j + ofs * i), \ 112540c7674eSSong Gao Vk->E2(j + ofs * i) % BIT); \ 1126d79fb8ddSSong Gao } \ 112740c7674eSSong Gao Vd->D(2 * i + 1) = 0; \ 112840c7674eSSong Gao } \ 1129d79fb8ddSSong Gao } 1130d79fb8ddSSong Gao 113140c7674eSSong Gao VSRLN(vsrln_b_h, 16, B, UH) 113240c7674eSSong Gao VSRLN(vsrln_h_w, 32, H, UW) 113340c7674eSSong Gao VSRLN(vsrln_w_d, 64, W, UD) 1134d79fb8ddSSong Gao 113540c7674eSSong Gao #define VSRAN(NAME, BIT, E1, E2, E3) \ 113604711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1137d79fb8ddSSong Gao { \ 113840c7674eSSong Gao int i, j, ofs; \ 113904711da1SSong Gao VReg *Vd = (VReg *)vd; \ 114004711da1SSong Gao VReg *Vj = (VReg *)vj; \ 114104711da1SSong Gao VReg *Vk = (VReg *)vk; \ 114240c7674eSSong Gao int oprsz = simd_oprsz(desc); \ 1143d79fb8ddSSong Gao \ 114440c7674eSSong Gao ofs = LSX_LEN / BIT; \ 114540c7674eSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 114640c7674eSSong Gao for (j = 0; j < ofs; j++) { \ 114740c7674eSSong Gao Vd->E1(j + ofs * 2 * i) = R_SHIFT(Vj->E2(j + ofs * i), \ 114840c7674eSSong Gao Vk->E3(j + ofs * i) % BIT); \ 1149d79fb8ddSSong Gao } \ 115040c7674eSSong Gao Vd->D(2 * i + 1) = 0; \ 115140c7674eSSong Gao } \ 1152d79fb8ddSSong Gao } 1153d79fb8ddSSong Gao 115440c7674eSSong Gao VSRAN(vsran_b_h, 16, B, H, UH) 115540c7674eSSong Gao VSRAN(vsran_h_w, 32, H, W, UW) 115640c7674eSSong Gao VSRAN(vsran_w_d, 64, W, D, UD) 1157d79fb8ddSSong Gao 115840c7674eSSong Gao #define VSRLNI(NAME, BIT, E1, E2) \ 1159329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 1160d79fb8ddSSong Gao { \ 116140c7674eSSong Gao int i, j, ofs; \ 116240c7674eSSong Gao VReg temp = {}; \ 1163329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1164329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 116540c7674eSSong Gao int oprsz = simd_oprsz(desc); \ 1166d79fb8ddSSong Gao \ 116740c7674eSSong Gao ofs = LSX_LEN / BIT; \ 116840c7674eSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 116940c7674eSSong Gao for (j = 0; j < ofs; j++) { \ 117040c7674eSSong Gao temp.E1(j + ofs * 2 * i) = R_SHIFT(Vj->E2(j + ofs * i), imm); \ 117140c7674eSSong Gao temp.E1(j + ofs * (2 * i + 1)) = R_SHIFT(Vd->E2(j + ofs * i), \ 117240c7674eSSong Gao imm); \ 117340c7674eSSong Gao } \ 1174d79fb8ddSSong Gao } \ 1175d79fb8ddSSong Gao *Vd = temp; \ 1176d79fb8ddSSong Gao } 1177d79fb8ddSSong Gao 1178329517d5SSong Gao void HELPER(vsrlni_d_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 1179d79fb8ddSSong Gao { 118040c7674eSSong Gao int i; 118140c7674eSSong Gao VReg temp = {}; 1182329517d5SSong Gao VReg *Vd = (VReg *)vd; 1183329517d5SSong Gao VReg *Vj = (VReg *)vj; 1184d79fb8ddSSong Gao 118540c7674eSSong Gao for (i = 0; i < 2; i++) { 118640c7674eSSong Gao temp.D(2 * i) = int128_getlo(int128_urshift(Vj->Q(i), imm % 128)); 118740c7674eSSong Gao temp.D(2 * i +1) = int128_getlo(int128_urshift(Vd->Q(i), imm % 128)); 118840c7674eSSong Gao } 1189d79fb8ddSSong Gao *Vd = temp; 1190d79fb8ddSSong Gao } 1191d79fb8ddSSong Gao 119240c7674eSSong Gao VSRLNI(vsrlni_b_h, 16, B, UH) 119340c7674eSSong Gao VSRLNI(vsrlni_h_w, 32, H, UW) 119440c7674eSSong Gao VSRLNI(vsrlni_w_d, 64, W, UD) 1195d79fb8ddSSong Gao 1196d79fb8ddSSong Gao #define VSRANI(NAME, BIT, E1, E2) \ 1197329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 1198d79fb8ddSSong Gao { \ 119940c7674eSSong Gao int i, j, ofs; \ 120040c7674eSSong Gao VReg temp = {}; \ 1201329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1202329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 120340c7674eSSong Gao int oprsz = simd_oprsz(desc); \ 1204d79fb8ddSSong Gao \ 120540c7674eSSong Gao ofs = LSX_LEN / BIT; \ 120640c7674eSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 120740c7674eSSong Gao for (j = 0; j < ofs; j++) { \ 120840c7674eSSong Gao temp.E1(j + ofs * 2 * i) = R_SHIFT(Vj->E2(j + ofs * i), imm); \ 120940c7674eSSong Gao temp.E1(j + ofs * (2 * i + 1)) = R_SHIFT(Vd->E2(j + ofs * i), \ 121040c7674eSSong Gao imm); \ 121140c7674eSSong Gao } \ 1212d79fb8ddSSong Gao } \ 1213d79fb8ddSSong Gao *Vd = temp; \ 1214d79fb8ddSSong Gao } 1215d79fb8ddSSong Gao 1216329517d5SSong Gao void HELPER(vsrani_d_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 1217d79fb8ddSSong Gao { 121840c7674eSSong Gao int i; 121940c7674eSSong Gao VReg temp = {}; 1220329517d5SSong Gao VReg *Vd = (VReg *)vd; 1221329517d5SSong Gao VReg *Vj = (VReg *)vj; 1222d79fb8ddSSong Gao 122340c7674eSSong Gao for (i = 0; i < 2; i++) { 122440c7674eSSong Gao temp.D(2 * i) = int128_getlo(int128_rshift(Vj->Q(i), imm % 128)); 122540c7674eSSong Gao temp.D(2 * i + 1) = int128_getlo(int128_rshift(Vd->Q(i), imm % 128)); 122640c7674eSSong Gao } 1227d79fb8ddSSong Gao *Vd = temp; 1228d79fb8ddSSong Gao } 1229d79fb8ddSSong Gao 1230d79fb8ddSSong Gao VSRANI(vsrani_b_h, 16, B, H) 1231d79fb8ddSSong Gao VSRANI(vsrani_h_w, 32, H, W) 1232d79fb8ddSSong Gao VSRANI(vsrani_w_d, 64, W, D) 1233a5200a17SSong Gao 1234c50ce38aSSong Gao #define VSRLRN(NAME, BIT, E1, E2, E3) \ 123504711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1236a5200a17SSong Gao { \ 1237c50ce38aSSong Gao int i, j, ofs; \ 123804711da1SSong Gao VReg *Vd = (VReg *)vd; \ 123904711da1SSong Gao VReg *Vj = (VReg *)vj; \ 124004711da1SSong Gao VReg *Vk = (VReg *)vk; \ 1241c50ce38aSSong Gao int oprsz = simd_oprsz(desc); \ 1242a5200a17SSong Gao \ 1243c50ce38aSSong Gao ofs = LSX_LEN / BIT; \ 1244c50ce38aSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 1245c50ce38aSSong Gao for (j = 0; j < ofs; j++) { \ 1246c50ce38aSSong Gao Vd->E1(j + ofs * 2 * i) = do_vsrlr_ ##E2(Vj->E2(j + ofs * i), \ 1247c50ce38aSSong Gao Vk->E3(j + ofs * i) % BIT); \ 1248a5200a17SSong Gao } \ 1249c50ce38aSSong Gao Vd->D(2 * i + 1) = 0; \ 1250c50ce38aSSong Gao } \ 1251a5200a17SSong Gao } 1252a5200a17SSong Gao 1253c50ce38aSSong Gao VSRLRN(vsrlrn_b_h, 16, B, H, UH) 1254c50ce38aSSong Gao VSRLRN(vsrlrn_h_w, 32, H, W, UW) 1255c50ce38aSSong Gao VSRLRN(vsrlrn_w_d, 64, W, D, UD) 1256a5200a17SSong Gao 1257c50ce38aSSong Gao #define VSRARN(NAME, BIT, E1, E2, E3) \ 125804711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1259a5200a17SSong Gao { \ 1260c50ce38aSSong Gao int i, j, ofs; \ 126104711da1SSong Gao VReg *Vd = (VReg *)vd; \ 126204711da1SSong Gao VReg *Vj = (VReg *)vj; \ 126304711da1SSong Gao VReg *Vk = (VReg *)vk; \ 1264c50ce38aSSong Gao int oprsz = simd_oprsz(desc); \ 1265a5200a17SSong Gao \ 1266c50ce38aSSong Gao ofs = LSX_LEN / BIT; \ 1267c50ce38aSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 1268c50ce38aSSong Gao for (j = 0; j < ofs; j++) { \ 1269c50ce38aSSong Gao Vd->E1(j + ofs * 2 * i) = do_vsrar_ ## E2(Vj->E2(j + ofs * i), \ 1270c50ce38aSSong Gao Vk->E3(j + ofs * i) % BIT); \ 1271a5200a17SSong Gao } \ 1272c50ce38aSSong Gao Vd->D(2 * i + 1) = 0; \ 1273c50ce38aSSong Gao } \ 1274a5200a17SSong Gao } 1275a5200a17SSong Gao 1276c50ce38aSSong Gao VSRARN(vsrarn_b_h, 16, B, H, UH) 1277c50ce38aSSong Gao VSRARN(vsrarn_h_w, 32, H, W, UW) 1278c50ce38aSSong Gao VSRARN(vsrarn_w_d, 64, W, D, UD) 1279a5200a17SSong Gao 1280a5200a17SSong Gao #define VSRLRNI(NAME, BIT, E1, E2) \ 1281329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 1282a5200a17SSong Gao { \ 1283c50ce38aSSong Gao int i, j, ofs; \ 1284c50ce38aSSong Gao VReg temp = {}; \ 1285329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1286329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 1287c50ce38aSSong Gao int oprsz = simd_oprsz(desc); \ 1288a5200a17SSong Gao \ 1289c50ce38aSSong Gao ofs = LSX_LEN / BIT; \ 1290c50ce38aSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 1291c50ce38aSSong Gao for (j = 0; j < ofs; j++) { \ 1292c50ce38aSSong Gao temp.E1(j + ofs * 2 * i) = do_vsrlr_ ## E2(Vj->E2(j + ofs * i), imm); \ 1293c50ce38aSSong Gao temp.E1(j + ofs * (2 * i + 1)) = do_vsrlr_ ## E2(Vd->E2(j + ofs * i), \ 1294c50ce38aSSong Gao imm); \ 1295c50ce38aSSong Gao } \ 1296a5200a17SSong Gao } \ 1297a5200a17SSong Gao *Vd = temp; \ 1298a5200a17SSong Gao } 1299a5200a17SSong Gao 1300329517d5SSong Gao void HELPER(vsrlrni_d_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 1301a5200a17SSong Gao { 1302c50ce38aSSong Gao int i; 1303c50ce38aSSong Gao VReg temp = {}; 1304329517d5SSong Gao VReg *Vd = (VReg *)vd; 1305329517d5SSong Gao VReg *Vj = (VReg *)vj; 1306c50ce38aSSong Gao Int128 r[4]; 1307c50ce38aSSong Gao int oprsz = simd_oprsz(desc); 1308a5200a17SSong Gao 1309c50ce38aSSong Gao for (i = 0; i < oprsz / 16; i++) { 1310a5200a17SSong Gao if (imm == 0) { 1311c50ce38aSSong Gao temp.D(2 * i) = int128_getlo(Vj->Q(i)); 1312c50ce38aSSong Gao temp.D(2 * i + 1) = int128_getlo(Vd->Q(i)); 1313a5200a17SSong Gao } else { 1314c50ce38aSSong Gao r[2 * i] = int128_and(int128_urshift(Vj->Q(i), (imm - 1)), 1315c50ce38aSSong Gao int128_one()); 1316c50ce38aSSong Gao r[2 * i + 1] = int128_and(int128_urshift(Vd->Q(i), (imm - 1)), 1317c50ce38aSSong Gao int128_one()); 1318c50ce38aSSong Gao temp.D(2 * i) = int128_getlo(int128_add(int128_urshift(Vj->Q(i), 1319c50ce38aSSong Gao imm), r[2 * i])); 1320c50ce38aSSong Gao temp.D(2 * i + 1) = int128_getlo(int128_add(int128_urshift(Vd->Q(i), 1321c50ce38aSSong Gao imm), r[ 2 * i + 1])); 1322c50ce38aSSong Gao } 1323a5200a17SSong Gao } 1324a5200a17SSong Gao *Vd = temp; 1325a5200a17SSong Gao } 1326a5200a17SSong Gao 1327a5200a17SSong Gao VSRLRNI(vsrlrni_b_h, 16, B, H) 1328a5200a17SSong Gao VSRLRNI(vsrlrni_h_w, 32, H, W) 1329a5200a17SSong Gao VSRLRNI(vsrlrni_w_d, 64, W, D) 1330a5200a17SSong Gao 1331a5200a17SSong Gao #define VSRARNI(NAME, BIT, E1, E2) \ 1332329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 1333a5200a17SSong Gao { \ 1334c50ce38aSSong Gao int i, j, ofs; \ 1335c50ce38aSSong Gao VReg temp = {}; \ 1336329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1337329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 1338c50ce38aSSong Gao int oprsz = simd_oprsz(desc); \ 1339a5200a17SSong Gao \ 1340c50ce38aSSong Gao ofs = LSX_LEN / BIT; \ 1341c50ce38aSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 1342c50ce38aSSong Gao for (j = 0; j < ofs; j++) { \ 1343c50ce38aSSong Gao temp.E1(j + ofs * 2 * i) = do_vsrar_ ## E2(Vj->E2(j + ofs * i), imm); \ 1344c50ce38aSSong Gao temp.E1(j + ofs * (2 * i + 1)) = do_vsrar_ ## E2(Vd->E2(j + ofs * i), \ 1345c50ce38aSSong Gao imm); \ 1346c50ce38aSSong Gao } \ 1347a5200a17SSong Gao } \ 1348a5200a17SSong Gao *Vd = temp; \ 1349a5200a17SSong Gao } 1350a5200a17SSong Gao 1351329517d5SSong Gao void HELPER(vsrarni_d_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 1352a5200a17SSong Gao { 1353c50ce38aSSong Gao int i; 1354c50ce38aSSong Gao VReg temp = {}; 1355329517d5SSong Gao VReg *Vd = (VReg *)vd; 1356329517d5SSong Gao VReg *Vj = (VReg *)vj; 1357c50ce38aSSong Gao Int128 r[4]; 1358c50ce38aSSong Gao int oprsz = simd_oprsz(desc); 1359a5200a17SSong Gao 1360c50ce38aSSong Gao for (i = 0; i < oprsz / 16; i++) { 1361a5200a17SSong Gao if (imm == 0) { 1362c50ce38aSSong Gao temp.D(2 * i) = int128_getlo(Vj->Q(i)); 1363c50ce38aSSong Gao temp.D(2 * i + 1) = int128_getlo(Vd->Q(i)); 1364a5200a17SSong Gao } else { 1365c50ce38aSSong Gao r[2 * i] = int128_and(int128_rshift(Vj->Q(i), (imm - 1)), 1366c50ce38aSSong Gao int128_one()); 1367c50ce38aSSong Gao r[2 * i + 1] = int128_and(int128_rshift(Vd->Q(i), (imm - 1)), 1368c50ce38aSSong Gao int128_one()); 1369c50ce38aSSong Gao temp.D(2 * i) = int128_getlo(int128_add(int128_rshift(Vj->Q(i), 1370c50ce38aSSong Gao imm), r[2 * i])); 1371c50ce38aSSong Gao temp.D(2 * i + 1) = int128_getlo(int128_add(int128_rshift(Vd->Q(i), 1372c50ce38aSSong Gao imm), r[2 * i + 1])); 1373c50ce38aSSong Gao } 1374a5200a17SSong Gao } 1375a5200a17SSong Gao *Vd = temp; 1376a5200a17SSong Gao } 1377a5200a17SSong Gao 1378a5200a17SSong Gao VSRARNI(vsrarni_b_h, 16, B, H) 1379a5200a17SSong Gao VSRARNI(vsrarni_h_w, 32, H, W) 1380a5200a17SSong Gao VSRARNI(vsrarni_w_d, 64, W, D) 138183b3815dSSong Gao 138283b3815dSSong Gao #define SSRLNS(NAME, T1, T2, T3) \ 138383b3815dSSong Gao static T1 do_ssrlns_ ## NAME(T2 e2, int sa, int sh) \ 138483b3815dSSong Gao { \ 138583b3815dSSong Gao T1 shft_res; \ 138683b3815dSSong Gao if (sa == 0) { \ 138783b3815dSSong Gao shft_res = e2; \ 138883b3815dSSong Gao } else { \ 138983b3815dSSong Gao shft_res = (((T1)e2) >> sa); \ 139083b3815dSSong Gao } \ 139183b3815dSSong Gao T3 mask; \ 139283b3815dSSong Gao mask = (1ull << sh) -1; \ 139383b3815dSSong Gao if (shft_res > mask) { \ 139483b3815dSSong Gao return mask; \ 139583b3815dSSong Gao } else { \ 139683b3815dSSong Gao return shft_res; \ 139783b3815dSSong Gao } \ 139883b3815dSSong Gao } 139983b3815dSSong Gao 140083b3815dSSong Gao SSRLNS(B, uint16_t, int16_t, uint8_t) 140183b3815dSSong Gao SSRLNS(H, uint32_t, int32_t, uint16_t) 140283b3815dSSong Gao SSRLNS(W, uint64_t, int64_t, uint32_t) 140383b3815dSSong Gao 14046256c8caSSong Gao #define VSSRLN(NAME, BIT, E1, E2, E3) \ 