nv50_ir_lowering_nvc0.cpp revision f667d15561820ee9dd8e836d43cce3ee52a4780e
1d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller/* 2d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller * Copyright 2011 Christoph Bumiller 3d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller * 4d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller * Permission is hereby granted, free of charge, to any person obtaining a 5d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller * copy of this software and associated documentation files (the "Software"), 6d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller * to deal in the Software without restriction, including without limitation 7d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller * and/or sell copies of the Software, and to permit persons to whom the 9d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller * Software is furnished to do so, subject to the following conditions: 10d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller * 11d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller * The above copyright notice and this permission notice shall be included in 12d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller * all copies or substantial portions of the Software. 13d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller * 14d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 173d8d5b298a268b119d840bc9bae0ee9e0c9244a9Kenneth Graunke * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR 183d8d5b298a268b119d840bc9bae0ee9e0c9244a9Kenneth Graunke * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 193d8d5b298a268b119d840bc9bae0ee9e0c9244a9Kenneth Graunke * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 203d8d5b298a268b119d840bc9bae0ee9e0c9244a9Kenneth Graunke * OTHER DEALINGS IN THE SOFTWARE. 21d2d19ea51fa3575a8d014a69a9b835c335728817Christoph Bumiller */ 2257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 235eb7ff1175a644ffe3b0f1a75cb235400355f9fbJohannes Obermayr#include "codegen/nv50_ir.h" 245eb7ff1175a644ffe3b0f1a75cb235400355f9fbJohannes Obermayr#include "codegen/nv50_ir_build_util.h" 2557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 265eb7ff1175a644ffe3b0f1a75cb235400355f9fbJohannes Obermayr#include "codegen/nv50_ir_target_nvc0.h" 273723ff52237194995d4f9f9fb5d66fb80110889eBen Skeggs#include "codegen/nv50_ir_lowering_nvc0.h" 2857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 2900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller#include <limits> 3000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller 3157594065c30feec9376be9b2132659f7d87362eeChristoph Bumillernamespace nv50_ir { 3257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 3357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller#define QOP_ADD 0 3457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller#define QOP_SUBR 1 3557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller#define QOP_SUB 2 3657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller#define QOP_MOV2 3 3757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 38717f55d79d9709a31e0f85a87f076ac13446701dChristoph Bumiller// UL UR LL LR 3957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller#define QUADOP(q, r, s, t) \ 40717f55d79d9709a31e0f85a87f076ac13446701dChristoph Bumiller ((QOP_##q << 6) | (QOP_##r << 4) | \ 41717f55d79d9709a31e0f85a87f076ac13446701dChristoph Bumiller (QOP_##s << 2) | (QOP_##t << 0)) 4257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 4357594065c30feec9376be9b2132659f7d87362eeChristoph Bumillervoid 4457594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LegalizeSSA::handleDIV(Instruction *i) 4557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 4657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller FlowInstruction *call; 4757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller int builtin; 4857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Value *def[2]; 4957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 5057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.setPosition(i, false); 5157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller def[0] = bld.mkMovToReg(0, i->getSrc(0))->getDef(0); 5257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller def[1] = bld.mkMovToReg(1, i->getSrc(1))->getDef(0); 5357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller switch (i->dType) { 5457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case TYPE_U32: builtin = NVC0_BUILTIN_DIV_U32; break; 5557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case TYPE_S32: builtin = NVC0_BUILTIN_DIV_S32; break; 5657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller default: 5757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return; 5857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 5957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller call = bld.mkFlow(OP_CALL, NULL, CC_ALWAYS, NULL); 6057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkMov(i->getDef(0), def[(i->op == OP_DIV) ? 0 : 1]); 6157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkClobber(FILE_GPR, (i->op == OP_DIV) ? 0xe : 0xd, 2); 6257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkClobber(FILE_PREDICATE, (i->dType == TYPE_S32) ? 0xf : 0x3, 0); 6357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 6457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller call->fixed = 1; 6557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller call->absolute = call->builtin = 1; 6657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller call->target.builtin = builtin; 6757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller delete_Instruction(prog, i); 6857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 6957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 7057594065c30feec9376be9b2132659f7d87362eeChristoph Bumillervoid 7157594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LegalizeSSA::handleRCPRSQ(Instruction *i) 7257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 73b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin assert(i->dType == TYPE_F64); 74b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin // There are instructions that will compute the high 32 bits of the 64-bit 75b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin // float. We will just stick 0 in the bottom 32 bits. 76b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin 77b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin bld.setPosition(i, false); 78b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin 79b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin // 1. Take the source and it up. 80b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin Value *src[2], *dst[2], *def = i->getDef(0); 81b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin bld.mkSplit(src, 4, i->getSrc(0)); 82b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin 83b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin // 2. We don't care about the low 32 bits of the destination. Stick a 0 in. 84b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin dst[0] = bld.loadImm(NULL, 0); 85b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin dst[1] = bld.getSSA(); 86b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin 87b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin // 3. The new version of the instruction takes the high 32 bits of the 88b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin // source and outputs the high 32 bits of the destination. 89b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin i->setSrc(0, src[1]); 90b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin i->setDef(0, dst[1]); 91b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin i->setType(TYPE_F32); 92b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin i->subOp = NV50_IR_SUBOP_RCPRSQ_64H; 93b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin 94b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin // 4. Recombine the two dst pieces back into the original destination. 95b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin bld.setPosition(i, true); 96b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin bld.mkOp2(OP_MERGE, TYPE_U64, def, dst[0], dst[1]); 9757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 9857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 996fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkinvoid 1006fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia MirkinNVC0LegalizeSSA::handleFTZ(Instruction *i) 1016fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin{ 1026fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin // Only want to flush float inputs 103d1eea18a595a468dbc2267a8d14197a3b1a5a4b6Ilia Mirkin assert(i->sType == TYPE_F32); 1046fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin 1056fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin // If we're already flushing denorms (and NaN's) to zero, no need for this. 1066fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin if (i->dnz) 1076fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin return; 1086fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin 1096fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin // Only certain classes of operations can flush 1106fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin OpClass cls = prog->getTarget()->getOpClass(i->op); 1116fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin if (cls != OPCLASS_ARITH && cls != OPCLASS_COMPARE && 1126fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin cls != OPCLASS_CONVERT) 1136fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin return; 1146fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin 1156fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin i->ftz = true; 1166fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin} 1176fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin 11857594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 11957594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LegalizeSSA::visit(Function *fn) 12057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 12157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.setProgram(fn->getProgram()); 12257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 12357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 12457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 12557594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 12657594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LegalizeSSA::visit(BasicBlock *bb) 12757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 12857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Instruction *next; 12957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller for (Instruction *i = bb->getEntry(); i; i = next) { 13057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller next = i->next; 131d1eea18a595a468dbc2267a8d14197a3b1a5a4b6Ilia Mirkin if (i->sType == TYPE_F32) { 1326fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin if (prog->getType() != Program::TYPE_COMPUTE) 1336fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin handleFTZ(i); 13457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller continue; 1356fe0d4f0354418c6e68dd352996e9891ddd4dfd6Ilia Mirkin } 13657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller switch (i->op) { 13757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_DIV: 13857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_MOD: 13957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller handleDIV(i); 14057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller break; 14157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_RCP: 14257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_RSQ: 14357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (i->dType == TYPE_F64) 14457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller handleRCPRSQ(i); 14557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller break; 14657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller default: 14757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller break; 14857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 14957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 15057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 15157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 15257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 153e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph BumillerNVC0LegalizePostRA::NVC0LegalizePostRA(const Program *prog) 15474be77a99e1196d07ebd941aee24313f7aa123c9Vinson Lee : rZero(NULL), 15574be77a99e1196d07ebd941aee24313f7aa123c9Vinson Lee carry(NULL), 15674be77a99e1196d07ebd941aee24313f7aa123c9Vinson Lee needTexBar(prog->getTarget()->getChipset() >= 0xe0) 157e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph Bumiller{ 158e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph Bumiller} 159e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph Bumiller 16057594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 16100fe442253744c4c4e7e68da44d6983da053968bChristoph BumillerNVC0LegalizePostRA::insnDominatedBy(const Instruction *later, 16200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller const Instruction *early) const 16300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller{ 16400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (early->bb == later->bb) 16500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller return early->serial < later->serial; 16600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller return later->bb->dominatedBy(early->bb); 16700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller} 16800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller 16900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumillervoid 17040c224a573f2b763046001e622aafca90f68c693Christoph BumillerNVC0LegalizePostRA::addTexUse(std::list<TexUse> &uses, 171a2af42c1d2dc91f4c31e25ff9fff15a89a9b6eadIlia Mirkin Instruction *usei, const Instruction *texi) 17240c224a573f2b763046001e622aafca90f68c693Christoph Bumiller{ 17340c224a573f2b763046001e622aafca90f68c693Christoph Bumiller bool add = true; 17440c224a573f2b763046001e622aafca90f68c693Christoph Bumiller for (std::list<TexUse>::iterator it = uses.begin(); 17540c224a573f2b763046001e622aafca90f68c693Christoph Bumiller it != uses.end();) { 17640c224a573f2b763046001e622aafca90f68c693Christoph Bumiller if (insnDominatedBy(usei, it->insn)) { 17740c224a573f2b763046001e622aafca90f68c693Christoph Bumiller add = false; 17840c224a573f2b763046001e622aafca90f68c693Christoph Bumiller break; 17940c224a573f2b763046001e622aafca90f68c693Christoph Bumiller } 18040c224a573f2b763046001e622aafca90f68c693Christoph Bumiller if (insnDominatedBy(it->insn, usei)) 18140c224a573f2b763046001e622aafca90f68c693Christoph Bumiller it = uses.erase(it); 18240c224a573f2b763046001e622aafca90f68c693Christoph Bumiller else 18340c224a573f2b763046001e622aafca90f68c693Christoph Bumiller ++it; 18440c224a573f2b763046001e622aafca90f68c693Christoph Bumiller } 18540c224a573f2b763046001e622aafca90f68c693Christoph Bumiller if (add) 186a2af42c1d2dc91f4c31e25ff9fff15a89a9b6eadIlia Mirkin uses.push_back(TexUse(usei, texi)); 18740c224a573f2b763046001e622aafca90f68c693Christoph Bumiller} 18840c224a573f2b763046001e622aafca90f68c693Christoph Bumiller 1897752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin// While it might be tempting to use the an algorithm that just looks at tex 1907752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin// uses, not all texture results are guaranteed to be used on all paths. In 1917752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin// the case where along some control flow path a texture result is never used, 1927752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin// we might reuse that register for something else, creating a 1937752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin// write-after-write hazard. So we have to manually look through all 1947752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin// instructions looking for ones that reference the registers in question. 19540c224a573f2b763046001e622aafca90f68c693Christoph Bumillervoid 1967752bbc44e78e982de3cd4c34862adc38a338234Ilia MirkinNVC0LegalizePostRA::findFirstUses( 1977752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin Instruction *texi, std::list<TexUse> &uses) 19840c224a573f2b763046001e622aafca90f68c693Christoph Bumiller{ 1997752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin int minGPR = texi->def(0).rep()->reg.data.id; 2007752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin int maxGPR = minGPR + texi->def(0).rep()->reg.size / 4 - 1; 20140c224a573f2b763046001e622aafca90f68c693Christoph Bumiller 2027752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin unordered_set<const BasicBlock *> visited; 2037752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin findFirstUsesBB(minGPR, maxGPR, texi->next, texi, uses, visited); 20440c224a573f2b763046001e622aafca90f68c693Christoph Bumiller} 20540c224a573f2b763046001e622aafca90f68c693Christoph Bumiller 20640c224a573f2b763046001e622aafca90f68c693Christoph Bumillervoid 2077752bbc44e78e982de3cd4c34862adc38a338234Ilia MirkinNVC0LegalizePostRA::findFirstUsesBB( 2087752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin int minGPR, int maxGPR, Instruction *start, 2097752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin const Instruction *texi, std::list<TexUse> &uses, 2107752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin unordered_set<const BasicBlock *> &visited) 21100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller{ 2127752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin const BasicBlock *bb = start->bb; 2137752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin 2147752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin // We don't process the whole bb the first time around. This is correct, 2157752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin // however we might be in a loop and hit this BB again, and need to process 2167752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin // the full thing. So only mark a bb as visited if we processed it from the 2177752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin // beginning. 2187752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin if (start == bb->getEntry()) { 2197752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin if (visited.find(bb) != visited.end()) 2207752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin return; 2217752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin visited.insert(bb); 2227752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin } 2237752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin 2247752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin for (Instruction *insn = start; insn != bb->getExit(); insn = insn->next) { 2257752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin if (insn->isNop()) 2267752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin continue; 2277752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin 2287752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin for (int d = 0; insn->defExists(d); ++d) { 2297752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin if (insn->def(d).getFile() != FILE_GPR || 2307752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin insn->def(d).rep()->reg.data.id < minGPR || 2317752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin insn->def(d).rep()->reg.data.id > maxGPR) 232c4bb436f7660c951cd27e52660cf825da68793e5Ilia Mirkin continue; 2337752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin addTexUse(uses, insn, texi); 2347752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin return; 2357752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin } 236c4bb436f7660c951cd27e52660cf825da68793e5Ilia Mirkin 2377752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin for (int s = 0; insn->srcExists(s); ++s) { 2387752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin if (insn->src(s).getFile() != FILE_GPR || 2397752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin insn->src(s).rep()->reg.data.id < minGPR || 2407752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin insn->src(s).rep()->reg.data.id > maxGPR) 2417752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin continue; 2427752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin addTexUse(uses, insn, texi); 2437752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin return; 24400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 24500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 2467752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin 2477752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin for (Graph::EdgeIterator ei = bb->cfg.outgoing(); !ei.end(); ei.next()) { 2487752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin findFirstUsesBB(minGPR, maxGPR, BasicBlock::get(ei.getNode())->getEntry(), 2497752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin texi, uses, visited); 2507752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin } 25100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller} 25200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller 25300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller// Texture barriers: 25400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller// This pass is a bit long and ugly and can probably be optimized. 