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) \ 451cd1c49adSSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t v) \ 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; \ 458cd1c49adSSong Gao \ 459cd1c49adSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 460cd1c49adSSong Gao Vd->E2(i) = ((T)Vj->E2(i)) * ((T)Vk->E2(i)) >> BIT; \ 461cd1c49adSSong Gao } \ 462cd1c49adSSong Gao } 463cd1c49adSSong Gao 464cd1c49adSSong Gao void HELPER(vmuh_d)(void *vd, void *vj, void *vk, uint32_t v) 465cd1c49adSSong Gao { 466cd1c49adSSong Gao uint64_t l, h1, h2; 467cd1c49adSSong Gao VReg *Vd = (VReg *)vd; 468cd1c49adSSong Gao VReg *Vj = (VReg *)vj; 469cd1c49adSSong Gao VReg *Vk = (VReg *)vk; 470cd1c49adSSong Gao 471cd1c49adSSong Gao muls64(&l, &h1, Vj->D(0), Vk->D(0)); 472cd1c49adSSong Gao muls64(&l, &h2, Vj->D(1), Vk->D(1)); 473cd1c49adSSong Gao 474cd1c49adSSong Gao Vd->D(0) = h1; 475cd1c49adSSong Gao Vd->D(1) = h2; 476cd1c49adSSong Gao } 477cd1c49adSSong Gao 478cd1c49adSSong Gao DO_VMUH(vmuh_b, 8, H, B, DO_MUH) 479cd1c49adSSong Gao DO_VMUH(vmuh_h, 16, W, H, DO_MUH) 480cd1c49adSSong Gao DO_VMUH(vmuh_w, 32, D, W, DO_MUH) 481cd1c49adSSong Gao 482cd1c49adSSong Gao void HELPER(vmuh_du)(void *vd, void *vj, void *vk, uint32_t v) 483cd1c49adSSong Gao { 484cd1c49adSSong Gao uint64_t l, h1, h2; 485cd1c49adSSong Gao VReg *Vd = (VReg *)vd; 486cd1c49adSSong Gao VReg *Vj = (VReg *)vj; 487cd1c49adSSong Gao VReg *Vk = (VReg *)vk; 488cd1c49adSSong Gao 489cd1c49adSSong Gao mulu64(&l, &h1, Vj->D(0), Vk->D(0)); 490cd1c49adSSong Gao mulu64(&l, &h2, Vj->D(1), Vk->D(1)); 491cd1c49adSSong Gao 492cd1c49adSSong Gao Vd->D(0) = h1; 493cd1c49adSSong Gao Vd->D(1) = h2; 494cd1c49adSSong Gao } 495cd1c49adSSong Gao 496cd1c49adSSong Gao DO_VMUH(vmuh_bu, 8, UH, UB, DO_MUH) 497cd1c49adSSong Gao DO_VMUH(vmuh_hu, 16, UW, UH, DO_MUH) 498cd1c49adSSong Gao DO_VMUH(vmuh_wu, 32, UD, UW, DO_MUH) 499cd1c49adSSong Gao 500cd1c49adSSong Gao #define DO_MUL(a, b) (a * b) 501cd1c49adSSong Gao 502cd1c49adSSong Gao DO_EVEN(vmulwev_h_b, 16, H, B, DO_MUL) 503cd1c49adSSong Gao DO_EVEN(vmulwev_w_h, 32, W, H, DO_MUL) 504cd1c49adSSong Gao DO_EVEN(vmulwev_d_w, 64, D, W, DO_MUL) 505cd1c49adSSong Gao 506cd1c49adSSong Gao DO_ODD(vmulwod_h_b, 16, H, B, DO_MUL) 507cd1c49adSSong Gao DO_ODD(vmulwod_w_h, 32, W, H, DO_MUL) 508cd1c49adSSong Gao DO_ODD(vmulwod_d_w, 64, D, W, DO_MUL) 509cd1c49adSSong Gao 510cd1c49adSSong Gao DO_EVEN(vmulwev_h_bu, 16, UH, UB, DO_MUL) 511cd1c49adSSong Gao DO_EVEN(vmulwev_w_hu, 32, UW, UH, DO_MUL) 512cd1c49adSSong Gao DO_EVEN(vmulwev_d_wu, 64, UD, UW, DO_MUL) 513cd1c49adSSong Gao 514cd1c49adSSong Gao DO_ODD(vmulwod_h_bu, 16, UH, UB, DO_MUL) 515cd1c49adSSong Gao DO_ODD(vmulwod_w_hu, 32, UW, UH, DO_MUL) 516cd1c49adSSong Gao DO_ODD(vmulwod_d_wu, 64, UD, UW, DO_MUL) 517cd1c49adSSong Gao 518cd1c49adSSong Gao DO_EVEN_U_S(vmulwev_h_bu_b, 16, H, UH, B, UB, DO_MUL) 519cd1c49adSSong Gao DO_EVEN_U_S(vmulwev_w_hu_h, 32, W, UW, H, UH, DO_MUL) 520cd1c49adSSong Gao DO_EVEN_U_S(vmulwev_d_wu_w, 64, D, UD, W, UW, DO_MUL) 521cd1c49adSSong Gao 522cd1c49adSSong Gao DO_ODD_U_S(vmulwod_h_bu_b, 16, H, UH, B, UB, DO_MUL) 523cd1c49adSSong Gao DO_ODD_U_S(vmulwod_w_hu_h, 32, W, UW, H, UH, DO_MUL) 524cd1c49adSSong Gao DO_ODD_U_S(vmulwod_d_wu_w, 64, D, UD, W, UW, DO_MUL) 525d3aec65bSSong Gao 526d3aec65bSSong Gao #define DO_MADD(a, b, c) (a + b * c) 527d3aec65bSSong Gao #define DO_MSUB(a, b, c) (a - b * c) 528d3aec65bSSong Gao 529d3aec65bSSong Gao #define VMADDSUB(NAME, BIT, E, DO_OP) \ 530d3aec65bSSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t v) \ 531d3aec65bSSong Gao { \ 532d3aec65bSSong Gao int i; \ 533d3aec65bSSong Gao VReg *Vd = (VReg *)vd; \ 534d3aec65bSSong Gao VReg *Vj = (VReg *)vj; \ 535d3aec65bSSong Gao VReg *Vk = (VReg *)vk; \ 536d3aec65bSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 537d3aec65bSSong Gao Vd->E(i) = DO_OP(Vd->E(i), Vj->E(i) ,Vk->E(i)); \ 538d3aec65bSSong Gao } \ 539d3aec65bSSong Gao } 540d3aec65bSSong Gao 541d3aec65bSSong Gao VMADDSUB(vmadd_b, 8, B, DO_MADD) 542d3aec65bSSong Gao VMADDSUB(vmadd_h, 16, H, DO_MADD) 543d3aec65bSSong Gao VMADDSUB(vmadd_w, 32, W, DO_MADD) 544d3aec65bSSong Gao VMADDSUB(vmadd_d, 64, D, DO_MADD) 545d3aec65bSSong Gao VMADDSUB(vmsub_b, 8, B, DO_MSUB) 546d3aec65bSSong Gao VMADDSUB(vmsub_h, 16, H, DO_MSUB) 547d3aec65bSSong Gao VMADDSUB(vmsub_w, 32, W, DO_MSUB) 548d3aec65bSSong Gao VMADDSUB(vmsub_d, 64, D, DO_MSUB) 549d3aec65bSSong Gao 550d3aec65bSSong Gao #define VMADDWEV(NAME, BIT, E1, E2, DO_OP) \ 551d3aec65bSSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t v) \ 552d3aec65bSSong Gao { \ 553d3aec65bSSong Gao int i; \ 554d3aec65bSSong Gao VReg *Vd = (VReg *)vd; \ 555d3aec65bSSong Gao VReg *Vj = (VReg *)vj; \ 556d3aec65bSSong Gao VReg *Vk = (VReg *)vk; \ 557d3aec65bSSong Gao typedef __typeof(Vd->E1(0)) TD; \ 558d3aec65bSSong Gao \ 559d3aec65bSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 560d3aec65bSSong Gao Vd->E1(i) += DO_OP((TD)Vj->E2(2 * i), (TD)Vk->E2(2 * i)); \ 561d3aec65bSSong Gao } \ 562d3aec65bSSong Gao } 563d3aec65bSSong Gao 564d3aec65bSSong Gao VMADDWEV(vmaddwev_h_b, 16, H, B, DO_MUL) 565d3aec65bSSong Gao VMADDWEV(vmaddwev_w_h, 32, W, H, DO_MUL) 566d3aec65bSSong Gao VMADDWEV(vmaddwev_d_w, 64, D, W, DO_MUL) 567d3aec65bSSong Gao VMADDWEV(vmaddwev_h_bu, 16, UH, UB, DO_MUL) 568d3aec65bSSong Gao VMADDWEV(vmaddwev_w_hu, 32, UW, UH, DO_MUL) 569d3aec65bSSong Gao VMADDWEV(vmaddwev_d_wu, 64, UD, UW, DO_MUL) 570d3aec65bSSong Gao 571d3aec65bSSong Gao #define VMADDWOD(NAME, BIT, E1, E2, DO_OP) \ 572d3aec65bSSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t v) \ 573d3aec65bSSong Gao { \ 574d3aec65bSSong Gao int i; \ 575d3aec65bSSong Gao VReg *Vd = (VReg *)vd; \ 576d3aec65bSSong Gao VReg *Vj = (VReg *)vj; \ 577d3aec65bSSong Gao VReg *Vk = (VReg *)vk; \ 578d3aec65bSSong Gao typedef __typeof(Vd->E1(0)) TD; \ 579d3aec65bSSong Gao \ 580d3aec65bSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 581d3aec65bSSong Gao Vd->E1(i) += DO_OP((TD)Vj->E2(2 * i + 1), \ 582d3aec65bSSong Gao (TD)Vk->E2(2 * i + 1)); \ 583d3aec65bSSong Gao } \ 584d3aec65bSSong Gao } 585d3aec65bSSong Gao 586d3aec65bSSong Gao VMADDWOD(vmaddwod_h_b, 16, H, B, DO_MUL) 587d3aec65bSSong Gao VMADDWOD(vmaddwod_w_h, 32, W, H, DO_MUL) 588d3aec65bSSong Gao VMADDWOD(vmaddwod_d_w, 64, D, W, DO_MUL) 589d3aec65bSSong Gao VMADDWOD(vmaddwod_h_bu, 16, UH, UB, DO_MUL) 590d3aec65bSSong Gao VMADDWOD(vmaddwod_w_hu, 32, UW, UH, DO_MUL) 591d3aec65bSSong Gao VMADDWOD(vmaddwod_d_wu, 64, UD, UW, DO_MUL) 592d3aec65bSSong Gao 593d3aec65bSSong Gao #define VMADDWEV_U_S(NAME, BIT, ES1, EU1, ES2, EU2, DO_OP) \ 594d3aec65bSSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t v) \ 595d3aec65bSSong Gao { \ 596d3aec65bSSong Gao int i; \ 597d3aec65bSSong Gao VReg *Vd = (VReg *)vd; \ 598d3aec65bSSong Gao VReg *Vj = (VReg *)vj; \ 599d3aec65bSSong Gao VReg *Vk = (VReg *)vk; \ 600d3aec65bSSong Gao typedef __typeof(Vd->ES1(0)) TS1; \ 601d3aec65bSSong Gao typedef __typeof(Vd->EU1(0)) TU1; \ 602d3aec65bSSong Gao \ 603d3aec65bSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 604d3aec65bSSong Gao Vd->ES1(i) += DO_OP((TU1)Vj->EU2(2 * i), \ 605d3aec65bSSong Gao (TS1)Vk->ES2(2 * i)); \ 606d3aec65bSSong Gao } \ 607d3aec65bSSong Gao } 608d3aec65bSSong Gao 609d3aec65bSSong Gao VMADDWEV_U_S(vmaddwev_h_bu_b, 16, H, UH, B, UB, DO_MUL) 610d3aec65bSSong Gao VMADDWEV_U_S(vmaddwev_w_hu_h, 32, W, UW, H, UH, DO_MUL) 611d3aec65bSSong Gao VMADDWEV_U_S(vmaddwev_d_wu_w, 64, D, UD, W, UW, DO_MUL) 612d3aec65bSSong Gao 613d3aec65bSSong Gao #define VMADDWOD_U_S(NAME, BIT, ES1, EU1, ES2, EU2, DO_OP) \ 614d3aec65bSSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t v) \ 615d3aec65bSSong Gao { \ 616d3aec65bSSong Gao int i; \ 617d3aec65bSSong Gao VReg *Vd = (VReg *)vd; \ 618d3aec65bSSong Gao VReg *Vj = (VReg *)vj; \ 619d3aec65bSSong Gao VReg *Vk = (VReg *)vk; \ 620d3aec65bSSong Gao typedef __typeof(Vd->ES1(0)) TS1; \ 621d3aec65bSSong Gao typedef __typeof(Vd->EU1(0)) TU1; \ 622d3aec65bSSong Gao \ 623d3aec65bSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 624d3aec65bSSong Gao Vd->ES1(i) += DO_OP((TU1)Vj->EU2(2 * i + 1), \ 625d3aec65bSSong Gao (TS1)Vk->ES2(2 * i + 1)); \ 626d3aec65bSSong Gao } \ 627d3aec65bSSong Gao } 628d3aec65bSSong Gao 629d3aec65bSSong Gao VMADDWOD_U_S(vmaddwod_h_bu_b, 16, H, UH, B, UB, DO_MUL) 630d3aec65bSSong Gao VMADDWOD_U_S(vmaddwod_w_hu_h, 32, W, UW, H, UH, DO_MUL) 631d3aec65bSSong Gao VMADDWOD_U_S(vmaddwod_d_wu_w, 64, D, UD, W, UW, DO_MUL) 6324cc4c0f7SSong Gao 6334cc4c0f7SSong Gao #define DO_DIVU(N, M) (unlikely(M == 0) ? 0 : N / M) 6344cc4c0f7SSong Gao #define DO_REMU(N, M) (unlikely(M == 0) ? 0 : N % M) 6354cc4c0f7SSong Gao #define DO_DIV(N, M) (unlikely(M == 0) ? 0 :\ 6364cc4c0f7SSong Gao unlikely((N == -N) && (M == (__typeof(N))(-1))) ? N : N / M) 6374cc4c0f7SSong Gao #define DO_REM(N, M) (unlikely(M == 0) ? 0 :\ 6384cc4c0f7SSong Gao unlikely((N == -N) && (M == (__typeof(N))(-1))) ? 0 : N % M) 6394cc4c0f7SSong Gao 6404cc4c0f7SSong Gao #define VDIV(NAME, BIT, E, DO_OP) \ 64104711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 6424cc4c0f7SSong Gao { \ 6434cc4c0f7SSong Gao int i; \ 64404711da1SSong Gao VReg *Vd = (VReg *)vd; \ 64504711da1SSong Gao VReg *Vj = (VReg *)vj; \ 64604711da1SSong Gao VReg *Vk = (VReg *)vk; \ 6474cc4c0f7SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 6484cc4c0f7SSong Gao Vd->E(i) = DO_OP(Vj->E(i), Vk->E(i)); \ 6494cc4c0f7SSong Gao } \ 6504cc4c0f7SSong Gao } 6514cc4c0f7SSong Gao 6524cc4c0f7SSong Gao VDIV(vdiv_b, 8, B, DO_DIV) 6534cc4c0f7SSong Gao VDIV(vdiv_h, 16, H, DO_DIV) 6544cc4c0f7SSong Gao VDIV(vdiv_w, 32, W, DO_DIV) 6554cc4c0f7SSong Gao VDIV(vdiv_d, 64, D, DO_DIV) 6564cc4c0f7SSong Gao VDIV(vdiv_bu, 8, UB, DO_DIVU) 6574cc4c0f7SSong Gao VDIV(vdiv_hu, 16, UH, DO_DIVU) 6584cc4c0f7SSong Gao VDIV(vdiv_wu, 32, UW, DO_DIVU) 6594cc4c0f7SSong Gao VDIV(vdiv_du, 64, UD, DO_DIVU) 6604cc4c0f7SSong Gao VDIV(vmod_b, 8, B, DO_REM) 6614cc4c0f7SSong Gao VDIV(vmod_h, 16, H, DO_REM) 6624cc4c0f7SSong Gao VDIV(vmod_w, 32, W, DO_REM) 6634cc4c0f7SSong Gao VDIV(vmod_d, 64, D, DO_REM) 6644cc4c0f7SSong Gao VDIV(vmod_bu, 8, UB, DO_REMU) 6654cc4c0f7SSong Gao VDIV(vmod_hu, 16, UH, DO_REMU) 6664cc4c0f7SSong Gao VDIV(vmod_wu, 32, UW, DO_REMU) 6674cc4c0f7SSong Gao VDIV(vmod_du, 64, UD, DO_REMU) 668cbe44190SSong Gao 669cbe44190SSong Gao #define VSAT_S(NAME, BIT, E) \ 670cbe44190SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t max, uint32_t v) \ 671cbe44190SSong Gao { \ 672cbe44190SSong Gao int i; \ 673cbe44190SSong Gao VReg *Vd = (VReg *)vd; \ 674cbe44190SSong Gao VReg *Vj = (VReg *)vj; \ 675cbe44190SSong Gao typedef __typeof(Vd->E(0)) TD; \ 676cbe44190SSong Gao \ 677cbe44190SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 678cbe44190SSong Gao Vd->E(i) = Vj->E(i) > (TD)max ? (TD)max : \ 679cbe44190SSong Gao Vj->E(i) < (TD)~max ? (TD)~max: Vj->E(i); \ 680cbe44190SSong Gao } \ 681cbe44190SSong Gao } 682cbe44190SSong Gao 683cbe44190SSong Gao VSAT_S(vsat_b, 8, B) 684cbe44190SSong Gao VSAT_S(vsat_h, 16, H) 685cbe44190SSong Gao VSAT_S(vsat_w, 32, W) 686cbe44190SSong Gao VSAT_S(vsat_d, 64, D) 687cbe44190SSong Gao 688cbe44190SSong Gao #define VSAT_U(NAME, BIT, E) \ 689cbe44190SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t max, uint32_t v) \ 690cbe44190SSong Gao { \ 691cbe44190SSong Gao int i; \ 692cbe44190SSong Gao VReg *Vd = (VReg *)vd; \ 693cbe44190SSong Gao VReg *Vj = (VReg *)vj; \ 694cbe44190SSong Gao typedef __typeof(Vd->E(0)) TD; \ 695cbe44190SSong Gao \ 696cbe44190SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 697cbe44190SSong Gao Vd->E(i) = Vj->E(i) > (TD)max ? (TD)max : Vj->E(i); \ 698cbe44190SSong Gao } \ 699cbe44190SSong Gao } 700cbe44190SSong Gao 701cbe44190SSong Gao VSAT_U(vsat_bu, 8, UB) 702cbe44190SSong Gao VSAT_U(vsat_hu, 16, UH) 703cbe44190SSong Gao VSAT_U(vsat_wu, 32, UW) 704cbe44190SSong Gao VSAT_U(vsat_du, 64, UD) 7053734ad93SSong Gao 7063734ad93SSong Gao #define VEXTH(NAME, BIT, E1, E2) \ 707ff27e335SSong Gao void HELPER(NAME)(void *vd, void *vj, uint32_t desc) \ 7083734ad93SSong Gao { \ 7093734ad93SSong Gao int i; \ 710ff27e335SSong Gao VReg *Vd = (VReg *)vd; \ 711ff27e335SSong Gao VReg *Vj = (VReg *)vj; \ 7123734ad93SSong Gao \ 7133734ad93SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 7143734ad93SSong Gao Vd->E1(i) = Vj->E2(i + LSX_LEN/BIT); \ 7153734ad93SSong Gao } \ 7163734ad93SSong Gao } 7173734ad93SSong Gao 718ff27e335SSong Gao void HELPER(vexth_q_d)(void *vd, void *vj, uint32_t desc) 7193734ad93SSong Gao { 720ff27e335SSong Gao VReg *Vd = (VReg *)vd; 721ff27e335SSong Gao VReg *Vj = (VReg *)vj; 7223734ad93SSong Gao 7233734ad93SSong Gao Vd->Q(0) = int128_makes64(Vj->D(1)); 7243734ad93SSong Gao } 7253734ad93SSong Gao 726ff27e335SSong Gao void HELPER(vexth_qu_du)(void *vd, void *vj, uint32_t desc) 7273734ad93SSong Gao { 728ff27e335SSong Gao VReg *Vd = (VReg *)vd; 729ff27e335SSong Gao VReg *Vj = (VReg *)vj; 7303734ad93SSong Gao 7313734ad93SSong Gao Vd->Q(0) = int128_make64((uint64_t)Vj->D(1)); 7323734ad93SSong Gao } 7333734ad93SSong Gao 7343734ad93SSong Gao VEXTH(vexth_h_b, 16, H, B) 7353734ad93SSong Gao VEXTH(vexth_w_h, 32, W, H) 7363734ad93SSong Gao VEXTH(vexth_d_w, 64, D, W) 7373734ad93SSong Gao VEXTH(vexth_hu_bu, 16, UH, UB) 7383734ad93SSong Gao VEXTH(vexth_wu_hu, 32, UW, UH) 7393734ad93SSong Gao VEXTH(vexth_du_wu, 64, UD, UW) 740f0e395dfSSong Gao 741f0e395dfSSong Gao #define DO_SIGNCOV(a, b) (a == 0 ? 