140504711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 140683b3815dSSong Gao { \ 14076256c8caSSong Gao int i, j, ofs; \ 140804711da1SSong Gao VReg *Vd = (VReg *)vd; \ 140904711da1SSong Gao VReg *Vj = (VReg *)vj; \ 141004711da1SSong Gao VReg *Vk = (VReg *)vk; \ 14116256c8caSSong Gao int oprsz = simd_oprsz(desc); \ 141283b3815dSSong Gao \ 14136256c8caSSong Gao ofs = LSX_LEN / BIT; \ 14146256c8caSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 14156256c8caSSong Gao for (j = 0; j < ofs; j++) { \ 14166256c8caSSong Gao Vd->E1(j + ofs * 2 * i) = do_ssrlns_ ## E1(Vj->E2(j + ofs * i), \ 14176256c8caSSong Gao Vk->E3(j + ofs * i) % BIT, \ 14186256c8caSSong Gao BIT / 2 - 1); \ 141983b3815dSSong Gao } \ 14206256c8caSSong Gao Vd->D(2 * i + 1) = 0; \ 14216256c8caSSong Gao } \ 142283b3815dSSong Gao } 142383b3815dSSong Gao 14246256c8caSSong Gao VSSRLN(vssrln_b_h, 16, B, H, UH) 14256256c8caSSong Gao VSSRLN(vssrln_h_w, 32, H, W, UW) 14266256c8caSSong Gao VSSRLN(vssrln_w_d, 64, W, D, UD) 142783b3815dSSong Gao 142883b3815dSSong Gao #define SSRANS(E, T1, T2) \ 142983b3815dSSong Gao static T1 do_ssrans_ ## E(T1 e2, int sa, int sh) \ 143083b3815dSSong Gao { \ 143183b3815dSSong Gao T1 shft_res; \ 143283b3815dSSong Gao if (sa == 0) { \ 143383b3815dSSong Gao shft_res = e2; \ 143483b3815dSSong Gao } else { \ 143583b3815dSSong Gao shft_res = e2 >> sa; \ 143683b3815dSSong Gao } \ 143783b3815dSSong Gao T2 mask; \ 143883b3815dSSong Gao mask = (1ll << sh) - 1; \ 143983b3815dSSong Gao if (shft_res > mask) { \ 144083b3815dSSong Gao return mask; \ 144183b3815dSSong Gao } else if (shft_res < -(mask + 1)) { \ 144283b3815dSSong Gao return ~mask; \ 144383b3815dSSong Gao } else { \ 144483b3815dSSong Gao return shft_res; \ 144583b3815dSSong Gao } \ 144683b3815dSSong Gao } 144783b3815dSSong Gao 144883b3815dSSong Gao SSRANS(B, int16_t, int8_t) 144983b3815dSSong Gao SSRANS(H, int32_t, int16_t) 145083b3815dSSong Gao SSRANS(W, int64_t, int32_t) 145183b3815dSSong Gao 14526256c8caSSong Gao #define VSSRAN(NAME, BIT, E1, E2, E3) \ 145304711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 145483b3815dSSong Gao { \ 14556256c8caSSong Gao int i, j, ofs; \ 145604711da1SSong Gao VReg *Vd = (VReg *)vd; \ 145704711da1SSong Gao VReg *Vj = (VReg *)vj; \ 145804711da1SSong Gao VReg *Vk = (VReg *)vk; \ 14596256c8caSSong Gao int oprsz = simd_oprsz(desc); \ 146083b3815dSSong Gao \ 14616256c8caSSong Gao ofs = LSX_LEN / BIT; \ 14626256c8caSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 14636256c8caSSong Gao for (j = 0; j < ofs; j++) { \ 14646256c8caSSong Gao Vd->E1(j + ofs * 2 * i) = do_ssrans_ ## E1(Vj->E2(j + ofs * i), \ 14656256c8caSSong Gao Vk->E3(j + ofs * i) % BIT, \ 14666256c8caSSong Gao BIT / 2 - 1); \ 146783b3815dSSong Gao } \ 14686256c8caSSong Gao Vd->D(2 * i + 1) = 0; \ 14696256c8caSSong Gao } \ 147083b3815dSSong Gao } 147183b3815dSSong Gao 14726256c8caSSong Gao VSSRAN(vssran_b_h, 16, B, H, UH) 14736256c8caSSong Gao VSSRAN(vssran_h_w, 32, H, W, UW) 14746256c8caSSong Gao VSSRAN(vssran_w_d, 64, W, D, UD) 147583b3815dSSong Gao 147683b3815dSSong Gao #define SSRLNU(E, T1, T2, T3) \ 147783b3815dSSong Gao static T1 do_ssrlnu_ ## E(T3 e2, int sa, int sh) \ 147883b3815dSSong Gao { \ 147983b3815dSSong Gao T1 shft_res; \ 148083b3815dSSong Gao if (sa == 0) { \ 148183b3815dSSong Gao shft_res = e2; \ 148283b3815dSSong Gao } else { \ 148383b3815dSSong Gao shft_res = (((T1)e2) >> sa); \ 148483b3815dSSong Gao } \ 148583b3815dSSong Gao T2 mask; \ 148683b3815dSSong Gao mask = (1ull << sh) - 1; \ 148783b3815dSSong Gao if (shft_res > mask) { \ 148883b3815dSSong Gao return mask; \ 148983b3815dSSong Gao } else { \ 149083b3815dSSong Gao return shft_res; \ 149183b3815dSSong Gao } \ 149283b3815dSSong Gao } 149383b3815dSSong Gao 149483b3815dSSong Gao SSRLNU(B, uint16_t, uint8_t, int16_t) 149583b3815dSSong Gao SSRLNU(H, uint32_t, uint16_t, int32_t) 149683b3815dSSong Gao SSRLNU(W, uint64_t, uint32_t, int64_t) 149783b3815dSSong Gao 14986256c8caSSong Gao #define VSSRLNU(NAME, BIT, E1, E2, E3) \ 149904711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 150083b3815dSSong Gao { \ 15016256c8caSSong Gao int i, j, ofs; \ 150204711da1SSong Gao VReg *Vd = (VReg *)vd; \ 150304711da1SSong Gao VReg *Vj = (VReg *)vj; \ 150404711da1SSong Gao VReg *Vk = (VReg *)vk; \ 15056256c8caSSong Gao int oprsz = simd_oprsz(desc); \ 150683b3815dSSong Gao \ 15076256c8caSSong Gao ofs = LSX_LEN / BIT; \ 15086256c8caSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 15096256c8caSSong Gao for (j = 0; j < ofs; j++) { \ 15106256c8caSSong Gao Vd->E1(j + ofs * 2 * i) = do_ssrlnu_ ## E1(Vj->E2(j + ofs * i), \ 15116256c8caSSong Gao Vk->E3(j + ofs * i) % BIT, \ 15126256c8caSSong Gao BIT / 2); \ 151383b3815dSSong Gao } \ 15146256c8caSSong Gao Vd->D(2 * i + 1) = 0; \ 15156256c8caSSong Gao } \ 151683b3815dSSong Gao } 151783b3815dSSong Gao 15186256c8caSSong Gao VSSRLNU(vssrln_bu_h, 16, B, H, UH) 15196256c8caSSong Gao VSSRLNU(vssrln_hu_w, 32, H, W, UW) 15206256c8caSSong Gao VSSRLNU(vssrln_wu_d, 64, W, D, UD) 152183b3815dSSong Gao 152283b3815dSSong Gao #define SSRANU(E, T1, T2, T3) \ 152383b3815dSSong Gao static T1 do_ssranu_ ## E(T3 e2, int sa, int sh) \ 152483b3815dSSong Gao { \ 152583b3815dSSong Gao T1 shft_res; \ 152683b3815dSSong Gao if (sa == 0) { \ 152783b3815dSSong Gao shft_res = e2; \ 152883b3815dSSong Gao } else { \ 152983b3815dSSong Gao shft_res = e2 >> sa; \ 153083b3815dSSong Gao } \ 153183b3815dSSong Gao if (e2 < 0) { \ 153283b3815dSSong Gao shft_res = 0; \ 153383b3815dSSong Gao } \ 153483b3815dSSong Gao T2 mask; \ 153583b3815dSSong Gao mask = (1ull << sh) - 1; \ 153683b3815dSSong Gao if (shft_res > mask) { \ 153783b3815dSSong Gao return mask; \ 153883b3815dSSong Gao } else { \ 153983b3815dSSong Gao return shft_res; \ 154083b3815dSSong Gao } \ 154183b3815dSSong Gao } 154283b3815dSSong Gao 154383b3815dSSong Gao SSRANU(B, uint16_t, uint8_t, int16_t) 154483b3815dSSong Gao SSRANU(H, uint32_t, uint16_t, int32_t) 154583b3815dSSong Gao SSRANU(W, uint64_t, uint32_t, int64_t) 154683b3815dSSong Gao 15476256c8caSSong Gao #define VSSRANU(NAME, BIT, E1, E2, E3) \ 154804711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 154983b3815dSSong Gao { \ 15506256c8caSSong Gao int i, j, ofs; \ 155104711da1SSong Gao VReg *Vd = (VReg *)vd; \ 155204711da1SSong Gao VReg *Vj = (VReg *)vj; \ 155304711da1SSong Gao VReg *Vk = (VReg *)vk; \ 15546256c8caSSong Gao int oprsz = simd_oprsz(desc); \ 155583b3815dSSong Gao \ 15566256c8caSSong Gao ofs = LSX_LEN / BIT; \ 15576256c8caSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 15586256c8caSSong Gao for (j = 0; j < ofs; j++) { \ 15596256c8caSSong Gao Vd->E1(j + ofs * 2 * i) = do_ssranu_ ## E1(Vj->E2(j + ofs * i), \ 15606256c8caSSong Gao Vk->E3(j + ofs * i) % BIT, \ 15616256c8caSSong Gao BIT / 2); \ 156283b3815dSSong Gao } \ 15636256c8caSSong Gao Vd->D(2 * i + 1) = 0; \ 15646256c8caSSong Gao } \ 156583b3815dSSong Gao } 156683b3815dSSong Gao 15676256c8caSSong Gao VSSRANU(vssran_bu_h, 16, B, H, UH) 15686256c8caSSong Gao VSSRANU(vssran_hu_w, 32, H, W, UW) 15696256c8caSSong Gao VSSRANU(vssran_wu_d, 64, W, D, UD) 157083b3815dSSong Gao 157183b3815dSSong Gao #define VSSRLNI(NAME, BIT, E1, E2) \ 1572329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 157383b3815dSSong Gao { \ 15746256c8caSSong Gao int i, j, ofs; \ 15756256c8caSSong Gao VReg temp = {}; \ 1576329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1577329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 15786256c8caSSong Gao int oprsz = simd_oprsz(desc); \ 157983b3815dSSong Gao \ 15806256c8caSSong Gao ofs = LSX_LEN / BIT; \ 15816256c8caSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 15826256c8caSSong Gao for (j = 0; j < ofs; j++) { \ 15836256c8caSSong Gao temp.E1(j + ofs * 2 * i) = do_ssrlns_ ## E1(Vj->E2(j + ofs * i), \ 15846256c8caSSong Gao imm, BIT / 2 - 1); \ 15856256c8caSSong Gao temp.E1(j + ofs * (2 * i + 1)) = do_ssrlns_ ## E1(Vd->E2(j + ofs * i), \ 15866256c8caSSong Gao imm, BIT / 2 - 1); \ 15876256c8caSSong Gao } \ 158883b3815dSSong Gao } \ 158983b3815dSSong Gao *Vd = temp; \ 159083b3815dSSong Gao } 159183b3815dSSong Gao 15926256c8caSSong Gao static void do_vssrlni_q(VReg *Vd, VReg *Vj, 15936256c8caSSong Gao uint64_t imm, int idx, Int128 mask) 159483b3815dSSong Gao { 15956256c8caSSong Gao Int128 shft_res1, shft_res2; 159683b3815dSSong Gao 159783b3815dSSong Gao if (imm == 0) { 15986256c8caSSong Gao shft_res1 = Vj->Q(idx); 15996256c8caSSong Gao shft_res2 = Vd->Q(idx); 160083b3815dSSong Gao } else { 16016256c8caSSong Gao shft_res1 = int128_urshift(Vj->Q(idx), imm); 16026256c8caSSong Gao shft_res2 = int128_urshift(Vd->Q(idx), imm); 160383b3815dSSong Gao } 160483b3815dSSong Gao 160583b3815dSSong Gao if (int128_ult(mask, shft_res1)) { 16066256c8caSSong Gao Vd->D(idx * 2) = int128_getlo(mask); 160783b3815dSSong Gao }else { 16086256c8caSSong Gao Vd->D(idx * 2) = int128_getlo(shft_res1); 160983b3815dSSong Gao } 161083b3815dSSong Gao 161183b3815dSSong Gao if (int128_ult(mask, shft_res2)) { 16126256c8caSSong Gao Vd->D(idx * 2 + 1) = int128_getlo(mask); 161383b3815dSSong Gao }else { 16146256c8caSSong Gao Vd->D(idx * 2 + 1) = int128_getlo(shft_res2); 16156256c8caSSong Gao } 16166256c8caSSong Gao } 16176256c8caSSong Gao 16186256c8caSSong Gao void HELPER(vssrlni_d_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 16196256c8caSSong Gao { 16206256c8caSSong Gao int i; 16216256c8caSSong Gao Int128 mask; 16226256c8caSSong Gao VReg *Vd = (VReg *)vd; 16236256c8caSSong Gao VReg *Vj = (VReg *)vj; 16246256c8caSSong Gao int oprsz = simd_oprsz(desc); 16256256c8caSSong Gao 16266256c8caSSong Gao mask = int128_sub(int128_lshift(int128_one(), 63), int128_one()); 16276256c8caSSong Gao 16286256c8caSSong Gao for (i = 0; i < oprsz / 16; i++) { 16296256c8caSSong Gao do_vssrlni_q(Vd, Vj, imm, i, mask); 163083b3815dSSong Gao } 163183b3815dSSong Gao } 163283b3815dSSong Gao 163383b3815dSSong Gao VSSRLNI(vssrlni_b_h, 16, B, H) 163483b3815dSSong Gao VSSRLNI(vssrlni_h_w, 32, H, W) 163583b3815dSSong Gao VSSRLNI(vssrlni_w_d, 64, W, D) 163683b3815dSSong Gao 163783b3815dSSong Gao #define VSSRANI(NAME, BIT, E1, E2) \ 1638329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 163983b3815dSSong Gao { \ 16406256c8caSSong Gao int i, j, ofs; \ 16416256c8caSSong Gao VReg temp = {}; \ 1642329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1643329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 16446256c8caSSong Gao int oprsz = simd_oprsz(desc); \ 164583b3815dSSong Gao \ 16466256c8caSSong Gao ofs = LSX_LEN / BIT; \ 16476256c8caSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 16486256c8caSSong Gao for (j = 0; j < ofs; j++) { \ 16496256c8caSSong Gao temp.E1(j + ofs * 2 * i) = do_ssrans_ ## E1(Vj->E2(j + ofs * i), \ 16506256c8caSSong Gao imm, BIT / 2 - 1); \ 16516256c8caSSong Gao temp.E1(j + ofs * (2 * i + 1)) = do_ssrans_ ## E1(Vd->E2(j + ofs * i), \ 16526256c8caSSong Gao imm, BIT / 2 - 1); \ 16536256c8caSSong Gao } \ 165483b3815dSSong Gao } \ 165583b3815dSSong Gao *Vd = temp; \ 165683b3815dSSong Gao } 165783b3815dSSong Gao 16586256c8caSSong Gao static void do_vssrani_d_q(VReg *Vd, VReg *Vj, 16596256c8caSSong Gao uint64_t imm, int idx, Int128 mask, Int128 min) 166083b3815dSSong Gao { 16616256c8caSSong Gao Int128 shft_res1, shft_res2; 166283b3815dSSong Gao 166383b3815dSSong Gao if (imm == 0) { 16646256c8caSSong Gao shft_res1 = Vj->Q(idx); 16656256c8caSSong Gao shft_res2 = Vd->Q(idx); 166683b3815dSSong Gao } else { 16676256c8caSSong Gao shft_res1 = int128_rshift(Vj->Q(idx), imm); 16686256c8caSSong Gao shft_res2 = int128_rshift(Vd->Q(idx), imm); 166983b3815dSSong Gao } 167083b3815dSSong Gao 167183b3815dSSong Gao if (int128_gt(shft_res1, mask)) { 16726256c8caSSong Gao Vd->D(idx * 2) = int128_getlo(mask); 167383b3815dSSong Gao } else if (int128_lt(shft_res1, int128_neg(min))) { 16746256c8caSSong Gao Vd->D(idx * 2) = int128_getlo(min); 167583b3815dSSong Gao } else { 16766256c8caSSong Gao Vd->D(idx * 2) = int128_getlo(shft_res1); 167783b3815dSSong Gao } 167883b3815dSSong Gao 167983b3815dSSong Gao if (int128_gt(shft_res2, mask)) { 16806256c8caSSong Gao Vd->D(idx * 2 + 1) = int128_getlo(mask); 168183b3815dSSong Gao } else if (int128_lt(shft_res2, int128_neg(min))) { 16826256c8caSSong Gao Vd->D(idx * 2 + 1) = int128_getlo(min); 168383b3815dSSong Gao } else { 16846256c8caSSong Gao Vd->D(idx * 2 + 1) = int128_getlo(shft_res2); 168583b3815dSSong Gao } 168683b3815dSSong Gao } 168783b3815dSSong Gao 16886256c8caSSong Gao void HELPER(vssrani_d_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 16896256c8caSSong Gao { 16906256c8caSSong Gao int i; 16916256c8caSSong Gao Int128 mask, min; 16926256c8caSSong Gao VReg *Vd = (VReg *)vd; 16936256c8caSSong Gao VReg *Vj = (VReg *)vj; 16946256c8caSSong Gao int oprsz = simd_oprsz(desc); 