25500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller// 25600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller// 1. obtain a list of TEXes and their outputs' first use(s) 25700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller// 2. calculate the barrier level of each first use (minimal number of TEXes, 25800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller// over all paths, between the TEX and the use in question) 25900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller// 3. for each barrier, if all paths from the source TEX to that barrier 26000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller// contain a barrier of lesser level, it can be culled 26100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumillerbool 26200fe442253744c4c4e7e68da44d6983da053968bChristoph BumillerNVC0LegalizePostRA::insertTextureBarriers(Function *fn) 26300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller{ 26400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller std::list<TexUse> *uses; 26500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller std::vector<Instruction *> texes; 26600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller std::vector<int> bbFirstTex; 26700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller std::vector<int> bbFirstUse; 26800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller std::vector<int> texCounts; 26900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller std::vector<TexUse> useVec; 27000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller ArrayList insns; 27100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller 27200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller fn->orderInstructions(insns); 27300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller 27400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller texCounts.resize(fn->allBBlocks.getSize(), 0); 27500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller bbFirstTex.resize(fn->allBBlocks.getSize(), insns.getSize()); 27600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller bbFirstUse.resize(fn->allBBlocks.getSize(), insns.getSize()); 27700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller 27800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller // tag BB CFG nodes by their id for later 27900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller for (ArrayList::Iterator i = fn->allBBlocks.iterator(); !i.end(); i.next()) { 28000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller BasicBlock *bb = reinterpret_cast<BasicBlock *>(i.get()); 28100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (bb) 28200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller bb->cfg.tag = bb->getId(); 28300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 28400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller 28500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller // gather the first uses for each TEX 28600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller for (int i = 0; i < insns.getSize(); ++i) { 28700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller Instruction *tex = reinterpret_cast<Instruction *>(insns.get(i)); 28800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (isTextureOp(tex->op)) { 28900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller texes.push_back(tex); 29000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (!texCounts.at(tex->bb->getId())) 29100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller bbFirstTex[tex->bb->getId()] = texes.size() - 1; 29200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller texCounts[tex->bb->getId()]++; 29300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 29400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 29500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller insns.clear(); 29600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (texes.empty()) 29700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller return false; 29800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller uses = new std::list<TexUse>[texes.size()]; 29900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (!uses) 30000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller return false; 301c4bb436f7660c951cd27e52660cf825da68793e5Ilia Mirkin for (size_t i = 0; i < texes.size(); ++i) { 3027752bbc44e78e982de3cd4c34862adc38a338234Ilia Mirkin findFirstUses(texes[i], uses[i]); 303c4bb436f7660c951cd27e52660cf825da68793e5Ilia Mirkin } 30400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller 30500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller // determine the barrier level at each use 30600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller for (size_t i = 0; i < texes.size(); ++i) { 30700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller for (std::list<TexUse>::iterator u = uses[i].begin(); u != uses[i].end(); 30800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller ++u) { 30900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller BasicBlock *tb = texes[i]->bb; 31000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller BasicBlock *ub = u->insn->bb; 31100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (tb == ub) { 31200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller u->level = 0; 31300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller for (size_t j = i + 1; j < texes.size() && 31400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller texes[j]->bb == tb && texes[j]->serial < u->insn->serial; 31500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller ++j) 31600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller u->level++; 31700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } else { 31800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller u->level = fn->cfg.findLightestPathWeight(&tb->cfg, 31900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller &ub->cfg, texCounts); 32000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (u->level < 0) { 32100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller WARN("Failed to find path TEX -> TEXBAR\n"); 32200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller u->level = 0; 32300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller continue; 32400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 32500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller // this counted all TEXes in the origin block, correct that 32600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller u->level -= i - bbFirstTex.at(tb->getId()) + 1 /* this TEX */; 32700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller // and did not count the TEXes in the destination block, add those 32800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller for (size_t j = bbFirstTex.at(ub->getId()); j < texes.size() && 32900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller texes[j]->bb == ub && texes[j]->serial < u->insn->serial; 33000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller ++j) 33100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller u->level++; 33200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 33300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller assert(u->level >= 0); 33400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller useVec.push_back(*u); 33500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 33600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 33700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller delete[] uses; 33800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller 33900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller // insert the barriers 34000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller for (size_t i = 0; i < useVec.size(); ++i) { 34100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller Instruction *prev = useVec[i].insn->prev; 34200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (useVec[i].level < 0) 34300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller continue; 34400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (prev && prev->op == OP_TEXBAR) { 34500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (prev->subOp > useVec[i].level) 34600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller prev->subOp = useVec[i].level; 34700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller prev->setSrc(prev->srcCount(), useVec[i].tex->getDef(0)); 34800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } else { 34900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller Instruction *bar = new_Instruction(func, OP_TEXBAR, TYPE_NONE); 35000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller bar->fixed = 1; 35100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller bar->subOp = useVec[i].level; 35200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller // make use explicit to ease latency calculation 35300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller bar->setSrc(bar->srcCount(), useVec[i].tex->getDef(0)); 35400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller useVec[i].insn->bb->insertBefore(useVec[i].insn, bar); 35500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 35600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 35700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller 3585966903c28a13f13923de308c5f5116a0d5c8cbdIlia Mirkin if (fn->getProgram()->optLevel < 3) 35900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller return true; 36000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller 36100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller std::vector<Limits> limitT, limitB, limitS; // entry, exit, single 36200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller 36300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller limitT.resize(fn->allBBlocks.getSize(), Limits(0, 0)); 36400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller limitB.resize(fn->allBBlocks.getSize(), Limits(0, 0)); 36500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller limitS.resize(fn->allBBlocks.getSize()); 36600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller 36700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller // cull unneeded barriers (should do that earlier, but for simplicity) 36840c224a573f2b763046001e622aafca90f68c693Christoph Bumiller IteratorRef bi = fn->cfg.iteratorCFG(); 36900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller // first calculate min/max outstanding TEXes for each BB 37000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller for (bi->reset(); !bi->end(); bi->next()) { 37100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get()); 37200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller BasicBlock *bb = BasicBlock::get(n); 37300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller int min = 0; 37400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller int max = std::numeric_limits<int>::max(); 37500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller for (Instruction *i = bb->getFirst(); i; i = i->next) { 37600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (isTextureOp(i->op)) { 37700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller min++; 37800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (max < std::numeric_limits<int>::max()) 37900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller max++; 38000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } else 38100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (i->op == OP_TEXBAR) { 38200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller min = MIN2(min, i->subOp); 38300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller max = MIN2(max, i->subOp); 38400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 38500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 38600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller // limits when looking at an isolated block 38700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller limitS[bb->getId()].min = min; 38800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller limitS[bb->getId()].max = max; 38900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 39000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller // propagate the min/max values 39100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller for (unsigned int l = 0; l <= fn->loopNestingBound; ++l) { 39200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller for (bi->reset(); !bi->end(); bi->next()) { 39300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get()); 39400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller BasicBlock *bb = BasicBlock::get(n); 39500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller const int bbId = bb->getId(); 39600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller for (Graph::EdgeIterator ei = n->incident(); !ei.end(); ei.next()) { 39700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller BasicBlock *in = BasicBlock::get(ei.getNode()); 39800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller const int inId = in->getId(); 39900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller limitT[bbId].min = MAX2(limitT[bbId].min, limitB[inId].min); 40000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller limitT[bbId].max = MAX2(limitT[bbId].max, limitB[inId].max); 40100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 40200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller // I just hope this is correct ... 40300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (limitS[bbId].max == std::numeric_limits<int>::max()) { 40400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller // no barrier 40500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller limitB[bbId].min = limitT[bbId].min + limitS[bbId].min; 40600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller limitB[bbId].max = limitT[bbId].max + limitS[bbId].min; 40700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } else { 40800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller // block contained a barrier 40900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller limitB[bbId].min = MIN2(limitS[bbId].max, 41000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller limitT[bbId].min + limitS[bbId].min); 41100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller limitB[bbId].max = MIN2(limitS[bbId].max, 41200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller limitT[bbId].max + limitS[bbId].min); 41300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 41400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 41500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 41600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller // finally delete unnecessary barriers 41700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller for (bi->reset(); !bi->end(); bi->next()) { 41800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller Graph::Node *n = reinterpret_cast<Graph::Node *>(bi->get()); 41900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller BasicBlock *bb = BasicBlock::get(n); 42000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller Instruction *prev = NULL; 42100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller Instruction *next; 42200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller int max = limitT[bb->getId()].max; 42300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller for (Instruction *i = bb->getFirst(); i; i = next) { 42400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller next = i->next; 42500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (i->op == OP_TEXBAR) { 42600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (i->subOp >= max) { 42700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller delete_Instruction(prog, i); 4287086636358b611a2bb124253e1fe870107e1cecbTiziano Bacocco i = NULL; 42900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } else { 43000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller max = i->subOp; 43100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (prev && prev->op == OP_TEXBAR && prev->subOp >= max) { 43200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller delete_Instruction(prog, prev); 43300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller prev = NULL; 43400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 43500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 43600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } else 43700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (isTextureOp(i->op)) { 43800fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller max++; 43900fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 4407086636358b611a2bb124253e1fe870107e1cecbTiziano Bacocco if (i && !i->isNop()) 44100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller prev = i; 44200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 44300fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller } 44400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller return true; 44500fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller} 44600fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller 44700fe442253744c4c4e7e68da44d6983da053968bChristoph Bumillerbool 44857594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LegalizePostRA::visit(Function *fn) 44957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 45000fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller if (needTexBar) 45100fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller insertTextureBarriers(fn); 45200fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller 4533433471e8b46dd9dd042a00f88ef9ad011a94aacChristoph Bumiller rZero = new_LValue(fn, FILE_GPR); 45499e4eba669f13a0dc80880f4f91e2338377c1667Christoph Bumiller carry = new_LValue(fn, FILE_FLAGS); 4554506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 4563433471e8b46dd9dd042a00f88ef9ad011a94aacChristoph Bumiller rZero->reg.data.id = prog->getTarget()->getFileSize(FILE_GPR); 45799e4eba669f13a0dc80880f4f91e2338377c1667Christoph Bumiller carry->reg.data.id = 0; 4584506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 45957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 46057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 46157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 46257594065c30feec9376be9b2132659f7d87362eeChristoph Bumillervoid 46357594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LegalizePostRA::replaceZero(Instruction *i) 46457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 46557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller for (int s = 0; i->srcExists(s); ++s) { 4664506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (s == 2 && i->op == OP_SUCLAMP) 4674506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller continue; 46857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller ImmediateValue *imm = i->getSrc(s)->asImm(); 46957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (imm && imm->reg.data.u64 == 0) 4703433471e8b46dd9dd042a00f88ef9ad011a94aacChristoph Bumiller i->setSrc(s, rZero); 47157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 47257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 47357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 47457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller// replace CONT with BRA for single unconditional continue 47557594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 47657594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LegalizePostRA::tryReplaceContWithBra(BasicBlock *bb) 47757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 47857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (bb->cfg.incidentCount() != 2 || bb->getEntry()->op != OP_PRECONT) 47957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return false; 48057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Graph::EdgeIterator ei = bb->cfg.incident(); 48157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (ei.getType() != Graph::Edge::BACK) 48257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller ei.next(); 48357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (ei.getType() != Graph::Edge::BACK) 48457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return false; 48557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller BasicBlock *contBB = BasicBlock::get(ei.getNode()); 48657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 48757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (!contBB->getExit() || contBB->getExit()->op != OP_CONT || 48857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller contBB->getExit()->getPredicate()) 48957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return false; 