0 : a < 0 ? -b : b) 742f0e395dfSSong Gao 743f0e395dfSSong Gao DO_3OP(vsigncov_b, 8, B, DO_SIGNCOV) 744f0e395dfSSong Gao DO_3OP(vsigncov_h, 16, H, DO_SIGNCOV) 745f0e395dfSSong Gao DO_3OP(vsigncov_w, 32, W, DO_SIGNCOV) 746f0e395dfSSong Gao DO_3OP(vsigncov_d, 64, D, DO_SIGNCOV) 747789f4a4cSSong Gao 748789f4a4cSSong Gao static uint64_t do_vmskltz_b(int64_t val) 749789f4a4cSSong Gao { 750789f4a4cSSong Gao uint64_t m = 0x8080808080808080ULL; 751789f4a4cSSong Gao uint64_t c = val & m; 752789f4a4cSSong Gao c |= c << 7; 753789f4a4cSSong Gao c |= c << 14; 754789f4a4cSSong Gao c |= c << 28; 755789f4a4cSSong Gao return c >> 56; 756789f4a4cSSong Gao } 757789f4a4cSSong Gao 758ff27e335SSong Gao void HELPER(vmskltz_b)(void *vd, void *vj, uint32_t desc) 759789f4a4cSSong Gao { 760789f4a4cSSong Gao uint16_t temp = 0; 761ff27e335SSong Gao VReg *Vd = (VReg *)vd; 762ff27e335SSong Gao VReg *Vj = (VReg *)vj; 763789f4a4cSSong Gao 764789f4a4cSSong Gao temp = do_vmskltz_b(Vj->D(0)); 765789f4a4cSSong Gao temp |= (do_vmskltz_b(Vj->D(1)) << 8); 766789f4a4cSSong Gao Vd->D(0) = temp; 767789f4a4cSSong Gao Vd->D(1) = 0; 768789f4a4cSSong Gao } 769789f4a4cSSong Gao 770789f4a4cSSong Gao static uint64_t do_vmskltz_h(int64_t val) 771789f4a4cSSong Gao { 772789f4a4cSSong Gao uint64_t m = 0x8000800080008000ULL; 773789f4a4cSSong Gao uint64_t c = val & m; 774789f4a4cSSong Gao c |= c << 15; 775789f4a4cSSong Gao c |= c << 30; 776789f4a4cSSong Gao return c >> 60; 777789f4a4cSSong Gao } 778789f4a4cSSong Gao 779ff27e335SSong Gao void HELPER(vmskltz_h)(void *vd, void *vj, uint32_t desc) 780789f4a4cSSong Gao { 781789f4a4cSSong Gao uint16_t temp = 0; 782ff27e335SSong Gao VReg *Vd = (VReg *)vd; 783ff27e335SSong Gao VReg *Vj = (VReg *)vj; 784789f4a4cSSong Gao 785789f4a4cSSong Gao temp = do_vmskltz_h(Vj->D(0)); 786789f4a4cSSong Gao temp |= (do_vmskltz_h(Vj->D(1)) << 4); 787789f4a4cSSong Gao Vd->D(0) = temp; 788789f4a4cSSong Gao Vd->D(1) = 0; 789789f4a4cSSong Gao } 790789f4a4cSSong Gao 791789f4a4cSSong Gao static uint64_t do_vmskltz_w(int64_t val) 792789f4a4cSSong Gao { 793789f4a4cSSong Gao uint64_t m = 0x8000000080000000ULL; 794789f4a4cSSong Gao uint64_t c = val & m; 795789f4a4cSSong Gao c |= c << 31; 796789f4a4cSSong Gao return c >> 62; 797789f4a4cSSong Gao } 798789f4a4cSSong Gao 799ff27e335SSong Gao void HELPER(vmskltz_w)(void *vd, void *vj, uint32_t desc) 800789f4a4cSSong Gao { 801789f4a4cSSong Gao uint16_t temp = 0; 802ff27e335SSong Gao VReg *Vd = (VReg *)vd; 803ff27e335SSong Gao VReg *Vj = (VReg *)vj; 804789f4a4cSSong Gao 805789f4a4cSSong Gao temp = do_vmskltz_w(Vj->D(0)); 806789f4a4cSSong Gao temp |= (do_vmskltz_w(Vj->D(1)) << 2); 807789f4a4cSSong Gao Vd->D(0) = temp; 808789f4a4cSSong Gao Vd->D(1) = 0; 809789f4a4cSSong Gao } 810789f4a4cSSong Gao 811789f4a4cSSong Gao static uint64_t do_vmskltz_d(int64_t val) 812789f4a4cSSong Gao { 813789f4a4cSSong Gao return (uint64_t)val >> 63; 814789f4a4cSSong Gao } 815ff27e335SSong Gao void HELPER(vmskltz_d)(void *vd, void *vj, uint32_t desc) 816789f4a4cSSong Gao { 817789f4a4cSSong Gao uint16_t temp = 0; 818ff27e335SSong Gao VReg *Vd = (VReg *)vd; 819ff27e335SSong Gao VReg *Vj = (VReg *)vj; 820789f4a4cSSong Gao 821789f4a4cSSong Gao temp = do_vmskltz_d(Vj->D(0)); 822789f4a4cSSong Gao temp |= (do_vmskltz_d(Vj->D(1)) << 1); 823789f4a4cSSong Gao Vd->D(0) = temp; 824789f4a4cSSong Gao Vd->D(1) = 0; 825789f4a4cSSong Gao } 826789f4a4cSSong Gao 827ff27e335SSong Gao void HELPER(vmskgez_b)(void *vd, void *vj, uint32_t desc) 828789f4a4cSSong Gao { 829789f4a4cSSong Gao uint16_t temp = 0; 830ff27e335SSong Gao VReg *Vd = (VReg *)vd; 831ff27e335SSong Gao VReg *Vj = (VReg *)vj; 832789f4a4cSSong Gao 833789f4a4cSSong Gao temp = do_vmskltz_b(Vj->D(0)); 834789f4a4cSSong Gao temp |= (do_vmskltz_b(Vj->D(1)) << 8); 835789f4a4cSSong Gao Vd->D(0) = (uint16_t)(~temp); 836789f4a4cSSong Gao Vd->D(1) = 0; 837789f4a4cSSong Gao } 838789f4a4cSSong Gao 839789f4a4cSSong Gao static uint64_t do_vmskez_b(uint64_t a) 840789f4a4cSSong Gao { 841789f4a4cSSong Gao uint64_t m = 0x7f7f7f7f7f7f7f7fULL; 842789f4a4cSSong Gao uint64_t c = ~(((a & m) + m) | a | m); 843789f4a4cSSong Gao c |= c << 7; 844789f4a4cSSong Gao c |= c << 14; 845789f4a4cSSong Gao c |= c << 28; 846789f4a4cSSong Gao return c >> 56; 847789f4a4cSSong Gao } 848789f4a4cSSong Gao 849ff27e335SSong Gao void HELPER(vmsknz_b)(void *vd, void *vj, uint32_t desc) 850789f4a4cSSong Gao { 851789f4a4cSSong Gao uint16_t temp = 0; 852ff27e335SSong Gao VReg *Vd = (VReg *)vd; 853ff27e335SSong Gao VReg *Vj = (VReg *)vj; 854789f4a4cSSong Gao 855789f4a4cSSong Gao temp = do_vmskez_b(Vj->D(0)); 856789f4a4cSSong Gao temp |= (do_vmskez_b(Vj->D(1)) << 8); 857789f4a4cSSong Gao Vd->D(0) = (uint16_t)(~temp); 858789f4a4cSSong Gao Vd->D(1) = 0; 859789f4a4cSSong Gao } 860f205a539SSong Gao 861f205a539SSong Gao void HELPER(vnori_b)(void *vd, void *vj, uint64_t imm, uint32_t v) 862f205a539SSong Gao { 863f205a539SSong Gao int i; 864f205a539SSong Gao VReg *Vd = (VReg *)vd; 865f205a539SSong Gao VReg *Vj = (VReg *)vj; 866f205a539SSong Gao 867f205a539SSong Gao for (i = 0; i < LSX_LEN/8; i++) { 868f205a539SSong Gao Vd->B(i) = ~(Vj->B(i) | (uint8_t)imm); 869f205a539SSong Gao } 870f205a539SSong Gao } 8719b21a7a5SSong Gao 8729b21a7a5SSong Gao #define VSLLWIL(NAME, BIT, E1, E2) \ 873329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 8749b21a7a5SSong Gao { \ 8759b21a7a5SSong Gao int i; \ 8769b21a7a5SSong Gao VReg temp; \ 877329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 878329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 8799b21a7a5SSong Gao typedef __typeof(temp.E1(0)) TD; \ 8809b21a7a5SSong Gao \ 8819b21a7a5SSong Gao temp.D(0) = 0; \ 8829b21a7a5SSong Gao temp.D(1) = 0; \ 8839b21a7a5SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 8849b21a7a5SSong Gao temp.E1(i) = (TD)Vj->E2(i) << (imm % BIT); \ 8859b21a7a5SSong Gao } \ 8869b21a7a5SSong Gao *Vd = temp; \ 8879b21a7a5SSong Gao } 8889b21a7a5SSong Gao 889ff27e335SSong Gao void HELPER(vextl_q_d)(void *vd, void *vj, uint32_t desc) 8909b21a7a5SSong Gao { 891ff27e335SSong Gao VReg *Vd = (VReg *)vd; 892ff27e335SSong Gao VReg *Vj = (VReg *)vj; 8939b21a7a5SSong Gao 8949b21a7a5SSong Gao Vd->Q(0) = int128_makes64(Vj->D(0)); 8959b21a7a5SSong Gao } 8969b21a7a5SSong Gao 897ff27e335SSong Gao void HELPER(vextl_qu_du)(void *vd, void *vj, uint32_t desc) 8989b21a7a5SSong Gao { 899ff27e335SSong Gao VReg *Vd = (VReg *)vd; 900ff27e335SSong Gao VReg *Vj = (VReg *)vj; 9019b21a7a5SSong Gao 9029b21a7a5SSong Gao Vd->Q(0) = int128_make64(Vj->D(0)); 9039b21a7a5SSong Gao } 9049b21a7a5SSong Gao 9059b21a7a5SSong Gao VSLLWIL(vsllwil_h_b, 16, H, B) 9069b21a7a5SSong Gao VSLLWIL(vsllwil_w_h, 32, W, H) 9079b21a7a5SSong Gao VSLLWIL(vsllwil_d_w, 64, D, W) 9089b21a7a5SSong Gao VSLLWIL(vsllwil_hu_bu, 16, UH, UB) 9099b21a7a5SSong Gao VSLLWIL(vsllwil_wu_hu, 32, UW, UH) 9109b21a7a5SSong Gao VSLLWIL(vsllwil_du_wu, 64, UD, UW) 911ecb93716SSong Gao 912ecb93716SSong Gao #define do_vsrlr(E, T) \ 913ecb93716SSong Gao static T do_vsrlr_ ##E(T s1, int sh) \ 914ecb93716SSong Gao { \ 915ecb93716SSong Gao if (sh == 0) { \ 916ecb93716SSong Gao return s1; \ 917ecb93716SSong Gao } else { \ 918ecb93716SSong Gao return (s1 >> sh) + ((s1 >> (sh - 1)) & 0x1); \ 919ecb93716SSong Gao } \ 920ecb93716SSong Gao } 921ecb93716SSong Gao 922ecb93716SSong Gao do_vsrlr(B, uint8_t) 923ecb93716SSong Gao do_vsrlr(H, uint16_t) 924ecb93716SSong Gao do_vsrlr(W, uint32_t) 925ecb93716SSong Gao do_vsrlr(D, uint64_t) 926ecb93716SSong Gao 927ecb93716SSong Gao #define VSRLR(NAME, BIT, T, E) \ 92804711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 929ecb93716SSong Gao { \ 930ecb93716SSong Gao int i; \ 93104711da1SSong Gao VReg *Vd = (VReg *)vd; \ 93204711da1SSong Gao VReg *Vj = (VReg *)vj; \ 93304711da1SSong Gao VReg *Vk = (VReg *)vk; \ 934ecb93716SSong Gao \ 935ecb93716SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 936ecb93716SSong Gao Vd->E(i) = do_vsrlr_ ## E(Vj->E(i), ((T)Vk->E(i))%BIT); \ 937ecb93716SSong Gao } \ 938ecb93716SSong Gao } 939ecb93716SSong Gao 940ecb93716SSong Gao VSRLR(vsrlr_b, 8, uint8_t, B) 941ecb93716SSong Gao VSRLR(vsrlr_h, 16, uint16_t, H) 942ecb93716SSong Gao VSRLR(vsrlr_w, 32, uint32_t, W) 943ecb93716SSong Gao VSRLR(vsrlr_d, 64, uint64_t, D) 944ecb93716SSong Gao 945ecb93716SSong Gao #define VSRLRI(NAME, BIT, E) \ 946329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 947ecb93716SSong Gao { \ 948ecb93716SSong Gao int i; \ 949329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 950329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 951ecb93716SSong Gao \ 952ecb93716SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 953ecb93716SSong Gao Vd->E(i) = do_vsrlr_ ## E(Vj->E(i), imm); \ 954ecb93716SSong Gao } \ 955ecb93716SSong Gao } 956ecb93716SSong Gao 957ecb93716SSong Gao VSRLRI(vsrlri_b, 8, B) 958ecb93716SSong Gao VSRLRI(vsrlri_h, 16, H) 959ecb93716SSong Gao VSRLRI(vsrlri_w, 32, W) 960ecb93716SSong Gao VSRLRI(vsrlri_d, 64, D) 961ecb93716SSong Gao 962ecb93716SSong Gao #define do_vsrar(E, T) \ 963ecb93716SSong Gao static T do_vsrar_ ##E(T s1, int sh) \ 964ecb93716SSong Gao { \ 965ecb93716SSong Gao if (sh == 0) { \ 966ecb93716SSong Gao return s1; \ 967ecb93716SSong Gao } else { \ 968ecb93716SSong Gao return (s1 >> sh) + ((s1 >> (sh - 1)) & 0x1); \ 969ecb93716SSong Gao } \ 970ecb93716SSong Gao } 971ecb93716SSong Gao 972ecb93716SSong Gao do_vsrar(B, int8_t) 973ecb93716SSong Gao do_vsrar(H, int16_t) 974ecb93716SSong Gao do_vsrar(W, int32_t) 975ecb93716SSong Gao do_vsrar(D, int64_t) 976ecb93716SSong Gao 977ecb93716SSong Gao #define VSRAR(NAME, BIT, T, E) \ 97804711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 979ecb93716SSong Gao { \ 980ecb93716SSong Gao int i; \ 98104711da1SSong Gao VReg *Vd = (VReg *)vd; \ 98204711da1SSong Gao VReg *Vj = (VReg *)vj; \ 98304711da1SSong Gao VReg *Vk = (VReg *)vk; \ 984ecb93716SSong Gao \ 985ecb93716SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 986ecb93716SSong Gao Vd->E(i) = do_vsrar_ ## E(Vj->E(i), ((T)Vk->E(i))%BIT); \ 987ecb93716SSong Gao } \ 988ecb93716SSong Gao } 989ecb93716SSong Gao 990ecb93716SSong Gao VSRAR(vsrar_b, 8, uint8_t, B) 991ecb93716SSong Gao VSRAR(vsrar_h, 16, uint16_t, H) 992ecb93716SSong Gao VSRAR(vsrar_w, 32, uint32_t, W) 993ecb93716SSong Gao VSRAR(vsrar_d, 64, uint64_t, D) 994ecb93716SSong Gao 995ecb93716SSong Gao #define VSRARI(NAME, BIT, E) \ 996329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 997ecb93716SSong Gao { \ 998ecb93716SSong Gao int i; \ 999329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1000329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 1001ecb93716SSong Gao \ 1002ecb93716SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 1003ecb93716SSong Gao Vd->E(i) = do_vsrar_ ## E(Vj->E(i), imm); \ 1004ecb93716SSong Gao } \ 1005ecb93716SSong Gao } 1006ecb93716SSong Gao 1007ecb93716SSong Gao VSRARI(vsrari_b, 8, B) 1008ecb93716SSong Gao VSRARI(vsrari_h, 16, H) 1009ecb93716SSong Gao VSRARI(vsrari_w, 32, W) 1010ecb93716SSong Gao VSRARI(vsrari_d, 64, D) 1011d79fb8ddSSong Gao 1012d79fb8ddSSong Gao #define R_SHIFT(a, b) (a >> b) 1013d79fb8ddSSong Gao 1014d79fb8ddSSong Gao #define VSRLN(NAME, BIT, T, E1, E2) \ 101504711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1016d79fb8ddSSong Gao { \ 1017d79fb8ddSSong Gao int i; \ 101804711da1SSong Gao VReg *Vd = (VReg *)vd; \ 101904711da1SSong Gao VReg *Vj = (VReg *)vj; \ 102004711da1SSong Gao VReg *Vk = (VReg *)vk; \ 1021d79fb8ddSSong Gao \ 1022d79fb8ddSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 1023d79fb8ddSSong Gao Vd->E1(i) = R_SHIFT((T)Vj->E2(i),((T)Vk->E2(i)) % BIT); \ 1024d79fb8ddSSong Gao } \ 1025d79fb8ddSSong Gao Vd->D(1) = 0; \ 1026d79fb8ddSSong Gao } 1027d79fb8ddSSong Gao 1028d79fb8ddSSong Gao VSRLN(vsrln_b_h, 16, uint16_t, B, H) 1029d79fb8ddSSong Gao VSRLN(vsrln_h_w, 32, uint32_t, H, W) 1030d79fb8ddSSong Gao VSRLN(vsrln_w_d, 64, uint64_t, W, D) 1031d79fb8ddSSong Gao 1032d79fb8ddSSong Gao #define VSRAN(NAME, BIT, T, E1, E2) \ 103304711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1034d79fb8ddSSong Gao { \ 1035d79fb8ddSSong Gao int i; \ 103604711da1SSong Gao VReg *Vd = (VReg *)vd; \ 103704711da1SSong Gao VReg *Vj = (VReg *)vj; \ 103804711da1SSong Gao VReg *Vk = (VReg *)vk; \ 1039d79fb8ddSSong Gao \ 1040d79fb8ddSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 1041d79fb8ddSSong Gao Vd->E1(i) = R_SHIFT(Vj->E2(i), ((T)Vk->E2(i)) % BIT); \ 1042d79fb8ddSSong Gao } \ 1043d79fb8ddSSong Gao Vd->D(1) = 0; \ 1044d79fb8ddSSong Gao } 1045d79fb8ddSSong Gao 1046d79fb8ddSSong Gao VSRAN(vsran_b_h, 16, uint16_t, B, H) 1047d79fb8ddSSong Gao VSRAN(vsran_h_w, 32, uint32_t, H, W) 1048d79fb8ddSSong Gao VSRAN(vsran_w_d, 64, uint64_t, W, D) 1049d79fb8ddSSong Gao 1050d79fb8ddSSong Gao #define VSRLNI(NAME, BIT, T, E1, E2) \ 1051329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 1052d79fb8ddSSong Gao { \ 1053d79fb8ddSSong Gao int i, max; \ 1054d79fb8ddSSong Gao VReg temp; \ 1055329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1056329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 1057d79fb8ddSSong Gao \ 1058d79fb8ddSSong Gao temp.D(0) = 0; \ 1059d79fb8ddSSong Gao temp.D(1) = 0; \ 1060d79fb8ddSSong Gao max = LSX_LEN/BIT; \ 1061d79fb8ddSSong Gao for (i = 0; i < max; i++) { \ 1062d79fb8ddSSong Gao temp.E1(i) = R_SHIFT((T)Vj->E2(i), imm); \ 1063d79fb8ddSSong Gao temp.E1(i + max) = R_SHIFT((T)Vd->E2(i), imm); \ 1064d79fb8ddSSong Gao } \ 1065d79fb8ddSSong Gao *Vd = temp; \ 1066d79fb8ddSSong Gao } 1067d79fb8ddSSong Gao 1068329517d5SSong Gao void HELPER(vsrlni_d_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 1069d79fb8ddSSong Gao { 1070d79fb8ddSSong Gao VReg temp; 1071329517d5SSong Gao VReg *Vd = (VReg *)vd; 1072329517d5SSong Gao VReg *Vj = (VReg *)vj; 1073d79fb8ddSSong Gao 1074d79fb8ddSSong Gao temp.D(0) = 0; 1075d79fb8ddSSong Gao temp.D(1) = 0; 1076d79fb8ddSSong Gao temp.D(0) = int128_getlo(int128_urshift(Vj->Q(0), imm % 128)); 1077d79fb8ddSSong Gao temp.D(1) = int128_getlo(int128_urshift(Vd->Q(0), imm % 128)); 1078d79fb8ddSSong Gao *Vd = temp; 1079d79fb8ddSSong Gao } 1080d79fb8ddSSong Gao 1081d79fb8ddSSong Gao VSRLNI(vsrlni_b_h, 16, uint16_t, B, H) 1082d79fb8ddSSong Gao VSRLNI(vsrlni_h_w, 32, uint32_t, H, W) 1083d79fb8ddSSong Gao VSRLNI(vsrlni_w_d, 64, uint64_t, W, D) 1084d79fb8ddSSong Gao 1085d79fb8ddSSong Gao #define VSRANI(NAME, BIT, E1, E2) \ 1086329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 1087d79fb8ddSSong Gao { \ 1088d79fb8ddSSong Gao int i, max; \ 1089d79fb8ddSSong Gao VReg temp; \ 1090329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1091329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 1092d79fb8ddSSong Gao \ 1093d79fb8ddSSong Gao temp.D(0) = 0; \ 1094d79fb8ddSSong Gao temp.D(1) = 0; \ 1095d79fb8ddSSong Gao max = LSX_LEN/BIT; \ 1096d79fb8ddSSong Gao for (i = 0; i < max; i++) { \ 1097d79fb8ddSSong Gao temp.E1(i) = R_SHIFT(Vj->E2(i), imm); \ 1098d79fb8ddSSong Gao temp.E1(i + max) = R_SHIFT(Vd->E2(i), imm); \ 1099d79fb8ddSSong Gao } \ 1100d79fb8ddSSong Gao *Vd = temp; \ 1101d79fb8ddSSong Gao } 1102d79fb8ddSSong Gao 1103329517d5SSong Gao void HELPER(vsrani_d_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 1104d79fb8ddSSong Gao { 1105d79fb8ddSSong Gao VReg temp; 1106329517d5SSong Gao VReg *Vd = (VReg *)vd; 1107329517d5SSong Gao VReg *Vj = (VReg *)vj; 1108d79fb8ddSSong Gao 1109d79fb8ddSSong Gao temp.D(0) = 0; 1110d79fb8ddSSong Gao temp.D(1) = 0; 1111d79fb8ddSSong Gao temp.D(0) = int128_getlo(int128_rshift(Vj->Q(0), imm % 128)); 1112d79fb8ddSSong Gao temp.D(1) = int128_getlo(int128_rshift(Vd->Q(0), imm % 128)); 1113d79fb8ddSSong Gao *Vd = temp; 1114d79fb8ddSSong Gao } 1115d79fb8ddSSong Gao 1116d79fb8ddSSong Gao VSRANI(vsrani_b_h, 16, B, H) 1117d79fb8ddSSong Gao VSRANI(vsrani_h_w, 32, H, W) 1118d79fb8ddSSong Gao VSRANI(vsrani_w_d, 64, W, D) 1119a5200a17SSong Gao 1120a5200a17SSong Gao #define VSRLRN(NAME, BIT, T, E1, E2) \ 112104711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1122a5200a17SSong Gao { \ 1123a5200a17SSong Gao int i; \ 112404711da1SSong