16956256c8caSSong Gao 16966256c8caSSong Gao mask = int128_sub(int128_lshift(int128_one(), 63), int128_one()); 16976256c8caSSong Gao min = int128_lshift(int128_one(), 63); 16986256c8caSSong Gao 16996256c8caSSong Gao for (i = 0; i < oprsz / 16; i++) { 17006256c8caSSong Gao do_vssrani_d_q(Vd, Vj, imm, i, mask, min); 17016256c8caSSong Gao } 17026256c8caSSong Gao } 17036256c8caSSong Gao 17046256c8caSSong Gao 170583b3815dSSong Gao VSSRANI(vssrani_b_h, 16, B, H) 170683b3815dSSong Gao VSSRANI(vssrani_h_w, 32, H, W) 170783b3815dSSong Gao VSSRANI(vssrani_w_d, 64, W, D) 170883b3815dSSong Gao 170983b3815dSSong Gao #define VSSRLNUI(NAME, BIT, E1, E2) \ 1710329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 171183b3815dSSong Gao { \ 17126256c8caSSong Gao int i, j, ofs; \ 17136256c8caSSong Gao VReg temp = {}; \ 1714329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1715329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 17166256c8caSSong Gao int oprsz = simd_oprsz(desc); \ 171783b3815dSSong Gao \ 17186256c8caSSong Gao ofs = LSX_LEN / BIT; \ 17196256c8caSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 17206256c8caSSong Gao for (j = 0; j < ofs; j++) { \ 17216256c8caSSong Gao temp.E1(j + ofs * 2 * i) = do_ssrlnu_ ## E1(Vj->E2(j + ofs * i), \ 17226256c8caSSong Gao imm, BIT / 2); \ 17236256c8caSSong Gao temp.E1(j + ofs * (2 * i + 1)) = do_ssrlnu_ ## E1(Vd->E2(j + ofs * i), \ 17246256c8caSSong Gao imm, BIT / 2); \ 17256256c8caSSong Gao } \ 172683b3815dSSong Gao } \ 172783b3815dSSong Gao *Vd = temp; \ 172883b3815dSSong Gao } 172983b3815dSSong Gao 1730329517d5SSong Gao void HELPER(vssrlni_du_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 173183b3815dSSong Gao { 17326256c8caSSong Gao int i; 17336256c8caSSong Gao Int128 mask; 1734329517d5SSong Gao VReg *Vd = (VReg *)vd; 1735329517d5SSong Gao VReg *Vj = (VReg *)vj; 17366256c8caSSong Gao int oprsz = simd_oprsz(desc); 173783b3815dSSong Gao 173883b3815dSSong Gao mask = int128_sub(int128_lshift(int128_one(), 64), int128_one()); 173983b3815dSSong Gao 17406256c8caSSong Gao for (i = 0; i < oprsz / 16; i++) { 17416256c8caSSong Gao do_vssrlni_q(Vd, Vj, imm, i, mask); 174283b3815dSSong Gao } 174383b3815dSSong Gao } 174483b3815dSSong Gao 174583b3815dSSong Gao VSSRLNUI(vssrlni_bu_h, 16, B, H) 174683b3815dSSong Gao VSSRLNUI(vssrlni_hu_w, 32, H, W) 174783b3815dSSong Gao VSSRLNUI(vssrlni_wu_d, 64, W, D) 174883b3815dSSong Gao 174983b3815dSSong Gao #define VSSRANUI(NAME, BIT, E1, E2) \ 1750329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 175183b3815dSSong Gao { \ 17526256c8caSSong Gao int i, j, ofs; \ 17536256c8caSSong Gao VReg temp = {}; \ 1754329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1755329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 17566256c8caSSong Gao int oprsz = simd_oprsz(desc); \ 175783b3815dSSong Gao \ 17586256c8caSSong Gao ofs = LSX_LEN / BIT; \ 17596256c8caSSong Gao for (i = 0; i < oprsz / 16; i++) { \ 17606256c8caSSong Gao for (j = 0; j < ofs; j++) { \ 17616256c8caSSong Gao temp.E1(j + ofs * 2 * i) = do_ssranu_ ## E1(Vj->E2(j + ofs * i), \ 17626256c8caSSong Gao imm, BIT / 2); \ 17636256c8caSSong Gao temp.E1(j + ofs * (2 * i + 1)) = do_ssranu_ ## E1(Vd->E2(j + ofs * i), \ 17646256c8caSSong Gao imm, BIT / 2); \ 17656256c8caSSong Gao } \ 176683b3815dSSong Gao } \ 176783b3815dSSong Gao *Vd = temp; \ 176883b3815dSSong Gao } 176983b3815dSSong Gao 17706256c8caSSong Gao static void do_vssrani_du_q(VReg *Vd, VReg *Vj, 17716256c8caSSong Gao uint64_t imm, int idx, Int128 mask) 177283b3815dSSong Gao { 17736256c8caSSong Gao Int128 shft_res1, shft_res2; 177483b3815dSSong Gao 177583b3815dSSong Gao if (imm == 0) { 17766256c8caSSong Gao shft_res1 = Vj->Q(idx); 17776256c8caSSong Gao shft_res2 = Vd->Q(idx); 177883b3815dSSong Gao } else { 17796256c8caSSong Gao shft_res1 = int128_rshift(Vj->Q(idx), imm); 17806256c8caSSong Gao shft_res2 = int128_rshift(Vd->Q(idx), imm); 178183b3815dSSong Gao } 178283b3815dSSong Gao 17836256c8caSSong Gao if (int128_lt(Vj->Q(idx), int128_zero())) { 178483b3815dSSong Gao shft_res1 = int128_zero(); 178583b3815dSSong Gao } 178683b3815dSSong Gao 17876256c8caSSong Gao if (int128_lt(Vd->Q(idx), int128_zero())) { 178883b3815dSSong Gao shft_res2 = int128_zero(); 178983b3815dSSong Gao } 179083b3815dSSong Gao if (int128_ult(mask, shft_res1)) { 17916256c8caSSong Gao Vd->D(idx * 2) = int128_getlo(mask); 179283b3815dSSong Gao }else { 17936256c8caSSong Gao Vd->D(idx * 2) = int128_getlo(shft_res1); 179483b3815dSSong Gao } 179583b3815dSSong Gao 179683b3815dSSong Gao if (int128_ult(mask, shft_res2)) { 17976256c8caSSong Gao Vd->D(idx * 2 + 1) = int128_getlo(mask); 179883b3815dSSong Gao }else { 17996256c8caSSong Gao Vd->D(idx * 2 + 1) = int128_getlo(shft_res2); 18006256c8caSSong Gao } 18016256c8caSSong Gao 18026256c8caSSong Gao } 18036256c8caSSong Gao 18046256c8caSSong Gao void HELPER(vssrani_du_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 18056256c8caSSong Gao { 18066256c8caSSong Gao int i; 18076256c8caSSong Gao Int128 mask; 18086256c8caSSong Gao VReg *Vd = (VReg *)vd; 18096256c8caSSong Gao VReg *Vj = (VReg *)vj; 18106256c8caSSong Gao int oprsz = simd_oprsz(desc); 18116256c8caSSong Gao 18126256c8caSSong Gao mask = int128_sub(int128_lshift(int128_one(), 64), int128_one()); 18136256c8caSSong Gao 18146256c8caSSong Gao for (i = 0; i < oprsz / 16; i++) { 18156256c8caSSong Gao do_vssrani_du_q(Vd, Vj, imm, i, mask); 181683b3815dSSong Gao } 181783b3815dSSong Gao } 181883b3815dSSong Gao 181983b3815dSSong Gao VSSRANUI(vssrani_bu_h, 16, B, H) 182083b3815dSSong Gao VSSRANUI(vssrani_hu_w, 32, H, W) 182183b3815dSSong Gao VSSRANUI(vssrani_wu_d, 64, W, D) 1822162cd32cSSong Gao 1823162cd32cSSong Gao #define SSRLRNS(E1, E2, T1, T2, T3) \ 1824162cd32cSSong Gao static T1 do_ssrlrns_ ## E1(T2 e2, int sa, int sh) \ 1825162cd32cSSong Gao { \ 1826162cd32cSSong Gao T1 shft_res; \ 1827162cd32cSSong Gao \ 1828162cd32cSSong Gao shft_res = do_vsrlr_ ## E2(e2, sa); \ 1829162cd32cSSong Gao T1 mask; \ 1830162cd32cSSong Gao mask = (1ull << sh) - 1; \ 1831162cd32cSSong Gao if (shft_res > mask) { \ 1832162cd32cSSong Gao return mask; \ 1833162cd32cSSong Gao } else { \ 1834162cd32cSSong Gao return shft_res; \ 1835162cd32cSSong Gao } \ 1836162cd32cSSong Gao } 1837162cd32cSSong Gao 1838162cd32cSSong Gao SSRLRNS(B, H, uint16_t, int16_t, uint8_t) 1839162cd32cSSong Gao SSRLRNS(H, W, uint32_t, int32_t, uint16_t) 1840162cd32cSSong Gao SSRLRNS(W, D, uint64_t, int64_t, uint32_t) 1841162cd32cSSong Gao 184277fca794SSong Gao #define VSSRLRN(NAME, BIT, E1, E2, E3) \ 184304711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1844162cd32cSSong Gao { \ 184577fca794SSong Gao int i, j, ofs; \ 184604711da1SSong Gao VReg *Vd = (VReg *)vd; \ 184704711da1SSong Gao VReg *Vj = (VReg *)vj; \ 184804711da1SSong Gao VReg *Vk = (VReg *)vk; \ 184977fca794SSong Gao int oprsz = simd_oprsz(desc); \ 1850162cd32cSSong Gao \ 185177fca794SSong Gao ofs = LSX_LEN / BIT; \ 185277fca794SSong Gao for (i = 0; i < oprsz / 16; i++) { \ 185377fca794SSong Gao for (j = 0; j < ofs; j++) { \ 185477fca794SSong Gao Vd->E1(j + ofs * 2 * i) = do_ssrlrns_ ## E1(Vj->E2(j + ofs * i), \ 185577fca794SSong Gao Vk->E3(j + ofs * i) % BIT, \ 185677fca794SSong Gao BIT / 2 - 1); \ 1857162cd32cSSong Gao } \ 185877fca794SSong Gao Vd->D(2 * i + 1) = 0; \ 185977fca794SSong Gao } \ 1860162cd32cSSong Gao } 1861162cd32cSSong Gao 186277fca794SSong Gao VSSRLRN(vssrlrn_b_h, 16, B, H, UH) 186377fca794SSong Gao VSSRLRN(vssrlrn_h_w, 32, H, W, UW) 186477fca794SSong Gao VSSRLRN(vssrlrn_w_d, 64, W, D, UD) 1865162cd32cSSong Gao 1866162cd32cSSong Gao #define SSRARNS(E1, E2, T1, T2) \ 1867162cd32cSSong Gao static T1 do_ssrarns_ ## E1(T1 e2, int sa, int sh) \ 1868162cd32cSSong Gao { \ 1869162cd32cSSong Gao T1 shft_res; \ 1870162cd32cSSong Gao \ 1871162cd32cSSong Gao shft_res = do_vsrar_ ## E2(e2, sa); \ 1872162cd32cSSong Gao T2 mask; \ 1873162cd32cSSong Gao mask = (1ll << sh) - 1; \ 1874162cd32cSSong Gao if (shft_res > mask) { \ 1875162cd32cSSong Gao return mask; \ 1876162cd32cSSong Gao } else if (shft_res < -(mask +1)) { \ 1877162cd32cSSong Gao return ~mask; \ 1878162cd32cSSong Gao } else { \ 1879162cd32cSSong Gao return shft_res; \ 1880162cd32cSSong Gao } \ 1881162cd32cSSong Gao } 1882162cd32cSSong Gao 1883162cd32cSSong Gao SSRARNS(B, H, int16_t, int8_t) 1884162cd32cSSong Gao SSRARNS(H, W, int32_t, int16_t) 1885162cd32cSSong Gao SSRARNS(W, D, int64_t, int32_t) 1886162cd32cSSong Gao 188777fca794SSong Gao #define VSSRARN(NAME, BIT, E1, E2, E3) \ 188804711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1889162cd32cSSong Gao { \ 189077fca794SSong Gao int i, j, ofs; \ 189104711da1SSong Gao VReg *Vd = (VReg *)vd; \ 189204711da1SSong Gao VReg *Vj = (VReg *)vj; \ 189304711da1SSong Gao VReg *Vk = (VReg *)vk; \ 189477fca794SSong Gao int oprsz = simd_oprsz(desc); \ 1895162cd32cSSong Gao \ 189677fca794SSong Gao ofs = LSX_LEN / BIT; \ 189777fca794SSong Gao for (i = 0; i < oprsz / 16; i++) { \ 189877fca794SSong Gao for (j = 0; j < ofs; j++) { \ 189977fca794SSong Gao Vd->E1(j + ofs * 2 * i) = do_ssrarns_ ## E1(Vj->E2(j + ofs * i), \ 190077fca794SSong Gao Vk->E3(j + ofs * i) % BIT, \ 190177fca794SSong Gao BIT/ 2 - 1); \ 1902162cd32cSSong Gao } \ 190377fca794SSong Gao Vd->D(2 * i + 1) = 0; \ 190477fca794SSong Gao } \ 1905162cd32cSSong Gao } 1906162cd32cSSong Gao 190777fca794SSong Gao VSSRARN(vssrarn_b_h, 16, B, H, UH) 190877fca794SSong Gao VSSRARN(vssrarn_h_w, 32, H, W, UW) 190977fca794SSong Gao VSSRARN(vssrarn_w_d, 64, W, D, UD) 1910162cd32cSSong Gao 1911162cd32cSSong Gao #define SSRLRNU(E1, E2, T1, T2, T3) \ 1912162cd32cSSong Gao static T1 do_ssrlrnu_ ## E1(T3 e2, int sa, int sh) \ 1913162cd32cSSong Gao { \ 1914162cd32cSSong Gao T1 shft_res; \ 1915162cd32cSSong Gao \ 1916162cd32cSSong Gao shft_res = do_vsrlr_ ## E2(e2, sa); \ 1917162cd32cSSong Gao \ 1918162cd32cSSong Gao T2 mask; \ 1919162cd32cSSong Gao mask = (1ull << sh) - 1; \ 1920162cd32cSSong Gao if (shft_res > mask) { \ 1921162cd32cSSong Gao return mask; \ 1922162cd32cSSong Gao } else { \ 1923162cd32cSSong Gao return shft_res; \ 1924162cd32cSSong Gao } \ 1925162cd32cSSong Gao } 1926162cd32cSSong Gao 1927162cd32cSSong Gao SSRLRNU(B, H, uint16_t, uint8_t, int16_t) 1928162cd32cSSong Gao SSRLRNU(H, W, uint32_t, uint16_t, int32_t) 1929162cd32cSSong Gao SSRLRNU(W, D, uint64_t, uint32_t, int64_t) 1930162cd32cSSong Gao 193177fca794SSong Gao #define VSSRLRNU(NAME, BIT, E1, E2, E3) \ 193204711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1933162cd32cSSong Gao { \ 193477fca794SSong Gao int i, j, ofs; \ 193504711da1SSong Gao VReg *Vd = (VReg *)vd; \ 193604711da1SSong Gao VReg *Vj = (VReg *)vj; \ 193704711da1SSong Gao VReg *Vk = (VReg *)vk; \ 193877fca794SSong Gao int oprsz = simd_oprsz(desc); \ 1939162cd32cSSong Gao \ 194077fca794SSong Gao ofs = LSX_LEN / BIT; \ 194177fca794SSong Gao for (i = 0; i < oprsz / 16; i++) { \ 194277fca794SSong Gao for (j = 0; j < ofs; j++) { \ 194377fca794SSong Gao Vd->E1(j + ofs * 2 * i) = do_ssrlrnu_ ## E1(Vj->E2(j + ofs * i), \ 194477fca794SSong Gao Vk->E3(j + ofs * i) % BIT, \ 194577fca794SSong Gao BIT / 2); \ 1946162cd32cSSong Gao } \ 194777fca794SSong Gao Vd->D(2 * i + 1) = 0; \ 194877fca794SSong Gao } \ 1949162cd32cSSong Gao } 1950162cd32cSSong Gao 195177fca794SSong Gao VSSRLRNU(vssrlrn_bu_h, 16, B, H, UH) 195277fca794SSong Gao VSSRLRNU(vssrlrn_hu_w, 32, H, W, UW) 195377fca794SSong Gao VSSRLRNU(vssrlrn_wu_d, 64, W, D, UD) 1954162cd32cSSong Gao 1955162cd32cSSong Gao #define SSRARNU(E1, E2, T1, T2, T3) \ 1956162cd32cSSong Gao static T1 do_ssrarnu_ ## E1(T3 e2, int sa, int sh) \ 1957162cd32cSSong Gao { \ 1958162cd32cSSong Gao T1 shft_res; \ 1959162cd32cSSong Gao \ 1960162cd32cSSong Gao if (e2 < 0) { \ 1961162cd32cSSong Gao shft_res = 0; \ 1962162cd32cSSong Gao } else { \ 1963162cd32cSSong Gao shft_res = do_vsrar_ ## E2(e2, sa); \ 1964162cd32cSSong Gao } \ 1965162cd32cSSong Gao T2 mask; \ 1966162cd32cSSong Gao mask = (1ull << sh) - 1; \ 1967162cd32cSSong Gao if (shft_res > mask) { \ 1968162cd32cSSong Gao return mask; \ 1969162cd32cSSong Gao } else { \ 1970162cd32cSSong Gao return shft_res; \ 1971162cd32cSSong Gao } \ 1972162cd32cSSong Gao } 1973162cd32cSSong Gao 1974162cd32cSSong Gao SSRARNU(B, H, uint16_t, uint8_t, int16_t) 1975162cd32cSSong Gao SSRARNU(H, W, uint32_t, uint16_t, int32_t) 1976162cd32cSSong Gao SSRARNU(W, D, uint64_t, uint32_t, int64_t) 1977162cd32cSSong Gao 197877fca794SSong Gao #define VSSRARNU(NAME, BIT, E1, E2, E3) \ 197904711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1980162cd32cSSong Gao { \ 198177fca794SSong Gao int i, j, ofs; \ 198204711da1SSong Gao VReg *Vd = (VReg *)vd; \ 198304711da1SSong Gao VReg *Vj = (VReg *)vj; \ 198404711da1SSong Gao VReg *Vk = (VReg *)vk; \ 198577fca794SSong Gao int oprsz = simd_oprsz(desc); \ 