49057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller contBB->getExit()->op = OP_BRA; 49157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bb->remove(bb->getEntry()); // delete PRECONT 49257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 49357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller ei.next(); 49457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller assert(ei.end() || ei.getType() != Graph::Edge::BACK); 49557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 49657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 49757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 49857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller// replace branches to join blocks with join ops 49957594065c30feec9376be9b2132659f7d87362eeChristoph Bumillervoid 50057594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LegalizePostRA::propagateJoin(BasicBlock *bb) 50157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 50257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (bb->getEntry()->op != OP_JOIN || bb->getEntry()->asFlow()->limit) 50357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return; 50457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next()) { 50557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller BasicBlock *in = BasicBlock::get(ei.getNode()); 50657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Instruction *exit = in->getExit(); 50757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (!exit) { 50857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller in->insertTail(new FlowInstruction(func, OP_JOIN, bb)); 50957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller // there should always be a terminator instruction 51057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller WARN("inserted missing terminator in BB:%i\n", in->getId()); 51157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } else 51257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (exit->op == OP_BRA) { 51357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller exit->op = OP_JOIN; 51457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller exit->asFlow()->limit = 1; // must-not-propagate marker 51557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 51657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 51757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bb->remove(bb->getEntry()); 51857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 51957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 52057594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 52157594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LegalizePostRA::visit(BasicBlock *bb) 52257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 52357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Instruction *i, *next; 52457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 52557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller // remove pseudo operations and non-fixed no-ops, split 64 bit operations 52657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller for (i = bb->getFirst(); i; i = next) { 52757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller next = i->next; 52857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (i->op == OP_EMIT || i->op == OP_RESTART) { 52957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (!i->getDef(0)->refCount()) 53057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->setDef(0, NULL); 5319362d4bc0a03860ec386156cf499e855a9c2d2a5Christoph Bumiller if (i->src(0).getFile() == FILE_IMMEDIATE) 5323433471e8b46dd9dd042a00f88ef9ad011a94aacChristoph Bumiller i->setSrc(0, rZero); // initial value must be 0 53307d3972b4927841bb892af16ff0389f8a241b24cBen Skeggs replaceZero(i); 53457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } else 53557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (i->isNop()) { 53657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bb->remove(i); 537313940b03cf7c857143b9e3ec0ab969ce4472c83Ilia Mirkin } else 538313940b03cf7c857143b9e3ec0ab969ce4472c83Ilia Mirkin if (i->op == OP_BAR && i->subOp == NV50_IR_SUBOP_BAR_SYNC && 539313940b03cf7c857143b9e3ec0ab969ce4472c83Ilia Mirkin prog->getType() != Program::TYPE_COMPUTE) { 540313940b03cf7c857143b9e3ec0ab969ce4472c83Ilia Mirkin // It seems like barriers are never required for tessellation since 541313940b03cf7c857143b9e3ec0ab969ce4472c83Ilia Mirkin // the warp size is 32, and there are always at most 32 tcs threads. 542313940b03cf7c857143b9e3ec0ab969ce4472c83Ilia Mirkin bb->remove(i); 54337b67db6ae34fb6586d640a7a1b6232f091dd812Ilia Mirkin } else 54437b67db6ae34fb6586d640a7a1b6232f091dd812Ilia Mirkin if (i->op == OP_LOAD && i->subOp == NV50_IR_SUBOP_LDC_IS) { 54537b67db6ae34fb6586d640a7a1b6232f091dd812Ilia Mirkin int offset = i->src(0).get()->reg.data.offset; 54637b67db6ae34fb6586d640a7a1b6232f091dd812Ilia Mirkin if (abs(offset) > 0x10000) 54737b67db6ae34fb6586d640a7a1b6232f091dd812Ilia Mirkin i->src(0).get()->reg.fileIndex += offset >> 16; 54837b67db6ae34fb6586d640a7a1b6232f091dd812Ilia Mirkin i->src(0).get()->reg.data.offset = (int)(short)offset; 54957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } else { 55099e4eba669f13a0dc80880f4f91e2338377c1667Christoph Bumiller // TODO: Move this to before register allocation for operations that 55199e4eba669f13a0dc80880f4f91e2338377c1667Christoph Bumiller // need the $c register ! 55299e4eba669f13a0dc80880f4f91e2338377c1667Christoph Bumiller if (typeSizeof(i->dType) == 8) { 55399e4eba669f13a0dc80880f4f91e2338377c1667Christoph Bumiller Instruction *hi; 55499e4eba669f13a0dc80880f4f91e2338377c1667Christoph Bumiller hi = BuildUtil::split64BitOpPostRA(func, i, rZero, carry); 55599e4eba669f13a0dc80880f4f91e2338377c1667Christoph Bumiller if (hi) 55699e4eba669f13a0dc80880f4f91e2338377c1667Christoph Bumiller next = hi; 55799e4eba669f13a0dc80880f4f91e2338377c1667Christoph Bumiller } 55899e4eba669f13a0dc80880f4f91e2338377c1667Christoph Bumiller 55957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (i->op != OP_MOV && i->op != OP_PFETCH) 56057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller replaceZero(i); 56157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 56257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 56357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (!bb->getEntry()) 56457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 56557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 56657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (!tryReplaceContWithBra(bb)) 56757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller propagateJoin(bb); 56857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 56957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 57057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 57157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 57257594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::NVC0LoweringPass(Program *prog) : targ(prog->getTarget()) 57357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 57457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.setProgram(prog); 5754506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller gMemBase = NULL; 57657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 57757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 57857594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 57957594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::visit(Function *fn) 58057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 58157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (prog->getType() == Program::TYPE_GEOMETRY) { 58257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller assert(!strncmp(fn->getName(), "MAIN", 4)); 58357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller // TODO: when we generate actual functions pass this value along somehow 58457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.setPosition(BasicBlock::get(fn->cfg.getRoot()), false); 58557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller gpEmitAddress = bld.loadImm(NULL, 0)->asLValue(); 5862ec5c8feb331af29548e98b0e78e810bbbc7009eChristoph Bumiller if (fn->cfgExit) { 5872ec5c8feb331af29548e98b0e78e810bbbc7009eChristoph Bumiller bld.setPosition(BasicBlock::get(fn->cfgExit)->getExit(), false); 5882ec5c8feb331af29548e98b0e78e810bbbc7009eChristoph Bumiller bld.mkMovToReg(0, gpEmitAddress); 5892ec5c8feb331af29548e98b0e78e810bbbc7009eChristoph Bumiller } 59057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 59157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 59257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 59357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 59457594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 59557594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::visit(BasicBlock *bb) 59657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 59757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 59857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 59957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 6007a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumillerinline Value * 6017a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph BumillerNVC0LoweringPass::loadTexHandle(Value *ptr, unsigned int slot) 6027a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller{ 603d86933e6f42b9c2f5bb617c66c91795c560a9abdSamuel Pitoiset uint8_t b = prog->driver->io.auxCBSlot; 6047a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller uint32_t off = prog->driver->io.texBindBase + slot * 4; 6057a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller return bld. 6067a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr); 6077a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller} 6087a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller 60957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller// move array source to first slot, convert to u16, add indirections 61057594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 61157594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::handleTEX(TexInstruction *i) 61257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 61337a08ddce54d28f90dc8db8e10792d0759938590Christoph Bumiller const int dim = i->tex.target.getDim() + i->tex.target.isCube(); 61437a08ddce54d28f90dc8db8e10792d0759938590Christoph Bumiller const int arg = i->tex.target.getArgCount(); 6154da54c91d24da891c56957f29274e7821c8254f6Christoph Bumiller const int lyr = arg - (i->tex.target.isMS() ? 2 : 1); 61619ba573a57ff6125a26ff9ae94cf43c36129645fIlia Mirkin const int chipset = prog->getTarget()->getChipset(); 61757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 6186eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin /* Only normalize in the non-explicit derivatives case. For explicit 6196eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin * derivatives, this is handled in handleManualTXD. 6206eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin */ 6216eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin if (i->tex.target.isCube() && i->dPdx[0].get() == NULL) { 6226eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin Value *src[3], *val; 6236eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin int c; 6246eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin for (c = 0; c < 3; ++c) 6256eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin src[c] = bld.mkOp1v(OP_ABS, TYPE_F32, bld.getSSA(), i->getSrc(c)); 6266eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin val = bld.getScratch(); 6276eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin bld.mkOp2(OP_MAX, TYPE_F32, val, src[0], src[1]); 6286eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin bld.mkOp2(OP_MAX, TYPE_F32, val, src[2], val); 6296eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin bld.mkOp1(OP_RCP, TYPE_F32, val, val); 6306eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin for (c = 0; c < 3; ++c) { 6316eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin i->setSrc(c, bld.mkOp2v(OP_MUL, TYPE_F32, bld.getSSA(), 6326eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin i->getSrc(c), val)); 6336eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin } 6346eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin } 6356eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin 636f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // Arguments to the TEX instruction are a little insane. Even though the 637f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // encoding is identical between SM20 and SM30, the arguments mean 638f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // different things between Fermi and Kepler+. A lot of arguments are 639f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // optional based on flags passed to the instruction. This summarizes the 640f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // order of things. 641f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // 642f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // Fermi: 643f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // array/indirect 644f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // coords 645f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // sample 646f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // lod bias 647f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // depth compare 648f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // offsets: 649f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // - tg4: 8 bits each, either 2 (1 offset reg) or 8 (2 offset reg) 650f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // - other: 4 bits each, single reg 651f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // 652f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // Kepler+: 653f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // indirect handle 654f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // array (+ offsets for txd in upper 16 bits) 655f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // coords 656f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // sample 657f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // lod bias 658f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // depth compare 659f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin // offsets (same as fermi, except txd which takes it with array) 6600532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin // 6610532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin // Maxwell (tex): 6620532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin // array 6630532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin // coords 6640532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin // indirect handle 6650532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin // sample 6660532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin // lod bias 6670532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin // depth compare 6680532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin // offsets 6690532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin // 6700532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin // Maxwell (txd): 6710532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin // indirect handle 6720532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin // coords 6730532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin // array + offsets 6740532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin // derivatives 675f525bd01d1430a5e33f57805f50fe4e89aa86ae8Ilia Mirkin 67619ba573a57ff6125a26ff9ae94cf43c36129645fIlia Mirkin if (chipset >= NVISA_GK104_CHIPSET) { 6777a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller if (i->tex.rIndirectSrc >= 0 || i->tex.sIndirectSrc >= 0) { 678b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin // XXX this ignores tsc, and assumes a 1:1 mapping 679b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin assert(i->tex.rIndirectSrc >= 0); 680b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin Value *hnd = loadTexHandle( 681b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(), 682b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin i->getIndirectR(), bld.mkImm(2)), 683b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin i->tex.r); 684b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin i->tex.r = 0xff; 685b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin i->tex.s = 0x1f; 686b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin i->setIndirectR(hnd); 687b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin i->setIndirectS(NULL); 68863b850403c90f33c295d3ad6be4ad749d4ea6274Ilia Mirkin } else if (i->tex.r == i->tex.s || i->op == OP_TXF) { 6894506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller i->tex.r += prog->driver->io.texBindBase / 4; 690e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph Bumiller i->tex.s = 0; // only a single cX[] value possible here 691e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph Bumiller } else { 6927a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller Value *hnd = bld.getScratch(); 6937a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller Value *rHnd = loadTexHandle(NULL, i->tex.r); 6947a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller Value *sHnd = loadTexHandle(NULL, i->tex.s); 6957a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller 6967a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller bld.mkOp3(OP_INSBF, TYPE_U32, hnd, rHnd, bld.mkImm(0x1400), sHnd); 6977a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller 6987a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller i->tex.r = 0; // not used for indirect tex 6997a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller i->tex.s = 0; 7007a91d3a2a4c4e7851fdb46465224213ce1874c9bChristoph Bumiller i->setIndirectR(hnd); 701e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph Bumiller } 702e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph Bumiller if (i->tex.target.isArray()) { 703e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph Bumiller LValue *layer = new_LValue(func, FILE_GPR); 7044da54c91d24da891c56957f29274e7821c8254f6Christoph Bumiller Value *src = i->getSrc(lyr); 705e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph Bumiller const int sat = (i->op == OP_TXF) ? 1 : 0; 706e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph Bumiller DataType sTy = (i->op == OP_TXF) ? TYPE_U32 : TYPE_F32; 707e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph Bumiller bld.mkCvt(OP_CVT, TYPE_U16, layer, sTy, src)->saturate = sat; 7080532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin if (i->op != OP_TXD || chipset < NVISA_GM107_CHIPSET) { 7090532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin for (int s = dim; s >= 1; --s) 7100532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin i->setSrc(s, i->getSrc(s - 1)); 7110532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin i->setSrc(0, layer); 7120532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin } else { 7130532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin i->setSrc(dim, layer); 7140532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin } 715e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph Bumiller } 716b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin // Move the indirect reference to the first place 7170532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin if (i->tex.rIndirectSrc >= 0 && ( 7180532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin i->op == OP_TXD || chipset < NVISA_GM107_CHIPSET)) { 719b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin Value *hnd = i->getIndirectR(); 720b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin 721b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin i->setIndirectR(NULL); 722b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin i->moveSources(0, 1); 723b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin i->setSrc(0, hnd); 724b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin i->tex.rIndirectSrc = 0; 725b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin i->tex.sIndirectSrc = -1; 726b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin } 727e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph Bumiller } else 728e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph Bumiller // (nvc0) generate and move the tsc/tic/array source to the front 72919ba573a57ff6125a26ff9ae94cf43c36129645fIlia Mirkin if (i->tex.target.isArray() || i->tex.rIndirectSrc >= 0 || i->tex.sIndirectSrc >= 0) { 73057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller LValue *src = new_LValue(func, FILE_GPR); // 0xttxsaaaa 73157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 7329807a8ddaf3d0b6d8bf8b3e7c0b01cc4c7db4f30Ilia Mirkin Value *ticRel = i->getIndirectR(); 7339807a8ddaf3d0b6d8bf8b3e7c0b01cc4c7db4f30Ilia Mirkin Value *tscRel = i->getIndirectS(); 7349807a8ddaf3d0b6d8bf8b3e7c0b01cc4c7db4f30Ilia Mirkin 735af3619e88043ce85560b8220dc16244f8898a926Ilia Mirkin if (ticRel) { 7369807a8ddaf3d0b6d8bf8b3e7c0b01cc4c7db4f30Ilia Mirkin i->setSrc(i->tex.rIndirectSrc, NULL); 737af3619e88043ce85560b8220dc16244f8898a926Ilia Mirkin if (i->tex.r) 738af3619e88043ce85560b8220dc16244f8898a926Ilia Mirkin ticRel = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(), 739af3619e88043ce85560b8220dc16244f8898a926Ilia Mirkin ticRel, bld.mkImm(i->tex.r)); 740af3619e88043ce85560b8220dc16244f8898a926Ilia Mirkin } 741af3619e88043ce85560b8220dc16244f8898a926Ilia Mirkin if (tscRel) { 7429807a8ddaf3d0b6d8bf8b3e7c0b01cc4c7db4f30Ilia Mirkin i->setSrc(i->tex.sIndirectSrc, NULL); 743af3619e88043ce85560b8220dc16244f8898a926Ilia Mirkin if (i->tex.s) 744af3619e88043ce85560b8220dc16244f8898a926Ilia Mirkin tscRel = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(), 745af3619e88043ce85560b8220dc16244f8898a926Ilia Mirkin tscRel, bld.mkImm(i->tex.s)); 746af3619e88043ce85560b8220dc16244f8898a926Ilia Mirkin } 7479807a8ddaf3d0b6d8bf8b3e7c0b01cc4c7db4f30Ilia Mirkin 7484da54c91d24da891c56957f29274e7821c8254f6Christoph Bumiller Value *arrayIndex = i->tex.target.isArray() ? i->getSrc(lyr) : NULL; 7497d98bfedd73d632041d27ff12ccf7c7be74a2dddIlia Mirkin if (arrayIndex) { 7507d98bfedd73d632041d27ff12ccf7c7be74a2dddIlia Mirkin for (int s = dim; s >= 1; --s) 7517d98bfedd73d632041d27ff12ccf7c7be74a2dddIlia Mirkin i->setSrc(s, i->getSrc(s - 1)); 7527d98bfedd73d632041d27ff12ccf7c7be74a2dddIlia Mirkin i->setSrc(0, arrayIndex); 7537d98bfedd73d632041d27ff12ccf7c7be74a2dddIlia Mirkin } else { 7547d98bfedd73d632041d27ff12ccf7c7be74a2dddIlia Mirkin i->moveSources(0, 1); 7557d98bfedd73d632041d27ff12ccf7c7be74a2dddIlia Mirkin } 75657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 757e4210a42bcfdb19336faa2ad4b807818c71a2982Christoph Bumiller if (arrayIndex) { 758e4210a42bcfdb19336faa2ad4b807818c71a2982Christoph Bumiller int sat = (i->op == OP_TXF) ? 