Gao VReg *Vd = (VReg *)vd; \ 112504711da1SSong Gao VReg *Vj = (VReg *)vj; \ 112604711da1SSong Gao VReg *Vk = (VReg *)vk; \ 1127a5200a17SSong Gao \ 1128a5200a17SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 1129a5200a17SSong Gao Vd->E1(i) = do_vsrlr_ ## E2(Vj->E2(i), ((T)Vk->E2(i))%BIT); \ 1130a5200a17SSong Gao } \ 1131a5200a17SSong Gao Vd->D(1) = 0; \ 1132a5200a17SSong Gao } 1133a5200a17SSong Gao 1134a5200a17SSong Gao VSRLRN(vsrlrn_b_h, 16, uint16_t, B, H) 1135a5200a17SSong Gao VSRLRN(vsrlrn_h_w, 32, uint32_t, H, W) 1136a5200a17SSong Gao VSRLRN(vsrlrn_w_d, 64, uint64_t, W, D) 1137a5200a17SSong Gao 1138a5200a17SSong Gao #define VSRARN(NAME, BIT, T, E1, E2) \ 113904711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1140a5200a17SSong Gao { \ 1141a5200a17SSong Gao int i; \ 114204711da1SSong Gao VReg *Vd = (VReg *)vd; \ 114304711da1SSong Gao VReg *Vj = (VReg *)vj; \ 114404711da1SSong Gao VReg *Vk = (VReg *)vk; \ 1145a5200a17SSong Gao \ 1146a5200a17SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 1147a5200a17SSong Gao Vd->E1(i) = do_vsrar_ ## E2(Vj->E2(i), ((T)Vk->E2(i))%BIT); \ 1148a5200a17SSong Gao } \ 1149a5200a17SSong Gao Vd->D(1) = 0; \ 1150a5200a17SSong Gao } 1151a5200a17SSong Gao 1152a5200a17SSong Gao VSRARN(vsrarn_b_h, 16, uint8_t, B, H) 1153a5200a17SSong Gao VSRARN(vsrarn_h_w, 32, uint16_t, H, W) 1154a5200a17SSong Gao VSRARN(vsrarn_w_d, 64, uint32_t, W, D) 1155a5200a17SSong Gao 1156a5200a17SSong Gao #define VSRLRNI(NAME, BIT, E1, E2) \ 1157329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 1158a5200a17SSong Gao { \ 1159a5200a17SSong Gao int i, max; \ 1160a5200a17SSong Gao VReg temp; \ 1161329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1162329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 1163a5200a17SSong Gao \ 1164a5200a17SSong Gao temp.D(0) = 0; \ 1165a5200a17SSong Gao temp.D(1) = 0; \ 1166a5200a17SSong Gao max = LSX_LEN/BIT; \ 1167a5200a17SSong Gao for (i = 0; i < max; i++) { \ 1168a5200a17SSong Gao temp.E1(i) = do_vsrlr_ ## E2(Vj->E2(i), imm); \ 1169a5200a17SSong Gao temp.E1(i + max) = do_vsrlr_ ## E2(Vd->E2(i), imm); \ 1170a5200a17SSong Gao } \ 1171a5200a17SSong Gao *Vd = temp; \ 1172a5200a17SSong Gao } 1173a5200a17SSong Gao 1174329517d5SSong Gao void HELPER(vsrlrni_d_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 1175a5200a17SSong Gao { 1176a5200a17SSong Gao VReg temp; 1177329517d5SSong Gao VReg *Vd = (VReg *)vd; 1178329517d5SSong Gao VReg *Vj = (VReg *)vj; 1179a5200a17SSong Gao Int128 r1, r2; 1180a5200a17SSong Gao 1181a5200a17SSong Gao if (imm == 0) { 1182a5200a17SSong Gao temp.D(0) = int128_getlo(Vj->Q(0)); 1183a5200a17SSong Gao temp.D(1) = int128_getlo(Vd->Q(0)); 1184a5200a17SSong Gao } else { 1185a5200a17SSong Gao r1 = int128_and(int128_urshift(Vj->Q(0), (imm -1)), int128_one()); 1186a5200a17SSong Gao r2 = int128_and(int128_urshift(Vd->Q(0), (imm -1)), int128_one()); 1187a5200a17SSong Gao 1188a5200a17SSong Gao temp.D(0) = int128_getlo(int128_add(int128_urshift(Vj->Q(0), imm), r1)); 1189a5200a17SSong Gao temp.D(1) = int128_getlo(int128_add(int128_urshift(Vd->Q(0), imm), r2)); 1190a5200a17SSong Gao } 1191a5200a17SSong Gao *Vd = temp; 1192a5200a17SSong Gao } 1193a5200a17SSong Gao 1194a5200a17SSong Gao VSRLRNI(vsrlrni_b_h, 16, B, H) 1195a5200a17SSong Gao VSRLRNI(vsrlrni_h_w, 32, H, W) 1196a5200a17SSong Gao VSRLRNI(vsrlrni_w_d, 64, W, D) 1197a5200a17SSong Gao 1198a5200a17SSong Gao #define VSRARNI(NAME, BIT, E1, E2) \ 1199329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 1200a5200a17SSong Gao { \ 1201a5200a17SSong Gao int i, max; \ 1202a5200a17SSong Gao VReg temp; \ 1203329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1204329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 1205a5200a17SSong Gao \ 1206a5200a17SSong Gao temp.D(0) = 0; \ 1207a5200a17SSong Gao temp.D(1) = 0; \ 1208a5200a17SSong Gao max = LSX_LEN/BIT; \ 1209a5200a17SSong Gao for (i = 0; i < max; i++) { \ 1210a5200a17SSong Gao temp.E1(i) = do_vsrar_ ## E2(Vj->E2(i), imm); \ 1211a5200a17SSong Gao temp.E1(i + max) = do_vsrar_ ## E2(Vd->E2(i), imm); \ 1212a5200a17SSong Gao } \ 1213a5200a17SSong Gao *Vd = temp; \ 1214a5200a17SSong Gao } 1215a5200a17SSong Gao 1216329517d5SSong Gao void HELPER(vsrarni_d_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 1217a5200a17SSong Gao { 1218a5200a17SSong Gao VReg temp; 1219329517d5SSong Gao VReg *Vd = (VReg *)vd; 1220329517d5SSong Gao VReg *Vj = (VReg *)vj; 1221a5200a17SSong Gao Int128 r1, r2; 1222a5200a17SSong Gao 1223a5200a17SSong Gao if (imm == 0) { 1224a5200a17SSong Gao temp.D(0) = int128_getlo(Vj->Q(0)); 1225a5200a17SSong Gao temp.D(1) = int128_getlo(Vd->Q(0)); 1226a5200a17SSong Gao } else { 1227a5200a17SSong Gao r1 = int128_and(int128_rshift(Vj->Q(0), (imm -1)), int128_one()); 1228a5200a17SSong Gao r2 = int128_and(int128_rshift(Vd->Q(0), (imm -1)), int128_one()); 1229a5200a17SSong Gao 1230a5200a17SSong Gao temp.D(0) = int128_getlo(int128_add(int128_rshift(Vj->Q(0), imm), r1)); 1231a5200a17SSong Gao temp.D(1) = int128_getlo(int128_add(int128_rshift(Vd->Q(0), imm), r2)); 1232a5200a17SSong Gao } 1233a5200a17SSong Gao *Vd = temp; 1234a5200a17SSong Gao } 1235a5200a17SSong Gao 1236a5200a17SSong Gao VSRARNI(vsrarni_b_h, 16, B, H) 1237a5200a17SSong Gao VSRARNI(vsrarni_h_w, 32, H, W) 1238a5200a17SSong Gao VSRARNI(vsrarni_w_d, 64, W, D) 123983b3815dSSong Gao 124083b3815dSSong Gao #define SSRLNS(NAME, T1, T2, T3) \ 124183b3815dSSong Gao static T1 do_ssrlns_ ## NAME(T2 e2, int sa, int sh) \ 124283b3815dSSong Gao { \ 124383b3815dSSong Gao T1 shft_res; \ 124483b3815dSSong Gao if (sa == 0) { \ 124583b3815dSSong Gao shft_res = e2; \ 124683b3815dSSong Gao } else { \ 124783b3815dSSong Gao shft_res = (((T1)e2) >> sa); \ 124883b3815dSSong Gao } \ 124983b3815dSSong Gao T3 mask; \ 125083b3815dSSong Gao mask = (1ull << sh) -1; \ 125183b3815dSSong Gao if (shft_res > mask) { \ 125283b3815dSSong Gao return mask; \ 125383b3815dSSong Gao } else { \ 125483b3815dSSong Gao return shft_res; \ 125583b3815dSSong Gao } \ 125683b3815dSSong Gao } 125783b3815dSSong Gao 125883b3815dSSong Gao SSRLNS(B, uint16_t, int16_t, uint8_t) 125983b3815dSSong Gao SSRLNS(H, uint32_t, int32_t, uint16_t) 126083b3815dSSong Gao SSRLNS(W, uint64_t, int64_t, uint32_t) 126183b3815dSSong Gao 126283b3815dSSong Gao #define VSSRLN(NAME, BIT, T, E1, E2) \ 126304711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 126483b3815dSSong Gao { \ 126583b3815dSSong Gao int i; \ 126604711da1SSong Gao VReg *Vd = (VReg *)vd; \ 126704711da1SSong Gao VReg *Vj = (VReg *)vj; \ 126804711da1SSong Gao VReg *Vk = (VReg *)vk; \ 126983b3815dSSong Gao \ 127083b3815dSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 127183b3815dSSong Gao Vd->E1(i) = do_ssrlns_ ## E1(Vj->E2(i), (T)Vk->E2(i)% BIT, BIT/2 -1); \ 127283b3815dSSong Gao } \ 127383b3815dSSong Gao Vd->D(1) = 0; \ 127483b3815dSSong Gao } 127583b3815dSSong Gao 127683b3815dSSong Gao VSSRLN(vssrln_b_h, 16, uint16_t, B, H) 127783b3815dSSong Gao VSSRLN(vssrln_h_w, 32, uint32_t, H, W) 127883b3815dSSong Gao VSSRLN(vssrln_w_d, 64, uint64_t, W, D) 127983b3815dSSong Gao 128083b3815dSSong Gao #define SSRANS(E, T1, T2) \ 128183b3815dSSong Gao static T1 do_ssrans_ ## E(T1 e2, int sa, int sh) \ 128283b3815dSSong Gao { \ 128383b3815dSSong Gao T1 shft_res; \ 128483b3815dSSong Gao if (sa == 0) { \ 128583b3815dSSong Gao shft_res = e2; \ 128683b3815dSSong Gao } else { \ 128783b3815dSSong Gao shft_res = e2 >> sa; \ 128883b3815dSSong Gao } \ 128983b3815dSSong Gao T2 mask; \ 129083b3815dSSong Gao mask = (1ll << sh) -1; \ 129183b3815dSSong Gao if (shft_res > mask) { \ 129283b3815dSSong Gao return mask; \ 129383b3815dSSong Gao } else if (shft_res < -(mask +1)) { \ 129483b3815dSSong Gao return ~mask; \ 129583b3815dSSong Gao } else { \ 129683b3815dSSong Gao return shft_res; \ 129783b3815dSSong Gao } \ 129883b3815dSSong Gao } 129983b3815dSSong Gao 130083b3815dSSong Gao SSRANS(B, int16_t, int8_t) 130183b3815dSSong Gao SSRANS(H, int32_t, int16_t) 130283b3815dSSong Gao SSRANS(W, int64_t, int32_t) 130383b3815dSSong Gao 130483b3815dSSong Gao #define VSSRAN(NAME, BIT, T, E1, E2) \ 130504711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 130683b3815dSSong Gao { \ 130783b3815dSSong Gao int i; \ 130804711da1SSong Gao VReg *Vd = (VReg *)vd; \ 130904711da1SSong Gao VReg *Vj = (VReg *)vj; \ 131004711da1SSong Gao VReg *Vk = (VReg *)vk; \ 131183b3815dSSong Gao \ 131283b3815dSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 131383b3815dSSong Gao Vd->E1(i) = do_ssrans_ ## E1(Vj->E2(i), (T)Vk->E2(i)%BIT, BIT/2 -1); \ 131483b3815dSSong Gao } \ 131583b3815dSSong Gao Vd->D(1) = 0; \ 131683b3815dSSong Gao } 131783b3815dSSong Gao 131883b3815dSSong Gao VSSRAN(vssran_b_h, 16, uint16_t, B, H) 131983b3815dSSong Gao VSSRAN(vssran_h_w, 32, uint32_t, H, W) 132083b3815dSSong Gao VSSRAN(vssran_w_d, 64, uint64_t, W, D) 132183b3815dSSong Gao 132283b3815dSSong Gao #define SSRLNU(E, T1, T2, T3) \ 132383b3815dSSong Gao static T1 do_ssrlnu_ ## E(T3 e2, int sa, int sh) \ 132483b3815dSSong Gao { \ 132583b3815dSSong Gao T1 shft_res; \ 132683b3815dSSong Gao if (sa == 0) { \ 132783b3815dSSong Gao shft_res = e2; \ 132883b3815dSSong Gao } else { \ 132983b3815dSSong Gao shft_res = (((T1)e2) >> sa); \ 133083b3815dSSong Gao } \ 133183b3815dSSong Gao T2 mask; \ 133283b3815dSSong Gao mask = (1ull << sh) -1; \ 133383b3815dSSong Gao if (shft_res > mask) { \ 133483b3815dSSong Gao return mask; \ 133583b3815dSSong Gao } else { \ 133683b3815dSSong Gao return shft_res; \ 133783b3815dSSong Gao } \ 133883b3815dSSong Gao } 133983b3815dSSong Gao 134083b3815dSSong Gao SSRLNU(B, uint16_t, uint8_t, int16_t) 134183b3815dSSong Gao SSRLNU(H, uint32_t, uint16_t, int32_t) 134283b3815dSSong Gao SSRLNU(W, uint64_t, uint32_t, int64_t) 134383b3815dSSong Gao 134483b3815dSSong Gao #define VSSRLNU(NAME, BIT, T, E1, E2) \ 134504711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 134683b3815dSSong Gao { \ 134783b3815dSSong Gao int i; \ 134804711da1SSong Gao VReg *Vd = (VReg *)vd; \ 134904711da1SSong Gao VReg *Vj = (VReg *)vj; \ 135004711da1SSong Gao VReg *Vk = (VReg *)vk; \ 135183b3815dSSong Gao \ 135283b3815dSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 135383b3815dSSong Gao Vd->E1(i) = do_ssrlnu_ ## E1(Vj->E2(i), (T)Vk->E2(i)%BIT, BIT/2); \ 135483b3815dSSong Gao } \ 135583b3815dSSong Gao Vd->D(1) = 0; \ 135683b3815dSSong Gao } 135783b3815dSSong Gao 135883b3815dSSong Gao VSSRLNU(vssrln_bu_h, 16, uint16_t, B, H) 135983b3815dSSong Gao VSSRLNU(vssrln_hu_w, 32, uint32_t, H, W) 136083b3815dSSong Gao VSSRLNU(vssrln_wu_d, 64, uint64_t, W, D) 136183b3815dSSong Gao 136283b3815dSSong Gao #define SSRANU(E, T1, T2, T3) \ 136383b3815dSSong Gao static T1 do_ssranu_ ## E(T3 e2, int sa, int sh) \ 136483b3815dSSong Gao { \ 136583b3815dSSong Gao T1 shft_res; \ 136683b3815dSSong Gao if (sa == 0) { \ 136783b3815dSSong Gao shft_res = e2; \ 136883b3815dSSong Gao } else { \ 136983b3815dSSong Gao shft_res = e2 >> sa; \ 137083b3815dSSong Gao } \ 137183b3815dSSong Gao if (e2 < 0) { \ 137283b3815dSSong Gao shft_res = 0; \ 137383b3815dSSong Gao } \ 137483b3815dSSong Gao T2 mask; \ 137583b3815dSSong Gao mask = (1ull << sh) -1; \ 137683b3815dSSong Gao if (shft_res > mask) { \ 137783b3815dSSong Gao return mask; \ 137883b3815dSSong Gao } else { \ 137983b3815dSSong Gao return shft_res; \ 138083b3815dSSong Gao } \ 138183b3815dSSong Gao } 138283b3815dSSong Gao 138383b3815dSSong Gao SSRANU(B, uint16_t, uint8_t, int16_t) 138483b3815dSSong Gao SSRANU(H, uint32_t, uint16_t, int32_t) 138583b3815dSSong Gao SSRANU(W, uint64_t, uint32_t, int64_t) 138683b3815dSSong Gao 138783b3815dSSong Gao #define VSSRANU(NAME, BIT, T, E1, E2) \ 138804711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 138983b3815dSSong Gao { \ 139083b3815dSSong Gao int i; \ 139104711da1SSong Gao VReg *Vd = (VReg *)vd; \ 139204711da1SSong Gao VReg *Vj = (VReg *)vj; \ 139304711da1SSong Gao VReg *Vk = (VReg *)vk; \ 139483b3815dSSong Gao \ 139583b3815dSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 139683b3815dSSong Gao Vd->E1(i) = do_ssranu_ ## E1(Vj->E2(i), (T)Vk->E2(i)%BIT, BIT/2); \ 139783b3815dSSong Gao } \ 139883b3815dSSong Gao Vd->D(1) = 0; \ 139983b3815dSSong Gao } 140083b3815dSSong Gao 140183b3815dSSong Gao VSSRANU(vssran_bu_h, 16, uint16_t, B, H) 140283b3815dSSong Gao VSSRANU(vssran_hu_w, 32, uint32_t, H, W) 140383b3815dSSong Gao VSSRANU(vssran_wu_d, 64, uint64_t, W, D) 140483b3815dSSong Gao 140583b3815dSSong Gao #define VSSRLNI(NAME, BIT, E1, E2) \ 1406329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 140783b3815dSSong Gao { \ 140883b3815dSSong Gao int i; \ 140983b3815dSSong Gao VReg temp; \ 1410329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1411329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 141283b3815dSSong Gao \ 141383b3815dSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 141483b3815dSSong Gao temp.E1(i) = do_ssrlns_ ## E1(Vj->E2(i), imm, BIT/2 -1); \ 141583b3815dSSong Gao temp.E1(i + LSX_LEN/BIT) = do_ssrlns_ ## E1(Vd->E2(i), imm, BIT/2 -1);\ 141683b3815dSSong Gao } \ 141783b3815dSSong Gao *Vd = temp; \ 141883b3815dSSong Gao } 141983b3815dSSong Gao 1420329517d5SSong Gao void HELPER(vssrlni_d_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 142183b3815dSSong Gao { 142283b3815dSSong Gao Int128 shft_res1, shft_res2, mask; 1423329517d5SSong Gao VReg *Vd = (VReg *)vd; 1424329517d5SSong Gao VReg *Vj = (VReg *)vj; 142583b3815dSSong Gao 142683b3815dSSong Gao if (imm == 0) { 142783b3815dSSong Gao shft_res1 = Vj->Q(0); 142883b3815dSSong Gao shft_res2 = Vd->Q(0); 142983b3815dSSong Gao } else { 143083b3815dSSong Gao shft_res1 = int128_urshift(Vj->Q(0), imm); 143183b3815dSSong Gao shft_res2 = int128_urshift(Vd->Q(0), imm); 143283b3815dSSong Gao } 143383b3815dSSong Gao mask = int128_sub(int128_lshift(int128_one(), 63), int128_one()); 143483b3815dSSong Gao 143583b3815dSSong Gao if (int128_ult(mask, shft_res1)) { 143683b3815dSSong Gao Vd->D(0) = int128_getlo(mask); 143783b3815dSSong Gao }else { 143883b3815dSSong Gao Vd->D(0) = int128_getlo(shft_res1); 143983b3815dSSong Gao } 144083b3815dSSong Gao 144183b3815dSSong Gao if (int128_ult(mask, shft_res2)) { 144283b3815dSSong Gao Vd->D(1) = int128_getlo(mask); 144383b3815dSSong Gao }else { 144483b3815dSSong Gao Vd->D(1) = int128_getlo(shft_res2); 144583b3815dSSong Gao } 144683b3815dSSong Gao } 144783b3815dSSong Gao 144883b3815dSSong Gao VSSRLNI(vssrlni_b_h, 16, B, H) 144983b3815dSSong Gao VSSRLNI(vssrlni_h_w, 32, H, W) 145083b3815dSSong Gao VSSRLNI(vssrlni_w_d, 64, W, D) 145183b3815dSSong Gao 145283b3815dSSong Gao #define VSSRANI(NAME, BIT, E1, E2) \ 1453329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 145483b3815dSSong Gao { \ 145583b3815dSSong Gao int i; \ 145683b3815dSSong