1986162cd32cSSong Gao \ 198777fca794SSong Gao ofs = LSX_LEN / BIT; \ 198877fca794SSong Gao for (i = 0; i < oprsz / 16; i++) { \ 198977fca794SSong Gao for (j = 0; j < ofs; j++) { \ 199077fca794SSong Gao Vd->E1(j + ofs * 2 * i) = do_ssrarnu_ ## E1(Vj->E2(j + ofs * i), \ 199177fca794SSong Gao Vk->E3(j + ofs * i) % BIT, \ 199277fca794SSong Gao BIT / 2); \ 1993162cd32cSSong Gao } \ 199477fca794SSong Gao Vd->D(2 * i + 1) = 0; \ 199577fca794SSong Gao } \ 1996162cd32cSSong Gao } 1997162cd32cSSong Gao 199877fca794SSong Gao VSSRARNU(vssrarn_bu_h, 16, B, H, UH) 199977fca794SSong Gao VSSRARNU(vssrarn_hu_w, 32, H, W, UW) 200077fca794SSong Gao VSSRARNU(vssrarn_wu_d, 64, W, D, UD) 2001162cd32cSSong Gao 2002162cd32cSSong Gao #define VSSRLRNI(NAME, BIT, E1, E2) \ 2003329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 2004162cd32cSSong Gao { \ 200577fca794SSong Gao int i, j, ofs; \ 200677fca794SSong Gao VReg temp = {}; \ 2007329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 2008329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 200977fca794SSong Gao int oprsz = simd_oprsz(desc); \ 2010162cd32cSSong Gao \ 201177fca794SSong Gao ofs = LSX_LEN / BIT; \ 201277fca794SSong Gao for (i = 0; i < oprsz / 16; i++) { \ 201377fca794SSong Gao for (j = 0; j < ofs; j++) { \ 201477fca794SSong Gao temp.E1(j + ofs * 2 * i) = do_ssrlrns_ ## E1(Vj->E2(j + ofs * i), \ 201577fca794SSong Gao imm, BIT / 2 - 1); \ 201677fca794SSong Gao temp.E1(j + ofs * (2 * i + 1)) = do_ssrlrns_ ## E1(Vd->E2(j + ofs * i), \ 201777fca794SSong Gao imm, BIT / 2 - 1); \ 201877fca794SSong Gao } \ 2019162cd32cSSong Gao } \ 2020162cd32cSSong Gao *Vd = temp; \ 2021162cd32cSSong Gao } 2022162cd32cSSong Gao 202377fca794SSong Gao static void do_vssrlrni_q(VReg *Vd, VReg * Vj, 202477fca794SSong Gao uint64_t imm, int idx, Int128 mask) 202577fca794SSong Gao { 202677fca794SSong Gao Int128 shft_res1, shft_res2, r1, r2; 202777fca794SSong Gao if (imm == 0) { 202877fca794SSong Gao shft_res1 = Vj->Q(idx); 202977fca794SSong Gao shft_res2 = Vd->Q(idx); 203077fca794SSong Gao } else { 203177fca794SSong Gao r1 = int128_and(int128_urshift(Vj->Q(idx), (imm - 1)), int128_one()); 203277fca794SSong Gao r2 = int128_and(int128_urshift(Vd->Q(idx), (imm - 1)), int128_one()); 203377fca794SSong Gao shft_res1 = (int128_add(int128_urshift(Vj->Q(idx), imm), r1)); 203477fca794SSong Gao shft_res2 = (int128_add(int128_urshift(Vd->Q(idx), imm), r2)); 203577fca794SSong Gao } 203677fca794SSong Gao 203777fca794SSong Gao if (int128_ult(mask, shft_res1)) { 203877fca794SSong Gao Vd->D(idx * 2) = int128_getlo(mask); 203977fca794SSong Gao }else { 204077fca794SSong Gao Vd->D(idx * 2) = int128_getlo(shft_res1); 204177fca794SSong Gao } 204277fca794SSong Gao 204377fca794SSong Gao if (int128_ult(mask, shft_res2)) { 204477fca794SSong Gao Vd->D(idx * 2 + 1) = int128_getlo(mask); 204577fca794SSong Gao }else { 204677fca794SSong Gao Vd->D(idx * 2 + 1) = int128_getlo(shft_res2); 204777fca794SSong Gao } 204877fca794SSong Gao } 204977fca794SSong Gao 205077fca794SSong Gao void HELPER(vssrlrni_d_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 205177fca794SSong Gao { 205277fca794SSong Gao int i; 205377fca794SSong Gao Int128 mask; 205477fca794SSong Gao VReg *Vd = (VReg *)vd; 205577fca794SSong Gao VReg *Vj = (VReg *)vj; 205677fca794SSong Gao int oprsz = simd_oprsz(desc); 205777fca794SSong Gao 205877fca794SSong Gao mask = int128_sub(int128_lshift(int128_one(), 63), int128_one()); 205977fca794SSong Gao 206077fca794SSong Gao for (i = 0; i < oprsz / 16; i++) { 206177fca794SSong Gao do_vssrlrni_q(Vd, Vj, imm, i, mask); 206277fca794SSong Gao } 2063162cd32cSSong Gao } 2064162cd32cSSong Gao 2065162cd32cSSong Gao VSSRLRNI(vssrlrni_b_h, 16, B, H) 2066162cd32cSSong Gao VSSRLRNI(vssrlrni_h_w, 32, H, W) 2067162cd32cSSong Gao VSSRLRNI(vssrlrni_w_d, 64, W, D) 2068162cd32cSSong Gao 2069162cd32cSSong Gao #define VSSRARNI(NAME, BIT, E1, E2) \ 2070329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 2071162cd32cSSong Gao { \ 207277fca794SSong Gao int i, j, ofs; \ 207377fca794SSong Gao VReg temp = {}; \ 2074329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 2075329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 207677fca794SSong Gao int oprsz = simd_oprsz(desc); \ 2077162cd32cSSong Gao \ 207877fca794SSong Gao ofs = LSX_LEN / BIT; \ 207977fca794SSong Gao for (i = 0; i < oprsz / 16; i++) { \ 208077fca794SSong Gao for (j = 0; j < ofs; j++) { \ 208177fca794SSong Gao temp.E1(j + ofs * 2 * i) = do_ssrarns_ ## E1(Vj->E2(j + ofs * i), \ 208277fca794SSong Gao imm, BIT / 2 - 1); \ 208377fca794SSong Gao temp.E1(j + ofs * (2 * i + 1)) = do_ssrarns_ ## E1(Vd->E2(j + ofs * i), \ 208477fca794SSong Gao imm, BIT / 2 - 1); \ 208577fca794SSong Gao } \ 2086162cd32cSSong Gao } \ 2087162cd32cSSong Gao *Vd = temp; \ 2088162cd32cSSong Gao } 2089162cd32cSSong Gao 209077fca794SSong Gao static void do_vssrarni_d_q(VReg *Vd, VReg *Vj, 209177fca794SSong Gao uint64_t imm, int idx, Int128 mask1, Int128 mask2) 2092162cd32cSSong Gao { 209377fca794SSong Gao Int128 shft_res1, shft_res2, r1, r2; 2094162cd32cSSong Gao 2095162cd32cSSong Gao if (imm == 0) { 209677fca794SSong Gao shft_res1 = Vj->Q(idx); 209777fca794SSong Gao shft_res2 = Vd->Q(idx); 2098162cd32cSSong Gao } else { 209977fca794SSong Gao r1 = int128_and(int128_rshift(Vj->Q(idx), (imm - 1)), int128_one()); 210077fca794SSong Gao r2 = int128_and(int128_rshift(Vd->Q(idx), (imm - 1)), int128_one()); 210177fca794SSong Gao shft_res1 = int128_add(int128_rshift(Vj->Q(idx), imm), r1); 210277fca794SSong Gao shft_res2 = int128_add(int128_rshift(Vd->Q(idx), imm), r2); 2103162cd32cSSong Gao } 210477fca794SSong Gao if (int128_gt(shft_res1, mask1)) { 210577fca794SSong Gao Vd->D(idx * 2) = int128_getlo(mask1); 210677fca794SSong Gao } else if (int128_lt(shft_res1, int128_neg(mask2))) { 210777fca794SSong Gao Vd->D(idx * 2) = int128_getlo(mask2); 210877fca794SSong Gao } else { 210977fca794SSong Gao Vd->D(idx * 2) = int128_getlo(shft_res1); 211077fca794SSong Gao } 211177fca794SSong Gao 211277fca794SSong Gao if (int128_gt(shft_res2, mask1)) { 211377fca794SSong Gao Vd->D(idx * 2 + 1) = int128_getlo(mask1); 211477fca794SSong Gao } else if (int128_lt(shft_res2, int128_neg(mask2))) { 211577fca794SSong Gao Vd->D(idx * 2 + 1) = int128_getlo(mask2); 211677fca794SSong Gao } else { 211777fca794SSong Gao Vd->D(idx * 2 + 1) = int128_getlo(shft_res2); 211877fca794SSong Gao } 211977fca794SSong Gao } 212077fca794SSong Gao 212177fca794SSong Gao void HELPER(vssrarni_d_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 212277fca794SSong Gao { 212377fca794SSong Gao int i; 212477fca794SSong Gao Int128 mask1, mask2; 212577fca794SSong Gao VReg *Vd = (VReg *)vd; 212677fca794SSong Gao VReg *Vj = (VReg *)vj; 212777fca794SSong Gao int oprsz = simd_oprsz(desc); 2128162cd32cSSong Gao 2129162cd32cSSong Gao mask1 = int128_sub(int128_lshift(int128_one(), 63), int128_one()); 2130162cd32cSSong Gao mask2 = int128_lshift(int128_one(), 63); 2131162cd32cSSong Gao 213277fca794SSong Gao for (i = 0; i < oprsz / 16; i++) { 213377fca794SSong Gao do_vssrarni_d_q(Vd, Vj, imm, i, mask1, mask2); 2134162cd32cSSong Gao } 2135162cd32cSSong Gao } 2136162cd32cSSong Gao 2137162cd32cSSong Gao VSSRARNI(vssrarni_b_h, 16, B, H) 2138162cd32cSSong Gao VSSRARNI(vssrarni_h_w, 32, H, W) 2139162cd32cSSong Gao VSSRARNI(vssrarni_w_d, 64, W, D) 2140162cd32cSSong Gao 2141162cd32cSSong Gao #define VSSRLRNUI(NAME, BIT, E1, E2) \ 2142329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 2143162cd32cSSong Gao { \ 214477fca794SSong Gao int i, j, ofs; \ 214577fca794SSong Gao VReg temp = {}; \ 2146329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 2147329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 214877fca794SSong Gao int oprsz = simd_oprsz(desc); \ 2149162cd32cSSong Gao \ 215077fca794SSong Gao ofs = LSX_LEN / BIT; \ 215177fca794SSong Gao for (i = 0; i < oprsz / 16; i++) { \ 215277fca794SSong Gao for (j = 0; j < ofs; j++) { \ 215377fca794SSong Gao temp.E1(j + ofs * 2 * i) = do_ssrlrnu_ ## E1(Vj->E2(j + ofs * i), \ 215477fca794SSong Gao imm, BIT / 2); \ 215577fca794SSong Gao temp.E1(j + ofs * (2 * i + 1)) = do_ssrlrnu_ ## E1(Vd->E2(j + ofs * i), \ 215677fca794SSong Gao imm, BIT / 2); \ 215777fca794SSong Gao } \ 2158162cd32cSSong Gao } \ 2159162cd32cSSong Gao *Vd = temp; \ 2160162cd32cSSong Gao } 2161162cd32cSSong Gao 216277fca794SSong Gao void HELPER(vssrlrni_du_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 216377fca794SSong Gao { 216477fca794SSong Gao int i; 216577fca794SSong Gao Int128 mask; 216677fca794SSong Gao VReg *Vd = (VReg *)vd; 216777fca794SSong Gao VReg *Vj = (VReg *)vj; 216877fca794SSong Gao int oprsz = simd_oprsz(desc); 216977fca794SSong Gao 217077fca794SSong Gao mask = int128_sub(int128_lshift(int128_one(), 64), int128_one()); 217177fca794SSong Gao 217277fca794SSong Gao for (i = 0; i < oprsz / 16; i++) { 217377fca794SSong Gao do_vssrlrni_q(Vd, Vj, imm, i, mask); 217477fca794SSong Gao } 217577fca794SSong Gao } 217677fca794SSong Gao 2177162cd32cSSong Gao VSSRLRNUI(vssrlrni_bu_h, 16, B, H) 2178162cd32cSSong Gao VSSRLRNUI(vssrlrni_hu_w, 32, H, W) 2179162cd32cSSong Gao VSSRLRNUI(vssrlrni_wu_d, 64, W, D) 2180162cd32cSSong Gao 2181162cd32cSSong Gao #define VSSRARNUI(NAME, BIT, E1, E2) \ 2182329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 2183162cd32cSSong Gao { \ 218477fca794SSong Gao int i, j, ofs; \ 218577fca794SSong Gao VReg temp = {}; \ 2186329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 2187329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 218877fca794SSong Gao int oprsz = simd_oprsz(desc); \ 2189162cd32cSSong Gao \ 219077fca794SSong Gao ofs = LSX_LEN / BIT; \ 219177fca794SSong Gao for (i = 0; i < oprsz / 16; i++) { \ 219277fca794SSong Gao for (j = 0; j < ofs; j++) { \ 219377fca794SSong Gao temp.E1(j + ofs * 2 * i) = do_ssrarnu_ ## E1(Vj->E2(j + ofs * i), \ 219477fca794SSong Gao imm, BIT / 2); \ 219577fca794SSong Gao temp.E1(j + ofs * (2 * i + 1)) = do_ssrarnu_ ## E1(Vd->E2(j + ofs * i), \ 219677fca794SSong Gao imm, BIT / 2); \ 219777fca794SSong Gao } \ 2198162cd32cSSong Gao } \ 2199162cd32cSSong Gao *Vd = temp; \ 2200162cd32cSSong Gao } 2201162cd32cSSong Gao 220277fca794SSong Gao static void do_vssrarni_du_q(VReg *Vd, VReg *Vj, 220377fca794SSong Gao uint64_t imm, int idx, Int128 mask1, Int128 mask2) 2204162cd32cSSong Gao { 220577fca794SSong Gao Int128 shft_res1, shft_res2, r1, r2; 2206162cd32cSSong Gao 2207162cd32cSSong Gao if (imm == 0) { 220877fca794SSong Gao shft_res1 = Vj->Q(idx); 220977fca794SSong Gao shft_res2 = Vd->Q(idx); 2210162cd32cSSong Gao } else { 221177fca794SSong Gao r1 = int128_and(int128_rshift(Vj->Q(idx), (imm - 1)), int128_one()); 221277fca794SSong Gao r2 = int128_and(int128_rshift(Vd->Q(idx), (imm - 1)), int128_one()); 221377fca794SSong Gao shft_res1 = int128_add(int128_rshift(Vj->Q(idx), imm), r1); 221477fca794SSong Gao shft_res2 = int128_add(int128_rshift(Vd->Q(idx), imm), r2); 2215162cd32cSSong Gao } 2216162cd32cSSong Gao 221777fca794SSong Gao if (int128_lt(Vj->Q(idx), int128_zero())) { 2218162cd32cSSong Gao shft_res1 = int128_zero(); 2219162cd32cSSong Gao } 222077fca794SSong Gao if (int128_lt(Vd->Q(idx), int128_zero())) { 2221162cd32cSSong Gao shft_res2 = int128_zero(); 2222162cd32cSSong Gao } 2223162cd32cSSong Gao 222477fca794SSong Gao if (int128_gt(shft_res1, mask1)) { 222577fca794SSong Gao Vd->D(idx * 2) = int128_getlo(mask1); 222677fca794SSong Gao } else if (int128_lt(shft_res1, int128_neg(mask2))) { 222777fca794SSong Gao Vd->D(idx * 2) = int128_getlo(mask2); 222877fca794SSong Gao } else { 222977fca794SSong Gao Vd->D(idx * 2) = int128_getlo(shft_res1); 223077fca794SSong Gao } 223177fca794SSong Gao 223277fca794SSong Gao if (int128_gt(shft_res2, mask1)) { 223377fca794SSong Gao Vd->D(idx * 2 + 1) = int128_getlo(mask1); 223477fca794SSong Gao } else if (int128_lt(shft_res2, int128_neg(mask2))) { 223577fca794SSong Gao Vd->D(idx * 2 + 1) = int128_getlo(mask2); 223677fca794SSong Gao } else { 223777fca794SSong Gao Vd->D(idx * 2 + 1) = int128_getlo(shft_res2); 223877fca794SSong Gao } 223977fca794SSong Gao } 224077fca794SSong Gao 224177fca794SSong Gao void HELPER(vssrarni_du_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 224277fca794SSong Gao { 224377fca794SSong Gao int i; 224477fca794SSong Gao Int128 mask1, mask2; 224577fca794SSong Gao VReg *Vd = (VReg *)vd; 224677fca794SSong Gao VReg *Vj = (VReg *)vj; 224777fca794SSong Gao int oprsz = simd_oprsz(desc); 224877fca794SSong Gao 2249162cd32cSSong Gao mask1 = int128_sub(int128_lshift(int128_one(), 64), int128_one()); 2250162cd32cSSong Gao mask2 = int128_lshift(int128_one(), 64); 2251162cd32cSSong Gao 225277fca794SSong Gao for (i = 0; i < oprsz / 16; i++) { 225377fca794SSong Gao do_vssrarni_du_q(Vd, Vj, imm, i, mask1, mask2); 