1f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org/* 2f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * Copyright 2011 Christoph Bumiller 3f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * 4f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * Permission is hereby granted, free of charge, to any person obtaining a 5f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * copy of this software and associated documentation files (the "Software"), 6f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * to deal in the Software without restriction, including without limitation 7f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * and/or sell copies of the Software, and to permit persons to whom the 9f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * Software is furnished to do so, subject to the following conditions: 10f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * 11f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * The above copyright notice and this permission notice shall be included in 12f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * all copies or substantial portions of the Software. 13f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * 14f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, 18f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF 19f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 20f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * SOFTWARE. 21f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org */ 22f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 23f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "nv50_ir_target_nv50.h" 24f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 25f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgnamespace nv50_ir { 26f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 27f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgTarget *getTargetNV50(unsigned int chipset) 28f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 29f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return new TargetNV50(chipset); 30f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 31f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 32f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgTargetNV50::TargetNV50(unsigned int card) : Target(true, false) 33f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 34f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org chipset = card; 35f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 36f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org wposMask = 0; 37f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (unsigned int i = 0; i <= SV_LAST; ++i) 38f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org sysvalLocation[i] = ~0; 39f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 40f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org initOpInfo(); 41f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 42f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 43f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#if 0 44f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org// BULTINS / LIBRARY FUNCTIONS: 45f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 46f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org// TODO 47f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgstatic const uint32_t nvc0_builtin_code[] = 48f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 49f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org}; 50f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 51f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgstatic const uint16_t nvc0_builtin_offsets[NV50_BUILTIN_COUNT] = 52f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 53f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org}; 54f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#endif 55f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 56f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgvoid 57f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgTargetNV50::getBuiltinCode(const uint32_t **code, uint32_t *size) const 58f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 59f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org *code = NULL; 60f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org *size = 0; 61f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 62f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 63f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orguint32_t 64f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgTargetNV50::getBuiltinOffset(int builtin) const 65f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 66f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 0; 67f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 68f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 69f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgstruct opProperties 70f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 71f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org operation op; 72f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org unsigned int mNeg : 4; 73f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org unsigned int mAbs : 4; 74f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org unsigned int mNot : 4; 75f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org unsigned int mSat : 4; 76f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org unsigned int fConst : 3; 77f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org unsigned int fShared : 3; 78f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org unsigned int fAttrib : 3; 79f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org unsigned int fImm : 3; 80f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org}; 81f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 82f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgstatic const struct opProperties _initProps[] = 83f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 84f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org // neg abs not sat c[] s[], a[], imm 85f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_ADD, 0x3, 0x0, 0x0, 0x8, 0x2, 0x1, 0x1, 0x2 }, 86f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_SUB, 0x3, 0x0, 0x0, 0x0, 0x2, 0x1, 0x1, 0x2 }, 87f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_MUL, 0x3, 0x0, 0x0, 0x0, 0x2, 0x1, 0x1, 0x2 }, 88f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_MAX, 0x3, 0x3, 0x0, 0x0, 0x2, 0x1, 0x1, 0x0 }, 89f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_MIN, 0x3, 0x3, 0x0, 0x0, 0x2, 0x1, 0x1, 0x0 }, 90f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_MAD, 0x7, 0x0, 0x0, 0x0, 0x6, 0x1, 0x1, 0x0 }, // special constraint 91f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_ABS, 0x0, 0x0, 0x0, 0x0, 0x0, 0x1, 0x1, 0x0 }, 92f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_NEG, 0x0, 0x1, 0x0, 0x0, 0x0, 0x1, 0x1, 0x0 }, 93f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_CVT, 0x1, 0x1, 0x0, 0x8, 0x0, 0x1, 0x1, 0x0 }, 94f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_AND, 0x0, 0x0, 0x3, 0x0, 0x0, 0x0, 0x0, 0x2 }, 95f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_OR, 0x0, 0x0, 0x3, 0x0, 0x0, 0x0, 0x0, 0x2 }, 96f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_XOR, 0x0, 0x0, 0x3, 0x0, 0x0, 0x0, 0x0, 0x2 }, 97f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_SHL, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x2 }, 98f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_SHR, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x2 }, 99f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_SET, 0x3, 0x3, 0x0, 0x0, 0x2, 0x1, 0x1, 0x0 }, 100f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_PREEX2, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 }, 101f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_PRESIN, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 }, 102f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_LG2, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 }, 103f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_RCP, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 }, 104f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_RSQ, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 }, 105f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_DFDX, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 }, 106f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { OP_DFDY, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 }, 107f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org}; 