Gao VReg temp; \ 1457329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1458329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 145983b3815dSSong Gao \ 146083b3815dSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 146183b3815dSSong Gao temp.E1(i) = do_ssrans_ ## E1(Vj->E2(i), imm, BIT/2 -1); \ 146283b3815dSSong Gao temp.E1(i + LSX_LEN/BIT) = do_ssrans_ ## E1(Vd->E2(i), imm, BIT/2 -1); \ 146383b3815dSSong Gao } \ 146483b3815dSSong Gao *Vd = temp; \ 146583b3815dSSong Gao } 146683b3815dSSong Gao 1467329517d5SSong Gao void HELPER(vssrani_d_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 146883b3815dSSong Gao { 146983b3815dSSong Gao Int128 shft_res1, shft_res2, mask, min; 1470329517d5SSong Gao VReg *Vd = (VReg *)vd; 1471329517d5SSong Gao VReg *Vj = (VReg *)vj; 147283b3815dSSong Gao 147383b3815dSSong Gao if (imm == 0) { 147483b3815dSSong Gao shft_res1 = Vj->Q(0); 147583b3815dSSong Gao shft_res2 = Vd->Q(0); 147683b3815dSSong Gao } else { 147783b3815dSSong Gao shft_res1 = int128_rshift(Vj->Q(0), imm); 147883b3815dSSong Gao shft_res2 = int128_rshift(Vd->Q(0), imm); 147983b3815dSSong Gao } 148083b3815dSSong Gao mask = int128_sub(int128_lshift(int128_one(), 63), int128_one()); 148183b3815dSSong Gao min = int128_lshift(int128_one(), 63); 148283b3815dSSong Gao 148383b3815dSSong Gao if (int128_gt(shft_res1, mask)) { 148483b3815dSSong Gao Vd->D(0) = int128_getlo(mask); 148583b3815dSSong Gao } else if (int128_lt(shft_res1, int128_neg(min))) { 148683b3815dSSong Gao Vd->D(0) = int128_getlo(min); 148783b3815dSSong Gao } else { 148883b3815dSSong Gao Vd->D(0) = int128_getlo(shft_res1); 148983b3815dSSong Gao } 149083b3815dSSong Gao 149183b3815dSSong Gao if (int128_gt(shft_res2, mask)) { 149283b3815dSSong Gao Vd->D(1) = int128_getlo(mask); 149383b3815dSSong Gao } else if (int128_lt(shft_res2, int128_neg(min))) { 149483b3815dSSong Gao Vd->D(1) = int128_getlo(min); 149583b3815dSSong Gao } else { 149683b3815dSSong Gao Vd->D(1) = int128_getlo(shft_res2); 149783b3815dSSong Gao } 149883b3815dSSong Gao } 149983b3815dSSong Gao 150083b3815dSSong Gao VSSRANI(vssrani_b_h, 16, B, H) 150183b3815dSSong Gao VSSRANI(vssrani_h_w, 32, H, W) 150283b3815dSSong Gao VSSRANI(vssrani_w_d, 64, W, D) 150383b3815dSSong Gao 150483b3815dSSong Gao #define VSSRLNUI(NAME, BIT, E1, E2) \ 1505329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 150683b3815dSSong Gao { \ 150783b3815dSSong Gao int i; \ 150883b3815dSSong Gao VReg temp; \ 1509329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1510329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 151183b3815dSSong Gao \ 151283b3815dSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 151383b3815dSSong Gao temp.E1(i) = do_ssrlnu_ ## E1(Vj->E2(i), imm, BIT/2); \ 151483b3815dSSong Gao temp.E1(i + LSX_LEN/BIT) = do_ssrlnu_ ## E1(Vd->E2(i), imm, BIT/2); \ 151583b3815dSSong Gao } \ 151683b3815dSSong Gao *Vd = temp; \ 151783b3815dSSong Gao } 151883b3815dSSong Gao 1519329517d5SSong Gao void HELPER(vssrlni_du_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 152083b3815dSSong Gao { 152183b3815dSSong Gao Int128 shft_res1, shft_res2, mask; 1522329517d5SSong Gao VReg *Vd = (VReg *)vd; 1523329517d5SSong Gao VReg *Vj = (VReg *)vj; 152483b3815dSSong Gao 152583b3815dSSong Gao if (imm == 0) { 152683b3815dSSong Gao shft_res1 = Vj->Q(0); 152783b3815dSSong Gao shft_res2 = Vd->Q(0); 152883b3815dSSong Gao } else { 152983b3815dSSong Gao shft_res1 = int128_urshift(Vj->Q(0), imm); 153083b3815dSSong Gao shft_res2 = int128_urshift(Vd->Q(0), imm); 153183b3815dSSong Gao } 153283b3815dSSong Gao mask = int128_sub(int128_lshift(int128_one(), 64), int128_one()); 153383b3815dSSong Gao 153483b3815dSSong Gao if (int128_ult(mask, shft_res1)) { 153583b3815dSSong Gao Vd->D(0) = int128_getlo(mask); 153683b3815dSSong Gao }else { 153783b3815dSSong Gao Vd->D(0) = int128_getlo(shft_res1); 153883b3815dSSong Gao } 153983b3815dSSong Gao 154083b3815dSSong Gao if (int128_ult(mask, shft_res2)) { 154183b3815dSSong Gao Vd->D(1) = int128_getlo(mask); 154283b3815dSSong Gao }else { 154383b3815dSSong Gao Vd->D(1) = int128_getlo(shft_res2); 154483b3815dSSong Gao } 154583b3815dSSong Gao } 154683b3815dSSong Gao 154783b3815dSSong Gao VSSRLNUI(vssrlni_bu_h, 16, B, H) 154883b3815dSSong Gao VSSRLNUI(vssrlni_hu_w, 32, H, W) 154983b3815dSSong Gao VSSRLNUI(vssrlni_wu_d, 64, W, D) 155083b3815dSSong Gao 155183b3815dSSong Gao #define VSSRANUI(NAME, BIT, E1, E2) \ 1552329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 155383b3815dSSong Gao { \ 155483b3815dSSong Gao int i; \ 155583b3815dSSong Gao VReg temp; \ 1556329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1557329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 155883b3815dSSong Gao \ 155983b3815dSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 156083b3815dSSong Gao temp.E1(i) = do_ssranu_ ## E1(Vj->E2(i), imm, BIT/2); \ 156183b3815dSSong Gao temp.E1(i + LSX_LEN/BIT) = do_ssranu_ ## E1(Vd->E2(i), imm, BIT/2); \ 156283b3815dSSong Gao } \ 156383b3815dSSong Gao *Vd = temp; \ 156483b3815dSSong Gao } 156583b3815dSSong Gao 1566329517d5SSong Gao void HELPER(vssrani_du_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 156783b3815dSSong Gao { 156883b3815dSSong Gao Int128 shft_res1, shft_res2, mask; 1569329517d5SSong Gao VReg *Vd = (VReg *)vd; 1570329517d5SSong Gao VReg *Vj = (VReg *)vj; 157183b3815dSSong Gao 157283b3815dSSong Gao if (imm == 0) { 157383b3815dSSong Gao shft_res1 = Vj->Q(0); 157483b3815dSSong Gao shft_res2 = Vd->Q(0); 157583b3815dSSong Gao } else { 157683b3815dSSong Gao shft_res1 = int128_rshift(Vj->Q(0), imm); 157783b3815dSSong Gao shft_res2 = int128_rshift(Vd->Q(0), imm); 157883b3815dSSong Gao } 157983b3815dSSong Gao 158083b3815dSSong Gao if (int128_lt(Vj->Q(0), int128_zero())) { 158183b3815dSSong Gao shft_res1 = int128_zero(); 158283b3815dSSong Gao } 158383b3815dSSong Gao 158483b3815dSSong Gao if (int128_lt(Vd->Q(0), int128_zero())) { 158583b3815dSSong Gao shft_res2 = int128_zero(); 158683b3815dSSong Gao } 158783b3815dSSong Gao 158883b3815dSSong Gao mask = int128_sub(int128_lshift(int128_one(), 64), int128_one()); 158983b3815dSSong Gao 159083b3815dSSong Gao if (int128_ult(mask, shft_res1)) { 159183b3815dSSong Gao Vd->D(0) = int128_getlo(mask); 159283b3815dSSong Gao }else { 159383b3815dSSong Gao Vd->D(0) = int128_getlo(shft_res1); 159483b3815dSSong Gao } 159583b3815dSSong Gao 159683b3815dSSong Gao if (int128_ult(mask, shft_res2)) { 159783b3815dSSong Gao Vd->D(1) = int128_getlo(mask); 159883b3815dSSong Gao }else { 159983b3815dSSong Gao Vd->D(1) = int128_getlo(shft_res2); 160083b3815dSSong Gao } 160183b3815dSSong Gao } 160283b3815dSSong Gao 160383b3815dSSong Gao VSSRANUI(vssrani_bu_h, 16, B, H) 160483b3815dSSong Gao VSSRANUI(vssrani_hu_w, 32, H, W) 160583b3815dSSong Gao VSSRANUI(vssrani_wu_d, 64, W, D) 1606162cd32cSSong Gao 1607162cd32cSSong Gao #define SSRLRNS(E1, E2, T1, T2, T3) \ 1608162cd32cSSong Gao static T1 do_ssrlrns_ ## E1(T2 e2, int sa, int sh) \ 1609162cd32cSSong Gao { \ 1610162cd32cSSong Gao T1 shft_res; \ 1611162cd32cSSong Gao \ 1612162cd32cSSong Gao shft_res = do_vsrlr_ ## E2(e2, sa); \ 1613162cd32cSSong Gao T1 mask; \ 1614162cd32cSSong Gao mask = (1ull << sh) -1; \ 1615162cd32cSSong Gao if (shft_res > mask) { \ 1616162cd32cSSong Gao return mask; \ 1617162cd32cSSong Gao } else { \ 1618162cd32cSSong Gao return shft_res; \ 1619162cd32cSSong Gao } \ 1620162cd32cSSong Gao } 1621162cd32cSSong Gao 1622162cd32cSSong Gao SSRLRNS(B, H, uint16_t, int16_t, uint8_t) 1623162cd32cSSong Gao SSRLRNS(H, W, uint32_t, int32_t, uint16_t) 1624162cd32cSSong Gao SSRLRNS(W, D, uint64_t, int64_t, uint32_t) 1625162cd32cSSong Gao 1626162cd32cSSong Gao #define VSSRLRN(NAME, BIT, T, E1, E2) \ 162704711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1628162cd32cSSong Gao { \ 1629162cd32cSSong Gao int i; \ 163004711da1SSong Gao VReg *Vd = (VReg *)vd; \ 163104711da1SSong Gao VReg *Vj = (VReg *)vj; \ 163204711da1SSong Gao VReg *Vk = (VReg *)vk; \ 1633162cd32cSSong Gao \ 1634162cd32cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 1635162cd32cSSong Gao Vd->E1(i) = do_ssrlrns_ ## E1(Vj->E2(i), (T)Vk->E2(i)%BIT, BIT/2 -1); \ 1636162cd32cSSong Gao } \ 1637162cd32cSSong Gao Vd->D(1) = 0; \ 1638162cd32cSSong Gao } 1639162cd32cSSong Gao 1640162cd32cSSong Gao VSSRLRN(vssrlrn_b_h, 16, uint16_t, B, H) 1641162cd32cSSong Gao VSSRLRN(vssrlrn_h_w, 32, uint32_t, H, W) 1642162cd32cSSong Gao VSSRLRN(vssrlrn_w_d, 64, uint64_t, W, D) 1643162cd32cSSong Gao 1644162cd32cSSong Gao #define SSRARNS(E1, E2, T1, T2) \ 1645162cd32cSSong Gao static T1 do_ssrarns_ ## E1(T1 e2, int sa, int sh) \ 1646162cd32cSSong Gao { \ 1647162cd32cSSong Gao T1 shft_res; \ 1648162cd32cSSong Gao \ 1649162cd32cSSong Gao shft_res = do_vsrar_ ## E2(e2, sa); \ 1650162cd32cSSong Gao T2 mask; \ 1651162cd32cSSong Gao mask = (1ll << sh) -1; \ 1652162cd32cSSong Gao if (shft_res > mask) { \ 1653162cd32cSSong Gao return mask; \ 1654162cd32cSSong Gao } else if (shft_res < -(mask +1)) { \ 1655162cd32cSSong Gao return ~mask; \ 1656162cd32cSSong Gao } else { \ 1657162cd32cSSong Gao return shft_res; \ 1658162cd32cSSong Gao } \ 1659162cd32cSSong Gao } 1660162cd32cSSong Gao 1661162cd32cSSong Gao SSRARNS(B, H, int16_t, int8_t) 1662162cd32cSSong Gao SSRARNS(H, W, int32_t, int16_t) 1663162cd32cSSong Gao SSRARNS(W, D, int64_t, int32_t) 1664162cd32cSSong Gao 1665162cd32cSSong Gao #define VSSRARN(NAME, BIT, T, E1, E2) \ 166604711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1667162cd32cSSong Gao { \ 1668162cd32cSSong Gao int i; \ 166904711da1SSong Gao VReg *Vd = (VReg *)vd; \ 167004711da1SSong Gao VReg *Vj = (VReg *)vj; \ 167104711da1SSong Gao VReg *Vk = (VReg *)vk; \ 1672162cd32cSSong Gao \ 1673162cd32cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 1674162cd32cSSong Gao Vd->E1(i) = do_ssrarns_ ## E1(Vj->E2(i), (T)Vk->E2(i)%BIT, BIT/2 -1); \ 1675162cd32cSSong Gao } \ 1676162cd32cSSong Gao Vd->D(1) = 0; \ 1677162cd32cSSong Gao } 1678162cd32cSSong Gao 1679162cd32cSSong Gao VSSRARN(vssrarn_b_h, 16, uint16_t, B, H) 1680162cd32cSSong Gao VSSRARN(vssrarn_h_w, 32, uint32_t, H, W) 1681162cd32cSSong Gao VSSRARN(vssrarn_w_d, 64, uint64_t, W, D) 1682162cd32cSSong Gao 1683162cd32cSSong Gao #define SSRLRNU(E1, E2, T1, T2, T3) \ 1684162cd32cSSong Gao static T1 do_ssrlrnu_ ## E1(T3 e2, int sa, int sh) \ 1685162cd32cSSong Gao { \ 1686162cd32cSSong Gao T1 shft_res; \ 1687162cd32cSSong Gao \ 1688162cd32cSSong Gao shft_res = do_vsrlr_ ## E2(e2, sa); \ 1689162cd32cSSong Gao \ 1690162cd32cSSong Gao T2 mask; \ 1691162cd32cSSong Gao mask = (1ull << sh) -1; \ 1692162cd32cSSong Gao if (shft_res > mask) { \ 1693162cd32cSSong Gao return mask; \ 1694162cd32cSSong Gao } else { \ 1695162cd32cSSong Gao return shft_res; \ 1696162cd32cSSong Gao } \ 1697162cd32cSSong Gao } 1698162cd32cSSong Gao 1699162cd32cSSong Gao SSRLRNU(B, H, uint16_t, uint8_t, int16_t) 1700162cd32cSSong Gao SSRLRNU(H, W, uint32_t, uint16_t, int32_t) 1701162cd32cSSong Gao SSRLRNU(W, D, uint64_t, uint32_t, int64_t) 1702162cd32cSSong Gao 1703162cd32cSSong Gao #define VSSRLRNU(NAME, BIT, T, E1, E2) \ 170404711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1705162cd32cSSong Gao { \ 1706162cd32cSSong Gao int i; \ 170704711da1SSong Gao VReg *Vd = (VReg *)vd; \ 170804711da1SSong Gao VReg *Vj = (VReg *)vj; \ 170904711da1SSong Gao VReg *Vk = (VReg *)vk; \ 1710162cd32cSSong Gao \ 1711162cd32cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 1712162cd32cSSong Gao Vd->E1(i) = do_ssrlrnu_ ## E1(Vj->E2(i), (T)Vk->E2(i)%BIT, BIT/2); \ 1713162cd32cSSong Gao } \ 1714162cd32cSSong Gao Vd->D(1) = 0; \ 1715162cd32cSSong Gao } 1716162cd32cSSong Gao 1717162cd32cSSong Gao VSSRLRNU(vssrlrn_bu_h, 16, uint16_t, B, H) 1718162cd32cSSong Gao VSSRLRNU(vssrlrn_hu_w, 32, uint32_t, H, W) 1719162cd32cSSong Gao VSSRLRNU(vssrlrn_wu_d, 64, uint64_t, W, D) 1720162cd32cSSong Gao 1721162cd32cSSong Gao #define SSRARNU(E1, E2, T1, T2, T3) \ 1722162cd32cSSong Gao static T1 do_ssrarnu_ ## E1(T3 e2, int sa, int sh) \ 1723162cd32cSSong Gao { \ 1724162cd32cSSong Gao T1 shft_res; \ 1725162cd32cSSong Gao \ 1726162cd32cSSong Gao if (e2 < 0) { \ 1727162cd32cSSong Gao shft_res = 0; \ 1728162cd32cSSong Gao } else { \ 1729162cd32cSSong Gao shft_res = do_vsrar_ ## E2(e2, sa); \ 1730162cd32cSSong Gao } \ 1731162cd32cSSong Gao T2 mask; \ 1732162cd32cSSong Gao mask = (1ull << sh) -1; \ 1733162cd32cSSong Gao if (shft_res > mask) { \ 1734162cd32cSSong Gao return mask; \ 1735162cd32cSSong Gao } else { \ 1736162cd32cSSong Gao return shft_res; \ 1737162cd32cSSong Gao } \ 1738162cd32cSSong Gao } 1739162cd32cSSong Gao 1740162cd32cSSong Gao SSRARNU(B, H, uint16_t, uint8_t, int16_t) 1741162cd32cSSong Gao SSRARNU(H, W, uint32_t, uint16_t, int32_t) 1742162cd32cSSong Gao SSRARNU(W, D, uint64_t, uint32_t, int64_t) 1743162cd32cSSong Gao 1744162cd32cSSong Gao #define VSSRARNU(NAME, BIT, T, E1, E2) \ 174504711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 1746162cd32cSSong Gao { \ 1747162cd32cSSong Gao int i; \ 174804711da1SSong Gao VReg *Vd = (VReg *)vd; \ 174904711da1SSong Gao VReg *Vj = (VReg *)vj; \ 175004711da1SSong Gao VReg *Vk = (VReg *)vk; \ 1751162cd32cSSong Gao \ 1752162cd32cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 1753162cd32cSSong Gao Vd->E1(i) = do_ssrarnu_ ## E1(Vj->E2(i), (T)Vk->E2(i)%BIT, BIT/2); \ 1754162cd32cSSong Gao } \ 1755162cd32cSSong Gao Vd->D(1) = 0; \ 1756162cd32cSSong Gao } 1757162cd32cSSong Gao 1758162cd32cSSong Gao VSSRARNU(vssrarn_bu_h, 16, uint16_t, B, H) 1759162cd32cSSong Gao VSSRARNU(vssrarn_hu_w, 32, uint32_t, H, W) 1760162cd32cSSong Gao VSSRARNU(vssrarn_wu_d, 64, uint64_t, W, D) 1761162cd32cSSong Gao 1762162cd32cSSong Gao #define VSSRLRNI(NAME, BIT, E1, E2) \ 1763329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 1764162cd32cSSong Gao { \ 1765162cd32cSSong Gao int i; \ 1766162cd32cSSong Gao VReg temp; \ 1767329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1768329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 1769162cd32cSSong Gao \ 1770162cd32cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 1771162cd32cSSong Gao temp.E1(i) = do_ssrlrns_ ## E1(Vj->E2(i), imm, BIT/2 -1); \ 1772162cd32cSSong Gao temp.E1(i + LSX_LEN/BIT) = do_ssrlrns_ ## E1(Vd->E2(i), imm, BIT/2 -1);\ 1773162cd32cSSong Gao } \ 1774162cd32cSSong Gao *Vd = temp; \ 1775162cd32cSSong Gao } 1776162cd32cSSong Gao 1777162cd32cSSong Gao #define VSSRLRNI_Q(NAME, sh) \ 1778329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 1779162cd32cSSong Gao { \ 1780162cd32cSSong Gao Int128 shft_res1, shft_res2, mask, r1, r2; \ 1781329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1782329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 1783162cd32cSSong Gao \ 1784162cd32cSSong Gao if (imm == 0) { \ 1785162cd32cSSong Gao shft_res1 = Vj->Q(0); \ 1786162cd32cSSong Gao shft_res2 = Vd->Q(0); \ 1787162cd32cSSong Gao } else { \ 1788162cd32cSSong Gao r1 = int128_and(int128_urshift(Vj->Q(0), (imm -1)), int128_one()); \ 1789162cd32cSSong Gao r2 = int128_and(int128_urshift(Vd->Q(0), (imm -1)), int128_one()); \ 1790162cd32cSSong Gao \ 1791162cd32cSSong Gao shft_res1 = (int128_add(int128_urshift(Vj->Q(0), imm), r1)); \ 1792162cd32cSSong Gao shft_res2 = (int128_add(int128_urshift(Vd->Q(0), imm), r2)); \ 1793162cd32cSSong Gao } \ 1794162cd32cSSong Gao \ 1795162cd32cSSong Gao mask = int128_sub(int128_lshift(int128_one(), sh), int128_one()); \ 1796162cd32cSSong Gao \ 1797162cd32cSSong Gao if (int128_ult(mask, shft_res1)) { \ 1798162cd32cSSong Gao Vd->D(0) = int128_getlo(mask); \ 1799162cd32cSSong Gao }else { \ 1800162cd32cSSong Gao Vd->D(0) = int128_getlo(shft_res1); \ 1801162cd32cSSong Gao } \ 1802162cd32cSSong Gao \ 1803162cd32cSSong Gao if (int128_ult(mask, shft_res2)) { \ 1804162cd32cSSong Gao Vd->D(1) = int128_getlo(mask); \ 1805162cd32cSSong Gao }else { \ 1806162cd32cSSong Gao Vd->D(1) = int128_getlo(shft_res2); \ 1807162cd32cSSong Gao } \ 1808162cd32cSSong Gao } 1809162cd32cSSong Gao 1810162cd32cSSong Gao VSSRLRNI(vssrlrni_b_h, 16, B, H) 1811162cd32cSSong Gao VSSRLRNI(vssrlrni_h_w, 32, H, W) 1812162cd32cSSong Gao VSSRLRNI(vssrlrni_w_d, 64, W, D) 1813162cd32cSSong Gao VSSRLRNI_Q(vssrlrni_d_q, 63) 1814162cd32cSSong Gao 1815162cd32cSSong Gao #define VSSRARNI(NAME, BIT, E1, E2) \ 1816329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 1817162cd32cSSong Gao { \ 1818162cd32cSSong Gao int i; \ 1819162cd32cSSong Gao VReg temp; \ 1820329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1821329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 1822162cd32cSSong Gao \ 1823162cd32cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 1824162cd32cSSong Gao temp.E1(i) = do_ssrarns_ ## E1(Vj->E2(i), imm, BIT/2 -1); \ 1825162cd32cSSong Gao temp.E1(i + LSX_LEN/BIT) = do_ssrarns_ ## E1(Vd->E2(i), imm, BIT/2 -1); \ 1826162cd32cSSong Gao } \ 1827162cd32cSSong Gao *Vd = temp; \ 1828162cd32cSSong Gao } 1829162cd32cSSong Gao 1830329517d5SSong Gao void HELPER(vssrarni_d_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 1831162cd32cSSong Gao { 1832162cd32cSSong Gao Int128 shft_res1, shft_res2, mask1, mask2, r1, r2; 1833329517d5SSong Gao VReg *Vd = (VReg *)vd; 1834329517d5SSong Gao VReg *Vj = (VReg *)vj; 1835162cd32cSSong Gao 1836162cd32cSSong Gao if (imm == 0) { 1837162cd32cSSong Gao shft_res1 = Vj->Q(0); 1838162cd32cSSong Gao shft_res2 = Vd->Q(0); 1839162cd32cSSong Gao } else { 1840162cd32cSSong Gao r1 = int128_and(int128_rshift(Vj->Q(0), (imm -1)), int128_one()); 1841162cd32cSSong Gao r2 = int128_and(int128_rshift(Vd->Q(0), (imm -1)), int128_one()); 1842162cd32cSSong Gao 1843162cd32cSSong Gao shft_res1 = int128_add(int128_rshift(Vj->Q(0), imm), r1); 1844162cd32cSSong Gao shft_res2 = int128_add(int128_rshift(Vd->Q(0), imm), r2); 1845162cd32cSSong Gao } 1846162cd32cSSong Gao 1847162cd32cSSong Gao mask1 = int128_sub(int128_lshift(int128_one(), 63), int128_one()); 1848162cd32cSSong Gao mask2 = int128_lshift(int128_one(), 63); 1849162cd32cSSong Gao 1850162cd32cSSong Gao if (int128_gt(shft_res1, mask1)) { 1851162cd32cSSong Gao Vd->D(0) = int128_getlo(mask1); 1852162cd32cSSong Gao } else if (int128_lt(shft_res1, int128_neg(mask2))) { 1853162cd32cSSong Gao Vd->D(0) = int128_getlo(mask2); 1854162cd32cSSong Gao } else { 1855162cd32cSSong Gao Vd->D(0) = int128_getlo(shft_res1); 1856162cd32cSSong Gao } 1857162cd32cSSong Gao 1858162cd32cSSong Gao if (int128_gt(shft_res2, mask1)) { 1859162cd32cSSong Gao Vd->D(1) = int128_getlo(mask1); 1860162cd32cSSong Gao } else if (int128_lt(shft_res2, int128_neg(mask2))) { 1861162cd32cSSong Gao Vd->D(1) = int128_getlo(mask2); 1862162cd32cSSong Gao } else { 1863162cd32cSSong Gao Vd->D(1) = int128_getlo(shft_res2); 1864162cd32cSSong Gao } 1865162cd32cSSong Gao } 1866162cd32cSSong Gao 1867162cd32cSSong Gao VSSRARNI(vssrarni_b_h, 16, B, H) 1868162cd32cSSong Gao VSSRARNI(vssrarni_h_w, 32, H, W) 1869162cd32cSSong Gao VSSRARNI(vssrarni_w_d, 64, W, D) 1870162cd32cSSong Gao 1871162cd32cSSong Gao #define VSSRLRNUI(NAME, BIT, E1, E2) \ 1872329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 1873162cd32cSSong Gao { \ 1874162cd32cSSong Gao int i; \ 1875162cd32cSSong Gao VReg temp; \ 1876329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1877329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 1878162cd32cSSong Gao \ 1879162cd32cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 1880162cd32cSSong Gao temp.E1(i) = do_ssrlrnu_ ## E1(Vj->E2(i), imm, BIT/2); \ 1881162cd32cSSong Gao temp.E1(i + LSX_LEN/BIT) = do_ssrlrnu_ ## E1(Vd->E2(i), imm, BIT/2); \ 1882162cd32cSSong Gao } \ 1883162cd32cSSong Gao *Vd = temp; \ 1884162cd32cSSong Gao } 1885162cd32cSSong Gao 1886162cd32cSSong Gao VSSRLRNUI(vssrlrni_bu_h, 16, B, H) 1887162cd32cSSong Gao VSSRLRNUI(vssrlrni_hu_w, 32, H, W) 1888162cd32cSSong Gao VSSRLRNUI(vssrlrni_wu_d, 64, W, D) 1889162cd32cSSong Gao VSSRLRNI_Q(vssrlrni_du_q, 64) 1890162cd32cSSong Gao 1891162cd32cSSong Gao #define VSSRARNUI(NAME, BIT, E1, E2) \ 1892329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 1893162cd32cSSong Gao { \ 1894162cd32cSSong Gao int i; \ 1895162cd32cSSong Gao VReg temp; \ 1896329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 1897329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 1898162cd32cSSong Gao \ 1899162cd32cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 1900162cd32cSSong Gao temp.E1(i) = do_ssrarnu_ ## E1(Vj->E2(i), imm, BIT/2); \ 1901162cd32cSSong Gao temp.E1(i + LSX_LEN/BIT) = do_ssrarnu_ ## E1(Vd->E2(i), imm, BIT/2); \ 1902162cd32cSSong Gao } \ 1903162cd32cSSong Gao *Vd = temp; \ 1904162cd32cSSong Gao } 1905162cd32cSSong Gao 1906329517d5SSong Gao void HELPER(vssrarni_du_q)(void *vd, void *vj, uint64_t imm, uint32_t desc) 1907162cd32cSSong Gao { 1908162cd32cSSong Gao Int128 shft_res1, shft_res2, mask1, mask2, r1, r2; 1909329517d5SSong Gao VReg *Vd = (VReg *)vd; 1910329517d5SSong Gao VReg *Vj = (VReg *)vj; 1911162cd32cSSong Gao 1912162cd32cSSong Gao if (imm == 0) { 1913162cd32cSSong Gao shft_res1 = Vj->Q(0); 1914162cd32cSSong Gao shft_res2 = Vd->Q(0); 1915162cd32cSSong Gao } else { 1916162cd32cSSong Gao r1 = int128_and(int128_rshift(Vj->Q(0), (imm -1)), int128_one()); 1917162cd32cSSong Gao r2 = int128_and(int128_rshift(Vd->Q(0), (imm -1)), int128_one()); 1918162cd32cSSong Gao 1919162cd32cSSong Gao shft_res1 = int128_add(int128_rshift(Vj->Q(0), imm), r1); 1920162cd32cSSong Gao shft_res2 = int128_add(int128_rshift(Vd->Q(0), imm), r2); 1921162cd32cSSong Gao } 1922162cd32cSSong Gao 1923162cd32cSSong Gao if (int128_lt(Vj->Q(0), int128_zero())) { 1924162cd32cSSong Gao shft_res1 = int128_zero(); 1925162cd32cSSong Gao } 1926162cd32cSSong Gao if (int128_lt(Vd->Q(0), int128_zero())) { 1927162cd32cSSong Gao shft_res2 = int128_zero(); 1928162cd32cSSong Gao } 1929162cd32cSSong Gao 1930162cd32cSSong Gao mask1 = int128_sub(int128_lshift(int128_one(), 64), int128_one()); 1931162cd32cSSong Gao mask2 = int128_lshift(int128_one(), 64); 1932162cd32cSSong Gao 1933162cd32cSSong Gao if (int128_gt(shft_res1, mask1)) { 1934162cd32cSSong Gao Vd->D(0) = int128_getlo(mask1); 1935162cd32cSSong Gao } else if (int128_lt(shft_res1, int128_neg(mask2))) { 1936162cd32cSSong Gao Vd->D(0) = int128_getlo(mask2); 1937162cd32cSSong Gao } else { 1938162cd32cSSong Gao Vd->D(0) = int128_getlo(shft_res1); 1939162cd32cSSong Gao } 1940162cd32cSSong Gao 1941162cd32cSSong Gao if (int128_gt(shft_res2, mask1)) { 1942162cd32cSSong Gao Vd->D(1) = int128_getlo(mask1); 1943162cd32cSSong Gao } else if (int128_lt(shft_res2, int128_neg(mask2))) { 1944162cd32cSSong Gao Vd->D(1) = int128_getlo(mask2); 1945162cd32cSSong Gao } else { 1946162cd32cSSong Gao Vd->D(1) = int128_getlo(shft_res2); 1947162cd32cSSong Gao } 1948162cd32cSSong Gao } 1949162cd32cSSong Gao 1950162cd32cSSong Gao VSSRARNUI(vssrarni_bu_h, 16, B, H) 1951162cd32cSSong Gao VSSRARNUI(vssrarni_hu_w, 32, H, W) 1952162cd32cSSong Gao VSSRARNUI(vssrarni_wu_d, 64, W, D) 19532e105e12SSong Gao 19542e105e12SSong Gao #define DO_2OP(NAME, BIT, E, DO_OP) \ 1955ff27e335SSong Gao void HELPER(NAME)(void *vd, void *vj, uint32_t desc) \ 19562e105e12SSong Gao { \ 19572e105e12SSong Gao int i; \ 1958ff27e335SSong Gao VReg *Vd = (VReg *)vd; \ 1959ff27e335SSong Gao VReg *Vj = (VReg *)vj; \ 19602e105e12SSong Gao \ 19612e105e12SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) \ 19622e105e12SSong Gao { \ 19632e105e12SSong Gao Vd->E(i) = DO_OP(Vj->E(i)); \ 19642e105e12SSong Gao } \ 19652e105e12SSong Gao } 19662e105e12SSong Gao 19672e105e12SSong Gao #define DO_CLO_B(N) (clz32(~N & 0xff) - 24) 19682e105e12SSong Gao #define DO_CLO_H(N) (clz32(~N & 0xffff) - 16) 19692e105e12SSong Gao #define DO_CLO_W(N) (clz32(~N)) 19702e105e12SSong Gao #define DO_CLO_D(N) (clz64(~N)) 19712e105e12SSong Gao #define DO_CLZ_B(N) (clz32(N) - 24) 19722e105e12SSong Gao #define DO_CLZ_H(N) (clz32(N) - 16) 19732e105e12SSong Gao #define DO_CLZ_W(N) (clz32(N)) 19742e105e12SSong Gao #define DO_CLZ_D(N) (clz64(N)) 19752e105e12SSong Gao 19762e105e12SSong Gao DO_2OP(vclo_b, 8, UB, DO_CLO_B) 19772e105e12SSong Gao DO_2OP(vclo_h, 16, UH, DO_CLO_H) 19782e105e12SSong Gao DO_2OP(vclo_w, 32, UW, DO_CLO_W) 19792e105e12SSong Gao DO_2OP(vclo_d, 64, UD, DO_CLO_D) 19802e105e12SSong Gao DO_2OP(vclz_b, 8, UB, DO_CLZ_B) 19812e105e12SSong Gao DO_2OP(vclz_h, 16, UH, DO_CLZ_H) 19822e105e12SSong Gao DO_2OP(vclz_w, 32, UW, DO_CLZ_W) 19832e105e12SSong Gao DO_2OP(vclz_d, 64, UD, DO_CLZ_D) 1984bb22ee57SSong Gao 1985bb22ee57SSong Gao #define VPCNT(NAME, BIT, E, FN) \ 1986ff27e335SSong Gao void HELPER(NAME)(void *vd, void *vj, uint32_t desc) \ 1987bb22ee57SSong Gao { \ 1988bb22ee57SSong Gao int i; \ 1989ff27e335SSong Gao VReg *Vd = (VReg *)vd; \ 1990ff27e335SSong Gao VReg *Vj = (VReg *)vj; \ 1991bb22ee57SSong Gao \ 1992bb22ee57SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) \ 1993bb22ee57SSong Gao { \ 1994bb22ee57SSong Gao Vd->E(i) = FN(Vj->E(i)); \ 1995bb22ee57SSong Gao } \ 1996bb22ee57SSong Gao } 1997bb22ee57SSong Gao 1998bb22ee57SSong Gao VPCNT(vpcnt_b, 8, UB, ctpop8) 1999bb22ee57SSong Gao VPCNT(vpcnt_h, 16, UH, ctpop16) 2000bb22ee57SSong Gao VPCNT(vpcnt_w, 32, UW, ctpop32) 2001bb22ee57SSong Gao VPCNT(vpcnt_d, 64, UD, ctpop64) 20020b1e6705SSong Gao 20030b1e6705SSong Gao #define DO_BITCLR(a, bit) (a & ~(1ull << bit)) 20040b1e6705SSong Gao #define DO_BITSET(a, bit) (a | 1ull << bit) 20050b1e6705SSong Gao #define DO_BITREV(a, bit) (a ^ (1ull << bit)) 20060b1e6705SSong Gao 20070b1e6705SSong Gao #define DO_BIT(NAME, BIT, E, DO_OP) \ 20080b1e6705SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t v) \ 20090b1e6705SSong Gao { \ 20100b1e6705SSong Gao int i; \ 20110b1e6705SSong Gao VReg *Vd = (VReg *)vd; \ 20120b1e6705SSong Gao VReg *Vj = (VReg *)vj; \ 20130b1e6705SSong Gao VReg *Vk = (VReg *)vk; \ 20140b1e6705SSong Gao \ 20150b1e6705SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 20160b1e6705SSong Gao Vd->E(i) = DO_OP(Vj->E(i), Vk->E(i)%BIT); \ 20170b1e6705SSong Gao } \ 20180b1e6705SSong Gao } 20190b1e6705SSong Gao 20200b1e6705SSong Gao DO_BIT(vbitclr_b, 8, UB, DO_BITCLR) 20210b1e6705SSong Gao DO_BIT(vbitclr_h, 16, UH, DO_BITCLR) 20220b1e6705SSong Gao DO_BIT(vbitclr_w, 32, UW, DO_BITCLR) 20230b1e6705SSong Gao DO_BIT(vbitclr_d, 64, UD, DO_BITCLR) 20240b1e6705SSong Gao DO_BIT(vbitset_b, 8, UB, DO_BITSET) 20250b1e6705SSong Gao DO_BIT(vbitset_h, 16, UH, DO_BITSET) 20260b1e6705SSong Gao DO_BIT(vbitset_w, 32, UW, DO_BITSET) 20270b1e6705SSong Gao DO_BIT(vbitset_d, 64, UD, DO_BITSET) 20280b1e6705SSong Gao DO_BIT(vbitrev_b, 8, UB, DO_BITREV) 20290b1e6705SSong Gao DO_BIT(vbitrev_h, 16, UH, DO_BITREV) 20300b1e6705SSong Gao DO_BIT(vbitrev_w, 32, UW, DO_BITREV) 20310b1e6705SSong Gao DO_BIT(vbitrev_d, 64, UD, DO_BITREV) 20320b1e6705SSong Gao 20330b1e6705SSong Gao #define DO_BITI(NAME, BIT, E, DO_OP) \ 20340b1e6705SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t v) \ 20350b1e6705SSong Gao { \ 20360b1e6705SSong Gao int i; \ 20370b1e6705SSong Gao VReg *Vd = (VReg *)vd; \ 20380b1e6705SSong Gao VReg *Vj = (VReg *)vj; \ 20390b1e6705SSong Gao \ 20400b1e6705SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 20410b1e6705SSong Gao Vd->E(i) = DO_OP(Vj->E(i), imm); \ 20420b1e6705SSong Gao } \ 20430b1e6705SSong Gao } 20440b1e6705SSong Gao 20450b1e6705SSong Gao DO_BITI(vbitclri_b, 8, UB, DO_BITCLR) 20460b1e6705SSong Gao DO_BITI(vbitclri_h, 16, UH, DO_BITCLR) 20470b1e6705SSong Gao DO_BITI(vbitclri_w, 32, UW, DO_BITCLR) 20480b1e6705SSong Gao DO_BITI(vbitclri_d, 64, UD, DO_BITCLR) 20490b1e6705SSong Gao DO_BITI(vbitseti_b, 8, UB, DO_BITSET) 20500b1e6705SSong Gao DO_BITI(vbitseti_h, 16, UH, DO_BITSET) 20510b1e6705SSong Gao DO_BITI(vbitseti_w, 32, UW, DO_BITSET) 20520b1e6705SSong Gao DO_BITI(vbitseti_d, 64, UD, DO_BITSET) 20530b1e6705SSong Gao DO_BITI(vbitrevi_b, 8, UB, DO_BITREV) 20540b1e6705SSong Gao DO_BITI(vbitrevi_h, 16, UH, DO_BITREV) 20550b1e6705SSong Gao DO_BITI(vbitrevi_w, 32, UW, DO_BITREV) 20560b1e6705SSong Gao DO_BITI(vbitrevi_d, 64, UD, DO_BITREV) 2057ac95a0b9SSong Gao 2058ac95a0b9SSong Gao #define VFRSTP(NAME, BIT, MASK, E) \ 205904711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 2060ac95a0b9SSong Gao { \ 2061ac95a0b9SSong Gao int i, m; \ 206204711da1SSong Gao VReg *Vd = (VReg *)vd; \ 206304711da1SSong Gao VReg *Vj = (VReg *)vj; \ 206404711da1SSong Gao VReg *Vk = (VReg *)vk; \ 2065ac95a0b9SSong Gao \ 2066ac95a0b9SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2067ac95a0b9SSong Gao if (Vj->E(i) < 0) { \ 2068ac95a0b9SSong Gao break; \ 2069ac95a0b9SSong Gao } \ 2070ac95a0b9SSong Gao } \ 2071ac95a0b9SSong Gao m = Vk->E(0) & MASK; \ 2072ac95a0b9SSong Gao Vd->E(m) = i; \ 2073ac95a0b9SSong Gao } 2074ac95a0b9SSong Gao 2075ac95a0b9SSong Gao VFRSTP(vfrstp_b, 8, 0xf, B) 2076ac95a0b9SSong Gao VFRSTP(vfrstp_h, 16, 0x7, H) 2077ac95a0b9SSong Gao 2078ac95a0b9SSong Gao #define VFRSTPI(NAME, BIT, E) \ 2079329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 2080ac95a0b9SSong Gao { \ 2081ac95a0b9SSong Gao int i, m; \ 2082329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 2083329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 2084ac95a0b9SSong Gao \ 2085ac95a0b9SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2086ac95a0b9SSong Gao if (Vj->E(i) < 0) { \ 2087ac95a0b9SSong Gao break; \ 2088ac95a0b9SSong Gao } \ 2089ac95a0b9SSong Gao } \ 2090ac95a0b9SSong Gao m = imm % (LSX_LEN/BIT); \ 2091ac95a0b9SSong Gao Vd->E(m) = i; \ 2092ac95a0b9SSong Gao } 2093ac95a0b9SSong Gao 2094ac95a0b9SSong Gao VFRSTPI(vfrstpi_b, 8, B) 2095ac95a0b9SSong Gao VFRSTPI(vfrstpi_h, 