2254162cd32cSSong Gao } 2255162cd32cSSong Gao } 2256162cd32cSSong Gao 2257162cd32cSSong Gao VSSRARNUI(vssrarni_bu_h, 16, B, H) 2258162cd32cSSong Gao VSSRARNUI(vssrarni_hu_w, 32, H, W) 2259162cd32cSSong Gao VSSRARNUI(vssrarni_wu_d, 64, W, D) 22602e105e12SSong Gao 22612e105e12SSong Gao #define DO_2OP(NAME, BIT, E, DO_OP) \ 2262ff27e335SSong Gao void HELPER(NAME)(void *vd, void *vj, uint32_t desc) \ 22632e105e12SSong Gao { \ 22642e105e12SSong Gao int i; \ 2265ff27e335SSong Gao VReg *Vd = (VReg *)vd; \ 2266ff27e335SSong Gao VReg *Vj = (VReg *)vj; \ 226712ad133fSSong Gao int oprsz = simd_oprsz(desc); \ 22682e105e12SSong Gao \ 226912ad133fSSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) \ 22702e105e12SSong Gao { \ 22712e105e12SSong Gao Vd->E(i) = DO_OP(Vj->E(i)); \ 22722e105e12SSong Gao } \ 22732e105e12SSong Gao } 22742e105e12SSong Gao 22752e105e12SSong Gao #define DO_CLO_B(N) (clz32(~N & 0xff) - 24) 22762e105e12SSong Gao #define DO_CLO_H(N) (clz32(~N & 0xffff) - 16) 22772e105e12SSong Gao #define DO_CLO_W(N) (clz32(~N)) 22782e105e12SSong Gao #define DO_CLO_D(N) (clz64(~N)) 22792e105e12SSong Gao #define DO_CLZ_B(N) (clz32(N) - 24) 22802e105e12SSong Gao #define DO_CLZ_H(N) (clz32(N) - 16) 22812e105e12SSong Gao #define DO_CLZ_W(N) (clz32(N)) 22822e105e12SSong Gao #define DO_CLZ_D(N) (clz64(N)) 22832e105e12SSong Gao 22842e105e12SSong Gao DO_2OP(vclo_b, 8, UB, DO_CLO_B) 22852e105e12SSong Gao DO_2OP(vclo_h, 16, UH, DO_CLO_H) 22862e105e12SSong Gao DO_2OP(vclo_w, 32, UW, DO_CLO_W) 22872e105e12SSong Gao DO_2OP(vclo_d, 64, UD, DO_CLO_D) 22882e105e12SSong Gao DO_2OP(vclz_b, 8, UB, DO_CLZ_B) 22892e105e12SSong Gao DO_2OP(vclz_h, 16, UH, DO_CLZ_H) 22902e105e12SSong Gao DO_2OP(vclz_w, 32, UW, DO_CLZ_W) 22912e105e12SSong Gao DO_2OP(vclz_d, 64, UD, DO_CLZ_D) 2292bb22ee57SSong Gao 2293bb22ee57SSong Gao #define VPCNT(NAME, BIT, E, FN) \ 2294ff27e335SSong Gao void HELPER(NAME)(void *vd, void *vj, uint32_t desc) \ 2295bb22ee57SSong Gao { \ 2296bb22ee57SSong Gao int i; \ 2297ff27e335SSong Gao VReg *Vd = (VReg *)vd; \ 2298ff27e335SSong Gao VReg *Vj = (VReg *)vj; \ 2299956dec74SSong Gao int oprsz = simd_oprsz(desc); \ 2300bb22ee57SSong Gao \ 2301956dec74SSong Gao for (i = 0; i < oprsz / (BIT / 8); i++) \ 2302bb22ee57SSong Gao { \ 2303bb22ee57SSong Gao Vd->E(i) = FN(Vj->E(i)); \ 2304bb22ee57SSong Gao } \ 2305bb22ee57SSong Gao } 2306bb22ee57SSong Gao 2307bb22ee57SSong Gao VPCNT(vpcnt_b, 8, UB, ctpop8) 2308bb22ee57SSong Gao VPCNT(vpcnt_h, 16, UH, ctpop16) 2309bb22ee57SSong Gao VPCNT(vpcnt_w, 32, UW, ctpop32) 2310bb22ee57SSong Gao VPCNT(vpcnt_d, 64, UD, ctpop64) 23110b1e6705SSong Gao 23120b1e6705SSong Gao #define DO_BITCLR(a, bit) (a & ~(1ull << bit)) 23130b1e6705SSong Gao #define DO_BITSET(a, bit) (a | 1ull << bit) 23140b1e6705SSong Gao #define DO_BITREV(a, bit) (a ^ (1ull << bit)) 23150b1e6705SSong Gao 23160b1e6705SSong Gao #define DO_BIT(NAME, BIT, E, DO_OP) \ 23170b1e6705SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t v) \ 23180b1e6705SSong Gao { \ 23190b1e6705SSong Gao int i; \ 23200b1e6705SSong Gao VReg *Vd = (VReg *)vd; \ 23210b1e6705SSong Gao VReg *Vj = (VReg *)vj; \ 23220b1e6705SSong Gao VReg *Vk = (VReg *)vk; \ 23230b1e6705SSong Gao \ 23240b1e6705SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 23250b1e6705SSong Gao Vd->E(i) = DO_OP(Vj->E(i), Vk->E(i)%BIT); \ 23260b1e6705SSong Gao } \ 23270b1e6705SSong Gao } 23280b1e6705SSong Gao 23290b1e6705SSong Gao DO_BIT(vbitclr_b, 8, UB, DO_BITCLR) 23300b1e6705SSong Gao DO_BIT(vbitclr_h, 16, UH, DO_BITCLR) 23310b1e6705SSong Gao DO_BIT(vbitclr_w, 32, UW, DO_BITCLR) 23320b1e6705SSong Gao DO_BIT(vbitclr_d, 64, UD, DO_BITCLR) 23330b1e6705SSong Gao DO_BIT(vbitset_b, 8, UB, DO_BITSET) 23340b1e6705SSong Gao DO_BIT(vbitset_h, 16, UH, DO_BITSET) 23350b1e6705SSong Gao DO_BIT(vbitset_w, 32, UW, DO_BITSET) 23360b1e6705SSong Gao DO_BIT(vbitset_d, 64, UD, DO_BITSET) 23370b1e6705SSong Gao DO_BIT(vbitrev_b, 8, UB, DO_BITREV) 23380b1e6705SSong Gao DO_BIT(vbitrev_h, 16, UH, DO_BITREV) 23390b1e6705SSong Gao DO_BIT(vbitrev_w, 32, UW, DO_BITREV) 23400b1e6705SSong Gao DO_BIT(vbitrev_d, 64, UD, DO_BITREV) 23410b1e6705SSong Gao 23420b1e6705SSong Gao #define DO_BITI(NAME, BIT, E, DO_OP) \ 23430b1e6705SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t v) \ 23440b1e6705SSong Gao { \ 23450b1e6705SSong Gao int i; \ 23460b1e6705SSong Gao VReg *Vd = (VReg *)vd; \ 23470b1e6705SSong Gao VReg *Vj = (VReg *)vj; \ 23480b1e6705SSong Gao \ 23490b1e6705SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 23500b1e6705SSong Gao Vd->E(i) = DO_OP(Vj->E(i), imm); \ 23510b1e6705SSong Gao } \ 23520b1e6705SSong Gao } 23530b1e6705SSong Gao 23540b1e6705SSong Gao DO_BITI(vbitclri_b, 8, UB, DO_BITCLR) 23550b1e6705SSong Gao DO_BITI(vbitclri_h, 16, UH, DO_BITCLR) 23560b1e6705SSong Gao DO_BITI(vbitclri_w, 32, UW, DO_BITCLR) 23570b1e6705SSong Gao DO_BITI(vbitclri_d, 64, UD, DO_BITCLR) 23580b1e6705SSong Gao DO_BITI(vbitseti_b, 8, UB, DO_BITSET) 23590b1e6705SSong Gao DO_BITI(vbitseti_h, 16, UH, DO_BITSET) 23600b1e6705SSong Gao DO_BITI(vbitseti_w, 32, UW, DO_BITSET) 23610b1e6705SSong Gao DO_BITI(vbitseti_d, 64, UD, DO_BITSET) 23620b1e6705SSong Gao DO_BITI(vbitrevi_b, 8, UB, DO_BITREV) 23630b1e6705SSong Gao DO_BITI(vbitrevi_h, 16, UH, DO_BITREV) 23640b1e6705SSong Gao DO_BITI(vbitrevi_w, 32, UW, DO_BITREV) 23650b1e6705SSong Gao DO_BITI(vbitrevi_d, 64, UD, DO_BITREV) 2366ac95a0b9SSong Gao 2367ac95a0b9SSong Gao #define VFRSTP(NAME, BIT, MASK, E) \ 236804711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 2369ac95a0b9SSong Gao { \ 2370ac95a0b9SSong Gao int i, m; \ 237104711da1SSong Gao VReg *Vd = (VReg *)vd; \ 237204711da1SSong Gao VReg *Vj = (VReg *)vj; \ 237304711da1SSong Gao VReg *Vk = (VReg *)vk; \ 2374ac95a0b9SSong Gao \ 2375ac95a0b9SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2376ac95a0b9SSong Gao if (Vj->E(i) < 0) { \ 2377ac95a0b9SSong Gao break; \ 2378ac95a0b9SSong Gao } \ 2379ac95a0b9SSong Gao } \ 2380ac95a0b9SSong Gao m = Vk->E(0) & MASK; \ 2381ac95a0b9SSong Gao Vd->E(m) = i; \ 2382ac95a0b9SSong Gao } 2383ac95a0b9SSong Gao 2384ac95a0b9SSong Gao VFRSTP(vfrstp_b, 8, 0xf, B) 2385ac95a0b9SSong Gao VFRSTP(vfrstp_h, 16, 0x7, H) 2386ac95a0b9SSong Gao 2387ac95a0b9SSong Gao #define VFRSTPI(NAME, BIT, E) \ 2388329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 2389ac95a0b9SSong Gao { \ 2390ac95a0b9SSong Gao int i, m; \ 2391329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 2392329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 2393ac95a0b9SSong Gao \ 2394ac95a0b9SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2395ac95a0b9SSong Gao if (Vj->E(i) < 0) { \ 2396ac95a0b9SSong Gao break; \ 2397ac95a0b9SSong Gao } \ 2398ac95a0b9SSong Gao } \ 2399ac95a0b9SSong Gao m = imm % (LSX_LEN/BIT); \ 2400ac95a0b9SSong Gao Vd->E(m) = i; \ 2401ac95a0b9SSong Gao } 2402ac95a0b9SSong Gao 2403ac95a0b9SSong Gao VFRSTPI(vfrstpi_b, 8, B) 2404ac95a0b9SSong Gao VFRSTPI(vfrstpi_h, 16, H) 2405aca67472SSong Gao 2406aca67472SSong Gao static void vec_update_fcsr0_mask(CPULoongArchState *env, 2407aca67472SSong Gao uintptr_t pc, int mask) 2408aca67472SSong Gao { 2409aca67472SSong Gao int flags = get_float_exception_flags(&env->fp_status); 2410aca67472SSong Gao 2411aca67472SSong Gao set_float_exception_flags(0, &env->fp_status); 2412aca67472SSong Gao 2413aca67472SSong Gao flags &= ~mask; 2414aca67472SSong Gao 2415aca67472SSong Gao if (flags) { 2416aca67472SSong Gao flags = ieee_ex_to_loongarch(flags); 2417aca67472SSong Gao UPDATE_FP_CAUSE(env->fcsr0, flags); 2418aca67472SSong Gao } 2419aca67472SSong Gao 2420aca67472SSong Gao if (GET_FP_ENABLES(env->fcsr0) & flags) { 2421aca67472SSong Gao do_raise_exception(env, EXCCODE_FPE, pc); 2422aca67472SSong Gao } else { 2423aca67472SSong Gao UPDATE_FP_FLAGS(env->fcsr0, flags); 2424aca67472SSong Gao } 2425aca67472SSong Gao } 2426aca67472SSong Gao 2427aca67472SSong Gao static void vec_update_fcsr0(CPULoongArchState *env, uintptr_t pc) 2428aca67472SSong Gao { 2429aca67472SSong Gao vec_update_fcsr0_mask(env, pc, 0); 2430aca67472SSong Gao } 2431aca67472SSong Gao 2432aca67472SSong Gao static inline void vec_clear_cause(CPULoongArchState *env) 2433aca67472SSong Gao { 2434aca67472SSong Gao SET_FP_CAUSE(env->fcsr0, 0); 2435aca67472SSong Gao } 2436aca67472SSong Gao 2437aca67472SSong Gao #define DO_3OP_F(NAME, BIT, E, FN) \ 24383b286753SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, \ 24393b286753SSong Gao CPULoongArchState *env, uint32_t desc) \ 2440aca67472SSong Gao { \ 2441aca67472SSong Gao int i; \ 24423b286753SSong Gao VReg *Vd = (VReg *)vd; \ 24433b286753SSong Gao VReg *Vj = (VReg *)vj; \ 24443b286753SSong Gao VReg *Vk = (VReg *)vk; \ 2445aca67472SSong Gao \ 2446aca67472SSong Gao vec_clear_cause(env); \ 2447aca67472SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2448aca67472SSong Gao Vd->E(i) = FN(Vj->E(i), Vk->E(i), &env->fp_status); \ 2449aca67472SSong Gao vec_update_fcsr0(env, GETPC()); \ 2450aca67472SSong Gao } \ 2451aca67472SSong Gao } 2452aca67472SSong Gao 2453aca67472SSong Gao DO_3OP_F(vfadd_s, 32, UW, float32_add) 2454aca67472SSong Gao DO_3OP_F(vfadd_d, 64, UD, float64_add) 2455aca67472SSong Gao DO_3OP_F(vfsub_s, 32, UW, float32_sub) 2456aca67472SSong Gao DO_3OP_F(vfsub_d, 64, UD, float64_sub) 2457aca67472SSong Gao DO_3OP_F(vfmul_s, 32, UW, float32_mul) 2458aca67472SSong Gao DO_3OP_F(vfmul_d, 64, UD, float64_mul) 2459aca67472SSong Gao DO_3OP_F(vfdiv_s, 32, UW, float32_div) 2460aca67472SSong Gao DO_3OP_F(vfdiv_d, 64, UD, float64_div) 2461aca67472SSong Gao DO_3OP_F(vfmax_s, 32, UW, float32_maxnum) 2462aca67472SSong Gao DO_3OP_F(vfmax_d, 64, UD, float64_maxnum) 2463aca67472SSong Gao DO_3OP_F(vfmin_s, 32, UW, float32_minnum) 2464aca67472SSong Gao DO_3OP_F(vfmin_d, 64, UD, float64_minnum) 2465aca67472SSong Gao DO_3OP_F(vfmaxa_s, 32, UW, float32_maxnummag) 2466aca67472SSong Gao DO_3OP_F(vfmaxa_d, 64, UD, float64_maxnummag) 2467aca67472SSong Gao DO_3OP_F(vfmina_s, 32, UW, float32_minnummag) 2468aca67472SSong Gao DO_3OP_F(vfmina_d, 64, UD, float64_minnummag) 2469aca67472SSong Gao 2470aca67472SSong Gao #define DO_4OP_F(NAME, BIT, E, FN, flags) \ 2471e2600dadSSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, void *va, \ 2472e2600dadSSong Gao CPULoongArchState *env, uint32_t desc) \ 2473aca67472SSong Gao { \ 2474aca67472SSong Gao int i; \ 2475e2600dadSSong Gao VReg *Vd = (VReg *)vd; \ 2476e2600dadSSong Gao VReg *Vj = (VReg *)vj; \ 2477e2600dadSSong Gao VReg *Vk = (VReg *)vk; \ 2478e2600dadSSong Gao VReg *Va = (VReg *)va; \ 2479aca67472SSong Gao \ 2480aca67472SSong Gao vec_clear_cause(env); \ 2481aca67472SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2482aca67472SSong Gao Vd->E(i) = FN(Vj->E(i), Vk->E(i), Va->E(i), flags, &env->fp_status); \ 2483aca67472SSong Gao vec_update_fcsr0(env, GETPC()); \ 2484aca67472SSong Gao } \ 2485aca67472SSong Gao } 2486aca67472SSong Gao 2487aca67472SSong Gao DO_4OP_F(vfmadd_s, 32, UW, float32_muladd, 0) 2488aca67472SSong Gao DO_4OP_F(vfmadd_d, 64, UD, float64_muladd, 0) 2489aca67472SSong Gao DO_4OP_F(vfmsub_s, 32, UW, float32_muladd, float_muladd_negate_c) 2490aca67472SSong Gao DO_4OP_F(vfmsub_d, 64, UD, float64_muladd, float_muladd_negate_c) 2491aca67472SSong Gao DO_4OP_F(vfnmadd_s, 32, UW, float32_muladd, float_muladd_negate_result) 2492aca67472SSong Gao DO_4OP_F(vfnmadd_d, 64, UD, float64_muladd, float_muladd_negate_result) 2493aca67472SSong Gao DO_4OP_F(vfnmsub_s, 32, UW, float32_muladd, 2494aca67472SSong Gao float_muladd_negate_c | float_muladd_negate_result) 2495aca67472SSong Gao DO_4OP_F(vfnmsub_d, 64, UD, float64_muladd, 2496aca67472SSong Gao float_muladd_negate_c | float_muladd_negate_result) 2497aca67472SSong Gao 2498aca67472SSong Gao #define DO_2OP_F(NAME, BIT, E, FN) \ 2499226bf881SSong Gao void HELPER(NAME)(void *vd, void *vj, \ 2500226bf881SSong Gao CPULoongArchState *env, uint32_t desc) \ 2501aca67472SSong Gao { \ 2502aca67472SSong Gao int i; \ 2503226bf881SSong Gao VReg *Vd = (VReg *)vd; \ 2504226bf881SSong Gao VReg *Vj = (VReg *)vj; \ 2505aca67472SSong Gao \ 2506aca67472SSong Gao vec_clear_cause(env); \ 2507aca67472SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2508aca67472SSong Gao Vd->E(i) = FN(env, Vj->E(i)); \ 2509aca67472SSong Gao } \ 2510aca67472SSong Gao } 2511aca67472SSong Gao 2512aca67472SSong Gao #define FLOGB(BIT, T) \ 2513aca67472SSong Gao static T do_flogb_## BIT(CPULoongArchState *env, T fj) \ 2514aca67472SSong Gao { \ 2515aca67472SSong Gao T fp, fd; \ 2516aca67472SSong Gao float_status *status = &env->fp_status; \ 2517aca67472SSong Gao FloatRoundMode old_mode = get_float_rounding_mode(status); \ 2518aca67472SSong Gao \ 2519aca67472SSong Gao set_float_rounding_mode(float_round_down, status); \ 2520aca67472SSong Gao fp = float ## BIT ##_log2(fj, status); \ 