108f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 109f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgvoid TargetNV50::initOpInfo() 110f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 111f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org unsigned int i, j; 112f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 113f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org static const uint32_t commutative[(OP_LAST + 31) / 32] = 114f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { 115f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org // ADD,MAD,MUL,AND,OR,XOR,MAX,MIN 116f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 0x0670ca00, 0x0000003f, 0x00000000 117f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org }; 118f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org static const uint32_t shortForm[(OP_LAST + 31) / 32] = 119f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { 120f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org // MOV,ADD,SUB,MUL,SAD,L/PINTERP,RCP,TEX,TXF 121f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 0x00010e40, 0x00000040, 0x00000498 122f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org }; 123f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org static const operation noDestList[] = 124f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { 125f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org OP_STORE, OP_WRSV, OP_EXPORT, OP_BRA, OP_CALL, OP_RET, OP_EXIT, 126f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org OP_DISCARD, OP_CONT, OP_BREAK, OP_PRECONT, OP_PREBREAK, OP_PRERET, 127f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org OP_JOIN, OP_JOINAT, OP_BRKPT, OP_MEMBAR, OP_EMIT, OP_RESTART, 128f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org OP_QUADON, OP_QUADPOP 129f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org }; 130f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org static const operation noPredList[] = 131f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { 132f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org OP_CALL, OP_PREBREAK, OP_PRERET, OP_QUADON, OP_QUADPOP, OP_JOINAT 133f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org }; 134f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 135f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (i = 0; i < DATA_FILE_COUNT; ++i) 136f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org nativeFileMap[i] = (DataFile)i; 137f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org nativeFileMap[FILE_PREDICATE] = FILE_FLAGS; 138f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 139f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (i = 0; i < OP_LAST; ++i) { 140f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].variants = NULL; 141f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].op = (operation)i; 142f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].srcTypes = 1 << (int)TYPE_F32; 143f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].dstTypes = 1 << (int)TYPE_F32; 144f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].immdBits = 0xffffffff; 145f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].srcNr = operationSrcNr[i]; 146f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 147f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (j = 0; j < opInfo[i].srcNr; ++j) { 148f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].srcMods[j] = 0; 149f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].srcFiles[j] = 1 << (int)FILE_GPR; 150f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 151f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].dstMods = 0; 152f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].dstFiles = 1 << (int)FILE_GPR; 153f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 154f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].hasDest = 1; 155f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].vector = (i >= OP_TEX && i <= OP_TEXCSAA); 156f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].commutative = (commutative[i / 32] >> (i % 32)) & 1; 157f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].pseudo = (i < OP_MOV); 158f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].predicate = !opInfo[i].pseudo; 159f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].flow = (i >= OP_BRA && i <= OP_JOIN); 160f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[i].minEncSize = (shortForm[i / 32] & (1 << (i % 32))) ? 4 : 8; 161f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 162f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (i = 0; i < sizeof(noDestList) / sizeof(noDestList[0]); ++i) 163f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[noDestList[i]].hasDest = 0; 164f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (i = 0; i < sizeof(noPredList) / sizeof(noPredList[0]); ++i) 165f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[noPredList[i]].predicate = 0; 166f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 167f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (i = 0; i < sizeof(_initProps) / sizeof(_initProps[0]); ++i) { 168f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org const struct opProperties *prop = &_initProps[i]; 169f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 170f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (int s = 0; s < 3; ++s) { 171f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (prop->mNeg & (1 << s)) 172f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_NEG; 173f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (prop->mAbs & (1 << s)) 174f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_ABS; 175f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (prop->mNot & (1 << s)) 176f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_NOT; 177f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (prop->fConst & (1 << s)) 178f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_MEMORY_CONST; 179f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (prop->fShared & (1 << s)) 180f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_MEMORY_SHARED; 181f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (prop->fAttrib & (1 << s)) 182f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_SHADER_INPUT; 183f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (prop->fImm & (1 << s)) 184f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_IMMEDIATE; 185f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 186f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (prop->mSat & 8) 187f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org opInfo[prop->op].dstMods = NV50_IR_MOD_SAT; 188f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 189f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 190f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 191f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgunsigned int 192f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgTargetNV50::getFileSize(DataFile file) const 193f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 194f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org switch (file) { 195f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_NULL: return 0; 196f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_GPR: return 256; // in 16-bit units ** 197f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_PREDICATE: return 0; 198f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_FLAGS: return 4; 199f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_ADDRESS: return 4; 200f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_IMMEDIATE: return 0; 201f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_MEMORY_CONST: return 65536; 202f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_SHADER_INPUT: return 0x200; 203f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_SHADER_OUTPUT: return 0x200; 204f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_MEMORY_GLOBAL: return 0xffffffff; 205f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_MEMORY_SHARED: return 16 << 10; 206f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_MEMORY_LOCAL: return 48 << 10; 207f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_SYSTEM_VALUE: return 16; 208f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org default: 209f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org assert(!"invalid file"); 210f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 0; 211f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 212f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org // ** only