16, H) 2096aca67472SSong Gao 2097aca67472SSong Gao static void vec_update_fcsr0_mask(CPULoongArchState *env, 2098aca67472SSong Gao uintptr_t pc, int mask) 2099aca67472SSong Gao { 2100aca67472SSong Gao int flags = get_float_exception_flags(&env->fp_status); 2101aca67472SSong Gao 2102aca67472SSong Gao set_float_exception_flags(0, &env->fp_status); 2103aca67472SSong Gao 2104aca67472SSong Gao flags &= ~mask; 2105aca67472SSong Gao 2106aca67472SSong Gao if (flags) { 2107aca67472SSong Gao flags = ieee_ex_to_loongarch(flags); 2108aca67472SSong Gao UPDATE_FP_CAUSE(env->fcsr0, flags); 2109aca67472SSong Gao } 2110aca67472SSong Gao 2111aca67472SSong Gao if (GET_FP_ENABLES(env->fcsr0) & flags) { 2112aca67472SSong Gao do_raise_exception(env, EXCCODE_FPE, pc); 2113aca67472SSong Gao } else { 2114aca67472SSong Gao UPDATE_FP_FLAGS(env->fcsr0, flags); 2115aca67472SSong Gao } 2116aca67472SSong Gao } 2117aca67472SSong Gao 2118aca67472SSong Gao static void vec_update_fcsr0(CPULoongArchState *env, uintptr_t pc) 2119aca67472SSong Gao { 2120aca67472SSong Gao vec_update_fcsr0_mask(env, pc, 0); 2121aca67472SSong Gao } 2122aca67472SSong Gao 2123aca67472SSong Gao static inline void vec_clear_cause(CPULoongArchState *env) 2124aca67472SSong Gao { 2125aca67472SSong Gao SET_FP_CAUSE(env->fcsr0, 0); 2126aca67472SSong Gao } 2127aca67472SSong Gao 2128aca67472SSong Gao #define DO_3OP_F(NAME, BIT, E, FN) \ 21293b286753SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, \ 21303b286753SSong Gao CPULoongArchState *env, uint32_t desc) \ 2131aca67472SSong Gao { \ 2132aca67472SSong Gao int i; \ 21333b286753SSong Gao VReg *Vd = (VReg *)vd; \ 21343b286753SSong Gao VReg *Vj = (VReg *)vj; \ 21353b286753SSong Gao VReg *Vk = (VReg *)vk; \ 2136aca67472SSong Gao \ 2137aca67472SSong Gao vec_clear_cause(env); \ 2138aca67472SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2139aca67472SSong Gao Vd->E(i) = FN(Vj->E(i), Vk->E(i), &env->fp_status); \ 2140aca67472SSong Gao vec_update_fcsr0(env, GETPC()); \ 2141aca67472SSong Gao } \ 2142aca67472SSong Gao } 2143aca67472SSong Gao 2144aca67472SSong Gao DO_3OP_F(vfadd_s, 32, UW, float32_add) 2145aca67472SSong Gao DO_3OP_F(vfadd_d, 64, UD, float64_add) 2146aca67472SSong Gao DO_3OP_F(vfsub_s, 32, UW, float32_sub) 2147aca67472SSong Gao DO_3OP_F(vfsub_d, 64, UD, float64_sub) 2148aca67472SSong Gao DO_3OP_F(vfmul_s, 32, UW, float32_mul) 2149aca67472SSong Gao DO_3OP_F(vfmul_d, 64, UD, float64_mul) 2150aca67472SSong Gao DO_3OP_F(vfdiv_s, 32, UW, float32_div) 2151aca67472SSong Gao DO_3OP_F(vfdiv_d, 64, UD, float64_div) 2152aca67472SSong Gao DO_3OP_F(vfmax_s, 32, UW, float32_maxnum) 2153aca67472SSong Gao DO_3OP_F(vfmax_d, 64, UD, float64_maxnum) 2154aca67472SSong Gao DO_3OP_F(vfmin_s, 32, UW, float32_minnum) 2155aca67472SSong Gao DO_3OP_F(vfmin_d, 64, UD, float64_minnum) 2156aca67472SSong Gao DO_3OP_F(vfmaxa_s, 32, UW, float32_maxnummag) 2157aca67472SSong Gao DO_3OP_F(vfmaxa_d, 64, UD, float64_maxnummag) 2158aca67472SSong Gao DO_3OP_F(vfmina_s, 32, UW, float32_minnummag) 2159aca67472SSong Gao DO_3OP_F(vfmina_d, 64, UD, float64_minnummag) 2160aca67472SSong Gao 2161aca67472SSong Gao #define DO_4OP_F(NAME, BIT, E, FN, flags) \ 2162e2600dadSSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, void *va, \ 2163e2600dadSSong Gao CPULoongArchState *env, uint32_t desc) \ 2164aca67472SSong Gao { \ 2165aca67472SSong Gao int i; \ 2166e2600dadSSong Gao VReg *Vd = (VReg *)vd; \ 2167e2600dadSSong Gao VReg *Vj = (VReg *)vj; \ 2168e2600dadSSong Gao VReg *Vk = (VReg *)vk; \ 2169e2600dadSSong Gao VReg *Va = (VReg *)va; \ 2170aca67472SSong Gao \ 2171aca67472SSong Gao vec_clear_cause(env); \ 2172aca67472SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2173aca67472SSong Gao Vd->E(i) = FN(Vj->E(i), Vk->E(i), Va->E(i), flags, &env->fp_status); \ 2174aca67472SSong Gao vec_update_fcsr0(env, GETPC()); \ 2175aca67472SSong Gao } \ 2176aca67472SSong Gao } 2177aca67472SSong Gao 2178aca67472SSong Gao DO_4OP_F(vfmadd_s, 32, UW, float32_muladd, 0) 2179aca67472SSong Gao DO_4OP_F(vfmadd_d, 64, UD, float64_muladd, 0) 2180aca67472SSong Gao DO_4OP_F(vfmsub_s, 32, UW, float32_muladd, float_muladd_negate_c) 2181aca67472SSong Gao DO_4OP_F(vfmsub_d, 64, UD, float64_muladd, float_muladd_negate_c) 2182aca67472SSong Gao DO_4OP_F(vfnmadd_s, 32, UW, float32_muladd, float_muladd_negate_result) 2183aca67472SSong Gao DO_4OP_F(vfnmadd_d, 64, UD, float64_muladd, float_muladd_negate_result) 2184aca67472SSong Gao DO_4OP_F(vfnmsub_s, 32, UW, float32_muladd, 2185aca67472SSong Gao float_muladd_negate_c | float_muladd_negate_result) 2186aca67472SSong Gao DO_4OP_F(vfnmsub_d, 64, UD, float64_muladd, 2187aca67472SSong Gao float_muladd_negate_c | float_muladd_negate_result) 2188aca67472SSong Gao 2189aca67472SSong Gao #define DO_2OP_F(NAME, BIT, E, FN) \ 2190226bf881SSong Gao void HELPER(NAME)(void *vd, void *vj, \ 2191226bf881SSong Gao CPULoongArchState *env, uint32_t desc) \ 2192aca67472SSong Gao { \ 2193aca67472SSong Gao int i; \ 2194226bf881SSong Gao VReg *Vd = (VReg *)vd; \ 2195226bf881SSong Gao VReg *Vj = (VReg *)vj; \ 2196aca67472SSong Gao \ 2197aca67472SSong Gao vec_clear_cause(env); \ 2198aca67472SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2199aca67472SSong Gao Vd->E(i) = FN(env, Vj->E(i)); \ 2200aca67472SSong Gao } \ 2201aca67472SSong Gao } 2202aca67472SSong Gao 2203aca67472SSong Gao #define FLOGB(BIT, T) \ 2204aca67472SSong Gao static T do_flogb_## BIT(CPULoongArchState *env, T fj) \ 2205aca67472SSong Gao { \ 2206aca67472SSong Gao T fp, fd; \ 2207aca67472SSong Gao float_status *status = &env->fp_status; \ 2208aca67472SSong Gao FloatRoundMode old_mode = get_float_rounding_mode(status); \ 2209aca67472SSong Gao \ 2210aca67472SSong Gao set_float_rounding_mode(float_round_down, status); \ 2211aca67472SSong Gao fp = float ## BIT ##_log2(fj, status); \ 2212aca67472SSong Gao fd = float ## BIT ##_round_to_int(fp, status); \ 2213aca67472SSong Gao set_float_rounding_mode(old_mode, status); \ 2214aca67472SSong Gao vec_update_fcsr0_mask(env, GETPC(), float_flag_inexact); \ 2215aca67472SSong Gao return fd; \ 2216aca67472SSong Gao } 2217aca67472SSong Gao 2218aca67472SSong Gao FLOGB(32, uint32_t) 2219aca67472SSong Gao FLOGB(64, uint64_t) 2220aca67472SSong Gao 2221aca67472SSong Gao #define FCLASS(NAME, BIT, E, FN) \ 2222226bf881SSong Gao void HELPER(NAME)(void *vd, void *vj, \ 2223226bf881SSong Gao CPULoongArchState *env, uint32_t desc) \ 2224aca67472SSong Gao { \ 2225aca67472SSong Gao int i; \ 2226226bf881SSong Gao VReg *Vd = (VReg *)vd; \ 2227226bf881SSong Gao VReg *Vj = (VReg *)vj; \ 2228aca67472SSong Gao \ 2229aca67472SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2230aca67472SSong Gao Vd->E(i) = FN(env, Vj->E(i)); \ 2231aca67472SSong Gao } \ 2232aca67472SSong Gao } 2233aca67472SSong Gao 2234aca67472SSong Gao FCLASS(vfclass_s, 32, UW, helper_fclass_s) 2235aca67472SSong Gao FCLASS(vfclass_d, 64, UD, helper_fclass_d) 2236aca67472SSong Gao 2237aca67472SSong Gao #define FSQRT(BIT, T) \ 2238aca67472SSong Gao static T do_fsqrt_## BIT(CPULoongArchState *env, T fj) \ 2239aca67472SSong Gao { \ 2240aca67472SSong Gao T fd; \ 2241aca67472SSong Gao fd = float ## BIT ##_sqrt(fj, &env->fp_status); \ 2242aca67472SSong Gao vec_update_fcsr0(env, GETPC()); \ 2243aca67472SSong Gao return fd; \ 2244aca67472SSong Gao } 2245aca67472SSong Gao 2246aca67472SSong Gao FSQRT(32, uint32_t) 2247aca67472SSong Gao FSQRT(64, uint64_t) 2248aca67472SSong Gao 2249aca67472SSong Gao #define FRECIP(BIT, T) \ 2250aca67472SSong Gao static T do_frecip_## BIT(CPULoongArchState *env, T fj) \ 2251aca67472SSong Gao { \ 2252aca67472SSong Gao T fd; \ 2253aca67472SSong Gao fd = float ## BIT ##_div(float ## BIT ##_one, fj, &env->fp_status); \ 2254aca67472SSong Gao vec_update_fcsr0(env, GETPC()); \ 2255aca67472SSong Gao return fd; \ 2256aca67472SSong Gao } 2257aca67472SSong Gao 2258aca67472SSong Gao FRECIP(32, uint32_t) 2259aca67472SSong Gao FRECIP(64, uint64_t) 2260aca67472SSong Gao 2261aca67472SSong Gao #define FRSQRT(BIT, T) \ 2262aca67472SSong Gao static T do_frsqrt_## BIT(CPULoongArchState *env, T fj) \ 2263aca67472SSong Gao { \ 2264aca67472SSong Gao T fd, fp; \ 2265aca67472SSong Gao fp = float ## BIT ##_sqrt(fj, &env->fp_status); \ 2266aca67472SSong Gao fd = float ## BIT ##_div(float ## BIT ##_one, fp, &env->fp_status); \ 2267aca67472SSong Gao vec_update_fcsr0(env, GETPC()); \ 2268aca67472SSong Gao return fd; \ 2269aca67472SSong Gao } 2270aca67472SSong Gao 2271aca67472SSong Gao FRSQRT(32, uint32_t) 2272aca67472SSong Gao FRSQRT(64, uint64_t) 2273aca67472SSong Gao 2274aca67472SSong Gao DO_2OP_F(vflogb_s, 32, UW, do_flogb_32) 2275aca67472SSong Gao DO_2OP_F(vflogb_d, 64, UD, do_flogb_64) 2276aca67472SSong Gao DO_2OP_F(vfsqrt_s, 32, UW, do_fsqrt_32) 2277aca67472SSong Gao DO_2OP_F(vfsqrt_d, 64, UD, do_fsqrt_64) 2278aca67472SSong Gao DO_2OP_F(vfrecip_s, 32, UW, do_frecip_32) 2279aca67472SSong Gao DO_2OP_F(vfrecip_d, 64, UD, do_frecip_64) 2280aca67472SSong Gao DO_2OP_F(vfrsqrt_s, 32, UW, do_frsqrt_32) 2281aca67472SSong Gao DO_2OP_F(vfrsqrt_d, 64, UD, do_frsqrt_64) 2282399665d2SSong Gao 2283399665d2SSong Gao static uint32_t float16_cvt_float32(uint16_t h, float_status *status) 2284399665d2SSong Gao { 2285399665d2SSong Gao return float16_to_float32(h, true, status); 2286399665d2SSong Gao } 2287399665d2SSong Gao static uint64_t float32_cvt_float64(uint32_t s, float_status *status) 2288399665d2SSong Gao { 2289399665d2SSong Gao return float32_to_float64(s, status); 2290399665d2SSong Gao } 2291399665d2SSong Gao 2292399665d2SSong Gao static uint16_t float32_cvt_float16(uint32_t s, float_status *status) 2293399665d2SSong Gao { 2294399665d2SSong Gao return float32_to_float16(s, true, status); 2295399665d2SSong Gao } 2296399665d2SSong Gao static uint32_t float64_cvt_float32(uint64_t d, float_status *status) 2297399665d2SSong Gao { 2298399665d2SSong Gao return float64_to_float32(d, status); 2299399665d2SSong Gao } 2300399665d2SSong Gao 2301226bf881SSong Gao void HELPER(vfcvtl_s_h)(void *vd, void *vj, 2302226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2303399665d2SSong Gao { 2304399665d2SSong Gao int i; 2305399665d2SSong Gao VReg temp; 2306226bf881SSong Gao VReg *Vd = (VReg *)vd; 2307226bf881SSong Gao VReg *Vj = (VReg *)vj; 2308399665d2SSong Gao 2309399665d2SSong Gao vec_clear_cause(env); 2310399665d2SSong Gao for (i = 0; i < LSX_LEN/32; i++) { 2311399665d2SSong Gao temp.UW(i) = float16_cvt_float32(Vj->UH(i), &env->fp_status); 2312399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2313399665d2SSong Gao } 2314399665d2SSong Gao *Vd = temp; 2315399665d2SSong Gao } 2316399665d2SSong Gao 2317226bf881SSong Gao void HELPER(vfcvtl_d_s)(void *vd, void *vj, 2318226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2319399665d2SSong Gao { 2320399665d2SSong Gao int i; 2321399665d2SSong Gao VReg temp; 2322226bf881SSong Gao VReg *Vd = (VReg *)vd; 2323226bf881SSong Gao VReg *Vj = (VReg *)vj; 2324399665d2SSong Gao 2325399665d2SSong Gao vec_clear_cause(env); 2326399665d2SSong Gao for (i = 0; i < LSX_LEN/64; i++) { 2327399665d2SSong Gao temp.UD(i) = float32_cvt_float64(Vj->UW(i), &env->fp_status); 2328399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2329399665d2SSong Gao } 2330399665d2SSong Gao *Vd = temp; 2331399665d2SSong Gao } 2332399665d2SSong Gao 2333226bf881SSong Gao void HELPER(vfcvth_s_h)(void *vd, void *vj, 2334226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2335399665d2SSong Gao { 2336399665d2SSong Gao int i; 2337399665d2SSong Gao VReg temp; 2338226bf881SSong Gao VReg *Vd = (VReg *)vd; 2339226bf881SSong Gao VReg *Vj = (VReg *)vj; 2340399665d2SSong Gao 2341399665d2SSong Gao vec_clear_cause(env); 2342399665d2SSong Gao for (i = 0; i < LSX_LEN/32; i++) { 2343399665d2SSong Gao temp.UW(i) = float16_cvt_float32(Vj->UH(i + 4), &env->fp_status); 2344399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2345399665d2SSong Gao } 2346399665d2SSong Gao *Vd = temp; 2347399665d2SSong Gao } 2348399665d2SSong Gao 2349226bf881SSong Gao void HELPER(vfcvth_d_s)(void *vd, void *vj, 2350226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2351399665d2SSong Gao { 2352399665d2SSong Gao int i; 2353399665d2SSong Gao VReg temp; 2354226bf881SSong Gao VReg *Vd = (VReg *)vd; 2355226bf881SSong Gao VReg *Vj = (VReg *)vj; 2356399665d2SSong Gao 2357399665d2SSong Gao vec_clear_cause(env); 2358399665d2SSong Gao for (i = 0; i < LSX_LEN/64; i++) { 2359399665d2SSong Gao temp.UD(i) = float32_cvt_float64(Vj->UW(i + 2), &env->fp_status); 2360399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2361399665d2SSong Gao } 2362399665d2SSong Gao *Vd = temp; 2363399665d2SSong Gao } 2364399665d2SSong Gao 23653b286753SSong Gao void HELPER(vfcvt_h_s)(void *vd, void *vj, void *vk, 23663b286753SSong Gao CPULoongArchState *env, uint32_t desc) 2367399665d2SSong Gao { 2368399665d2SSong Gao int i; 2369399665d2SSong Gao VReg temp; 23703b286753SSong Gao VReg *Vd = (VReg *)vd; 23713b286753SSong Gao VReg *Vj = (VReg *)vj; 23723b286753SSong Gao VReg *Vk = (VReg *)vk; 2373399665d2SSong Gao 2374399665d2SSong Gao vec_clear_cause(env); 2375399665d2SSong Gao for(i = 0; i < LSX_LEN/32; i++) { 2376399665d2SSong Gao temp.UH(i + 4) = float32_cvt_float16(Vj->UW(i), &env->fp_status); 2377399665d2SSong Gao temp.UH(i) = float32_cvt_float16(Vk->UW(i), &env->fp_status); 2378399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2379399665d2SSong Gao } 2380399665d2SSong Gao *Vd = temp; 2381399665d2SSong Gao } 2382399665d2SSong Gao 23833b286753SSong Gao void HELPER(vfcvt_s_d)(void *vd, void *vj, void *vk, 23843b286753SSong Gao CPULoongArchState *env, uint32_t desc) 2385399665d2SSong Gao { 2386399665d2SSong Gao int i; 2387399665d2SSong Gao VReg temp; 23883b286753SSong Gao VReg *Vd = (VReg *)vd; 23893b286753SSong Gao VReg *Vj = (VReg *)vj; 23903b286753SSong Gao VReg *Vk = (VReg *)vk; 2391399665d2SSong Gao 2392399665d2SSong Gao vec_clear_cause(env); 2393399665d2SSong Gao for(i = 0; i < LSX_LEN/64; i++) { 2394399665d2SSong Gao temp.UW(i + 2) = float64_cvt_float32(Vj->UD(i), &env->fp_status); 2395399665d2SSong Gao temp.UW(i) = float64_cvt_float32(Vk->UD(i), &env->fp_status); 2396399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2397399665d2SSong Gao } 2398399665d2SSong Gao *Vd = temp; 2399399665d2SSong Gao } 2400399665d2SSong Gao 2401226bf881SSong Gao void HELPER(vfrint_s)(void *vd, void *vj, 2402226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2403399665d2SSong Gao { 2404399665d2SSong Gao int i; 2405226bf881SSong Gao VReg *Vd = (VReg *)vd; 2406226bf881SSong Gao VReg *Vj = (VReg *)vj; 2407399665d2SSong Gao 2408399665d2SSong Gao vec_clear_cause(env); 2409399665d2SSong Gao for (i = 0; i < 4; i++) { 2410399665d2SSong Gao