1 : 0; 759e4210a42bcfdb19336faa2ad4b807818c71a2982Christoph Bumiller DataType sTy = (i->op == OP_TXF) ? TYPE_U32 : TYPE_F32; 760e4210a42bcfdb19336faa2ad4b807818c71a2982Christoph Bumiller bld.mkCvt(OP_CVT, TYPE_U16, src, sTy, arrayIndex)->saturate = sat; 761e4210a42bcfdb19336faa2ad4b807818c71a2982Christoph Bumiller } else { 76257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.loadImm(src, 0); 763e4210a42bcfdb19336faa2ad4b807818c71a2982Christoph Bumiller } 76457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 7659807a8ddaf3d0b6d8bf8b3e7c0b01cc4c7db4f30Ilia Mirkin if (ticRel) 76657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkOp3(OP_INSBF, TYPE_U32, src, ticRel, bld.mkImm(0x0917), src); 7679807a8ddaf3d0b6d8bf8b3e7c0b01cc4c7db4f30Ilia Mirkin if (tscRel) 76857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkOp3(OP_INSBF, TYPE_U32, src, tscRel, bld.mkImm(0x0710), src); 76957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 77057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->setSrc(0, src); 77157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 77257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 77319ba573a57ff6125a26ff9ae94cf43c36129645fIlia Mirkin // For nvc0, the sample id has to be in the second operand, as the offset 77419ba573a57ff6125a26ff9ae94cf43c36129645fIlia Mirkin // does. Right now we don't know how to pass both in, and this case can't 77519ba573a57ff6125a26ff9ae94cf43c36129645fIlia Mirkin // happen with OpenGL. On nve0, the sample id is part of the texture 77619ba573a57ff6125a26ff9ae94cf43c36129645fIlia Mirkin // coordinate argument. 77719ba573a57ff6125a26ff9ae94cf43c36129645fIlia Mirkin assert(chipset >= NVISA_GK104_CHIPSET || 77819ba573a57ff6125a26ff9ae94cf43c36129645fIlia Mirkin !i->tex.useOffsets || !i->tex.target.isMS()); 77919ba573a57ff6125a26ff9ae94cf43c36129645fIlia Mirkin 780f782d6e792db2ed7773a2d22866dbcdb1e4062eeIlia Mirkin // offset is between lod and dc 78157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (i->tex.useOffsets) { 78257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller int n, c; 78371c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller int s = i->srcCount(0xff, true); 7848aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin if (i->op != OP_TXD || chipset < NVISA_GK104_CHIPSET) { 7858aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin if (i->tex.target.isShadow()) 7868aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin s--; 7878aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin if (i->srcExists(s)) // move potential predicate out of the way 7888aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin i->moveSources(s, 1); 7898aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin if (i->tex.useOffsets == 4 && i->srcExists(s + 1)) 7908aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin i->moveSources(s + 1, 1); 7918aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin } 792f6579e4b17a6010fadb464b5179dea5779c74968Ilia Mirkin if (i->op == OP_TXG) { 793f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin // Either there is 1 offset, which goes into the 2 low bytes of the 794f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin // first source, or there are 4 offsets, which go into 2 sources (8 795f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin // values, 1 byte each). 796f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin Value *offs[2] = {NULL, NULL}; 797f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin for (n = 0; n < i->tex.useOffsets; n++) { 798f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin for (c = 0; c < 2; ++c) { 799f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin if ((n % 2) == 0 && c == 0) 800f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin offs[n / 2] = i->offset[n][c].get(); 801f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin else 802f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin bld.mkOp3(OP_INSBF, TYPE_U32, 803f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin offs[n / 2], 804f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin i->offset[n][c].get(), 805f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin bld.mkImm(0x800 | ((n * 16 + c * 8) % 32)), 806f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin offs[n / 2]); 807f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin } 808f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin } 809f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin i->setSrc(s, offs[0]); 810f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin if (offs[1]) 811f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin i->setSrc(s + 1, offs[1]); 812f6579e4b17a6010fadb464b5179dea5779c74968Ilia Mirkin } else { 813f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin unsigned imm = 0; 814f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin assert(i->tex.useOffsets == 1); 815f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin for (c = 0; c < 3; ++c) { 816f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin ImmediateValue val; 817fb1afd1ea5fd25d82c75c5c3a2aba0bcb53b6d47Ilia Mirkin if (!i->offset[0][c].getImmediate(val)) 818fb1afd1ea5fd25d82c75c5c3a2aba0bcb53b6d47Ilia Mirkin assert(!"non-immediate offset passed to non-TXG"); 819f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin imm |= (val.reg.data.u32 & 0xf) << (c * 4); 820f3aa999383074d666d6e3f3506e66b0c937904caIlia Mirkin } 8218aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin if (i->op == OP_TXD && chipset >= NVISA_GK104_CHIPSET) { 8228aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin // The offset goes into the upper 16 bits of the array index. So 8238aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin // create it if it's not already there, and INSBF it if it already 8248aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin // is. 825b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin s = (i->tex.rIndirectSrc >= 0) ? 1 : 0; 8260532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin if (chipset >= NVISA_GM107_CHIPSET) 8270532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin s += dim; 8288aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin if (i->tex.target.isArray()) { 8290532a5fd00cdddda0fd1727fb519cb4312f47e83Ilia Mirkin bld.mkOp3(OP_INSBF, TYPE_U32, i->getSrc(s), 8308aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin bld.loadImm(NULL, imm), bld.mkImm(0xc10), 831b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin i->getSrc(s)); 8328aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin } else { 833b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin i->moveSources(s, 1); 834b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin i->setSrc(s, bld.loadImm(NULL, imm << 16)); 8358aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin } 8368aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin } else { 8378aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin i->setSrc(s, bld.loadImm(NULL, imm)); 8388aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin } 839f6579e4b17a6010fadb464b5179dea5779c74968Ilia Mirkin } 84057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 84157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 84219ba573a57ff6125a26ff9ae94cf43c36129645fIlia Mirkin if (chipset >= NVISA_GK104_CHIPSET) { 84371c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller // 84471c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller // If TEX requires more than 4 sources, the 2nd register tuple must be 84571c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller // aligned to 4, even if it consists of just a single 4-byte register. 84671c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller // 84771c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller // XXX HACK: We insert 0 sources to avoid the 5 or 6 regs case. 84871c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller // 84971c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller int s = i->srcCount(0xff, true); 85071c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller if (s > 4 && s < 7) { 85171c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller if (i->srcExists(s)) // move potential predicate out of the way 85271c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller i->moveSources(s, 7 - s); 85371c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller while (s < 7) 85471c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller i->setSrc(s++, bld.loadImm(NULL, 0)); 85571c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller } 85671c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller } 85771c1c8a9b89ca1ecca1857c53cd8c648c9c9a871Christoph Bumiller 85857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 85957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 86057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 86157594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 86257594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::handleManualTXD(TexInstruction *i) 86357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 86457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller static const uint8_t qOps[4][2] = 86557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller { 86657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller { QUADOP(MOV2, ADD, MOV2, ADD), QUADOP(MOV2, MOV2, ADD, ADD) }, // l0 86757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller { QUADOP(SUBR, MOV2, SUBR, MOV2), QUADOP(MOV2, MOV2, ADD, ADD) }, // l1 86857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller { QUADOP(MOV2, ADD, MOV2, ADD), QUADOP(SUBR, SUBR, MOV2, MOV2) }, // l2 86957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller { QUADOP(SUBR, MOV2, SUBR, MOV2), QUADOP(SUBR, SUBR, MOV2, MOV2) }, // l3 87057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller }; 87157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Value *def[4][4]; 87257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Value *crd[3]; 87357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Instruction *tex; 87457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Value *zero = bld.loadImm(bld.getSSA(), 0); 87557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller int l, c; 87669e8b476d07544d6ef06414a1a78ce5c04761fdbIlia Mirkin const int dim = i->tex.target.getDim() + i->tex.target.isCube(); 877f667d15561820ee9dd8e836d43cce3ee52a4780eIlia Mirkin 878f667d15561820ee9dd8e836d43cce3ee52a4780eIlia Mirkin // This function is invoked after handleTEX lowering, so we have to expect 879f667d15561820ee9dd8e836d43cce3ee52a4780eIlia Mirkin // the arguments in the order that the hw wants them. For Fermi, array and 880f667d15561820ee9dd8e836d43cce3ee52a4780eIlia Mirkin // indirect are both in the leading arg, while for Kepler, array and 881f667d15561820ee9dd8e836d43cce3ee52a4780eIlia Mirkin // indirect are separate (and both precede the coordinates). Maxwell is 882f667d15561820ee9dd8e836d43cce3ee52a4780eIlia Mirkin // handled in a separate function. 883f667d15561820ee9dd8e836d43cce3ee52a4780eIlia Mirkin unsigned array; 884f667d15561820ee9dd8e836d43cce3ee52a4780eIlia Mirkin if (targ->getChipset() < NVISA_GK104_CHIPSET) 885f667d15561820ee9dd8e836d43cce3ee52a4780eIlia Mirkin array = i->tex.target.isArray() || i->tex.rIndirectSrc >= 0; 886f667d15561820ee9dd8e836d43cce3ee52a4780eIlia Mirkin else 887f667d15561820ee9dd8e836d43cce3ee52a4780eIlia Mirkin array = i->tex.target.isArray() + (i->tex.rIndirectSrc >= 0); 88857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 88957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->op = OP_TEX; // no need to clone dPdx/dPdy later 89057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 89157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller for (c = 0; c < dim; ++c) 89257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller crd[c] = bld.getScratch(); 89357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 89457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkOp(OP_QUADON, TYPE_NONE, NULL); 89557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller for (l = 0; l < 4; ++l) { 8966eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin Value *src[3], *val; 89757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller // mov coordinates from lane l to all lanes 89857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller for (c = 0; c < dim; ++c) 899afea9bae67208cdb00b27a60c9cb013bf7d6de52Ilia Mirkin bld.mkQuadop(0x00, crd[c], l, i->getSrc(c + array), zero); 90057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller // add dPdx from lane l to lanes dx 90157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller for (c = 0; c < dim; ++c) 90257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkQuadop(qOps[l][0], crd[c], l, i->dPdx[c].get(), crd[c]); 90357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller // add dPdy from lane l to lanes dy 90457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller for (c = 0; c < dim; ++c) 90557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkQuadop(qOps[l][1], crd[c], l, i->dPdy[c].get(), crd[c]); 9066eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin // normalize cube coordinates 9076eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin if (i->tex.target.isCube()) { 9086eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin for (c = 0; c < 3; ++c) 9096eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin src[c] = bld.mkOp1v(OP_ABS, TYPE_F32, bld.getSSA(), crd[c]); 9106eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin val = bld.getScratch(); 9116eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin bld.mkOp2(OP_MAX, TYPE_F32, val, src[0], src[1]); 9126eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin bld.mkOp2(OP_MAX, TYPE_F32, val, src[2], val); 9136eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin bld.mkOp1(OP_RCP, TYPE_F32, val, val); 9146eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin for (c = 0; c < 3; ++c) 9156eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin src[c] = bld.mkOp2v(OP_MUL, TYPE_F32, bld.getSSA(), crd[c], val); 9166eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin } else { 9176eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin for (c = 0; c < dim; ++c) 9186eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin src[c] = crd[c]; 9196eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin } 92057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller // texture 921a05e6a3fa28168d58a13cfb07f7a664e84b925aeFrancisco Jerez bld.insert(tex = cloneForward(func, i)); 92257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller for (c = 0; c < dim; ++c) 9236eeb284e4f74a2fe5ae6cba90f97f219935e24dfIlia Mirkin tex->setSrc(c + array, src[c]); 92457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller // save results 92557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller for (c = 0; i->defExists(c); ++c) { 92657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Instruction *mov; 92757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller def[c][l] = bld.getSSA(); 92857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller mov = bld.mkMov(def[c][l], tex->getDef(c)); 92957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller mov->fixed = 1; 93057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller mov->lanes = 1 << l; 93157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 93257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 93357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkOp(OP_QUADPOP, TYPE_NONE, NULL); 93457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 93557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller for (c = 0; i->defExists(c); ++c) { 93657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Instruction *u = bld.mkOp(OP_UNION, TYPE_U32, i->getDef(c)); 93757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller for (l = 0; l < 4; ++l) 93857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller u->setSrc(l, def[c][l]); 93957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 94057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 94157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->bb->remove(i); 94257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 94357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 94457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 94557594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 94657594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::handleTXD(TexInstruction *txd) 94757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 94869e8b476d07544d6ef06414a1a78ce5c04761fdbIlia Mirkin int dim = txd->tex.target.getDim() + txd->tex.target.isCube(); 9498aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin unsigned arg = txd->tex.target.getArgCount(); 9508aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin unsigned expected_args = arg; 9518aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin const int chipset = prog->getTarget()->getChipset(); 9528aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin 9538aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin if (chipset >= NVISA_GK104_CHIPSET) { 9548aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin if (!txd->tex.target.isArray() && txd->tex.useOffsets) 9558aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin expected_args++; 956b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin if (txd->tex.rIndirectSrc >= 0 || txd->tex.sIndirectSrc >= 0) 957b3cbd862242e0ff75584fef706f2b2a3da8e49f2Ilia Mirkin expected_args++; 9588aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin } else { 9598aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin if (txd->tex.useOffsets) 9608aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin expected_args++; 9617f937875c0289c2ffc2dc8306add72d5de7951efIlia Mirkin if (!txd->tex.target.isArray() && ( 9627f937875c0289c2ffc2dc8306add72d5de7951efIlia Mirkin txd->tex.rIndirectSrc >= 0 || txd->tex.sIndirectSrc >= 0)) 9637f937875c0289c2ffc2dc8306add72d5de7951efIlia Mirkin expected_args++; 9648aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin } 9658aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin 9668aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin if (expected_args > 4 || 9678aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin dim > 2 || 96869e8b476d07544d6ef06414a1a78ce5c04761fdbIlia Mirkin txd->tex.target.isShadow()) 9698aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin txd->op = OP_TEX; 97057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 97157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller handleTEX(txd); 97238a20281fcc2ed244aea0aaa268035533f48a183Christoph Bumiller while (txd->srcExists(arg)) 97357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller ++arg; 97457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 9759c930639d9f6d713ccfd16b390a41a9f584f348cChristoph Bumiller txd->tex.derivAll = true; 9768aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin if (txd->op == OP_TEX) 97757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return handleManualTXD(txd); 97857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 9798aa34dc9cb1f4b1b17e49da98e54066832afc98eIlia Mirkin assert(arg == expected_args); 98057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller for (int c = 0; c < dim; ++c) { 