first 128 units encodable for 16-bit regs 213f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 214f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 215f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgunsigned int 216f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgTargetNV50::getFileUnit(DataFile file) const 217f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 218f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (file == FILE_GPR || file == FILE_ADDRESS) 219f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 1; 220f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (file == FILE_SYSTEM_VALUE) 221f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 2; 222f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 0; 223f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 224f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 225f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orguint32_t 226f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgTargetNV50::getSVAddress(DataFile shaderFile, const Symbol *sym) const 227f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 228f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org switch (sym->reg.data.sv.sv) { 229f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case SV_FACE: 230f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 0x3fc; 231f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case SV_POSITION: 232f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { 233f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org uint32_t addr = sysvalLocation[sym->reg.data.sv.sv]; 234f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (int c = 0; c < sym->reg.data.sv.index; ++c) 235f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (wposMask & (1 << c)) 236f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org addr += 4; 237f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return addr; 238f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 239f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case SV_NCTAID: 240f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 0x8 + 2 * sym->reg.data.sv.index; 241f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case SV_CTAID: 242f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 0xc + 2 * sym->reg.data.sv.index; 243f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case SV_NTID: 244f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 0x2 + 2 * sym->reg.data.sv.index; 245f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case SV_TID: 246f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 0; 247f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org default: 248f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return sysvalLocation[sym->reg.data.sv.sv]; 249f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 250f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 251f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 252f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org// long: rrr, arr, rcr, acr, rrc, arc, gcr, grr 253f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org// short: rr, ar, rc, gr 254f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org// immd: ri, gi 255f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgbool 256f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgTargetNV50::insnCanLoad(const Instruction *i, int s, 257f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org const Instruction *ld) const 258f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 259f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org DataFile sf = ld->src(0).getFile(); 260f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 261f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (sf == FILE_IMMEDIATE && (i->predSrc >= 0 || i->flagsDef >= 0)) 262f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 263f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (s >= opInfo[i->op].srcNr) 264f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 265f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (!(opInfo[i->op].srcFiles[s] & (1 << (int)sf))) 266f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 267f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (s == 2 && i->src(1).getFile() != FILE_GPR) 268f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 269f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 270f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org // NOTE: don't rely on flagsDef 271f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (int d = 0; i->defExists(d); ++d) 272f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (i->def(d).getFile() == FILE_FLAGS) 273f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 274f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 275f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org unsigned mode = 0; 276f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 277f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (int z = 0; z < Target::operationSrcNr[i->op]; ++z) { 278f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org DataFile zf = (z == s) ? sf : i->src(z).getFile(); 279f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org switch (zf) { 280f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_GPR: 281f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org break; 282f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_MEMORY_SHARED: 283f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_SHADER_INPUT: 284f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org mode |= 1 << (z * 2); 285f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org break; 286f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_MEMORY_CONST: 287f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org mode |= 2 << (z * 2); 288f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org break; 289f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_IMMEDIATE: 290f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org mode |= 3 << (z * 2); 291f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org default: 292f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org break; 293f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 294f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 295f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 296f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org switch (mode) { 297f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case 0x00: 298f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case 0x01: 299f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case 0x03: 300f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case 0x08: 301f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case 0x09: 302f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case 0x0c: 303f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case 0x20: 304f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case 0x21: 305f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org break; 306f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case 0x0d: 307f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (ld->bb->getProgram()->getType() != Program::TYPE_GEOMETRY) 308f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 309f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org default: 310f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 311f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 312f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 313f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org uint8_t ldSize; 314f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 315f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if ((i->op == OP_MUL || i->op == OP_MAD) && !isFloatType(i->dType)) { 316f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org // 32-bit MUL will be split into 16-bit MULs 317f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (ld->src(0).isIndirect(0)) 318f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 319f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (sf == FILE_IMMEDIATE) 320f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 321f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ldSize = 2; 322f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } else { 323f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ldSize = typeSizeof(ld->dType); 324f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 325f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 326f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (sf == FILE_IMMEDIATE) 