Vd->W(i) = float32_round_to_int(Vj->UW(i), &env->fp_status); 2411399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2412399665d2SSong Gao } 2413399665d2SSong Gao } 2414399665d2SSong Gao 2415226bf881SSong Gao void HELPER(vfrint_d)(void *vd, void *vj, 2416226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2417399665d2SSong Gao { 2418399665d2SSong Gao int i; 2419226bf881SSong Gao VReg *Vd = (VReg *)vd; 2420226bf881SSong Gao VReg *Vj = (VReg *)vj; 2421399665d2SSong Gao 2422399665d2SSong Gao vec_clear_cause(env); 2423399665d2SSong Gao for (i = 0; i < 2; i++) { 2424399665d2SSong Gao Vd->D(i) = float64_round_to_int(Vj->UD(i), &env->fp_status); 2425399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2426399665d2SSong Gao } 2427399665d2SSong Gao } 2428399665d2SSong Gao 2429399665d2SSong Gao #define FCVT_2OP(NAME, BIT, E, MODE) \ 2430226bf881SSong Gao void HELPER(NAME)(void *vd, void *vj, \ 2431226bf881SSong Gao CPULoongArchState *env, uint32_t desc) \ 2432399665d2SSong Gao { \ 2433399665d2SSong Gao int i; \ 2434226bf881SSong Gao VReg *Vd = (VReg *)vd; \ 2435226bf881SSong Gao VReg *Vj = (VReg *)vj; \ 2436399665d2SSong Gao \ 2437399665d2SSong Gao vec_clear_cause(env); \ 2438399665d2SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2439399665d2SSong Gao FloatRoundMode old_mode = get_float_rounding_mode(&env->fp_status); \ 2440399665d2SSong Gao set_float_rounding_mode(MODE, &env->fp_status); \ 2441399665d2SSong Gao Vd->E(i) = float## BIT ## _round_to_int(Vj->E(i), &env->fp_status); \ 2442399665d2SSong Gao set_float_rounding_mode(old_mode, &env->fp_status); \ 2443399665d2SSong Gao vec_update_fcsr0(env, GETPC()); \ 2444399665d2SSong Gao } \ 2445399665d2SSong Gao } 2446399665d2SSong Gao 2447399665d2SSong Gao FCVT_2OP(vfrintrne_s, 32, UW, float_round_nearest_even) 2448399665d2SSong Gao FCVT_2OP(vfrintrne_d, 64, UD, float_round_nearest_even) 2449399665d2SSong Gao FCVT_2OP(vfrintrz_s, 32, UW, float_round_to_zero) 2450399665d2SSong Gao FCVT_2OP(vfrintrz_d, 64, UD, float_round_to_zero) 2451399665d2SSong Gao FCVT_2OP(vfrintrp_s, 32, UW, float_round_up) 2452399665d2SSong Gao FCVT_2OP(vfrintrp_d, 64, UD, float_round_up) 2453399665d2SSong Gao FCVT_2OP(vfrintrm_s, 32, UW, float_round_down) 2454399665d2SSong Gao FCVT_2OP(vfrintrm_d, 64, UD, float_round_down) 2455399665d2SSong Gao 2456399665d2SSong Gao #define FTINT(NAME, FMT1, FMT2, T1, T2, MODE) \ 2457399665d2SSong Gao static T2 do_ftint ## NAME(CPULoongArchState *env, T1 fj) \ 2458399665d2SSong Gao { \ 2459399665d2SSong Gao T2 fd; \ 2460399665d2SSong Gao FloatRoundMode old_mode = get_float_rounding_mode(&env->fp_status); \ 2461399665d2SSong Gao \ 2462399665d2SSong Gao set_float_rounding_mode(MODE, &env->fp_status); \ 2463399665d2SSong Gao fd = do_## FMT1 ##_to_## FMT2(env, fj); \ 2464399665d2SSong Gao set_float_rounding_mode(old_mode, &env->fp_status); \ 2465399665d2SSong Gao return fd; \ 2466399665d2SSong Gao } 2467399665d2SSong Gao 2468399665d2SSong Gao #define DO_FTINT(FMT1, FMT2, T1, T2) \ 2469399665d2SSong Gao static T2 do_## FMT1 ##_to_## FMT2(CPULoongArchState *env, T1 fj) \ 2470399665d2SSong Gao { \ 2471399665d2SSong Gao T2 fd; \ 2472399665d2SSong Gao \ 2473399665d2SSong Gao fd = FMT1 ##_to_## FMT2(fj, &env->fp_status); \ 2474399665d2SSong Gao if (get_float_exception_flags(&env->fp_status) & (float_flag_invalid)) { \ 2475399665d2SSong Gao if (FMT1 ##_is_any_nan(fj)) { \ 2476399665d2SSong Gao fd = 0; \ 2477399665d2SSong Gao } \ 2478399665d2SSong Gao } \ 2479399665d2SSong Gao vec_update_fcsr0(env, GETPC()); \ 2480399665d2SSong Gao return fd; \ 2481399665d2SSong Gao } 2482399665d2SSong Gao 2483399665d2SSong Gao DO_FTINT(float32, int32, uint32_t, uint32_t) 2484399665d2SSong Gao DO_FTINT(float64, int64, uint64_t, uint64_t) 2485399665d2SSong Gao DO_FTINT(float32, uint32, uint32_t, uint32_t) 2486399665d2SSong Gao DO_FTINT(float64, uint64, uint64_t, uint64_t) 2487399665d2SSong Gao DO_FTINT(float64, int32, uint64_t, uint32_t) 2488399665d2SSong Gao DO_FTINT(float32, int64, uint32_t, uint64_t) 2489399665d2SSong Gao 2490399665d2SSong Gao FTINT(rne_w_s, float32, int32, uint32_t, uint32_t, float_round_nearest_even) 2491399665d2SSong Gao FTINT(rne_l_d, float64, int64, uint64_t, uint64_t, float_round_nearest_even) 2492399665d2SSong Gao FTINT(rp_w_s, float32, int32, uint32_t, uint32_t, float_round_up) 2493399665d2SSong Gao FTINT(rp_l_d, float64, int64, uint64_t, uint64_t, float_round_up) 2494399665d2SSong Gao FTINT(rz_w_s, float32, int32, uint32_t, uint32_t, float_round_to_zero) 2495399665d2SSong Gao FTINT(rz_l_d, float64, int64, uint64_t, uint64_t, float_round_to_zero) 2496399665d2SSong Gao FTINT(rm_w_s, float32, int32, uint32_t, uint32_t, float_round_down) 2497399665d2SSong Gao FTINT(rm_l_d, float64, int64, uint64_t, uint64_t, float_round_down) 2498399665d2SSong Gao 2499399665d2SSong Gao DO_2OP_F(vftintrne_w_s, 32, UW, do_ftintrne_w_s) 2500399665d2SSong Gao DO_2OP_F(vftintrne_l_d, 64, UD, do_ftintrne_l_d) 2501399665d2SSong Gao DO_2OP_F(vftintrp_w_s, 32, UW, do_ftintrp_w_s) 2502399665d2SSong Gao DO_2OP_F(vftintrp_l_d, 64, UD, do_ftintrp_l_d) 2503399665d2SSong Gao DO_2OP_F(vftintrz_w_s, 32, UW, do_ftintrz_w_s) 2504399665d2SSong Gao DO_2OP_F(vftintrz_l_d, 64, UD, do_ftintrz_l_d) 2505399665d2SSong Gao DO_2OP_F(vftintrm_w_s, 32, UW, do_ftintrm_w_s) 2506399665d2SSong Gao DO_2OP_F(vftintrm_l_d, 64, UD, do_ftintrm_l_d) 2507399665d2SSong Gao DO_2OP_F(vftint_w_s, 32, UW, do_float32_to_int32) 2508399665d2SSong Gao DO_2OP_F(vftint_l_d, 64, UD, do_float64_to_int64) 2509399665d2SSong Gao 2510399665d2SSong Gao FTINT(rz_wu_s, float32, uint32, uint32_t, uint32_t, float_round_to_zero) 2511399665d2SSong Gao FTINT(rz_lu_d, float64, uint64, uint64_t, uint64_t, float_round_to_zero) 2512399665d2SSong Gao 2513399665d2SSong Gao DO_2OP_F(vftintrz_wu_s, 32, UW, do_ftintrz_wu_s) 2514399665d2SSong Gao DO_2OP_F(vftintrz_lu_d, 64, UD, do_ftintrz_lu_d) 2515399665d2SSong Gao DO_2OP_F(vftint_wu_s, 32, UW, do_float32_to_uint32) 2516399665d2SSong Gao DO_2OP_F(vftint_lu_d, 64, UD, do_float64_to_uint64) 2517399665d2SSong Gao 2518399665d2SSong Gao FTINT(rm_w_d, float64, int32, uint64_t, uint32_t, float_round_down) 2519399665d2SSong Gao FTINT(rp_w_d, float64, int32, uint64_t, uint32_t, float_round_up) 2520399665d2SSong Gao FTINT(rz_w_d, float64, int32, uint64_t, uint32_t, float_round_to_zero) 2521399665d2SSong Gao FTINT(rne_w_d, float64, int32, uint64_t, uint32_t, float_round_nearest_even) 2522399665d2SSong Gao 2523399665d2SSong Gao #define FTINT_W_D(NAME, FN) \ 25243b286753SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, \ 25253b286753SSong Gao CPULoongArchState *env, uint32_t desc) \ 2526399665d2SSong Gao { \ 2527399665d2SSong Gao int i; \ 2528399665d2SSong Gao VReg temp; \ 25293b286753SSong Gao VReg *Vd = (VReg *)vd; \ 25303b286753SSong Gao VReg *Vj = (VReg *)vj; \ 25313b286753SSong Gao VReg *Vk = (VReg *)vk; \ 2532399665d2SSong Gao \ 2533399665d2SSong Gao vec_clear_cause(env); \ 2534399665d2SSong Gao for (i = 0; i < 2; i++) { \ 2535399665d2SSong Gao temp.W(i + 2) = FN(env, Vj->UD(i)); \ 2536399665d2SSong Gao temp.W(i) = FN(env, Vk->UD(i)); \ 2537399665d2SSong Gao } \ 2538399665d2SSong Gao *Vd = temp; \ 2539399665d2SSong Gao } 2540399665d2SSong Gao 2541399665d2SSong Gao FTINT_W_D(vftint_w_d, do_float64_to_int32) 2542399665d2SSong Gao FTINT_W_D(vftintrm_w_d, do_ftintrm_w_d) 2543399665d2SSong Gao FTINT_W_D(vftintrp_w_d, do_ftintrp_w_d) 2544399665d2SSong Gao FTINT_W_D(vftintrz_w_d, do_ftintrz_w_d) 2545399665d2SSong Gao FTINT_W_D(vftintrne_w_d, do_ftintrne_w_d) 2546399665d2SSong Gao 2547399665d2SSong Gao FTINT(rml_l_s, float32, int64, uint32_t, uint64_t, float_round_down) 2548399665d2SSong Gao FTINT(rpl_l_s, float32, int64, uint32_t, uint64_t, float_round_up) 2549399665d2SSong Gao FTINT(rzl_l_s, float32, int64, uint32_t, uint64_t, float_round_to_zero) 2550399665d2SSong Gao FTINT(rnel_l_s, float32, int64, uint32_t, uint64_t, float_round_nearest_even) 2551399665d2SSong Gao FTINT(rmh_l_s, float32, int64, uint32_t, uint64_t, float_round_down) 2552399665d2SSong Gao FTINT(rph_l_s, float32, int64, uint32_t, uint64_t, float_round_up) 2553399665d2SSong Gao FTINT(rzh_l_s, float32, int64, uint32_t, uint64_t, float_round_to_zero) 2554399665d2SSong Gao FTINT(rneh_l_s, float32, int64, uint32_t, uint64_t, float_round_nearest_even) 2555399665d2SSong Gao 2556399665d2SSong Gao #define FTINTL_L_S(NAME, FN) \ 2557226bf881SSong Gao void HELPER(NAME)(void *vd, void *vj, \ 2558226bf881SSong Gao CPULoongArchState *env, uint32_t desc) \ 2559399665d2SSong Gao { \ 2560399665d2SSong Gao int i; \ 2561399665d2SSong Gao VReg temp; \ 2562226bf881SSong Gao VReg *Vd = (VReg *)vd; \ 2563226bf881SSong Gao VReg *Vj = (VReg *)vj; \ 2564399665d2SSong Gao \ 2565399665d2SSong Gao vec_clear_cause(env); \ 2566399665d2SSong Gao for (i = 0; i < 2; i++) { \ 2567399665d2SSong Gao temp.D(i) = FN(env, Vj->UW(i)); \ 2568399665d2SSong Gao } \ 2569399665d2SSong Gao *Vd = temp; \ 2570399665d2SSong Gao } 2571399665d2SSong Gao 2572399665d2SSong Gao FTINTL_L_S(vftintl_l_s, do_float32_to_int64) 2573399665d2SSong Gao FTINTL_L_S(vftintrml_l_s, do_ftintrml_l_s) 2574399665d2SSong Gao FTINTL_L_S(vftintrpl_l_s, do_ftintrpl_l_s) 2575399665d2SSong Gao FTINTL_L_S(vftintrzl_l_s, do_ftintrzl_l_s) 2576399665d2SSong Gao FTINTL_L_S(vftintrnel_l_s, do_ftintrnel_l_s) 2577399665d2SSong Gao 2578399665d2SSong Gao #define FTINTH_L_S(NAME, FN) \ 2579226bf881SSong Gao void HELPER(NAME)(void *vd, void *vj, \ 2580226bf881SSong Gao CPULoongArchState *env, uint32_t desc) \ 2581399665d2SSong Gao { \ 2582399665d2SSong Gao int i; \ 2583399665d2SSong Gao VReg temp; \ 2584226bf881SSong Gao VReg *Vd = (VReg *)vd; \ 2585226bf881SSong Gao VReg *Vj = (VReg *)vj; \ 2586399665d2SSong Gao \ 2587399665d2SSong Gao vec_clear_cause(env); \ 2588399665d2SSong Gao for (i = 0; i < 2; i++) { \ 2589399665d2SSong Gao temp.D(i) = FN(env, Vj->UW(i + 2)); \ 2590399665d2SSong Gao } \ 2591399665d2SSong Gao *Vd = temp; \ 2592399665d2SSong Gao } 2593399665d2SSong Gao 2594399665d2SSong Gao FTINTH_L_S(vftinth_l_s, do_float32_to_int64) 2595399665d2SSong Gao FTINTH_L_S(vftintrmh_l_s, do_ftintrmh_l_s) 2596399665d2SSong Gao FTINTH_L_S(vftintrph_l_s, do_ftintrph_l_s) 2597399665d2SSong Gao FTINTH_L_S(vftintrzh_l_s, do_ftintrzh_l_s) 2598399665d2SSong Gao FTINTH_L_S(vftintrneh_l_s, do_ftintrneh_l_s) 2599399665d2SSong Gao 2600399665d2SSong Gao #define FFINT(NAME, FMT1, FMT2, T1, T2) \ 2601399665d2SSong Gao static T2 do_ffint_ ## NAME(CPULoongArchState *env, T1 fj) \ 2602399665d2SSong Gao { \ 2603399665d2SSong Gao T2 fd; \ 2604399665d2SSong Gao \ 2605399665d2SSong Gao fd = FMT1 ##_to_## FMT2(fj, &env->fp_status); \ 2606399665d2SSong Gao vec_update_fcsr0(env, GETPC()); \ 2607399665d2SSong Gao return fd; \ 2608399665d2SSong Gao } 2609399665d2SSong Gao 2610399665d2SSong Gao FFINT(s_w, int32, float32, int32_t, uint32_t) 2611399665d2SSong Gao FFINT(d_l, int64, float64, int64_t, uint64_t) 2612399665d2SSong Gao FFINT(s_wu, uint32, float32, uint32_t, uint32_t) 2613399665d2SSong Gao FFINT(d_lu, uint64, float64, uint64_t, uint64_t) 2614399665d2SSong Gao 2615399665d2SSong Gao DO_2OP_F(vffint_s_w, 32, W, do_ffint_s_w) 2616399665d2SSong Gao DO_2OP_F(vffint_d_l, 64, D, do_ffint_d_l) 2617399665d2SSong Gao DO_2OP_F(vffint_s_wu, 32, UW, do_ffint_s_wu) 2618399665d2SSong Gao DO_2OP_F(vffint_d_lu, 64, UD, do_ffint_d_lu) 2619399665d2SSong Gao 2620226bf881SSong Gao void HELPER(vffintl_d_w)(void *vd, void *vj, 2621226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2622399665d2SSong Gao { 2623399665d2SSong Gao int i; 2624399665d2SSong Gao VReg temp; 2625226bf881SSong Gao VReg *Vd = (VReg *)vd; 2626226bf881SSong Gao VReg *Vj = (VReg *)vj; 2627399665d2SSong Gao 2628399665d2SSong Gao vec_clear_cause(env); 2629399665d2SSong Gao for (i = 0; i < 2; i++) { 2630399665d2SSong Gao temp.D(i) = int32_to_float64(Vj->W(i), &env->fp_status); 2631399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2632399665d2SSong Gao } 2633399665d2SSong Gao *Vd = temp; 2634399665d2SSong Gao } 2635399665d2SSong Gao 2636226bf881SSong Gao void HELPER(vffinth_d_w)(void *vd, void *vj, 2637226bf881SSong Gao CPULoongArchState *env, uint32_t desc) 2638399665d2SSong Gao { 2639399665d2SSong Gao int i; 2640399665d2SSong Gao VReg temp; 2641226bf881SSong Gao VReg *Vd = (VReg *)vd; 2642226bf881SSong Gao VReg *Vj = (VReg *)vj; 2643399665d2SSong Gao 2644399665d2SSong Gao vec_clear_cause(env); 2645399665d2SSong Gao for (i = 0; i < 2; i++) { 2646399665d2SSong Gao temp.D(i) = int32_to_float64(Vj->W(i + 2), &env->fp_status); 2647399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2648399665d2SSong Gao } 2649399665d2SSong Gao *Vd = temp; 2650399665d2SSong Gao } 2651399665d2SSong Gao 26523b286753SSong Gao void HELPER(vffint_s_l)(void *vd, void *vj, void *vk, 26533b286753SSong Gao CPULoongArchState *env, uint32_t desc) 2654399665d2SSong Gao { 2655399665d2SSong Gao int i; 2656399665d2SSong Gao VReg temp; 26573b286753SSong Gao VReg *Vd = (VReg *)vd; 26583b286753SSong Gao VReg *Vj = (VReg *)vj; 26593b286753SSong Gao VReg *Vk = (VReg *)vk; 2660399665d2SSong Gao 2661399665d2SSong Gao vec_clear_cause(env); 2662399665d2SSong Gao for (i = 0; i < 2; i++) { 2663399665d2SSong Gao temp.W(i + 2) = int64_to_float32(Vj->D(i), &env->fp_status); 2664399665d2SSong Gao temp.W(i) = int64_to_float32(Vk->D(i), &env->fp_status); 2665399665d2SSong Gao vec_update_fcsr0(env, GETPC()); 2666399665d2SSong Gao } 2667399665d2SSong Gao *Vd = temp; 2668399665d2SSong Gao } 2669f435e1e5SSong Gao 2670f435e1e5SSong Gao #define VSEQ(a, b) (a == b ? -1 : 0) 2671f435e1e5SSong Gao #define VSLE(a, b) (a <= b ? -1 : 0) 2672f435e1e5SSong Gao #define VSLT(a, b) (a < b ? -1 : 0) 2673f435e1e5SSong Gao 2674f435e1e5SSong Gao #define VCMPI(NAME, BIT, E, DO_OP) \ 2675f435e1e5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t v) \ 2676f435e1e5SSong Gao { \ 2677f435e1e5SSong Gao int i; \ 2678f435e1e5SSong Gao VReg *Vd = (VReg *)vd; \ 2679f435e1e5SSong Gao VReg *Vj = (VReg *)vj; \ 2680f435e1e5SSong Gao typedef __typeof(Vd->E(0)) TD; \ 2681f435e1e5SSong Gao \ 2682f435e1e5SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2683f435e1e5SSong Gao Vd->E(i) = DO_OP(Vj->E(i), (TD)imm); \ 2684f435e1e5SSong Gao } \ 2685f435e1e5SSong Gao } 2686f435e1e5SSong Gao 2687f435e1e5SSong Gao VCMPI(vseqi_b, 8, B, VSEQ) 2688f435e1e5SSong Gao VCMPI(vseqi_h, 16, H, VSEQ) 2689f435e1e5SSong Gao VCMPI(vseqi_w, 32, W, VSEQ) 2690f435e1e5SSong Gao VCMPI(vseqi_d, 64, D, VSEQ) 2691f435e1e5SSong Gao VCMPI(vslei_b, 8, B, VSLE) 2692f435e1e5SSong Gao VCMPI(vslei_h, 16, H, VSLE) 2693f435e1e5SSong Gao VCMPI(vslei_w, 32, W, VSLE) 2694f435e1e5SSong Gao VCMPI(vslei_d, 64, D, VSLE) 2695f435e1e5SSong Gao VCMPI(vslei_bu, 8, UB, VSLE) 2696f435e1e5SSong Gao VCMPI(vslei_hu, 16, UH, VSLE) 