2521aca67472SSong Gao fd = float ## BIT ##_round_to_int(fp, status); \ 2522aca67472SSong Gao set_float_rounding_mode(old_mode, status); \ 2523aca67472SSong Gao vec_update_fcsr0_mask(env, GETPC(), float_flag_inexact); \ 2524aca67472SSong Gao return fd; \ 2525aca67472SSong Gao } 2526aca67472SSong Gao 2527aca67472SSong Gao FLOGB(32, uint32_t) 2528aca67472SSong Gao FLOGB(64, uint64_t) 2529aca67472SSong Gao 2530aca67472SSong Gao #define FCLASS(NAME, BIT, E, FN) \ 2531226bf881SSong Gao void HELPER(NAME)(void *vd, void *vj, \ 2532226bf881SSong Gao CPULoongArchState *env, uint32_t desc) \ 2533aca67472SSong Gao { \ 2534aca67472SSong Gao int i; \ 2535226bf881SSong Gao VReg *Vd = (VReg *)vd; \ 2536226bf881SSong Gao VReg *Vj = (VReg *)vj; \ 2537aca67472SSong Gao \ 2538aca67472SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2539aca67472SSong Gao Vd->E(i) = FN(env, Vj->E(i)); \ 2540aca67472SSong Gao } \ 2541aca67472SSong Gao } 2542aca67472SSong Gao 2543aca67472SSong Gao FCLASS(vfclass_s, 32, UW, helper_fclass_s) 2544aca67472SSong Gao FCLASS(vfclass_d, 64, UD, helper_fclass_d) 2545aca67472SSong Gao 2546aca67472SSong Gao #define FSQRT(BIT, T) \ 2547aca67472SSong Gao static T do_fsqrt_## BIT(CPULoongArchState *env, T fj) \ 2548aca67472SSong Gao { \ 2549aca67472SSong Gao T fd; \ 2550aca67472SSong Gao fd = float ## BIT ##_sqrt(fj, &env->fp_status); \ 2551aca67472SSong Gao vec_update_fcsr0(env, GETPC()); \ 2552aca67472SSong Gao return fd; \ 2553aca67472SSong Gao } 2554aca67472SSong Gao 2555aca67472SSong Gao FSQRT(32, uint32_t) 2556aca67472SSong Gao FSQRT(64, uint64_t) 2557aca67472SSong Gao 2558aca67472SSong Gao #define FRECIP(BIT, T) \ 2559aca67472SSong Gao static T do_frecip_## BIT(CPULoongArchState *env, T fj) \ 2560aca67472SSong Gao { \ 2561aca67472SSong Gao T fd; \ 2562aca67472SSong Gao fd = float ## BIT ##_div(float ## BIT ##_one, fj, &env->fp_status); \ 2563aca67472SSong Gao vec_update_fcsr0(env, GETPC()); \ 2564aca67472SSong Gao return fd; \ 2565aca67472SSong Gao } 2566aca67472SSong Gao 2567aca67472SSong Gao FRECIP(32, uint32_t) 2568aca67472SSong Gao FRECIP(64, uint64_t) 2569aca67472SSong Gao 2570aca67472SSong Gao #define FRSQRT(BIT, T) \ 2571aca67472SSong Gao static T do_frsqrt_## BIT(CPULoongArchState *env, T fj) \ 2572aca67472SSong Gao { \ 2573aca67472SSong Gao T fd, fp; \ 2574aca67472SSong Gao fp = float ## BIT ##_sqrt(fj, &env->fp_status); \ 2575aca67472SSong Gao fd = float ## BIT ##_div(float ## BIT ##_one, fp, &env->fp_status); \ 2576aca67472SSong Gao vec_update_fcsr0(env, GETPC()); \ 2577aca67472SSong Gao return fd; \ 2578aca67472SSong Gao } 2579aca67472SSong Gao 2580aca67472SSong Gao FRSQRT(32, uint32_t) 2581aca67472SSong Gao FRSQRT(64, uint64_t) 2582aca67472SSong Gao 2583aca67472SSong Gao DO_2OP_F(vflogb_s, 32, UW, do_flogb_32) 2584aca67472SSong Gao DO_2OP_F(vflogb_d, 64, UD, do_flogb_64) 2585aca67472SSong Gao DO_2OP_F(vfsqrt_s, 32, UW, do_fsqrt_32) 2586aca67472SSong Gao DO_2OP_F(vfsqrt_d, 64, UD, do_fsqrt_64) 2587aca67472SSong Gao DO_2OP_F(vfrecip_s, 32, UW, do_frecip_32) 2588aca67472SSong Gao DO_2OP_F(vfrecip_d, 64, UD, do_frecip_64) 2589aca67472SSong Gao DO_2OP_F(vfrsqrt_s, 32, UW, do_frsqrt_32) 2590aca67472SSong Gao DO_2OP_F(vfrsqrt_d, 64, UD, do_frsqrt_64) 2591399665d2SSong Gao 2592399665d2SSong Gao static uint32_t float16_cvt_float32(uint16_t h, float_status *status) 2593399665d2SSong Gao { 2594399665d2SSong Gao return float16_to_float32(h, true, status); 2595399665d2SSong Gao } 2596399665d2SSong Gao static uint64_t float32_cvt_float64(uint32_t s, float_status *status) 2597399665d2SSong Gao { 2598399665d2SSong Gao return float32_to_float64(s, status); 2599399665d2SSong Gao } 2600399665d2SSong Gao 2601399665d2SSong Gao static uint16_t float32_cvt_float16(uint32_t s, float_status *status) 2602399665d2SSong Gao { 2603399665d2SSong Gao return float32_to_float16(s, true, status); 2604399665d2SSong Gao } 2605399665d2SSong Gao static uint32_t float64_cvt_float32(uint64_t d, float_status *status) 2606399665d2SSong Gao { 2607399665d2SSong Gao return float64_to_float32(d, status); 2608399665d2SSong Gao } 2609399665d2SSong Gao 2610226bf881SSong Gao void HELPER(vfcvtl_s_h)(void *vd, void *vj, 2611226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2612399665d2SSong Gao { 2613399665d2SSong Gao int i; 2614399665d2SSong Gao VReg temp; 2615226bf881SSong Gao VReg *Vd = (VReg *)vd; 2616226bf881SSong Gao VReg *Vj = (VReg *)vj; 2617399665d2SSong Gao 2618399665d2SSong Gao vec_clear_cause(env); 2619399665d2SSong Gao for (i = 0; i < LSX_LEN/32; i++) { 2620399665d2SSong Gao temp.UW(i) = float16_cvt_float32(Vj->UH(i), &env->fp_status); 2621399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2622399665d2SSong Gao } 2623399665d2SSong Gao *Vd = temp; 2624399665d2SSong Gao } 2625399665d2SSong Gao 2626226bf881SSong Gao void HELPER(vfcvtl_d_s)(void *vd, void *vj, 2627226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2628399665d2SSong Gao { 2629399665d2SSong Gao int i; 2630399665d2SSong Gao VReg temp; 2631226bf881SSong Gao VReg *Vd = (VReg *)vd; 2632226bf881SSong Gao VReg *Vj = (VReg *)vj; 2633399665d2SSong Gao 2634399665d2SSong Gao vec_clear_cause(env); 2635399665d2SSong Gao for (i = 0; i < LSX_LEN/64; i++) { 2636399665d2SSong Gao temp.UD(i) = float32_cvt_float64(Vj->UW(i), &env->fp_status); 2637399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2638399665d2SSong Gao } 2639399665d2SSong Gao *Vd = temp; 2640399665d2SSong Gao } 2641399665d2SSong Gao 2642226bf881SSong Gao void HELPER(vfcvth_s_h)(void *vd, void *vj, 2643226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2644399665d2SSong Gao { 2645399665d2SSong Gao int i; 2646399665d2SSong Gao VReg temp; 2647226bf881SSong Gao VReg *Vd = (VReg *)vd; 2648226bf881SSong Gao VReg *Vj = (VReg *)vj; 2649399665d2SSong Gao 2650399665d2SSong Gao vec_clear_cause(env); 2651399665d2SSong Gao for (i = 0; i < LSX_LEN/32; i++) { 2652399665d2SSong Gao temp.UW(i) = float16_cvt_float32(Vj->UH(i + 4), &env->fp_status); 2653399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2654399665d2SSong Gao } 2655399665d2SSong Gao *Vd = temp; 2656399665d2SSong Gao } 2657399665d2SSong Gao 2658226bf881SSong Gao void HELPER(vfcvth_d_s)(void *vd, void *vj, 2659226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2660399665d2SSong Gao { 2661399665d2SSong Gao int i; 2662399665d2SSong Gao VReg temp; 2663226bf881SSong Gao VReg *Vd = (VReg *)vd; 2664226bf881SSong Gao VReg *Vj = (VReg *)vj; 2665399665d2SSong Gao 2666399665d2SSong Gao vec_clear_cause(env); 2667399665d2SSong Gao for (i = 0; i < LSX_LEN/64; i++) { 2668399665d2SSong Gao temp.UD(i) = float32_cvt_float64(Vj->UW(i + 2), &env->fp_status); 2669399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2670399665d2SSong Gao } 2671399665d2SSong Gao *Vd = temp; 2672399665d2SSong Gao } 2673399665d2SSong Gao 26743b286753SSong Gao void HELPER(vfcvt_h_s)(void *vd, void *vj, void *vk, 26753b286753SSong Gao CPULoongArchState *env, uint32_t desc) 2676399665d2SSong Gao { 2677399665d2SSong Gao int i; 2678399665d2SSong Gao VReg temp; 26793b286753SSong Gao VReg *Vd = (VReg *)vd; 26803b286753SSong Gao VReg *Vj = (VReg *)vj; 26813b286753SSong Gao VReg *Vk = (VReg *)vk; 2682399665d2SSong Gao 2683399665d2SSong Gao vec_clear_cause(env); 2684399665d2SSong Gao for(i = 0; i < LSX_LEN/32; i++) { 2685399665d2SSong Gao temp.UH(i + 4) = float32_cvt_float16(Vj->UW(i), &env->fp_status); 2686399665d2SSong Gao temp.UH(i) = float32_cvt_float16(Vk->UW(i), &env->fp_status); 2687399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2688399665d2SSong Gao } 2689399665d2SSong Gao *Vd = temp; 2690399665d2SSong Gao } 2691399665d2SSong Gao 26923b286753SSong Gao void HELPER(vfcvt_s_d)(void *vd, void *vj, void *vk, 26933b286753SSong Gao CPULoongArchState *env, uint32_t desc) 2694399665d2SSong Gao { 2695399665d2SSong Gao int i; 2696399665d2SSong Gao VReg temp; 26973b286753SSong Gao VReg *Vd = (VReg *)vd; 26983b286753SSong Gao VReg *Vj = (VReg *)vj; 26993b286753SSong Gao VReg *Vk = (VReg *)vk; 2700399665d2SSong Gao 2701399665d2SSong Gao vec_clear_cause(env); 2702399665d2SSong Gao for(i = 0; i < LSX_LEN/64; i++) { 2703399665d2SSong Gao temp.UW(i + 2) = float64_cvt_float32(Vj->UD(i), &env->fp_status); 2704399665d2SSong Gao temp.UW(i) = float64_cvt_float32(Vk->UD(i), &env->fp_status); 2705399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2706399665d2SSong Gao } 2707399665d2SSong Gao *Vd = temp; 2708399665d2SSong Gao } 2709399665d2SSong Gao 2710226bf881SSong Gao void HELPER(vfrint_s)(void *vd, void *vj, 2711226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2712399665d2SSong Gao { 2713399665d2SSong Gao int i; 2714226bf881SSong Gao VReg *Vd = (VReg *)vd; 2715226bf881SSong Gao VReg *Vj = (VReg *)vj; 2716399665d2SSong Gao 2717399665d2SSong Gao vec_clear_cause(env); 2718399665d2SSong Gao for (i = 0; i < 4; i++) { 2719399665d2SSong Gao Vd->W(i) = float32_round_to_int(Vj->UW(i), &env->fp_status); 2720399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2721399665d2SSong Gao } 2722399665d2SSong Gao } 2723399665d2SSong Gao 2724226bf881SSong Gao void HELPER(vfrint_d)(void *vd, void *vj, 2725226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2726399665d2SSong Gao { 2727399665d2SSong Gao int i; 2728226bf881SSong Gao VReg *Vd = (VReg *)vd; 2729226bf881SSong Gao VReg *Vj = (VReg *)vj; 2730399665d2SSong Gao 2731399665d2SSong Gao vec_clear_cause(env); 2732399665d2SSong Gao for (i = 0; i < 2; i++) { 2733399665d2SSong Gao Vd->D(i) = float64_round_to_int(Vj->UD(i), &env->fp_status); 2734399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2735399665d2SSong Gao } 2736399665d2SSong Gao } 2737399665d2SSong Gao 2738399665d2SSong Gao #define FCVT_2OP(NAME, BIT, E, MODE) \ 2739226bf881SSong Gao void HELPER(NAME)(void *vd, void *vj, \ 2740226bf881SSong Gao CPULoongArchState *env, uint32_t desc) \ 2741399665d2SSong Gao { \ 2742399665d2SSong Gao int i; \ 2743226bf881SSong Gao VReg *Vd = (VReg *)vd; \ 2744226bf881SSong Gao VReg *Vj = (VReg *)vj; \ 2745399665d2SSong Gao \ 2746399665d2SSong Gao vec_clear_cause(env); \ 2747399665d2SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2748399665d2SSong Gao FloatRoundMode old_mode = get_float_rounding_mode(&env->fp_status); \ 2749399665d2SSong Gao set_float_rounding_mode(MODE, &env->fp_status); \ 2750399665d2SSong Gao Vd->E(i) = float## BIT ## _round_to_int(Vj->E(i), &env->fp_status); \ 2751399665d2SSong Gao set_float_rounding_mode(old_mode, &env->fp_status); \ 2752399665d2SSong Gao vec_update_fcsr0(env, GETPC()); \ 2753399665d2SSong Gao } \ 2754399665d2SSong Gao } 2755399665d2SSong Gao 2756399665d2SSong Gao FCVT_2OP(vfrintrne_s, 32, UW, float_round_nearest_even) 2757399665d2SSong Gao FCVT_2OP(vfrintrne_d, 64, UD, float_round_nearest_even) 2758399665d2SSong Gao FCVT_2OP(vfrintrz_s, 32, UW, float_round_to_zero) 2759399665d2SSong Gao FCVT_2OP(vfrintrz_d, 64, UD, float_round_to_zero) 2760399665d2SSong Gao FCVT_2OP(vfrintrp_s, 32, UW, float_round_up) 2761399665d2SSong Gao FCVT_2OP(vfrintrp_d, 64, UD, float_round_up) 2762399665d2SSong Gao FCVT_2OP(vfrintrm_s, 32, UW, float_round_down) 2763399665d2SSong Gao FCVT_2OP(vfrintrm_d, 64, UD, float_round_down) 2764399665d2SSong Gao 2765399665d2SSong Gao #define FTINT(NAME, FMT1, FMT2, T1, T2, MODE) \ 2766399665d2SSong Gao static T2 do_ftint ## NAME(CPULoongArchState *env, T1 fj) \ 2767399665d2SSong Gao { \ 2768399665d2SSong Gao T2 fd; \ 2769399665d2SSong Gao FloatRoundMode old_mode = get_float_rounding_mode(&env->fp_status); \ 2770399665d2SSong Gao \ 2771399665d2SSong Gao set_float_rounding_mode(MODE, &env->fp_status); \ 2772399665d2SSong Gao fd = do_## FMT1 ##_to_## FMT2(env, fj); \ 2773399665d2SSong Gao set_float_rounding_mode(old_mode, &env->fp_status); \ 2774399665d2SSong Gao return fd; \ 2775399665d2SSong Gao } 2776399665d2SSong Gao 2777399665d2SSong Gao #define DO_FTINT(FMT1, FMT2, T1, T2) \ 2778399665d2SSong Gao static T2 do_## FMT1 ##_to_## FMT2(CPULoongArchState *env, T1 fj) \ 2779399665d2SSong Gao { \ 2780399665d2SSong Gao T2 fd; \ 2781399665d2SSong Gao \ 2782399665d2SSong Gao fd = FMT1 ##_to_## FMT2(fj, &env->fp_status); \ 2783399665d2SSong Gao if (get_float_exception_flags(&env->fp_status) & (float_flag_invalid)) { \ 2784399665d2SSong Gao if (FMT1 ##_is_any_nan(fj)) { \ 2785399665d2SSong Gao fd = 0; \ 2786399665d2SSong Gao } \ 2787399665d2SSong Gao } \ 2788399665d2SSong Gao vec_update_fcsr0(env, GETPC()); \ 2789399665d2SSong Gao return fd; \ 2790399665d2SSong Gao } 2791399665d2SSong Gao 2792399665d2SSong Gao DO_FTINT(float32, int32, uint32_t, uint32_t) 2793399665d2SSong Gao DO_FTINT(float64, int64, uint64_t, uint64_t) 2794399665d2SSong Gao DO_FTINT(float32, uint32, uint32_t, uint32_t) 2795399665d2SSong Gao DO_FTINT(float64, uint64, uint64_t, uint64_t) 2796399665d2SSong Gao