98138a20281fcc2ed244aea0aaa268035533f48a183Christoph Bumiller txd->setSrc(arg + c * 2 + 0, txd->dPdx[c]); 98238a20281fcc2ed244aea0aaa268035533f48a183Christoph Bumiller txd->setSrc(arg + c * 2 + 1, txd->dPdy[c]); 9839362d4bc0a03860ec386156cf499e855a9c2d2a5Christoph Bumiller txd->dPdx[c].set(NULL); 9849362d4bc0a03860ec386156cf499e855a9c2d2a5Christoph Bumiller txd->dPdy[c].set(NULL); 98557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 98657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 98757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 98857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 98957594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 99030cb66cd745fc793a2349f1d17046c50cd51c558Christoph BumillerNVC0LoweringPass::handleTXQ(TexInstruction *txq) 99130cb66cd745fc793a2349f1d17046c50cd51c558Christoph Bumiller{ 9925877a594d54fdd2b3aa329f4d35b3491a7ee8a33Ilia Mirkin const int chipset = prog->getTarget()->getChipset(); 9935877a594d54fdd2b3aa329f4d35b3491a7ee8a33Ilia Mirkin if (chipset >= NVISA_GK104_CHIPSET && txq->tex.rIndirectSrc < 0) 9945877a594d54fdd2b3aa329f4d35b3491a7ee8a33Ilia Mirkin txq->tex.r += prog->driver->io.texBindBase / 4; 9955877a594d54fdd2b3aa329f4d35b3491a7ee8a33Ilia Mirkin 996346ce0b98832e33d5411200002571b3edea9e2bbIlia Mirkin if (txq->tex.rIndirectSrc < 0) 99720e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin return true; 99820e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin 99920e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin Value *ticRel = txq->getIndirectR(); 100020e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin 100120e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin txq->setIndirectS(NULL); 100220e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin txq->tex.sIndirectSrc = -1; 100320e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin 1004346ce0b98832e33d5411200002571b3edea9e2bbIlia Mirkin assert(ticRel); 1005346ce0b98832e33d5411200002571b3edea9e2bbIlia Mirkin 100620e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin if (chipset < NVISA_GK104_CHIPSET) { 100720e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin LValue *src = new_LValue(func, FILE_GPR); // 0xttxsaaaa 100820e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin 1009346ce0b98832e33d5411200002571b3edea9e2bbIlia Mirkin txq->setSrc(txq->tex.rIndirectSrc, NULL); 1010346ce0b98832e33d5411200002571b3edea9e2bbIlia Mirkin if (txq->tex.r) 1011346ce0b98832e33d5411200002571b3edea9e2bbIlia Mirkin ticRel = bld.mkOp2v(OP_ADD, TYPE_U32, bld.getScratch(), 1012346ce0b98832e33d5411200002571b3edea9e2bbIlia Mirkin ticRel, bld.mkImm(txq->tex.r)); 101320e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin 1014346ce0b98832e33d5411200002571b3edea9e2bbIlia Mirkin bld.mkOp2(OP_SHL, TYPE_U32, src, ticRel, bld.mkImm(0x17)); 101520e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin 101620e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin txq->moveSources(0, 1); 101720e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin txq->setSrc(0, src); 101820e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin } else { 101920e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin Value *hnd = loadTexHandle( 102020e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(), 102120e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin txq->getIndirectR(), bld.mkImm(2)), 102220e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin txq->tex.r); 102320e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin txq->tex.r = 0xff; 102420e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin txq->tex.s = 0x1f; 102520e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin 1026b346a84e270a50f0a8f1a6e474a51da04dd72f0eIlia Mirkin txq->setIndirectR(NULL); 1027b346a84e270a50f0a8f1a6e474a51da04dd72f0eIlia Mirkin txq->moveSources(0, 1); 1028b346a84e270a50f0a8f1a6e474a51da04dd72f0eIlia Mirkin txq->setSrc(0, hnd); 1029b346a84e270a50f0a8f1a6e474a51da04dd72f0eIlia Mirkin txq->tex.rIndirectSrc = 0; 103020e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin } 103120e484afa4874e87cd18daffd66286bb893cf3fbIlia Mirkin 103230cb66cd745fc793a2349f1d17046c50cd51c558Christoph Bumiller return true; 103330cb66cd745fc793a2349f1d17046c50cd51c558Christoph Bumiller} 103430cb66cd745fc793a2349f1d17046c50cd51c558Christoph Bumiller 103530cb66cd745fc793a2349f1d17046c50cd51c558Christoph Bumillerbool 1036423f64e83ab5b1ea7de475ae80300a8408522743Ilia MirkinNVC0LoweringPass::handleTXLQ(TexInstruction *i) 1037423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin{ 1038423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin /* The outputs are inverted compared to what the TGSI instruction 1039423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin * expects. Take that into account in the mask. 1040423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin */ 1041423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin assert((i->tex.mask & ~3) == 0); 1042423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin if (i->tex.mask == 1) 1043423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin i->tex.mask = 2; 1044423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin else if (i->tex.mask == 2) 1045423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin i->tex.mask = 1; 1046423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin handleTEX(i); 1047423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin bld.setPosition(i, true); 1048423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin 1049423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin /* The returned values are not quite what we want: 1050423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin * (a) convert from s16/u16 to f32 1051423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin * (b) multiply by 1/256 1052423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin */ 1053423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin for (int def = 0; def < 2; ++def) { 1054423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin if (!i->defExists(def)) 1055423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin continue; 1056423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin enum DataType type = TYPE_S16; 1057423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin if (i->tex.mask == 2 || def > 0) 1058423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin type = TYPE_U16; 1059423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin bld.mkCvt(OP_CVT, TYPE_F32, i->getDef(def), type, i->getDef(def)); 1060423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin bld.mkOp2(OP_MUL, TYPE_F32, i->getDef(def), 1061423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin i->getDef(def), bld.loadImm(NULL, 1.0f / 256)); 1062423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin } 1063423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin if (i->tex.mask == 3) { 1064423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin LValue *t = new_LValue(func, FILE_GPR); 1065423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin bld.mkMov(t, i->getDef(0)); 1066423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin bld.mkMov(i->getDef(0), i->getDef(1)); 1067423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin bld.mkMov(i->getDef(1), t); 1068423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin } 1069423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin return true; 1070423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin} 1071423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin 1072423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkinbool 10732c4eeb0b5cf17caa06cb3fa46d4f64e6a8005d23Ilia MirkinNVC0LoweringPass::handleSUQ(Instruction *suq) 10742c4eeb0b5cf17caa06cb3fa46d4f64e6a8005d23Ilia Mirkin{ 10752c4eeb0b5cf17caa06cb3fa46d4f64e6a8005d23Ilia Mirkin suq->op = OP_MOV; 10767b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin suq->setSrc(0, loadResLength32(suq->getIndirect(0, 1), 10777b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin suq->getSrc(0)->reg.fileIndex * 16)); 10787b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin suq->setIndirect(0, 0, NULL); 10797b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin suq->setIndirect(0, 1, NULL); 10802c4eeb0b5cf17caa06cb3fa46d4f64e6a8005d23Ilia Mirkin return true; 10812c4eeb0b5cf17caa06cb3fa46d4f64e6a8005d23Ilia Mirkin} 10822c4eeb0b5cf17caa06cb3fa46d4f64e6a8005d23Ilia Mirkin 108314a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoisetvoid 108414a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel PitoisetNVC0LoweringPass::handleSharedATOM(Instruction *atom) 108514a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset{ 108614a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset assert(atom->src(0).getFile() == FILE_MEMORY_SHARED); 108714a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 108814a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset BasicBlock *currBB = atom->bb; 108914a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset BasicBlock *tryLockAndSetBB = atom->bb->splitBefore(atom, false); 109014a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset BasicBlock *joinBB = atom->bb->splitAfter(atom); 109114a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 109214a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset bld.setPosition(currBB, true); 109314a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset assert(!currBB->joinAt); 109414a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset currBB->joinAt = bld.mkFlow(OP_JOINAT, joinBB, CC_ALWAYS, NULL); 109514a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 109614a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset bld.mkFlow(OP_BRA, tryLockAndSetBB, CC_ALWAYS, NULL); 109714a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset currBB->cfg.attach(&tryLockAndSetBB->cfg, Graph::Edge::TREE); 109814a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 109914a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset bld.setPosition(tryLockAndSetBB, true); 110014a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 110114a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset Instruction *ld = 110214a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset bld.mkLoad(TYPE_U32, atom->getDef(0), 110314a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset bld.mkSymbol(FILE_MEMORY_SHARED, 0, TYPE_U32, 0), NULL); 110414a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset ld->setDef(1, bld.getSSA(1, FILE_PREDICATE)); 110514a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset ld->subOp = NV50_IR_SUBOP_LOAD_LOCKED; 110614a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 110714a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset Value *stVal; 110814a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset if (atom->subOp == NV50_IR_SUBOP_ATOM_EXCH) { 110914a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset // Read the old value, and write the new one. 111014a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset stVal = atom->getSrc(1); 111114a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset } else if (atom->subOp == NV50_IR_SUBOP_ATOM_CAS) { 111214a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset CmpInstruction *set = 111314a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, bld.getSSA(1, FILE_PREDICATE), 111414a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset TYPE_U32, ld->getDef(0), atom->getSrc(1)); 111514a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset set->setPredicate(CC_P, ld->getDef(1)); 111614a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 11176526225f888a08b301e8c39ec70b4e739081e490Samuel Pitoiset Instruction *selp = 11186526225f888a08b301e8c39ec70b4e739081e490Samuel Pitoiset bld.mkOp3(OP_SELP, TYPE_U32, bld.getSSA(), ld->getDef(0), 11196526225f888a08b301e8c39ec70b4e739081e490Samuel Pitoiset atom->getSrc(2), set->getDef(0)); 11206526225f888a08b301e8c39ec70b4e739081e490Samuel Pitoiset selp->src(2).mod = Modifier(NV50_IR_MOD_NOT); 112114a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset selp->setPredicate(CC_P, ld->getDef(1)); 112214a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 112314a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset stVal = selp->getDef(0); 112414a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset } else { 112514a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset operation op; 112614a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 112714a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset switch (atom->subOp) { 112814a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset case NV50_IR_SUBOP_ATOM_ADD: 112914a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset op = OP_ADD; 113014a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset break; 113114a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset case NV50_IR_SUBOP_ATOM_AND: 113214a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset op = OP_AND; 113314a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset break; 113414a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset case NV50_IR_SUBOP_ATOM_OR: 113514a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset op = OP_OR; 113614a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset break; 113714a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset case NV50_IR_SUBOP_ATOM_XOR: 113814a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset op = OP_XOR; 113914a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset break; 114014a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset case NV50_IR_SUBOP_ATOM_MIN: 114114a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset op = OP_MIN; 114214a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset break; 114314a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset case NV50_IR_SUBOP_ATOM_MAX: 114414a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset op = OP_MAX; 114514a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset break; 114614a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset default: 114714a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset assert(0); 1148e05492fd7f0e1a9454482a9174f5870b8cb5a41eSamuel Pitoiset return; 114914a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset } 115014a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 115114a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset Instruction *i = 115214a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset bld.mkOp2(op, atom->dType, bld.getSSA(), ld->getDef(0), 115314a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset atom->getSrc(1)); 115414a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset i->setPredicate(CC_P, ld->getDef(1)); 115514a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 115614a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset stVal = i->getDef(0); 115714a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset } 115814a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 115914a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset Instruction *st = 116014a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset bld.mkStore(OP_STORE, TYPE_U32, 116114a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset bld.mkSymbol(FILE_MEMORY_SHARED, 0, TYPE_U32, 0), 116214a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset NULL, stVal); 116314a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset st->setPredicate(CC_P, ld->getDef(1)); 116414a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset st->subOp = NV50_IR_SUBOP_STORE_UNLOCKED; 116514a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 116614a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset // Loop until the lock is acquired. 116714a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset bld.mkFlow(OP_BRA, tryLockAndSetBB, CC_NOT_P, ld->getDef(1)); 116814a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset tryLockAndSetBB->cfg.attach(&tryLockAndSetBB->cfg, Graph::Edge::BACK); 116914a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset tryLockAndSetBB->cfg.attach(&joinBB->cfg, Graph::Edge::CROSS); 117014a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset bld.mkFlow(OP_BRA, joinBB, CC_ALWAYS, NULL); 117114a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 117214a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset bld.remove(atom); 117314a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 117414a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset bld.setPosition(joinBB, false); 117514a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset bld.mkFlow(OP_JOIN, NULL, CC_ALWAYS, NULL)->fixed = 1; 117614a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset} 117714a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 11782c4eeb0b5cf17caa06cb3fa46d4f64e6a8005d23Ilia Mirkinbool 1179c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph BumillerNVC0LoweringPass::handleATOM(Instruction *atom) 1180c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller{ 1181c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller SVSemantic sv; 11827b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin Value *ptr = atom->getIndirect(0, 0), *ind = atom->getIndirect(0, 1), *base; 1183c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller 1184c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller switch (atom->src(0).getFile()) { 1185c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller case FILE_MEMORY_LOCAL: 1186c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller sv = SV_LBASE; 1187c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller break; 1188c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller case FILE_MEMORY_SHARED: 118914a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset handleSharedATOM(atom); 119014a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset return true; 1191c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller default: 1192c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller assert(atom->src(0).getFile() == FILE_MEMORY_GLOBAL); 11937b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin base = loadResInfo64(ind, atom->getSrc(0)->reg.fileIndex * 16); 1194c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin assert(base->reg.size == 8); 1195c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin if (ptr) 1196c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin base = bld.mkOp2v(OP_ADD, TYPE_U64, base, base, ptr); 1197c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin assert(base->reg.size == 8); 1198c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin atom->setIndirect(0, 0, base); 1199c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller return true; 1200c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller } 1201c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin base = 1202c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller bld.mkOp1v(OP_RDSV, TYPE_U32, bld.getScratch(), bld.mkSysVal(sv, 0)); 1203c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller 1204c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller atom->setSrc(0, cloneShallow(func, atom->getSrc(0))); 1205c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller atom->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL; 1206c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller if (ptr) 1207c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller base = bld.mkOp2v(OP_ADD, TYPE_U32, base, base, ptr); 12087b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin atom->setIndirect(0, 1, NULL); 1209c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller atom->setIndirect(0, 0, base); 1210c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller 1211c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller return true; 1212c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller} 1213c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller 121475f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumillerbool 121575f1f852b00ad0d766684d01695322b93a2acd55Christoph BumillerNVC0LoweringPass::handleCasExch(Instruction *cas, bool needCctl) 121675f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller{ 121714a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset if (cas->src(0).getFile() == FILE_MEMORY_SHARED) { 121814a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset // ATOM_CAS and ATOM_EXCH are handled in handleSharedATOM(). 