327f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return true; 328f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 329f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 330f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org // Check if memory access is encodable: 331f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 332f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (ldSize < 4 && sf == FILE_SHADER_INPUT) // no < 4-byte aligned a[] access 333f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 334f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (ld->getSrc(0)->reg.data.offset > (int32_t)(127 * ldSize)) 335f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 336f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 337f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (ld->src(0).isIndirect(0)) { 338f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (int z = 0; i->srcExists(z); ++z) 339f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (i->src(z).isIndirect(0)) 340f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 341f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 342f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org // s[] access only possible in CP, $aX always applies 343f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (sf == FILE_MEMORY_SHARED) 344f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return true; 345f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (!ld->bb) // can't check type ... 346f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 347f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org Program::Type pt = ld->bb->getProgram()->getType(); 348f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 349f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org // $aX applies to c[] only in VP, FP, GP if p[] is not accessed 350f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (pt == Program::TYPE_COMPUTE) 351f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 352f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (pt == Program::TYPE_GEOMETRY) { 353f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (sf == FILE_MEMORY_CONST) 354f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return i->src(s).getFile() != FILE_SHADER_INPUT; 355f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return sf == FILE_SHADER_INPUT; 356f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 357f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return sf == FILE_MEMORY_CONST; 358f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 359f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return true; 360f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 361f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 362f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgbool 363f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgTargetNV50::isAccessSupported(DataFile file, DataType ty) const 364f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 365f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (ty == TYPE_B96 || ty == TYPE_NONE) 366f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 367f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (typeSizeof(ty) > 4) 368f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return (file == FILE_MEMORY_LOCAL) || (file == FILE_MEMORY_GLOBAL); 369f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return true; 370f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 371f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 372f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgbool 373f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgTargetNV50::isOpSupported(operation op, DataType ty) const 374f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 375f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (ty == TYPE_F64 && chipset < 0xa0) 376f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 377f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 378f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org switch (op) { 379f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_PRERET: 380f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return chipset >= 0xa0; 381f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_TXG: 382f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return chipset >= 0xa3; 383f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_POW: 384f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_SQRT: 385f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_DIV: 386f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_MOD: 387f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_SET_AND: 388f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_SET_OR: 389f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_SET_XOR: 390f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_SLCT: 391f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_SELP: 392f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_POPCNT: 393f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_INSBF: 394f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_EXTBF: 395f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_EXIT: // want exit modifier instead (on NOP if required) 396f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 397f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_SAD: 398f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return ty == TYPE_S32; 399f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org default: 400f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return true; 401f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 402f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 403f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 404f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgbool 405f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgTargetNV50::isModSupported(const Instruction *insn, int s, Modifier mod) const 406f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 407f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (!isFloatType(insn->dType)) { 408f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org switch (insn->op) { 409f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_ABS: 410f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_NEG: 411f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_CVT: 412f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_CEIL: 413f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_FLOOR: 414f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_TRUNC: 415f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_AND: 416f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_OR: 417f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_XOR: 418f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org break; 419f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_ADD: 420f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (insn->src(s ? 0 : 1).mod.neg()) 421f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 422f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org break; 423f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_SUB: 424f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (s == 0) 425f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return insn->src(1).mod.neg() ? false : true; 426f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org break; 427f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_SET: 428f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (insn->sType != TYPE_F32) 429f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 430f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org break; 431f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org default: 432f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 433f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 434f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 435f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (s > 3) 436f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 437f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return (mod & Modifier(opInfo[insn->op].srcMods[s])) == mod; 438f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 439f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 440f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgbool 441f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgTargetNV50::mayPredicate(const Instruction *insn, const Value *pred) const 442f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 443f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (insn->getPredicate() || insn->flagsSrc >= 