DO_FTINT(float64, int32, uint64_t, uint32_t) 2797399665d2SSong Gao DO_FTINT(float32, int64, uint32_t, uint64_t) 2798399665d2SSong Gao 2799399665d2SSong Gao FTINT(rne_w_s, float32, int32, uint32_t, uint32_t, float_round_nearest_even) 2800399665d2SSong Gao FTINT(rne_l_d, float64, int64, uint64_t, uint64_t, float_round_nearest_even) 2801399665d2SSong Gao FTINT(rp_w_s, float32, int32, uint32_t, uint32_t, float_round_up) 2802399665d2SSong Gao FTINT(rp_l_d, float64, int64, uint64_t, uint64_t, float_round_up) 2803399665d2SSong Gao FTINT(rz_w_s, float32, int32, uint32_t, uint32_t, float_round_to_zero) 2804399665d2SSong Gao FTINT(rz_l_d, float64, int64, uint64_t, uint64_t, float_round_to_zero) 2805399665d2SSong Gao FTINT(rm_w_s, float32, int32, uint32_t, uint32_t, float_round_down) 2806399665d2SSong Gao FTINT(rm_l_d, float64, int64, uint64_t, uint64_t, float_round_down) 2807399665d2SSong Gao 2808399665d2SSong Gao DO_2OP_F(vftintrne_w_s, 32, UW, do_ftintrne_w_s) 2809399665d2SSong Gao DO_2OP_F(vftintrne_l_d, 64, UD, do_ftintrne_l_d) 2810399665d2SSong Gao DO_2OP_F(vftintrp_w_s, 32, UW, do_ftintrp_w_s) 2811399665d2SSong Gao DO_2OP_F(vftintrp_l_d, 64, UD, do_ftintrp_l_d) 2812399665d2SSong Gao DO_2OP_F(vftintrz_w_s, 32, UW, do_ftintrz_w_s) 2813399665d2SSong Gao DO_2OP_F(vftintrz_l_d, 64, UD, do_ftintrz_l_d) 2814399665d2SSong Gao DO_2OP_F(vftintrm_w_s, 32, UW, do_ftintrm_w_s) 2815399665d2SSong Gao DO_2OP_F(vftintrm_l_d, 64, UD, do_ftintrm_l_d) 2816399665d2SSong Gao DO_2OP_F(vftint_w_s, 32, UW, do_float32_to_int32) 2817399665d2SSong Gao DO_2OP_F(vftint_l_d, 64, UD, do_float64_to_int64) 2818399665d2SSong Gao 2819399665d2SSong Gao FTINT(rz_wu_s, float32, uint32, uint32_t, uint32_t, float_round_to_zero) 2820399665d2SSong Gao FTINT(rz_lu_d, float64, uint64, uint64_t, uint64_t, float_round_to_zero) 2821399665d2SSong Gao 2822399665d2SSong Gao DO_2OP_F(vftintrz_wu_s, 32, UW, do_ftintrz_wu_s) 2823399665d2SSong Gao DO_2OP_F(vftintrz_lu_d, 64, UD, do_ftintrz_lu_d) 2824399665d2SSong Gao DO_2OP_F(vftint_wu_s, 32, UW, do_float32_to_uint32) 2825399665d2SSong Gao DO_2OP_F(vftint_lu_d, 64, UD, do_float64_to_uint64) 2826399665d2SSong Gao 2827399665d2SSong Gao FTINT(rm_w_d, float64, int32, uint64_t, uint32_t, float_round_down) 2828399665d2SSong Gao FTINT(rp_w_d, float64, int32, uint64_t, uint32_t, float_round_up) 2829399665d2SSong Gao FTINT(rz_w_d, float64, int32, uint64_t, uint32_t, float_round_to_zero) 2830399665d2SSong Gao FTINT(rne_w_d, float64, int32, uint64_t, uint32_t, float_round_nearest_even) 2831399665d2SSong Gao 2832399665d2SSong Gao #define FTINT_W_D(NAME, FN) \ 28333b286753SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, \ 28343b286753SSong Gao CPULoongArchState *env, uint32_t desc) \ 2835399665d2SSong Gao { \ 2836399665d2SSong Gao int i; \ 2837399665d2SSong Gao VReg temp; \ 28383b286753SSong Gao VReg *Vd = (VReg *)vd; \ 28393b286753SSong Gao VReg *Vj = (VReg *)vj; \ 28403b286753SSong Gao VReg *Vk = (VReg *)vk; \ 2841399665d2SSong Gao \ 2842399665d2SSong Gao vec_clear_cause(env); \ 2843399665d2SSong Gao for (i = 0; i < 2; i++) { \ 2844399665d2SSong Gao temp.W(i + 2) = FN(env, Vj->UD(i)); \ 2845399665d2SSong Gao temp.W(i) = FN(env, Vk->UD(i)); \ 2846399665d2SSong Gao } \ 2847399665d2SSong Gao *Vd = temp; \ 2848399665d2SSong Gao } 2849399665d2SSong Gao 2850399665d2SSong Gao FTINT_W_D(vftint_w_d, do_float64_to_int32) 2851399665d2SSong Gao FTINT_W_D(vftintrm_w_d, do_ftintrm_w_d) 2852399665d2SSong Gao FTINT_W_D(vftintrp_w_d, do_ftintrp_w_d) 2853399665d2SSong Gao FTINT_W_D(vftintrz_w_d, do_ftintrz_w_d) 2854399665d2SSong Gao FTINT_W_D(vftintrne_w_d, do_ftintrne_w_d) 2855399665d2SSong Gao 2856399665d2SSong Gao FTINT(rml_l_s, float32, int64, uint32_t, uint64_t, float_round_down) 2857399665d2SSong Gao FTINT(rpl_l_s, float32, int64, uint32_t, uint64_t, float_round_up) 2858399665d2SSong Gao FTINT(rzl_l_s, float32, int64, uint32_t, uint64_t, float_round_to_zero) 2859399665d2SSong Gao FTINT(rnel_l_s, float32, int64, uint32_t, uint64_t, float_round_nearest_even) 2860399665d2SSong Gao FTINT(rmh_l_s, float32, int64, uint32_t, uint64_t, float_round_down) 2861399665d2SSong Gao FTINT(rph_l_s, float32, int64, uint32_t, uint64_t, float_round_up) 2862399665d2SSong Gao FTINT(rzh_l_s, float32, int64, uint32_t, uint64_t, float_round_to_zero) 2863399665d2SSong Gao FTINT(rneh_l_s, float32, int64, uint32_t, uint64_t, float_round_nearest_even) 2864399665d2SSong Gao 2865399665d2SSong Gao #define FTINTL_L_S(NAME, FN) \ 2866226bf881SSong Gao void HELPER(NAME)(void *vd, void *vj, \ 2867226bf881SSong Gao CPULoongArchState *env, uint32_t desc) \ 2868399665d2SSong Gao { \ 2869399665d2SSong Gao int i; \ 2870399665d2SSong Gao VReg temp; \ 2871226bf881SSong Gao VReg *Vd = (VReg *)vd; \ 2872226bf881SSong Gao VReg *Vj = (VReg *)vj; \ 2873399665d2SSong Gao \ 2874399665d2SSong Gao vec_clear_cause(env); \ 2875399665d2SSong Gao for (i = 0; i < 2; i++) { \ 2876399665d2SSong Gao temp.D(i) = FN(env, Vj->UW(i)); \ 2877399665d2SSong Gao } \ 2878399665d2SSong Gao *Vd = temp; \ 2879399665d2SSong Gao } 2880399665d2SSong Gao 2881399665d2SSong Gao FTINTL_L_S(vftintl_l_s, do_float32_to_int64) 2882399665d2SSong Gao FTINTL_L_S(vftintrml_l_s, do_ftintrml_l_s) 2883399665d2SSong Gao FTINTL_L_S(vftintrpl_l_s, do_ftintrpl_l_s) 2884399665d2SSong Gao FTINTL_L_S(vftintrzl_l_s, do_ftintrzl_l_s) 2885399665d2SSong Gao FTINTL_L_S(vftintrnel_l_s, do_ftintrnel_l_s) 2886399665d2SSong Gao 2887399665d2SSong Gao #define FTINTH_L_S(NAME, FN) \ 2888226bf881SSong Gao void HELPER(NAME)(void *vd, void *vj, \ 2889226bf881SSong Gao CPULoongArchState *env, uint32_t desc) \ 2890399665d2SSong Gao { \ 2891399665d2SSong Gao int i; \ 2892399665d2SSong Gao VReg temp; \ 2893226bf881SSong Gao VReg *Vd = (VReg *)vd; \ 2894226bf881SSong Gao VReg *Vj = (VReg *)vj; \ 2895399665d2SSong Gao \ 2896399665d2SSong Gao vec_clear_cause(env); \ 2897399665d2SSong Gao for (i = 0; i < 2; i++) { \ 2898399665d2SSong Gao temp.D(i) = FN(env, Vj->UW(i + 2)); \ 2899399665d2SSong Gao } \ 2900399665d2SSong Gao *Vd = temp; \ 2901399665d2SSong Gao } 2902399665d2SSong Gao 2903399665d2SSong Gao FTINTH_L_S(vftinth_l_s, do_float32_to_int64) 2904399665d2SSong Gao FTINTH_L_S(vftintrmh_l_s, do_ftintrmh_l_s) 2905399665d2SSong Gao FTINTH_L_S(vftintrph_l_s, do_ftintrph_l_s) 2906399665d2SSong Gao FTINTH_L_S(vftintrzh_l_s, do_ftintrzh_l_s) 2907399665d2SSong Gao FTINTH_L_S(vftintrneh_l_s, do_ftintrneh_l_s) 2908399665d2SSong Gao 2909399665d2SSong Gao #define FFINT(NAME, FMT1, FMT2, T1, T2) \ 2910399665d2SSong Gao static T2 do_ffint_ ## NAME(CPULoongArchState *env, T1 fj) \ 2911399665d2SSong Gao { \ 2912399665d2SSong Gao T2 fd; \ 2913399665d2SSong Gao \ 2914399665d2SSong Gao fd = FMT1 ##_to_## FMT2(fj, &env->fp_status); \ 2915399665d2SSong Gao vec_update_fcsr0(env, GETPC()); \ 2916399665d2SSong Gao return fd; \ 2917399665d2SSong Gao } 2918399665d2SSong Gao 2919399665d2SSong Gao FFINT(s_w, int32, float32, int32_t, uint32_t) 2920399665d2SSong Gao FFINT(d_l, int64, float64, int64_t, uint64_t) 2921399665d2SSong Gao FFINT(s_wu, uint32, float32, uint32_t, uint32_t) 2922399665d2SSong Gao FFINT(d_lu, uint64, float64, uint64_t, uint64_t) 2923399665d2SSong Gao 2924399665d2SSong Gao DO_2OP_F(vffint_s_w, 32, W, do_ffint_s_w) 2925399665d2SSong Gao DO_2OP_F(vffint_d_l, 64, D, do_ffint_d_l) 2926399665d2SSong Gao DO_2OP_F(vffint_s_wu, 32, UW, do_ffint_s_wu) 2927399665d2SSong Gao DO_2OP_F(vffint_d_lu, 64, UD, do_ffint_d_lu) 2928399665d2SSong Gao 2929226bf881SSong Gao void HELPER(vffintl_d_w)(void *vd, void *vj, 2930226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2931399665d2SSong Gao { 2932399665d2SSong Gao int i; 2933399665d2SSong Gao VReg temp; 2934226bf881SSong Gao VReg *Vd = (VReg *)vd; 2935226bf881SSong Gao VReg *Vj = (VReg *)vj; 2936399665d2SSong Gao 2937399665d2SSong Gao vec_clear_cause(env); 2938399665d2SSong Gao for (i = 0; i < 2; i++) { 2939399665d2SSong Gao temp.D(i) = int32_to_float64(Vj->W(i), &env->fp_status); 2940399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2941399665d2SSong Gao } 2942399665d2SSong Gao *Vd = temp; 2943399665d2SSong Gao } 2944399665d2SSong Gao 2945226bf881SSong Gao void HELPER(vffinth_d_w)(void *vd, void *vj, 2946226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2947399665d2SSong Gao { 2948399665d2SSong Gao int i; 2949399665d2SSong Gao VReg temp; 2950226bf881SSong Gao VReg *Vd = (VReg *)vd; 2951226bf881SSong Gao VReg *Vj = (VReg *)vj; 2952399665d2SSong Gao 2953399665d2SSong Gao vec_clear_cause(env); 2954399665d2SSong Gao for (i = 0; i < 2; i++) { 2955399665d2SSong Gao temp.D(i) = int32_to_float64(Vj->W(i + 2), &env->fp_status); 2956399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2957399665d2SSong Gao } 2958399665d2SSong Gao *Vd = temp; 2959399665d2SSong Gao } 2960399665d2SSong Gao 29613b286753SSong Gao void HELPER(vffint_s_l)(void *vd, void *vj, void *vk, 29623b286753SSong Gao CPULoongArchState *env, uint32_t desc) 2963399665d2SSong Gao { 2964399665d2SSong Gao int i; 2965399665d2SSong Gao VReg temp; 29663b286753SSong Gao VReg *Vd = (VReg *)vd; 29673b286753SSong Gao VReg *Vj = (VReg *)vj; 29683b286753SSong Gao VReg *Vk = (VReg *)vk; 2969399665d2SSong Gao 2970399665d2SSong Gao vec_clear_cause(env); 2971399665d2SSong Gao for (i = 0; i < 2; i++) { 2972399665d2SSong Gao temp.W(i + 2) = int64_to_float32(Vj->D(i), &env->fp_status); 2973399665d2SSong Gao temp.W(i) = int64_to_float32(Vk->D(i), &env->fp_status); 2974399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2975399665d2SSong Gao } 2976399665d2SSong Gao *Vd = temp; 2977399665d2SSong Gao } 2978f435e1e5SSong Gao 2979f435e1e5SSong Gao #define VSEQ(a, b) (a == b ? -1 : 0) 2980f435e1e5SSong Gao #define VSLE(a, b) (a <= b ? -1 : 0) 2981f435e1e5SSong Gao #define VSLT(a, b) (a < b ? -1 : 0) 2982f435e1e5SSong Gao 2983f435e1e5SSong Gao #define VCMPI(NAME, BIT, E, DO_OP) \ 2984f435e1e5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t v) \ 2985f435e1e5SSong Gao { \ 2986f435e1e5SSong Gao int i; \ 2987f435e1e5SSong Gao VReg *Vd = (VReg *)vd; \ 2988f435e1e5SSong Gao VReg *Vj = (VReg *)vj; \ 2989f435e1e5SSong Gao typedef __typeof(Vd->E(0)) TD; \ 2990f435e1e5SSong Gao \ 2991f435e1e5SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2992f435e1e5SSong Gao Vd->E(i) = DO_OP(Vj->E(i), (TD)imm); \ 2993f435e1e5SSong Gao } \ 2994f435e1e5SSong Gao } 2995f435e1e5SSong Gao 2996f435e1e5SSong Gao VCMPI(vseqi_b, 8, B, VSEQ) 2997f435e1e5SSong Gao VCMPI(vseqi_h, 16, H, VSEQ) 2998f435e1e5SSong Gao VCMPI(vseqi_w, 32, W, VSEQ) 2999f435e1e5SSong Gao VCMPI(vseqi_d, 64, D, VSEQ) 3000f435e1e5SSong Gao VCMPI(vslei_b, 8, B, VSLE) 3001f435e1e5SSong Gao VCMPI(vslei_h, 16, H, VSLE) 3002f435e1e5SSong Gao VCMPI(vslei_w, 32, W, VSLE) 3003f435e1e5SSong Gao VCMPI(vslei_d, 64, D, VSLE) 3004f435e1e5SSong Gao VCMPI(vslei_bu, 8, UB, VSLE) 3005f435e1e5SSong Gao VCMPI(vslei_hu, 16, UH, VSLE) 3006f435e1e5SSong Gao VCMPI(vslei_wu, 32, UW, VSLE) 3007f435e1e5SSong Gao VCMPI(vslei_du, 64, UD, VSLE) 3008f435e1e5SSong Gao VCMPI(vslti_b, 8, B, VSLT) 3009f435e1e5SSong Gao VCMPI(vslti_h, 16, H, VSLT) 3010f435e1e5SSong Gao VCMPI(vslti_w, 32, W, VSLT) 3011f435e1e5SSong Gao VCMPI(vslti_d, 64, D, VSLT) 3012f435e1e5SSong Gao VCMPI(vslti_bu, 8, UB, VSLT) 3013f435e1e5SSong Gao VCMPI(vslti_hu, 16, UH, VSLT) 3014f435e1e5SSong Gao VCMPI(vslti_wu, 32, UW, VSLT) 3015f435e1e5SSong Gao VCMPI(vslti_du, 64, UD, VSLT) 3016386c4e86SSong Gao 3017386c4e86SSong Gao static uint64_t vfcmp_common(CPULoongArchState *env, 3018386c4e86SSong Gao FloatRelation cmp, uint32_t flags) 3019386c4e86SSong Gao { 3020386c4e86SSong Gao uint64_t ret = 0; 3021386c4e86SSong Gao 3022386c4e86SSong Gao switch (cmp) { 3023386c4e86SSong Gao case float_relation_less: 3024386c4e86SSong Gao ret = (flags & FCMP_LT); 3025386c4e86SSong Gao break; 3026386c4e86SSong Gao case float_relation_equal: 3027386c4e86SSong Gao ret = (flags & FCMP_EQ); 3028386c4e86SSong Gao break; 3029386c4e86SSong Gao case float_relation_greater: 3030386c4e86SSong Gao ret = (flags & FCMP_GT); 3031386c4e86SSong Gao break; 3032386c4e86SSong Gao case float_relation_unordered: 3033386c4e86SSong Gao ret = (flags & FCMP_UN); 3034386c4e86SSong Gao break; 3035386c4e86SSong Gao default: 3036386c4e86SSong Gao g_assert_not_reached(); 3037386c4e86SSong Gao } 3038386c4e86SSong Gao 3039386c4e86SSong Gao if (ret) { 3040386c4e86SSong Gao ret = -1; 3041386c4e86SSong Gao } 3042386c4e86SSong Gao 3043386c4e86SSong Gao return ret; 3044386c4e86SSong Gao } 3045386c4e86SSong Gao 3046386c4e86SSong Gao #define VFCMP(NAME, BIT, E, FN) \ 3047386c4e86SSong Gao void HELPER(NAME)(CPULoongArchState *env, \ 3048386c4e86SSong Gao uint32_t vd, uint32_t vj, uint32_t vk, uint32_t flags) \ 3049386c4e86SSong Gao { \ 3050386c4e86SSong Gao int i; \ 3051386c4e86SSong Gao VReg t; \ 3052386c4e86SSong Gao