121914a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset return false; 122014a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset } 122114a810e9d0dbf52e547ae6e16a68487d7cb92004Samuel Pitoiset 122275f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller if (cas->subOp != NV50_IR_SUBOP_ATOM_CAS && 122375f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller cas->subOp != NV50_IR_SUBOP_ATOM_EXCH) 122475f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller return false; 122575f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller bld.setPosition(cas, true); 122675f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller 122775f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller if (needCctl) { 122875f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller Instruction *cctl = bld.mkOp1(OP_CCTL, TYPE_NONE, NULL, cas->getSrc(0)); 122975f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller cctl->setIndirect(0, 0, cas->getIndirect(0, 0)); 123075f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller cctl->fixed = 1; 123175f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller cctl->subOp = NV50_IR_SUBOP_CCTL_IV; 123275f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller if (cas->isPredicated()) 123375f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller cctl->setPredicate(cas->cc, cas->getPredicate()); 123475f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller } 123575f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller 1236df043f076464d817a9d88c4c43757e65b6eae3f9Ilia Mirkin if (cas->subOp == NV50_IR_SUBOP_ATOM_CAS) { 123775f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller // CAS is crazy. It's 2nd source is a double reg, and the 3rd source 123875f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller // should be set to the high part of the double reg or bad things will 123975f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller // happen elsewhere in the universe. 124075f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller // Also, it sometimes returns the new value instead of the old one 124175f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller // under mysterious circumstances. 124275f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller Value *dreg = bld.getSSA(8); 124375f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller bld.setPosition(cas, false); 124475f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller bld.mkOp2(OP_MERGE, TYPE_U64, dreg, cas->getSrc(1), cas->getSrc(2)); 124575f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller cas->setSrc(1, dreg); 1246df043f076464d817a9d88c4c43757e65b6eae3f9Ilia Mirkin cas->setSrc(2, dreg); 124775f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller } 124875f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller 124975f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller return true; 125075f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller} 125175f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller 12524506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumillerinline Value * 12534506ed28de7f9d76bbc99c0758a7891b84528729Christoph BumillerNVC0LoweringPass::loadResInfo32(Value *ptr, uint32_t off) 12544506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller{ 1255d86933e6f42b9c2f5bb617c66c91795c560a9abdSamuel Pitoiset uint8_t b = prog->driver->io.auxCBSlot; 12564506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller off += prog->driver->io.suInfoBase; 12574506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller return bld. 12584506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr); 12594506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller} 12604506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 12614506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumillerinline Value * 1262c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia MirkinNVC0LoweringPass::loadResInfo64(Value *ptr, uint32_t off) 1263c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin{ 1264d86933e6f42b9c2f5bb617c66c91795c560a9abdSamuel Pitoiset uint8_t b = prog->driver->io.auxCBSlot; 1265c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin off += prog->driver->io.suInfoBase; 1266c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin 12677b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin if (ptr) 12687b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getScratch(), ptr, bld.mkImm(4)); 12697b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin 1270c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin return bld. 1271c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin mkLoadv(TYPE_U64, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U64, off), ptr); 1272c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin} 1273c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin 1274c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkininline Value * 1275c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia MirkinNVC0LoweringPass::loadResLength32(Value *ptr, uint32_t off) 1276c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin{ 1277d86933e6f42b9c2f5bb617c66c91795c560a9abdSamuel Pitoiset uint8_t b = prog->driver->io.auxCBSlot; 1278c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin off += prog->driver->io.suInfoBase; 1279c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin 12807b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin if (ptr) 12817b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getScratch(), ptr, bld.mkImm(4)); 12827b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin 1283c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin return bld. 1284c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U64, off + 8), ptr); 1285c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin} 1286c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin 1287c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkininline Value * 12884506ed28de7f9d76bbc99c0758a7891b84528729Christoph BumillerNVC0LoweringPass::loadMsInfo32(Value *ptr, uint32_t off) 12894506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller{ 12904506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller uint8_t b = prog->driver->io.msInfoCBSlot; 12914506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller off += prog->driver->io.msInfoBase; 12924506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller return bld. 12934506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U32, off), ptr); 12944506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller} 12954506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 12964506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller/* On nvc0, surface info is obtained via the surface binding points passed 12974506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller * to the SULD/SUST instructions. 12984506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller * On nve4, surface info is stored in c[] and is used by various special 12994506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller * instructions, e.g. for clamping coordiantes or generating an address. 13004506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller * They couldn't just have added an equivalent to TIC now, couldn't they ? 13014506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller */ 13024506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_ADDR 0x00 13034506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_FMT 0x04 13044506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_DIM_X 0x08 13054506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_PITCH 0x0c 13064506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_DIM_Y 0x10 13074506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_ARRAY 0x14 13084506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_DIM_Z 0x18 13094506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_UNK1C 0x1c 13104506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_WIDTH 0x20 13114506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_HEIGHT 0x24 13124506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_DEPTH 0x28 13134506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_TARGET 0x2c 13144506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_CALL 0x30 13154506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_RAW_X 0x34 13164506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_MS_X 0x38 13174506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_MS_Y 0x3c 13184506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 13194506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO__STRIDE 0x40 13204506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 13214506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_DIM(i) (0x08 + (i) * 8) 13224506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_SIZE(i) (0x20 + (i) * 4) 13234506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller#define NVE4_SU_INFO_MS(i) (0x38 + (i) * 4) 13244506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 13254506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumillerstatic inline uint16_t getSuClampSubOp(const TexInstruction *su, int c) 13264506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller{ 13274506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller switch (su->tex.target.getEnum()) { 13284506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case TEX_TARGET_BUFFER: return NV50_IR_SUBOP_SUCLAMP_PL(0, 1); 13294506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case TEX_TARGET_RECT: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2); 13304506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case TEX_TARGET_1D: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2); 13314506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case TEX_TARGET_1D_ARRAY: return (c == 1) ? 13324506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller NV50_IR_SUBOP_SUCLAMP_PL(0, 2) : 13334506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller NV50_IR_SUBOP_SUCLAMP_SD(0, 2); 13344506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case TEX_TARGET_2D: return NV50_IR_SUBOP_SUCLAMP_BL(0, 2); 13354506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case TEX_TARGET_2D_MS: return NV50_IR_SUBOP_SUCLAMP_BL(0, 2); 13364506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case TEX_TARGET_2D_ARRAY: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2); 13374506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case TEX_TARGET_2D_MS_ARRAY: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2); 13384506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case TEX_TARGET_3D: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2); 13394506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case TEX_TARGET_CUBE: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2); 13404506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case TEX_TARGET_CUBE_ARRAY: return NV50_IR_SUBOP_SUCLAMP_SD(0, 2); 13414506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller default: 13424506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller assert(0); 13434506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller return 0; 13444506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 13454506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller} 13464506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 13474506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumillervoid 13484506ed28de7f9d76bbc99c0758a7891b84528729Christoph BumillerNVC0LoweringPass::adjustCoordinatesMS(TexInstruction *tex) 13494506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller{ 13504506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller const uint16_t base = tex->tex.r * NVE4_SU_INFO__STRIDE; 13514506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller const int arg = tex->tex.target.getArgCount(); 13524506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 13534506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (tex->tex.target == TEX_TARGET_2D_MS) 13544506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller tex->tex.target = TEX_TARGET_2D; 13554506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller else 13564506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (tex->tex.target == TEX_TARGET_2D_MS_ARRAY) 13574506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller tex->tex.target = TEX_TARGET_2D_ARRAY; 13584506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller else 13594506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller return; 13604506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 13614506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *x = tex->getSrc(0); 13624506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *y = tex->getSrc(1); 13634506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *s = tex->getSrc(arg - 1); 13644506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 13654506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *tx = bld.getSSA(), *ty = bld.getSSA(), *ts = bld.getSSA(); 13664506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 13674506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *ms_x = loadResInfo32(NULL, base + NVE4_SU_INFO_MS(0)); 13684506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *ms_y = loadResInfo32(NULL, base + NVE4_SU_INFO_MS(1)); 13694506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 13704506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp2(OP_SHL, TYPE_U32, tx, x, ms_x); 13714506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp2(OP_SHL, TYPE_U32, ty, y, ms_y); 13724506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 13734506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller s = bld.mkOp2v(OP_AND, TYPE_U32, ts, s, bld.loadImm(NULL, 0x7)); 13744506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller s = bld.mkOp2v(OP_SHL, TYPE_U32, ts, ts, bld.mkImm(3)); 13754506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 13764506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *dx = loadMsInfo32(ts, 0x0); 13774506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *dy = loadMsInfo32(ts, 0x4); 13784506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 13794506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp2(OP_ADD, TYPE_U32, tx, tx, dx); 13804506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp2(OP_ADD, TYPE_U32, ty, ty, dy); 13814506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 13824506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller tex->setSrc(0, tx); 13834506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller tex->setSrc(1, ty); 13844506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller tex->moveSources(arg, -1); 13854506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller} 13864506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 13874506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller// Sets 64-bit "generic address", predicate and format sources for SULD/SUST. 13884506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller// They're computed from the coordinates using the surface info in c[] space. 13894506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumillervoid 13904506ed28de7f9d76bbc99c0758a7891b84528729Christoph BumillerNVC0LoweringPass::processSurfaceCoordsNVE4(TexInstruction *su) 13914506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller{ 13924506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Instruction *insn; 13934506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller const bool atom = su->op == OP_SUREDB || su->op == OP_SUREDP; 13944506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller const bool raw = 13954506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller su->op == OP_SULDB || su->op == OP_SUSTB || su->op == OP_SUREDB; 13964506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller const int idx = su->tex.r; 13974506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller const int dim = su->tex.target.getDim(); 13984506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller const int arg = dim + (su->tex.target.isArray() ? 1 : 0); 13994506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller const uint16_t base = idx * NVE4_SU_INFO__STRIDE; 14004506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller int c; 14014506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *zero = bld.mkImm(0); 14024506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *p1 = NULL; 14034506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *v; 14044506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *src[3]; 14054506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *bf, *eau, *off; 14064506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *addr, *pred; 14074506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 14084506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller off = bld.getScratch(4); 14094506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bf = bld.getScratch(4); 14104506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller addr = bld.getSSA(8); 14114506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller pred = bld.getScratch(1, FILE_PREDICATE); 14124506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 14134506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.setPosition(su, false); 14144506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 14154506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller adjustCoordinatesMS(su); 14164506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 14174506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // calculate clamped coordinates 14184506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller for (c = 0; c < arg; ++c) { 14194506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller src[c] = bld.getScratch(); 14204506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (c == 0 && raw) 14214506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller v = loadResInfo32(NULL, base + NVE4_SU_INFO_RAW_X); 14224506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller else 14234506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller v = loadResInfo32(NULL, base + NVE4_SU_INFO_DIM(c)); 14244506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp3(OP_SUCLAMP, TYPE_S32, src[c], su->getSrc(c), v, zero) 14254506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller ->subOp = getSuClampSubOp(su, c); 14264506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 14274506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller for (; c < 3; ++c) 14284506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller src[c] = zero; 14294506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 14304506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // set predicate output 14314506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (su->tex.target == TEX_TARGET_BUFFER) { 14324506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller src[0]->getInsn()->setFlagsDef(1, pred); 14334506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } else 14344506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (su->tex.target.isArray()) { 14354506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller p1 = bld.getSSA(1, FILE_PREDICATE); 14364506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller src[dim]->getInsn()->setFlagsDef(1, p1); 14374506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 14384506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 14394506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // calculate pixel offset 14404506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (dim == 1) { 14414506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (su->tex.target != TEX_TARGET_BUFFER) 14424506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp2(OP_AND, TYPE_U32, off, src[0], bld.loadImm(NULL, 0xffff)); 14434506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } else 14444506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (dim == 3) { 14454506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller v = loadResInfo32(NULL, base + NVE4_SU_INFO_UNK1C); 14464506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp3(OP_MADSP, TYPE_U32, off, src[2], v, src[1]) 14474506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller ->subOp = NV50_IR_SUBOP_MADSP(4,2,8); // u16l u16l u16l 14484506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 14494506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller v = loadResInfo32(NULL, base + NVE4_SU_INFO_PITCH); 14504506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp3(OP_MADSP, TYPE_U32, off, off, v, src[0]) 14514506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller ->subOp = NV50_IR_SUBOP_MADSP(0,2,8); // u32 u16l u16l 14524506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } else { 14534506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller assert(dim == 2); 14544506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller v = loadResInfo32(NULL, base + NVE4_SU_INFO_PITCH); 14554506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp3(OP_MADSP, TYPE_U32, off, src[1], v, src[0]) 14564506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller ->subOp = su->tex.target.isArray() ? 