2697f435e1e5SSong Gao VCMPI(vslei_wu, 32, UW, VSLE) 2698f435e1e5SSong Gao VCMPI(vslei_du, 64, UD, VSLE) 2699f435e1e5SSong Gao VCMPI(vslti_b, 8, B, VSLT) 2700f435e1e5SSong Gao VCMPI(vslti_h, 16, H, VSLT) 2701f435e1e5SSong Gao VCMPI(vslti_w, 32, W, VSLT) 2702f435e1e5SSong Gao VCMPI(vslti_d, 64, D, VSLT) 2703f435e1e5SSong Gao VCMPI(vslti_bu, 8, UB, VSLT) 2704f435e1e5SSong Gao VCMPI(vslti_hu, 16, UH, VSLT) 2705f435e1e5SSong Gao VCMPI(vslti_wu, 32, UW, VSLT) 2706f435e1e5SSong Gao VCMPI(vslti_du, 64, UD, VSLT) 2707386c4e86SSong Gao 2708386c4e86SSong Gao static uint64_t vfcmp_common(CPULoongArchState *env, 2709386c4e86SSong Gao FloatRelation cmp, uint32_t flags) 2710386c4e86SSong Gao { 2711386c4e86SSong Gao uint64_t ret = 0; 2712386c4e86SSong Gao 2713386c4e86SSong Gao switch (cmp) { 2714386c4e86SSong Gao case float_relation_less: 2715386c4e86SSong Gao ret = (flags & FCMP_LT); 2716386c4e86SSong Gao break; 2717386c4e86SSong Gao case float_relation_equal: 2718386c4e86SSong Gao ret = (flags & FCMP_EQ); 2719386c4e86SSong Gao break; 2720386c4e86SSong Gao case float_relation_greater: 2721386c4e86SSong Gao ret = (flags & FCMP_GT); 2722386c4e86SSong Gao break; 2723386c4e86SSong Gao case float_relation_unordered: 2724386c4e86SSong Gao ret = (flags & FCMP_UN); 2725386c4e86SSong Gao break; 2726386c4e86SSong Gao default: 2727386c4e86SSong Gao g_assert_not_reached(); 2728386c4e86SSong Gao } 2729386c4e86SSong Gao 2730386c4e86SSong Gao if (ret) { 2731386c4e86SSong Gao ret = -1; 2732386c4e86SSong Gao } 2733386c4e86SSong Gao 2734386c4e86SSong Gao return ret; 2735386c4e86SSong Gao } 2736386c4e86SSong Gao 2737386c4e86SSong Gao #define VFCMP(NAME, BIT, E, FN) \ 2738386c4e86SSong Gao void HELPER(NAME)(CPULoongArchState *env, \ 2739386c4e86SSong Gao uint32_t vd, uint32_t vj, uint32_t vk, uint32_t flags) \ 2740386c4e86SSong Gao { \ 2741386c4e86SSong Gao int i; \ 2742386c4e86SSong Gao VReg t; \ 2743386c4e86SSong Gao VReg *Vd = &(env->fpr[vd].vreg); \ 2744386c4e86SSong Gao VReg *Vj = &(env->fpr[vj].vreg); \ 2745386c4e86SSong Gao VReg *Vk = &(env->fpr[vk].vreg); \ 2746386c4e86SSong Gao \ 2747386c4e86SSong Gao vec_clear_cause(env); \ 2748386c4e86SSong Gao for (i = 0; i < LSX_LEN/BIT ; i++) { \ 2749386c4e86SSong Gao FloatRelation cmp; \ 2750386c4e86SSong Gao cmp = FN(Vj->E(i), Vk->E(i), &env->fp_status); \ 2751386c4e86SSong Gao t.E(i) = vfcmp_common(env, cmp, flags); \ 2752386c4e86SSong Gao vec_update_fcsr0(env, GETPC()); \ 2753386c4e86SSong Gao } \ 2754386c4e86SSong Gao *Vd = t; \ 2755386c4e86SSong Gao } 2756386c4e86SSong Gao 2757386c4e86SSong Gao VFCMP(vfcmp_c_s, 32, UW, float32_compare_quiet) 2758386c4e86SSong Gao VFCMP(vfcmp_s_s, 32, UW, float32_compare) 2759386c4e86SSong Gao VFCMP(vfcmp_c_d, 64, UD, float64_compare_quiet) 2760386c4e86SSong Gao VFCMP(vfcmp_s_d, 64, UD, float64_compare) 2761d0dfa19aSSong Gao 2762d0dfa19aSSong Gao void HELPER(vbitseli_b)(void *vd, void *vj, uint64_t imm, uint32_t v) 2763d0dfa19aSSong Gao { 2764d0dfa19aSSong Gao int i; 2765d0dfa19aSSong Gao VReg *Vd = (VReg *)vd; 2766d0dfa19aSSong Gao VReg *Vj = (VReg *)vj; 2767d0dfa19aSSong Gao 2768d0dfa19aSSong Gao for (i = 0; i < 16; i++) { 2769d0dfa19aSSong Gao Vd->B(i) = (~Vd->B(i) & Vj->B(i)) | (Vd->B(i) & imm); 2770d0dfa19aSSong Gao } 2771d0dfa19aSSong Gao } 2772d0dfa19aSSong Gao 2773d0dfa19aSSong Gao /* Copy from target/arm/tcg/sve_helper.c */ 2774d0dfa19aSSong Gao static inline bool do_match2(uint64_t n, uint64_t m0, uint64_t m1, int esz) 2775d0dfa19aSSong Gao { 2776d0dfa19aSSong Gao uint64_t bits = 8 << esz; 2777d0dfa19aSSong Gao uint64_t ones = dup_const(esz, 1); 2778d0dfa19aSSong Gao uint64_t signs = ones << (bits - 1); 2779d0dfa19aSSong Gao uint64_t cmp0, cmp1; 2780d0dfa19aSSong Gao 2781d0dfa19aSSong Gao cmp1 = dup_const(esz, n); 2782d0dfa19aSSong Gao cmp0 = cmp1 ^ m0; 2783d0dfa19aSSong Gao cmp1 = cmp1 ^ m1; 2784d0dfa19aSSong Gao cmp0 = (cmp0 - ones) & ~cmp0; 2785d0dfa19aSSong Gao cmp1 = (cmp1 - ones) & ~cmp1; 2786d0dfa19aSSong Gao return (cmp0 | cmp1) & signs; 2787d0dfa19aSSong Gao } 2788d0dfa19aSSong Gao 2789d0dfa19aSSong Gao #define SETANYEQZ(NAME, MO) \ 2790d0dfa19aSSong Gao void HELPER(NAME)(CPULoongArchState *env, uint32_t cd, uint32_t vj) \ 2791d0dfa19aSSong Gao { \ 2792d0dfa19aSSong Gao VReg *Vj = &(env->fpr[vj].vreg); \ 2793d0dfa19aSSong Gao \ 2794d0dfa19aSSong Gao env->cf[cd & 0x7] = do_match2(0, Vj->D(0), Vj->D(1), MO); \ 2795d0dfa19aSSong Gao } 2796d0dfa19aSSong Gao SETANYEQZ(vsetanyeqz_b, MO_8) 2797d0dfa19aSSong Gao SETANYEQZ(vsetanyeqz_h, MO_16) 2798d0dfa19aSSong Gao SETANYEQZ(vsetanyeqz_w, MO_32) 2799d0dfa19aSSong Gao SETANYEQZ(vsetanyeqz_d, MO_64) 2800d0dfa19aSSong Gao 2801d0dfa19aSSong Gao #define SETALLNEZ(NAME, MO) \ 2802d0dfa19aSSong Gao void HELPER(NAME)(CPULoongArchState *env, uint32_t cd, uint32_t vj) \ 2803d0dfa19aSSong Gao { \ 2804d0dfa19aSSong Gao VReg *Vj = &(env->fpr[vj].vreg); \ 2805d0dfa19aSSong Gao \ 2806d0dfa19aSSong Gao env->cf[cd & 0x7]= !do_match2(0, Vj->D(0), Vj->D(1), MO); \ 2807d0dfa19aSSong Gao } 2808d0dfa19aSSong Gao SETALLNEZ(vsetallnez_b, MO_8) 2809d0dfa19aSSong Gao SETALLNEZ(vsetallnez_h, MO_16) 2810d0dfa19aSSong Gao SETALLNEZ(vsetallnez_w, MO_32) 2811d0dfa19aSSong Gao SETALLNEZ(vsetallnez_d, MO_64) 2812d5e5563cSSong Gao 2813d5e5563cSSong Gao #define VPACKEV(NAME, BIT, E) \ 281404711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 2815d5e5563cSSong Gao { \ 2816d5e5563cSSong Gao int i; \ 2817d5e5563cSSong Gao VReg temp; \ 281804711da1SSong Gao VReg *Vd = (VReg *)vd; \ 281904711da1SSong Gao VReg *Vj = (VReg *)vj; \ 282004711da1SSong Gao VReg *Vk = (VReg *)vk; \ 2821d5e5563cSSong Gao \ 2822d5e5563cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2823d5e5563cSSong Gao temp.E(2 * i + 1) = Vj->E(2 * i); \ 2824d5e5563cSSong Gao temp.E(2 *i) = Vk->E(2 * i); \ 2825d5e5563cSSong Gao } \ 2826d5e5563cSSong Gao *Vd = temp; \ 2827d5e5563cSSong Gao } 2828d5e5563cSSong Gao 2829d5e5563cSSong Gao VPACKEV(vpackev_b, 16, B) 2830d5e5563cSSong Gao VPACKEV(vpackev_h, 32, H) 2831d5e5563cSSong Gao VPACKEV(vpackev_w, 64, W) 2832d5e5563cSSong Gao VPACKEV(vpackev_d, 128, D) 2833d5e5563cSSong Gao 2834d5e5563cSSong Gao #define VPACKOD(NAME, BIT, E) \ 283504711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 2836d5e5563cSSong Gao { \ 2837d5e5563cSSong Gao int i; \ 2838d5e5563cSSong Gao VReg temp; \ 283904711da1SSong Gao VReg *Vd = (VReg *)vd; \ 284004711da1SSong Gao VReg *Vj = (VReg *)vj; \ 284104711da1SSong Gao VReg *Vk = (VReg *)vk; \ 2842d5e5563cSSong Gao \ 2843d5e5563cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2844d5e5563cSSong Gao temp.E(2 * i + 1) = Vj->E(2 * i + 1); \ 2845d5e5563cSSong Gao temp.E(2 * i) = Vk->E(2 * i + 1); \ 2846d5e5563cSSong Gao } \ 2847d5e5563cSSong Gao *Vd = temp; \ 2848d5e5563cSSong Gao } 2849d5e5563cSSong Gao 2850d5e5563cSSong Gao VPACKOD(vpackod_b, 16, B) 2851d5e5563cSSong Gao VPACKOD(vpackod_h, 32, H) 2852d5e5563cSSong Gao VPACKOD(vpackod_w, 64, W) 2853d5e5563cSSong Gao VPACKOD(vpackod_d, 128, D) 2854d5e5563cSSong Gao 2855d5e5563cSSong Gao #define VPICKEV(NAME, BIT, E) \ 285604711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 2857d5e5563cSSong Gao { \ 2858d5e5563cSSong Gao int i; \ 2859d5e5563cSSong Gao VReg temp; \ 286004711da1SSong Gao VReg *Vd = (VReg *)vd; \ 286104711da1SSong Gao VReg *Vj = (VReg *)vj; \ 286204711da1SSong Gao VReg *Vk = (VReg *)vk; \ 2863d5e5563cSSong Gao \ 2864d5e5563cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2865d5e5563cSSong Gao temp.E(i + LSX_LEN/BIT) = Vj->E(2 * i); \ 2866d5e5563cSSong Gao temp.E(i) = Vk->E(2 * i); \ 2867d5e5563cSSong Gao } \ 2868d5e5563cSSong Gao *Vd = temp; \ 2869d5e5563cSSong Gao } 2870d5e5563cSSong Gao 2871d5e5563cSSong Gao VPICKEV(vpickev_b, 16, B) 2872d5e5563cSSong Gao VPICKEV(vpickev_h, 32, H) 2873d5e5563cSSong Gao VPICKEV(vpickev_w, 64, W) 2874d5e5563cSSong Gao VPICKEV(vpickev_d, 128, D) 2875d5e5563cSSong Gao 2876d5e5563cSSong Gao #define VPICKOD(NAME, BIT, E) \ 287704711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 2878d5e5563cSSong Gao { \ 2879d5e5563cSSong Gao int i; \ 2880d5e5563cSSong Gao VReg temp; \ 288104711da1SSong Gao VReg *Vd = (VReg *)vd; \ 288204711da1SSong Gao VReg *Vj = (VReg *)vj; \ 288304711da1SSong Gao VReg *Vk = (VReg *)vk; \ 2884d5e5563cSSong Gao \ 2885d5e5563cSSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2886d5e5563cSSong Gao temp.E(i + LSX_LEN/BIT) = Vj->E(2 * i + 1); \ 2887d5e5563cSSong Gao temp.E(i) = Vk->E(2 * i + 1); \ 2888d5e5563cSSong Gao } \ 2889d5e5563cSSong Gao *Vd = temp; \ 2890d5e5563cSSong Gao } 2891d5e5563cSSong Gao 2892d5e5563cSSong Gao VPICKOD(vpickod_b, 16, B) 2893d5e5563cSSong Gao VPICKOD(vpickod_h, 32, H) 2894d5e5563cSSong Gao VPICKOD(vpickod_w, 64, W) 2895d5e5563cSSong Gao VPICKOD(vpickod_d, 128, D) 2896e93dd431SSong Gao 2897e93dd431SSong Gao #define VILVL(NAME, BIT, E) \ 289804711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 2899e93dd431SSong Gao { \ 2900e93dd431SSong Gao int i; \ 2901e93dd431SSong Gao VReg temp; \ 290204711da1SSong Gao VReg *Vd = (VReg *)vd; \ 290304711da1SSong Gao VReg *Vj = (VReg *)vj; \ 290404711da1SSong Gao VReg *Vk = (VReg *)vk; \ 2905e93dd431SSong Gao \ 2906e93dd431SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2907e93dd431SSong Gao temp.E(2 * i + 1) = Vj->E(i); \ 2908e93dd431SSong Gao temp.E(2 * i) = Vk->E(i); \ 2909e93dd431SSong Gao } \ 2910e93dd431SSong Gao *Vd = temp; \ 2911e93dd431SSong Gao } 2912e93dd431SSong Gao 2913e93dd431SSong Gao VILVL(vilvl_b, 16, B) 2914e93dd431SSong Gao VILVL(vilvl_h, 32, H) 2915e93dd431SSong Gao VILVL(vilvl_w, 64, W) 2916e93dd431SSong Gao VILVL(vilvl_d, 128, D) 2917e93dd431SSong Gao 2918e93dd431SSong Gao #define VILVH(NAME, BIT, E) \ 291904711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 2920e93dd431SSong Gao { \ 2921e93dd431SSong Gao int i; \ 2922e93dd431SSong Gao VReg temp; \ 292304711da1SSong Gao VReg *Vd = (VReg *)vd; \ 292404711da1SSong Gao VReg *Vj = (VReg *)vj; \ 292504711da1SSong Gao VReg *Vk = (VReg *)vk; \ 2926e93dd431SSong Gao \ 2927e93dd431SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2928e93dd431SSong Gao temp.E(2 * i + 1) = Vj->E(i + LSX_LEN/BIT); \ 2929e93dd431SSong Gao temp.E(2 * i) = Vk->E(i + LSX_LEN/BIT); \ 2930e93dd431SSong Gao } \ 2931e93dd431SSong Gao *Vd = temp; \ 2932e93dd431SSong Gao } 2933e93dd431SSong Gao 2934e93dd431SSong Gao VILVH(vilvh_b, 16, B) 2935e93dd431SSong Gao VILVH(vilvh_h, 32, H) 2936e93dd431SSong Gao VILVH(vilvh_w, 64, W) 2937e93dd431SSong Gao VILVH(vilvh_d, 128, D) 2938e93dd431SSong Gao 2939eb48ab22SSong Gao void HELPER(vshuf_b)(void *vd, void *vj, void *vk, void *va, uint32_t desc) 2940e93dd431SSong Gao { 2941e93dd431SSong Gao int i, m; 2942e93dd431SSong Gao VReg temp; 2943eb48ab22SSong Gao VReg *Vd = (VReg *)vd; 2944eb48ab22SSong Gao VReg *Vj = (VReg *)vj; 2945eb48ab22SSong Gao VReg *Vk = (VReg *)vk; 2946eb48ab22SSong Gao VReg *Va = (VReg *)va; 2947e93dd431SSong Gao 2948e93dd431SSong Gao m = LSX_LEN/8; 2949e93dd431SSong Gao for (i = 0; i < m ; i++) { 2950e93dd431SSong Gao uint64_t k = (uint8_t)Va->B(i) % (2 * m); 2951e93dd431SSong Gao temp.B(i) = k < m ? Vk->B(k) : Vj->B(k - m); 2952e93dd431SSong Gao } 2953e93dd431SSong Gao *Vd = temp; 2954e93dd431SSong Gao } 2955e93dd431SSong Gao 2956e93dd431SSong Gao #define VSHUF(NAME, BIT, E) \ 295704711da1SSong Gao void HELPER(NAME)(void *vd, void *vj, void *vk, uint32_t desc) \ 2958e93dd431SSong Gao { \ 2959e93dd431SSong Gao int i, m; \ 2960e93dd431SSong Gao VReg temp; \ 296104711da1SSong Gao VReg *Vd = (VReg *)vd; \ 296204711da1SSong Gao VReg *Vj = (VReg *)vj; \ 296304711da1SSong Gao VReg *Vk = (VReg *)vk; \ 2964e93dd431SSong Gao \ 2965e93dd431SSong Gao m = LSX_LEN/BIT; \ 2966e93dd431SSong Gao for (i = 0; i < m; i++) { \ 2967e93dd431SSong Gao uint64_t k = ((uint8_t) Vd->E(i)) % (2 * m); \ 2968e93dd431SSong Gao temp.E(i) = k < m ? Vk->E(k) : Vj->E(k - m); \ 2969e93dd431SSong Gao } \ 2970e93dd431SSong Gao *Vd = temp; \ 2971e93dd431SSong Gao } 2972e93dd431SSong Gao 2973e93dd431SSong Gao VSHUF(vshuf_h, 16, H) 2974e93dd431SSong Gao VSHUF(vshuf_w, 32, W) 2975e93dd431SSong Gao VSHUF(vshuf_d, 64, D) 2976e93dd431SSong Gao 2977e93dd431SSong Gao #define VSHUF4I(NAME, BIT, E) \ 2978329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 2979e93dd431SSong Gao { \ 2980e93dd431SSong Gao int i; \ 2981e93dd431SSong Gao VReg temp; \ 2982329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 2983329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 2984e93dd431SSong Gao \ 2985e93dd431SSong Gao for (i = 0; i < LSX_LEN/BIT; i++) { \ 2986e93dd431SSong Gao temp.E(i) = Vj->E(((i) & 0xfc) + (((imm) >> \ 2987e93dd431SSong Gao (2 * ((i) & 0x03))) & 0x03)); \ 2988e93dd431SSong Gao } \ 2989e93dd431SSong Gao *Vd = temp; \ 2990e93dd431SSong Gao } 2991e93dd431SSong Gao 2992e93dd431SSong Gao VSHUF4I(vshuf4i_b, 8, B) 2993e93dd431SSong Gao VSHUF4I(vshuf4i_h, 16, H) 2994e93dd431SSong Gao VSHUF4I(vshuf4i_w, 32, W) 2995e93dd431SSong Gao 2996329517d5SSong Gao void HELPER(vshuf4i_d)(void *vd, void *vj, uint64_t imm, uint32_t desc) 2997e93dd431SSong Gao { 2998329517d5SSong Gao VReg *Vd = (VReg *)vd; 2999329517d5SSong Gao VReg *Vj = (VReg *)vj; 3000e93dd431SSong Gao 3001e93dd431SSong Gao VReg temp; 3002e93dd431SSong Gao temp.D(0) = (imm & 2 ? Vj : Vd)->D(imm & 1); 3003e93dd431SSong Gao temp.D(1) = (imm & 8 ? Vj : Vd)->D((imm >> 2) & 1); 3004e93dd431SSong Gao *Vd = temp; 3005e93dd431SSong Gao } 3006e93dd431SSong Gao 3007329517d5SSong Gao void HELPER(vpermi_w)(void *vd, void *vj, uint64_t imm, uint32_t desc) 3008e93dd431SSong Gao { 3009e93dd431SSong Gao VReg temp; 3010329517d5SSong Gao VReg *Vd = (VReg *)vd; 3011329517d5SSong Gao VReg *Vj = (VReg *)vj; 3012e93dd431SSong Gao 3013e93dd431SSong Gao temp.W(0) = Vj->W(imm & 0x3); 3014e93dd431SSong Gao temp.W(1) = Vj->W((imm >> 2) & 0x3); 3015e93dd431SSong Gao temp.W(2) = Vd->W((imm >> 4) & 0x3); 3016e93dd431SSong Gao temp.W(3) = Vd->W((imm >> 6) & 0x3); 3017e93dd431SSong Gao *Vd = temp; 3018e93dd431SSong Gao } 3019e93dd431SSong Gao 3020e93dd431SSong Gao #define VEXTRINS(NAME, BIT, E, MASK) \ 3021329517d5SSong Gao void HELPER(NAME)(void *vd, void *vj, uint64_t imm, uint32_t desc) \ 3022e93dd431SSong Gao { \ 3023e93dd431SSong Gao int ins, extr; \ 3024329517d5SSong Gao VReg *Vd = (VReg *)vd; \ 3025329517d5SSong Gao VReg *Vj = (VReg *)vj; \ 3026e93dd431SSong Gao \ 3027e93dd431SSong Gao ins = (imm >> 4) & MASK; \ 3028e93dd431SSong Gao extr = imm & MASK; \ 3029e93dd431SSong Gao Vd->E(ins) = Vj->E(extr); \ 3030e93dd431SSong Gao } 3031e93dd431SSong Gao 3032e93dd431SSong Gao VEXTRINS(vextrins_b, 8, B, 0xf) 3033e93dd431SSong Gao VEXTRINS(vextrins_h, 16, H, 0x7) 3034e93dd431SSong Gao VEXTRINS(vextrins_w, 32, W, 0x3) 3035e93dd431SSong Gao VEXTRINS(vextrins_d, 64, D, 0x1) 3036