0) 444f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 445f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (int s = 0; insn->srcExists(s); ++s) 446f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (insn->src(s).getFile() == FILE_IMMEDIATE) 447f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 448f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return opInfo[insn->op].predicate; 449f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 450f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 451f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgbool 452f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgTargetNV50::isSatSupported(const Instruction *insn) const 453f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 454f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (insn->op == OP_CVT) 455f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return true; 456f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (insn->dType != TYPE_F32) 457f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 458f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return opInfo[insn->op].dstMods & NV50_IR_MOD_SAT; 459f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 460f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 461f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgint TargetNV50::getLatency(const Instruction *i) const 462f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 463f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org // TODO: tune these values 464f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (i->op == OP_LOAD) { 465f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org switch (i->src(0).getFile()) { 466f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_MEMORY_LOCAL: 467f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case FILE_MEMORY_GLOBAL: 468f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 100; // really 400 to 800 469f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org default: 470f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 22; 471f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 472f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 473f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 22; 474f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 475f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 476f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org// These are "inverse" throughput values, i.e. the number of cycles required 477f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org// to issue a specific instruction for a full warp (32 threads). 478f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org// 479f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org// Assuming we have more than 1 warp in flight, a higher issue latency results 480f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org// in a lower result latency since the MP will have spent more time with other 481f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org// warps. 482f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org// This also helps to determine the number of cycles between instructions in 483f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org// a single warp. 484f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org// 485f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgint TargetNV50::getThroughput(const Instruction *i) const 486f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 487f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org // TODO: tune these values 488f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (i->dType == TYPE_F32) { 489f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org switch (i->op) { 490f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_RCP: 491f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_RSQ: 492f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_LG2: 493f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_SIN: 494f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_COS: 495f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_PRESIN: 496f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case OP_PREEX2: 497f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 16; 498f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org default: 499f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 4; 500f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 501f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } else 502f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (i->dType == TYPE_U32 || i->dType == TYPE_S32) { 503f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 4; 504f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } else 505f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (i->dType == TYPE_F64) { 506f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 32; 507f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } else { 508f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return 1; 509f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 510f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 511f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 512f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgstatic void 513f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgrecordLocation(uint16_t *locs, uint8_t *masks, 514f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org const struct nv50_ir_varying *var) 515f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 516f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org uint16_t addr = var->slot[0] * 4; 517f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 518f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org switch (var->sn) { 519f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case TGSI_SEMANTIC_POSITION: locs[SV_POSITION] = addr; break; 520f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case TGSI_SEMANTIC_INSTANCEID: locs[SV_INSTANCE_ID] = addr; break; 521f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case TGSI_SEMANTIC_VERTEXID: locs[SV_VERTEX_ID] = addr; break; 522f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case TGSI_SEMANTIC_PRIMID: locs[SV_PRIMITIVE_ID] = addr; break; 523f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case NV50_SEMANTIC_LAYER: locs[SV_LAYER] = addr; break; 524f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case NV50_SEMANTIC_VIEWPORTINDEX: locs[SV_VIEWPORT_INDEX] = addr; break; 525f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org default: 526f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org break; 527f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 528f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (var->sn == TGSI_SEMANTIC_POSITION && masks) 529f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org masks[0] = var->mask; 530f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 531f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 532f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgvoid 533f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgTargetNV50::parseDriverInfo(const struct nv50_ir_prog_info *info) 534f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 535f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org unsigned int i; 536f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (i = 0; i < info->numOutputs; ++i) 537f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org recordLocation(sysvalLocation, NULL, &info->out[i]); 538f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (i = 0; i < info->numInputs; ++i) 539f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org recordLocation(sysvalLocation, &wposMask, &info->in[i]); 540f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (i = 0; i < info->numSysVals; ++i) 541f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org recordLocation(sysvalLocation, NULL, &info->sv[i]); 542f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 543f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (sysvalLocation[SV_POSITION] >= 0x200) { 544f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org // not assigned by driver, but we need it internally 545f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org wposMask = 0x8; 546f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org sysvalLocation[SV_POSITION] = 0; 547f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 548f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 549f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 550f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} // namespace nv50_ir 551