14574506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller NV50_IR_SUBOP_MADSP_SD : NV50_IR_SUBOP_MADSP(4,2,8); // u16l u16l u16l 14584506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 14594506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 14604506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // calculate effective address part 1 14614506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (su->tex.target == TEX_TARGET_BUFFER) { 14624506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (raw) { 14634506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bf = src[0]; 14644506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } else { 14654506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller v = loadResInfo32(NULL, base + NVE4_SU_INFO_FMT); 14664506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp3(OP_VSHL, TYPE_U32, bf, src[0], v, zero) 14674506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller ->subOp = NV50_IR_SUBOP_V1(7,6,8|2); 14684506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 14694506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } else { 14704506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *y = src[1]; 14714506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *z = src[2]; 14724506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller uint16_t subOp = 0; 14734506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 14744506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller switch (dim) { 14754506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case 1: 14764506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller y = zero; 14774506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller z = zero; 14784506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller break; 14794506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case 2: 14804506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller z = off; 14814506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (!su->tex.target.isArray()) { 14824506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller z = loadResInfo32(NULL, base + NVE4_SU_INFO_UNK1C); 14834506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller subOp = NV50_IR_SUBOP_SUBFM_3D; 14844506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 14854506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller break; 14864506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller default: 14874506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller subOp = NV50_IR_SUBOP_SUBFM_3D; 14884506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller assert(dim == 3); 14894506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller break; 14904506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 14914506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller insn = bld.mkOp3(OP_SUBFM, TYPE_U32, bf, src[0], y, z); 14924506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller insn->subOp = subOp; 14934506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller insn->setFlagsDef(1, pred); 14944506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 14954506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 14964506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // part 2 14974506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller v = loadResInfo32(NULL, base + NVE4_SU_INFO_ADDR); 14984506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 14994506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (su->tex.target == TEX_TARGET_BUFFER) { 15004506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller eau = v; 15014506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } else { 15024506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller eau = bld.mkOp3v(OP_SUEAU, TYPE_U32, bld.getScratch(4), off, bf, v); 15034506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 15044506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // add array layer offset 15054506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (su->tex.target.isArray()) { 15064506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller v = loadResInfo32(NULL, base + NVE4_SU_INFO_ARRAY); 15074506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (dim == 1) 15084506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp3(OP_MADSP, TYPE_U32, eau, src[1], v, eau) 15094506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller ->subOp = NV50_IR_SUBOP_MADSP(4,0,0); // u16 u24 u32 15104506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller else 15114506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp3(OP_MADSP, TYPE_U32, eau, v, src[2], eau) 15124506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller ->subOp = NV50_IR_SUBOP_MADSP(0,0,0); // u32 u24 u32 15134506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // combine predicates 15144506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller assert(p1); 15154506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp2(OP_OR, TYPE_U8, pred, pred, p1); 15164506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 15174506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 15184506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (atom) { 15194506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *lo = bf; 15204506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (su->tex.target == TEX_TARGET_BUFFER) { 15214506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller lo = zero; 15224506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkMov(off, bf); 15234506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 15244506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // bf == g[] address & 0xff 15254506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // eau == g[] address >> 8 15264506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp3(OP_PERMT, TYPE_U32, bf, lo, bld.loadImm(NULL, 0x6540), eau); 15274506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp3(OP_PERMT, TYPE_U32, eau, zero, bld.loadImm(NULL, 0x0007), eau); 15284506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } else 15294506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (su->op == OP_SULDP && su->tex.target == TEX_TARGET_BUFFER) { 15304506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // Convert from u32 to u8 address format, which is what the library code 15314506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // doing SULDP currently uses. 15324506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // XXX: can SUEAU do this ? 15334506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // XXX: does it matter that we don't mask high bytes in bf ? 15344506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // Grrr. 15354506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp2(OP_SHR, TYPE_U32, off, bf, bld.mkImm(8)); 15364506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp2(OP_ADD, TYPE_U32, eau, eau, off); 15374506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 15384506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 15394506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp2(OP_MERGE, TYPE_U64, addr, bf, eau); 15404506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 15414506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (atom && su->tex.target == TEX_TARGET_BUFFER) 15424506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp2(OP_ADD, TYPE_U64, addr, addr, off); 15434506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 15444506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // let's just set it 0 for raw access and hope it works 15454506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller v = raw ? 15464506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkImm(0) : loadResInfo32(NULL, base + NVE4_SU_INFO_FMT); 15474506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 15484506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // get rid of old coordinate sources, make space for fmt info and predicate 15494506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller su->moveSources(arg, 3 - arg); 15504506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // set 64 bit address and 32-bit format sources 15514506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller su->setSrc(0, addr); 15524506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller su->setSrc(1, v); 15534506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller su->setSrc(2, pred); 15544506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller} 15554506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 15564506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumillervoid 15574506ed28de7f9d76bbc99c0758a7891b84528729Christoph BumillerNVC0LoweringPass::handleSurfaceOpNVE4(TexInstruction *su) 15584506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller{ 15594506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller processSurfaceCoordsNVE4(su); 15604506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 15614506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // Who do we hate more ? The person who decided that nvc0's SULD doesn't 15624506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // have to support conversion or the person who decided that, in OpenCL, 15634506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // you don't have to specify the format here like you do in OpenGL ? 15644506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 15654506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (su->op == OP_SULDP) { 15664506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // We don't patch shaders. Ever. 15674506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // You get an indirect call to our library blob here. 15684506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller // But at least it's uniform. 15694506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller FlowInstruction *call; 15704506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller LValue *p[3]; 15714506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller LValue *r[5]; 15724506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller uint16_t base = su->tex.r * NVE4_SU_INFO__STRIDE + NVE4_SU_INFO_CALL; 15734506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 15744506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller for (int i = 0; i < 4; ++i) 15754506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller (r[i] = bld.getScratch(4, FILE_GPR))->reg.data.id = i; 15764506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller for (int i = 0; i < 3; ++i) 15774506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller (p[i] = bld.getScratch(1, FILE_PREDICATE))->reg.data.id = i; 15784506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller (r[4] = bld.getScratch(8, FILE_GPR))->reg.data.id = 4; 15794506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 15804506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkMov(p[1], bld.mkImm((su->cache == CACHE_CA) ? 1 : 0), TYPE_U8); 15814506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkMov(p[2], bld.mkImm((su->cache == CACHE_CG) ? 1 : 0), TYPE_U8); 15824506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkMov(p[0], su->getSrc(2), TYPE_U8); 15834506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkMov(r[4], su->getSrc(0), TYPE_U64); 15844506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkMov(r[2], su->getSrc(1), TYPE_U32); 15854506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 15864506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller call = bld.mkFlow(OP_CALL, NULL, su->cc, su->getPredicate()); 15874506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 15884506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller call->indirect = 1; 15894506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller call->absolute = 1; 15904506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller call->setSrc(0, bld.mkSymbol(FILE_MEMORY_CONST, 1591d86933e6f42b9c2f5bb617c66c91795c560a9abdSamuel Pitoiset prog->driver->io.auxCBSlot, TYPE_U32, 15924506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller prog->driver->io.suInfoBase + base)); 15934506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller call->setSrc(1, r[2]); 15944506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller call->setSrc(2, r[4]); 15954506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller for (int i = 0; i < 3; ++i) 15964506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller call->setSrc(3 + i, p[i]); 15974506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller for (int i = 0; i < 4; ++i) { 15984506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller call->setDef(i, r[i]); 15994506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkMov(su->getDef(i), r[i]); 16004506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 16014506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller call->setDef(4, p[1]); 16024506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller delete_Instruction(bld.getProgram(), su); 16034506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 16044506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 16054506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (su->op == OP_SUREDB || su->op == OP_SUREDP) { 160675f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller // FIXME: for out of bounds access, destination value will be undefined ! 16074506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Value *pred = su->getSrc(2); 16084506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller CondCode cc = CC_NOT_P; 16094506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (su->getPredicate()) { 16104506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller pred = bld.getScratch(1, FILE_PREDICATE); 16114506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller cc = su->cc; 16124506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (cc == CC_NOT_P) { 16134506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp2(OP_OR, TYPE_U8, pred, su->getPredicate(), su->getSrc(2)); 16144506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } else { 16154506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller bld.mkOp2(OP_AND, TYPE_U8, pred, su->getPredicate(), su->getSrc(2)); 16164506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller pred->getInsn()->src(1).mod = Modifier(NV50_IR_MOD_NOT); 16174506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 16184506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 16194506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller Instruction *red = bld.mkOp(OP_ATOM, su->dType, su->getDef(0)); 16204506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller red->subOp = su->subOp; 16214506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (!gMemBase) 16224506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller gMemBase = bld.mkSymbol(FILE_MEMORY_GLOBAL, 0, TYPE_U32, 0); 16234506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller red->setSrc(0, gMemBase); 16244506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller red->setSrc(1, su->getSrc(3)); 16254506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (su->subOp == NV50_IR_SUBOP_ATOM_CAS) 16264506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller red->setSrc(2, su->getSrc(4)); 16274506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller red->setIndirect(0, 0, su->getSrc(0)); 16284506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller red->setPredicate(cc, pred); 16294506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller delete_Instruction(bld.getProgram(), su); 163075f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller handleCasExch(red, true); 16314506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } else { 16324506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller su->sType = (su->tex.target == TEX_TARGET_BUFFER) ? TYPE_U32 : TYPE_U8; 16334506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller } 16344506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller} 16354506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller 1636c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumillerbool 163757594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::handleWRSV(Instruction *i) 163857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 163957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Instruction *st; 164057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Symbol *sym; 164157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller uint32_t addr; 164257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 164357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller // must replace, $sreg are not writeable 164457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller addr = targ->getSVAddress(FILE_SHADER_OUTPUT, i->getSrc(0)->asSym()); 164557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (addr >= 0x400) 164657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return false; 164757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller sym = bld.mkSymbol(FILE_SHADER_OUTPUT, 0, i->sType, addr); 164857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 164957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller st = bld.mkStore(OP_EXPORT, i->dType, sym, i->getIndirect(0, 0), 165057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->getSrc(1)); 165157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller st->perPatch = i->perPatch; 165257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 165357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.getBB()->remove(i); 165457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 165557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 165657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 165757594065c30feec9376be9b2132659f7d87362eeChristoph Bumillervoid 165857594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::readTessCoord(LValue *dst, int c) 165957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 166057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Value *laneid = bld.getSSA(); 166157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Value *x, *y; 166257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 166357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkOp1(OP_RDSV, TYPE_U32, laneid, bld.mkSysVal(SV_LANEID, 0)); 166457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 166557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (c == 0) { 166657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller x = dst; 166757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller y = NULL; 166857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } else 166957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (c == 1) { 167057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller x = NULL; 167157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller y = dst; 167257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } else { 167357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller assert(c == 2); 167457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller x = bld.getSSA(); 167557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller y = bld.getSSA(); 167657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 167757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (x) 167857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkFetch(x, TYPE_F32, FILE_SHADER_OUTPUT, 0x2f0, NULL, laneid); 167957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (y) 16803fc2818f2b9e8a19e5349442e50dcee4858452c6Christoph Bumiller bld.mkFetch(y, TYPE_F32, FILE_SHADER_OUTPUT, 0x2f4, NULL, laneid); 168157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 168257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (c == 2) { 168357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkOp2(OP_ADD, TYPE_F32, dst, x, y); 168457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkOp2(OP_SUB, TYPE_F32, dst, bld.loadImm(NULL, 1.0f), dst); 168557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 168657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 168757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 168857594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 168957594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::handleRDSV(Instruction *i) 169057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 169157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Symbol *sym = i->getSrc(0)->asSym(); 1692ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller const SVSemantic sv = sym->reg.data.sv.sv; 169357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Value *vtx = NULL; 169457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Instruction *ld; 169557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller uint32_t addr = targ->getSVAddress(FILE_SHADER_INPUT, sym); 169657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 1697ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller if (addr >= 0x400) { 1698ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller // mov $sreg 1699ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller if (sym->reg.data.sv.index == 3) { 1700ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller // TGSI backend may use 4th component of TID,NTID,CTAID,NCTAID 1701ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller i->op = OP_MOV; 1702ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller i->setSrc(0, bld.mkImm((sv == SV_NTID || sv == SV_NCTAID) ? 