VReg *Vd = &(env->fpr[vd].vreg); \ 3053386c4e86SSong Gao VReg *Vj = &(env->fpr[vj].vreg); \ 3054386c4e86SSong Gao VReg *Vk = &(env->fpr[vk].vreg); \ 3055386c4e86SSong Gao \ 3056386c4e86SSong Gao vec_clear_cause(env); \ 3057386c4e86SSong Gao for (i = 0; i < LSX_LEN/BIT ; i++) { \ 3058386c4e86SSong Gao FloatRelation cmp; \ 3059386c4e86SSong Gao cmp = FN(Vj->E(i), Vk->E(i), &env->fp_status); \ 3060386c4e86SSong Gao t.E(i) = vfcmp_common(env, cmp, flags); \ 3061386c4e86SSong Gao vec_update_fcsr0(env, GETPC()); \ 3062386c4e86SSong Gao } \ 3063386c4e86SSong Gao *Vd = t; \ 3064386c4e86SSong Gao } 3065386c4e86SSong Gao 3066386c4e86SSong Gao VFCMP(vfcmp_c_s, 32, UW, float32_compare_quiet) 3067386c4e86SSong Gao VFCMP(vfcmp_s_s, 32, UW, float32_compare) 3068386c4e86SSong Gao VFCMP(vfcmp_c_d, 64, UD, float64_compare_quiet) 3069386c4e86SSong Gao VFCMP(vfcmp_s_d, 64, UD, float64_compare) 3070d0dfa19aSSong Gao 3071d0dfa19aSSong Gao void HELPER(vbitseli_b)(void *vd, void *vj, uint64_t imm, uint32_t v) 3072d0dfa19aSSong Gao { 3073d0dfa19aSSong Gao int i; 3074d0dfa19aSSong Gao VReg *Vd = (VReg *)vd; 3075d0dfa19aSSong Gao VReg *Vj = (VReg *)vj; 3076d0dfa19aSSong Gao 3077d0dfa19aSSong Gao for (i = 0; i < 16; i++) { 3078d0dfa19aSSong Gao Vd->B(i) = (~Vd->B(i) & Vj->B(i)) | (Vd->B(i) & imm); 3079d0dfa19aSSong Gao } 3080d0dfa19aSSong Gao } 3081d0dfa19aSSong Gao 3082d0dfa19aSSong Gao /* Copy from target/arm/tcg/sve_helper.c */ 3083d0dfa19aSSong Gao static inline bool do_match2(uint64_t n, uint64_t m0, uint64_t m1, int esz) 3084d0dfa19aSSong Gao { 3085d0dfa19aSSong Gao uint64_t bits = 8 << esz; 3086d0dfa19aSSong Gao uint64_t ones = dup_const(esz, 1); 3087d0dfa19aSSong Gao uint64_t signs = ones << (bits - 1); 3088d0dfa19aSSong Gao uint64_t cmp0, cmp1; 3089d0dfa19aSSong Gao 3090d0dfa19aSSong Gao cmp1 = dup_const(esz, n); 3091d0dfa19aSSong Gao cmp0 = cmp1 ^ m0; 3092d0dfa19aSSong Gao cmp1 = cmp1 ^ m1; 3093d0dfa19aSSong Gao cmp0 = (cmp0 - ones) & ~cmp0; 3094d0dfa19aSSong Gao cmp1 = (cmp1 - ones) & ~cmp1; 3095d0dfa19aSSong Gao return (cmp0 | cmp1) & signs; 3096d0dfa19aSSong Gao } 3097d0dfa19aSSong Gao 3098d0dfa19aSSong Gao #define SETANYEQZ(NAME, MO) \ 3099d0dfa19aSSong Gao void HELPER(NAME)(CPULoongArchState *env, uint32_t cd, uint32_t vj) \ 3100d0dfa19aSSong Gao { \ 3101d0dfa19aSSong Gao VReg *Vj = &(env->fpr[vj].vreg); \ 3102d0dfa19aSSong Gao \ 3103d0dfa19aSSong Gao env->cf[cd & 0x7] = do_match2(0, Vj->D(0), Vj->D(1), MO); \ 3104d0dfa19aSSong Gao } 3105d0dfa19aSSong Gao SETANYEQZ(vsetanyeqz_b, MO_8) 3106d0dfa19aSSong Gao SETANYEQZ(vsetanyeqz_h, MO_16) 3107d0dfa19aSSong Gao SETANYEQZ(vsetanyeqz_w, MO_32) 3108d0dfa19aSSong Gao SETANYEQZ(vsetanyeqz_d, MO_64) 3109d0dfa19aSSong Gao 3110d0dfa19aSSong Gao #define SETALLNEZ(NAME, MO) \ 3111d0dfa19aSSong Gao void HELPER(NAME)(CPULoongArchState *env, uint32_t cd, uint32_t vj) \ 3112d0dfa19aSSong Gao { \ 3113d0dfa19aSSong Gao VReg *Vj = &(env->fpr[vj].vreg); \ 3114d0dfa19aSSong Gao \ 3115d0dfa19aSSong Gao env->cf[cd & 0x7]= !do_match2(0, Vj->D(0), Vj->D(1), MO); \ 3116d0dfa19aSSong Gao } 3117d0dfa19aSSong Gao SETALLNEZ(vsetallnez_b, MO_8) 3118d0dfa19aSSong Gao SETALLNEZ(vsetallnez_h, MO_16) 3119d0dfa19aSSong Gao SETALLNEZ(vsetallnez_w, MO_32) 3120d0dfa19aSSong Gao SETALLNEZ(vsetallnez_d, MO_64) 3121d5e5563cSSong Gao 3122d5e5563cSSong Gao #define VPACKEV(NAME, BIT, E) \ 312304711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 3124d5e5563cSSong Gao { \ 3125d5e5563cSSong Gao int i; \ 3126d5e5563cSSong Gao VReg temp; \ 312704711da1SSong Gao VReg *Vd = (VReg *)vd; \ 312804711da1SSong Gao VReg *Vj = (VReg *)vj; \ 312904711da1SSong Gao VReg *Vk = (VReg *)vk; \ 3130d5e5563cSSong Gao \ 3131d5e5563cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 3132d5e5563cSSong Gao temp.E(2 * i + 1) = Vj->E(2 * i); \ 3133d5e5563cSSong Gao temp.E(2 *i) = Vk->E(2 * i); \ 3134d5e5563cSSong Gao } \ 3135d5e5563cSSong Gao *Vd = temp; \ 3136d5e5563cSSong Gao } 3137d5e5563cSSong Gao 3138d5e5563cSSong Gao VPACKEV(vpackev_b, 16, B) 3139d5e5563cSSong Gao VPACKEV(vpackev_h, 32, H) 3140d5e5563cSSong Gao VPACKEV(vpackev_w, 64, W) 3141d5e5563cSSong Gao VPACKEV(vpackev_d, 128, D) 3142d5e5563cSSong Gao 3143d5e5563cSSong Gao #define VPACKOD(NAME, BIT, E) \ 314404711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 3145d5e5563cSSong Gao { \ 3146d5e5563cSSong Gao int i; \ 3147d5e5563cSSong Gao VReg temp; \ 314804711da1SSong Gao VReg *Vd = (VReg *)vd; \ 314904711da1SSong Gao VReg *Vj = (VReg *)vj; \ 315004711da1SSong Gao VReg *Vk = (VReg *)vk; \ 3151d5e5563cSSong Gao \ 3152d5e5563cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 3153d5e5563cSSong Gao temp.E(2 * i + 1) = Vj->E(2 * i + 1); \ 3154d5e5563cSSong Gao temp.E(2 * i) = Vk->E(2 * i + 1); \ 3155d5e5563cSSong Gao } \ 3156d5e5563cSSong Gao *Vd = temp; \ 3157d5e5563cSSong Gao } 3158d5e5563cSSong Gao 3159d5e5563cSSong Gao VPACKOD(vpackod_b, 16, B) 3160d5e5563cSSong Gao VPACKOD(vpackod_h, 32, H) 3161d5e5563cSSong Gao VPACKOD(vpackod_w, 64, W) 3162d5e5563cSSong Gao VPACKOD(vpackod_d, 128, D) 3163d5e5563cSSong Gao 3164d5e5563cSSong Gao #define VPICKEV(NAME, BIT, E) \ 316504711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 3166d5e5563cSSong Gao { \ 3167d5e5563cSSong Gao int i; \ 3168d5e5563cSSong Gao VReg temp; \ 316904711da1SSong Gao VReg *Vd = (VReg *)vd; \ 317004711da1SSong Gao VReg *Vj = (VReg *)vj; \ 317104711da1SSong Gao VReg *Vk = (VReg *)vk; \ 3172d5e5563cSSong Gao \ 3173d5e5563cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 3174d5e5563cSSong Gao temp.E(i + LSX_LEN/BIT) = Vj->E(2 * i); \ 3175d5e5563cSSong Gao temp.E(i) = Vk->E(2 * i); \ 3176d5e5563cSSong Gao } \ 3177d5e5563cSSong Gao *Vd = temp; \ 3178d5e5563cSSong Gao } 3179d5e5563cSSong Gao 3180d5e5563cSSong Gao VPICKEV(vpickev_b, 16, B) 3181d5e5563cSSong Gao VPICKEV(vpickev_h, 32, H) 3182d5e5563cSSong Gao VPICKEV(vpickev_w, 64, W) 3183d5e5563cSSong Gao VPICKEV(vpickev_d, 128, D) 3184d5e5563cSSong Gao 3185d5e5563cSSong Gao #define VPICKOD(NAME, BIT, E) \ 318604711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 3187d5e5563cSSong Gao { \ 3188d5e5563cSSong Gao int i; \ 3189d5e5563cSSong Gao VReg temp; \ 319004711da1SSong Gao VReg *Vd = (VReg *)vd; \ 319104711da1SSong Gao VReg *Vj = (VReg *)vj; \ 319204711da1SSong Gao VReg *Vk = (VReg *)vk; \ 3193d5e5563cSSong Gao \ 3194d5e5563cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 3195d5e5563cSSong Gao temp.E(i + LSX_LEN/BIT) = Vj->E(2 * i + 1); \ 3196d5e5563cSSong Gao temp.E(i) = Vk->E(2 * i + 1); \ 3197d5e5563cSSong Gao } \ 3198d5e5563cSSong Gao *Vd = temp; \ 3199d5e5563cSSong Gao } 3200d5e5563cSSong Gao 3201d5e5563cSSong Gao VPICKOD(vpickod_b, 16, B) 3202d5e5563cSSong Gao VPICKOD(vpickod_h, 32, H) 3203d5e5563cSSong Gao VPICKOD(vpickod_w, 64, W) 3204d5e5563cSSong Gao VPICKOD(vpickod_d, 128, D) 3205e93dd431SSong Gao 3206e93dd431SSong Gao #define VILVL(NAME, BIT, E) \ 320704711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 3208e93dd431SSong Gao { \ 3209e93dd431SSong Gao int i; \ 3210e93dd431SSong Gao VReg temp; \ 321104711da1SSong Gao VReg *Vd = (VReg *)vd; \ 321204711da1SSong Gao VReg *Vj = (VReg *)vj; \ 321304711da1SSong Gao VReg *Vk = (VReg *)vk; \ 3214e93dd431SSong Gao \ 3215e93dd431SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 3216e93dd431SSong Gao temp.E(2 * i + 1) = Vj->E(i); \ 3217e93dd431SSong Gao temp.E(2 * i) = Vk->E(i); \ 3218e93dd431SSong Gao } \ 3219e93dd431SSong Gao *Vd = temp; \ 3220e93dd431SSong Gao } 3221e93dd431SSong Gao 3222e93dd431SSong Gao VILVL(vilvl_b, 16, B) 3223e93dd431SSong Gao VILVL(vilvl_h, 32, H) 3224e93dd431SSong Gao VILVL(vilvl_w, 64, W) 3225e93dd431SSong Gao VILVL(vilvl_d, 128, D) 3226e93dd431SSong Gao 3227e93dd431SSong Gao #define VILVH(NAME, BIT, E) \ 322804711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 3229e93dd431SSong Gao { \ 3230e93dd431SSong Gao int i; \ 3231e93dd431SSong Gao VReg temp; \ 323204711da1SSong Gao VReg *Vd = (VReg *)vd; \ 323304711da1SSong Gao VReg *Vj = (VReg *)vj; \ 323404711da1SSong Gao VReg *Vk = (VReg *)vk; \ 3235e93dd431SSong Gao \ 3236e93dd431SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 3237e93dd431SSong Gao temp.E(2 * i + 1) = Vj->E(i + LSX_LEN/BIT); \ 3238e93dd431SSong Gao temp.E(2 * i) = Vk->E(i + LSX_LEN/BIT); \ 3239e93dd431SSong Gao } \ 3240e93dd431SSong Gao *Vd = temp; \ 3241e93dd431SSong Gao } 3242e93dd431SSong Gao 3243e93dd431SSong Gao VILVH(vilvh_b, 16, B) 3244e93dd431SSong Gao VILVH(vilvh_h, 32, H) 3245e93dd431SSong Gao VILVH(vilvh_w, 64, W) 3246e93dd431SSong Gao VILVH(vilvh_d, 128, D) 3247e93dd431SSong Gao 3248eb48ab22SSong Gao void HELPER(vshuf_b)(void *vd, void *vj, void *vk, void *va, uint32_t desc) 3249e93dd431SSong Gao { 3250e93dd431SSong Gao int i, m; 3251e93dd431SSong Gao VReg temp; 3252eb48ab22SSong Gao VReg *Vd = (VReg *)vd; 3253eb48ab22SSong Gao VReg *Vj = (VReg *)vj; 3254eb48ab22SSong Gao VReg *Vk = (VReg *)vk; 3255eb48ab22SSong Gao VReg *Va = (VReg *)va; 3256e93dd431SSong Gao 3257e93dd431SSong Gao m = LSX_LEN/8; 3258e93dd431SSong Gao for (i = 0; i < m ; i++) { 3259e93dd431SSong Gao uint64_t k = (uint8_t)Va->B(i) % (2 * m); 3260e93dd431SSong Gao temp.B(i) = k < m ? Vk->B(k) : Vj->B(k - m); 3261e93dd431SSong Gao } 3262e93dd431SSong Gao *Vd = temp; 3263e93dd431SSong Gao } 3264e93dd431SSong Gao 3265e93dd431SSong Gao #define VSHUF(NAME, BIT, E) \ 326604711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 3267e93dd431SSong Gao { \ 3268e93dd431SSong Gao int i, m; \ 3269e93dd431SSong Gao VReg temp; \ 327004711da1SSong Gao VReg *Vd = (VReg *)vd; \ 327104711da1SSong Gao VReg *Vj = (VReg *)vj; \ 327204711da1SSong Gao VReg *Vk = (VReg *)vk; \ 3273e93dd431SSong Gao \ 3274e93dd431SSong Gao m = LSX_LEN/BIT; \ 3275e93dd431SSong Gao for (i = 0; i < m; i++) { \ 3276e93dd431SSong Gao uint64_t k = ((uint8_t) Vd->E(i)) % (2 * m); \ 3277e93dd431SSong Gao temp.E(i) = k < m ? Vk->E(k) : Vj->E(k - m); \ 3278e93dd431SSong Gao } \ 3279e93dd431SSong Gao *Vd = temp; \ 3280e93dd431SSong Gao } 3281e93dd431SSong Gao 3282e93dd431SSong Gao VSHUF(vshuf_h, 16, H) 3283e93dd431SSong Gao VSHUF(vshuf_w, 32, W) 3284e93dd431SSong Gao VSHUF(vshuf_d, 64, D) 3285e93dd431SSong Gao 3286e93dd431SSong Gao #define VSHUF4I(NAME, BIT, E) \ 3287329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 3288e93dd431SSong Gao { \ 3289e93dd431SSong Gao int i; \ 3290e93dd431SSong Gao VReg temp; \ 3291329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 3292329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 3293e93dd431SSong Gao \ 3294e93dd431SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 3295e93dd431SSong Gao temp.E(i) = Vj->E(((i) & 0xfc) + (((imm) >> \ 3296e93dd431SSong Gao (2 * ((i) & 0x03))) & 0x03)); \ 3297e93dd431SSong Gao } \ 3298e93dd431SSong Gao *Vd = temp; \ 3299e93dd431SSong Gao } 3300e93dd431SSong Gao 3301e93dd431SSong Gao VSHUF4I(vshuf4i_b, 8, B) 3302e93dd431SSong Gao VSHUF4I(vshuf4i_h, 16, H) 3303e93dd431SSong Gao VSHUF4I(vshuf4i_w, 32, W) 3304e93dd431SSong Gao 3305329517d5SSong Gao void HELPER(vshuf4i_d)(void *vd, void *vj, uint64_t imm, uint32_t desc) 3306e93dd431SSong Gao { 3307329517d5SSong Gao VReg *Vd = (VReg *)vd; 3308329517d5SSong Gao VReg *Vj = (VReg *)vj; 3309e93dd431SSong Gao 3310e93dd431SSong Gao VReg temp; 3311e93dd431SSong Gao temp.D(0) = (imm & 2 ? Vj : Vd)->D(imm & 1); 3312e93dd431SSong Gao temp.D(1) = (imm & 8 ? Vj : Vd)->D((imm >> 2) & 1); 3313e93dd431SSong Gao *Vd = temp; 3314e93dd431SSong Gao } 3315e93dd431SSong Gao 3316329517d5SSong Gao void HELPER(vpermi_w)(void *vd, void *vj, uint64_t imm, uint32_t desc) 3317e93dd431SSong Gao { 3318e93dd431SSong Gao VReg temp; 3319329517d5SSong Gao VReg *Vd = (VReg *)vd; 3320329517d5SSong Gao VReg *Vj = (VReg *)vj; 3321e93dd431SSong Gao 3322e93dd431SSong Gao temp.W(0) = Vj->W(imm & 0x3); 3323e93dd431SSong Gao temp.W(1) = Vj->W((imm >> 2) & 0x3); 3324e93dd431SSong Gao temp.W(2) = Vd->W((imm >> 4) & 0x3); 3325e93dd431SSong Gao temp.W(3) = Vd->W((imm >> 6) & 0x3); 3326e93dd431SSong Gao *Vd = temp; 3327e93dd431SSong Gao } 3328e93dd431SSong Gao 3329e93dd431SSong Gao #define VEXTRINS(NAME, BIT, E, MASK) \ 3330329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 3331e93dd431SSong Gao { \ 3332e93dd431SSong Gao int ins, extr; \ 3333329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 3334329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 3335e93dd431SSong Gao \ 3336e93dd431SSong Gao ins = (imm >> 4) & MASK; \ 3337e93dd431SSong Gao extr = imm & MASK; \ 3338e93dd431SSong Gao Vd->E(ins) = Vj->E(extr); \ 3339e93dd431SSong Gao } 3340e93dd431SSong Gao 3341e93dd431SSong Gao VEXTRINS(vextrins_b, 8, B, 0xf) 3342e93dd431SSong Gao VEXTRINS(vextrins_h, 16, H, 0x7) 3343e93dd431SSong Gao VEXTRINS(vextrins_w, 32, W, 0x3) 3344e93dd431SSong Gao VEXTRINS(vextrins_d, 64, D, 0x1) 3345