1 : 0)); 1703ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller } 17047cf2bffe8254de6808202d866598ec4c9afe1a51Ilia Mirkin if (sv == SV_VERTEX_COUNT) { 17057cf2bffe8254de6808202d866598ec4c9afe1a51Ilia Mirkin bld.setPosition(i, true); 17067cf2bffe8254de6808202d866598ec4c9afe1a51Ilia Mirkin bld.mkOp2(OP_EXTBF, TYPE_U32, i->getDef(0), i->getDef(0), bld.mkImm(0x808)); 17077cf2bffe8254de6808202d866598ec4c9afe1a51Ilia Mirkin } 170857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 1709ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller } 171057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 1711ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller switch (sv) { 171257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case SV_POSITION: 171357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller assert(prog->getType() == Program::TYPE_FRAGMENT); 1714acaed8f41d3cf57092f3fe3a607b8069c72b57f1Ilia Mirkin if (i->srcExists(1)) { 1715acaed8f41d3cf57092f3fe3a607b8069c72b57f1Ilia Mirkin // Pass offset through to the interpolation logic 1716acaed8f41d3cf57092f3fe3a607b8069c72b57f1Ilia Mirkin ld = bld.mkInterp(NV50_IR_INTERP_LINEAR | NV50_IR_INTERP_OFFSET, 1717acaed8f41d3cf57092f3fe3a607b8069c72b57f1Ilia Mirkin i->getDef(0), addr, NULL); 1718acaed8f41d3cf57092f3fe3a607b8069c72b57f1Ilia Mirkin ld->setSrc(1, i->getSrc(1)); 1719acaed8f41d3cf57092f3fe3a607b8069c72b57f1Ilia Mirkin } else { 1720acaed8f41d3cf57092f3fe3a607b8069c72b57f1Ilia Mirkin bld.mkInterp(NV50_IR_INTERP_LINEAR, i->getDef(0), addr, NULL); 1721acaed8f41d3cf57092f3fe3a607b8069c72b57f1Ilia Mirkin } 172252c8c52b222e1fdb4c1f4ca3dedde9cd7b9c321fChristoph Bumiller break; 172352c8c52b222e1fdb4c1f4ca3dedde9cd7b9c321fChristoph Bumiller case SV_FACE: 172452c8c52b222e1fdb4c1f4ca3dedde9cd7b9c321fChristoph Bumiller { 172552c8c52b222e1fdb4c1f4ca3dedde9cd7b9c321fChristoph Bumiller Value *face = i->getDef(0); 172652c8c52b222e1fdb4c1f4ca3dedde9cd7b9c321fChristoph Bumiller bld.mkInterp(NV50_IR_INTERP_FLAT, face, addr, NULL); 172752c8c52b222e1fdb4c1f4ca3dedde9cd7b9c321fChristoph Bumiller if (i->dType == TYPE_F32) { 1728354206f407fffd5f0b553dcbcc46b178d0b22c47Ilia Mirkin bld.mkOp2(OP_OR, TYPE_U32, face, face, bld.mkImm(0x00000001)); 1729354206f407fffd5f0b553dcbcc46b178d0b22c47Ilia Mirkin bld.mkOp1(OP_NEG, TYPE_S32, face, face); 1730354206f407fffd5f0b553dcbcc46b178d0b22c47Ilia Mirkin bld.mkCvt(OP_CVT, TYPE_F32, face, TYPE_S32, face); 173152c8c52b222e1fdb4c1f4ca3dedde9cd7b9c321fChristoph Bumiller } 173252c8c52b222e1fdb4c1f4ca3dedde9cd7b9c321fChristoph Bumiller } 173357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller break; 173457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case SV_TESS_COORD: 173557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller assert(prog->getType() == Program::TYPE_TESSELLATION_EVAL); 173657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller readTessCoord(i->getDef(0)->asLValue(), i->getSrc(0)->reg.data.sv.index); 173757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller break; 1738ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller case SV_NTID: 1739ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller case SV_NCTAID: 1740ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller case SV_GRIDID: 1741ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller assert(targ->getChipset() >= NVISA_GK104_CHIPSET); // mov $sreg otherwise 1742ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller if (sym->reg.data.sv.index == 3) { 1743ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller i->op = OP_MOV; 1744ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller i->setSrc(0, bld.mkImm(sv == SV_GRIDID ? 0 : 1)); 1745ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller return true; 1746ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller } 1747ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller addr += prog->driver->prop.cp.gridInfoBase; 1748ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller bld.mkLoad(TYPE_U32, i->getDef(0), 174926cc411db87f924003f227874d7a047dd8b5e5a4Samuel Pitoiset bld.mkSymbol(FILE_MEMORY_CONST, prog->driver->io.auxCBSlot, 175026cc411db87f924003f227874d7a047dd8b5e5a4Samuel Pitoiset TYPE_U32, addr), NULL); 1751ae59a7d35d2b6f23634617dc91e1baf85c9d6d81Christoph Bumiller break; 1752af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin case SV_SAMPLE_INDEX: 1753af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin // TODO: Properly pass source as an address in the PIX address space 1754af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin // (which can be of the form [r0+offset]). But this is currently 1755af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin // unnecessary. 1756af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0)); 1757af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin ld->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID; 1758af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin break; 1759af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin case SV_SAMPLE_POS: { 1760af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin Value *off = new_LValue(func, FILE_GPR); 1761af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0)); 1762af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin ld->subOp = NV50_IR_SUBOP_PIXLD_SAMPLEID; 1763af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin bld.mkOp2(OP_SHL, TYPE_U32, off, i->getDef(0), bld.mkImm(3)); 1764af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin bld.mkLoad(TYPE_F32, 1765af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin i->getDef(0), 1766af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin bld.mkSymbol( 1767d86933e6f42b9c2f5bb617c66c91795c560a9abdSamuel Pitoiset FILE_MEMORY_CONST, prog->driver->io.auxCBSlot, 1768af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin TYPE_U32, prog->driver->io.sampleInfoBase + 1769af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin 4 * sym->reg.data.sv.index), 1770af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin off); 1771af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin break; 1772af38ef907c89ecb1125bf258cafa0793f79a5eb7Ilia Mirkin } 1773b3a2398aded19e25124a4a1d228eb3843827f6b2Ilia Mirkin case SV_SAMPLE_MASK: 1774b3a2398aded19e25124a4a1d228eb3843827f6b2Ilia Mirkin ld = bld.mkOp1(OP_PIXLD, TYPE_U32, i->getDef(0), bld.mkImm(0)); 1775b3a2398aded19e25124a4a1d228eb3843827f6b2Ilia Mirkin ld->subOp = NV50_IR_SUBOP_PIXLD_COVMASK; 1776b3a2398aded19e25124a4a1d228eb3843827f6b2Ilia Mirkin break; 1777517a93b346e720082e22e358b63b5dbc5c42aa09Ilia Mirkin case SV_BASEVERTEX: 1778517a93b346e720082e22e358b63b5dbc5c42aa09Ilia Mirkin case SV_BASEINSTANCE: 1779517a93b346e720082e22e358b63b5dbc5c42aa09Ilia Mirkin case SV_DRAWID: 1780517a93b346e720082e22e358b63b5dbc5c42aa09Ilia Mirkin ld = bld.mkLoad(TYPE_U32, i->getDef(0), 1781517a93b346e720082e22e358b63b5dbc5c42aa09Ilia Mirkin bld.mkSymbol(FILE_MEMORY_CONST, 1782517a93b346e720082e22e358b63b5dbc5c42aa09Ilia Mirkin prog->driver->io.auxCBSlot, 1783517a93b346e720082e22e358b63b5dbc5c42aa09Ilia Mirkin TYPE_U32, 1784517a93b346e720082e22e358b63b5dbc5c42aa09Ilia Mirkin prog->driver->io.drawInfoBase + 1785517a93b346e720082e22e358b63b5dbc5c42aa09Ilia Mirkin 4 * (sv - SV_BASEVERTEX)), 1786517a93b346e720082e22e358b63b5dbc5c42aa09Ilia Mirkin NULL); 1787517a93b346e720082e22e358b63b5dbc5c42aa09Ilia Mirkin break; 178857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller default: 17897e0036a49258326cc2d875f2960d18c6b3665036Ilia Mirkin if (prog->getType() == Program::TYPE_TESSELLATION_EVAL && !i->perPatch) 179057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller vtx = bld.mkOp1v(OP_PFETCH, TYPE_U32, bld.getSSA(), bld.mkImm(0)); 179157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller ld = bld.mkFetch(i->getDef(0), i->dType, 179257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller FILE_SHADER_INPUT, addr, i->getIndirect(0, 0), vtx); 179357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller ld->perPatch = i->perPatch; 179457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller break; 179557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 179657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.getBB()->remove(i); 179757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 179857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 179957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 180057594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 180157594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::handleDIV(Instruction *i) 180257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 180357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (!isFloatType(i->dType)) 180457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 1805b5f2c0505fd4f66422e034b041cdf0bc3dc46e99Christoph Bumiller bld.setPosition(i, false); 1806b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin Instruction *rcp = bld.mkOp1(OP_RCP, i->dType, bld.getSSA(typeSizeof(i->dType)), i->getSrc(1)); 180757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->op = OP_MUL; 180857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->setSrc(1, rcp->getDef(0)); 180957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 181057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 181157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 181257594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 181357594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::handleMOD(Instruction *i) 181457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 1815b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin if (!isFloatType(i->dType)) 181657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 1817b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin LValue *value = bld.getScratch(typeSizeof(i->dType)); 1818b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin bld.mkOp1(OP_RCP, i->dType, value, i->getSrc(1)); 1819b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin bld.mkOp2(OP_MUL, i->dType, value, i->getSrc(0), value); 1820b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin bld.mkOp1(OP_TRUNC, i->dType, value, value); 1821b87b498b88c51fb8c80901b8f581331d3fbcd972Ilia Mirkin bld.mkOp2(OP_MUL, i->dType, value, i->getSrc(1), value); 182257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->op = OP_SUB; 182357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->setSrc(1, value); 182457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 182557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 182657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 182757594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 182857594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::handleSQRT(Instruction *i) 182957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 1830c1e4a6bfbf015801c6a8b0ae694482421a22c2d9Ilia Mirkin if (i->dType == TYPE_F64) { 1831c1e4a6bfbf015801c6a8b0ae694482421a22c2d9Ilia Mirkin Value *pred = bld.getSSA(1, FILE_PREDICATE); 18329184d9a0bbe8a8b88d676a20f95d66ceee9eaf21Pierre Moreau Value *zero = bld.loadImm(NULL, 0.0); 1833c1e4a6bfbf015801c6a8b0ae694482421a22c2d9Ilia Mirkin Value *dst = bld.getSSA(8); 1834c1e4a6bfbf015801c6a8b0ae694482421a22c2d9Ilia Mirkin bld.mkOp1(OP_RSQ, i->dType, dst, i->getSrc(0)); 1835c1e4a6bfbf015801c6a8b0ae694482421a22c2d9Ilia Mirkin bld.mkCmp(OP_SET, CC_LE, i->dType, pred, i->dType, i->getSrc(0), zero); 1836c1e4a6bfbf015801c6a8b0ae694482421a22c2d9Ilia Mirkin bld.mkOp3(OP_SELP, TYPE_U64, dst, zero, dst, pred); 1837c1e4a6bfbf015801c6a8b0ae694482421a22c2d9Ilia Mirkin i->op = OP_MUL; 1838c1e4a6bfbf015801c6a8b0ae694482421a22c2d9Ilia Mirkin i->setSrc(1, dst); 1839c1e4a6bfbf015801c6a8b0ae694482421a22c2d9Ilia Mirkin // TODO: Handle this properly with a library function 1840c1e4a6bfbf015801c6a8b0ae694482421a22c2d9Ilia Mirkin } else { 1841c1e4a6bfbf015801c6a8b0ae694482421a22c2d9Ilia Mirkin bld.setPosition(i, true); 1842c1e4a6bfbf015801c6a8b0ae694482421a22c2d9Ilia Mirkin i->op = OP_RSQ; 1843c1e4a6bfbf015801c6a8b0ae694482421a22c2d9Ilia Mirkin bld.mkOp1(OP_RCP, i->dType, i->getDef(0), i->getDef(0)); 1844c1e4a6bfbf015801c6a8b0ae694482421a22c2d9Ilia Mirkin } 184557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 184657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 184757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 184857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 184957594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 185057594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::handlePOW(Instruction *i) 185157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 185257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller LValue *val = bld.getScratch(); 185357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 185457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkOp1(OP_LG2, TYPE_F32, val, i->getSrc(0)); 185557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkOp2(OP_MUL, TYPE_F32, val, i->getSrc(1), val)->dnz = 1; 185657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkOp1(OP_PREEX2, TYPE_F32, val, val); 185757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 185857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->op = OP_EX2; 185957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->setSrc(0, val); 186057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->setSrc(1, NULL); 186157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 186257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 186357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 186457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 186557594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 186657594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::handleEXPORT(Instruction *i) 186757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 186857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (prog->getType() == Program::TYPE_FRAGMENT) { 186957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller int id = i->getSrc(0)->reg.data.offset / 4; 187057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 18719362d4bc0a03860ec386156cf499e855a9c2d2a5Christoph Bumiller if (i->src(0).isIndirect(0)) // TODO, ugly 187257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return false; 187357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->op = OP_MOV; 187400fe442253744c4c4e7e68da44d6983da053968bChristoph Bumiller i->subOp = NV50_IR_SUBOP_MOV_FINAL; 18759362d4bc0a03860ec386156cf499e855a9c2d2a5Christoph Bumiller i->src(0).set(i->src(1)); 187657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->setSrc(1, NULL); 187757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->setDef(0, new_LValue(func, FILE_GPR)); 187857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->getDef(0)->reg.data.id = id; 187957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 188057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller prog->maxGPR = MAX2(prog->maxGPR, id); 188157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } else 188257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (prog->getType() == Program::TYPE_GEOMETRY) { 188357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->setIndirect(0, 1, gpEmitAddress); 188457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 188557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 188657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 188757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 188857594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 188957594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::handleOUT(Instruction *i) 189057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 18912f2467cb23ce19770c95ce1f004dc11750dffc6dIlia Mirkin Instruction *prev = i->prev; 18922f2467cb23ce19770c95ce1f004dc11750dffc6dIlia Mirkin ImmediateValue stream, prevStream; 18932f2467cb23ce19770c95ce1f004dc11750dffc6dIlia Mirkin 18942f2467cb23ce19770c95ce1f004dc11750dffc6dIlia Mirkin // Only merge if the stream ids match. Also, note that the previous 18952f2467cb23ce19770c95ce1f004dc11750dffc6dIlia Mirkin // instruction would have already been lowered, so we take arg1 from it. 18962f2467cb23ce19770c95ce1f004dc11750dffc6dIlia Mirkin if (i->op == OP_RESTART && prev && prev->op == OP_EMIT && 18972f2467cb23ce19770c95ce1f004dc11750dffc6dIlia Mirkin i->src(0).getImmediate(stream) && 18982f2467cb23ce19770c95ce1f004dc11750dffc6dIlia Mirkin prev->src(1).getImmediate(prevStream) && 18992f2467cb23ce19770c95ce1f004dc11750dffc6dIlia Mirkin stream.reg.data.u32 == prevStream.reg.data.u32) { 190057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->prev->subOp = NV50_IR_SUBOP_EMIT_RESTART; 190157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller delete_Instruction(prog, i); 190257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } else { 190357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller assert(gpEmitAddress); 190457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->setDef(0, gpEmitAddress); 19052f2467cb23ce19770c95ce1f004dc11750dffc6dIlia Mirkin i->setSrc(1, i->getSrc(0)); 190657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->setSrc(0, gpEmitAddress); 190757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 190857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return true; 190957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 191057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 191157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller// Generate a binary predicate if an instruction is predicated by 191257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller// e.g. an f32 value. 191357594065c30feec9376be9b2132659f7d87362eeChristoph Bumillervoid 191457594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::checkPredicate(Instruction *insn) 191557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 191657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Value *pred = insn->getPredicate(); 191757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller Value *pdst; 191857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 191957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (!pred || pred->reg.file == FILE_PREDICATE) 192057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return; 192157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller pdst = new_LValue(func, FILE_PREDICATE); 192257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 192357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller // CAUTION: don't use pdst->getInsn, the definition might not be unique, 192457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller // delay turning PSET(FSET(x,y),0) into PSET(x,y) to a later pass 192557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 1926bbe3d6dc29f218e4d790e5ea359d3c6736e94226Dave Airlie bld.mkCmp(OP_SET, CC_NEU, insn->dType, pdst, insn->dType, bld.mkImm(0), pred); 192757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 192857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller insn->setPredicate(insn->cc, pdst); 192957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 193057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 193157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller// 193257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller// - add quadop dance for texturing 193357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller// - put FP outputs in GPRs 193457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller// - convert instruction sequences 193557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller// 193657594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 193757594065c30feec9376be9b2132659f7d87362eeChristoph BumillerNVC0LoweringPass::visit(Instruction *i) 193857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 19399d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin bool ret = true; 1940405bd00f3c98cb78d1dda1f3bf5d74155b18cd57Christoph Bumiller bld.setPosition(i, false); 194157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 194257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (i->cc != CC_ALWAYS) 194357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller checkPredicate(i); 194457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 194557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller switch (i->op) { 194657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_TEX: 194757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_TXB: 194857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_TXL: 194957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_TXF: 195057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_TXG: 195157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return handleTEX(i->asTex()); 195257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_TXD: 195357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return handleTXD(i->asTex()); 1954423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin case OP_TXLQ: 1955423f64e83ab5b1ea7de475ae80300a8408522743Ilia Mirkin return handleTXLQ(i->asTex()); 195630cb66cd745fc793a2349f1d17046c50cd51c558Christoph Bumiller case OP_TXQ: 195730cb66cd745fc793a2349f1d17046c50cd51c558Christoph Bumiller return handleTXQ(i->asTex()); 195857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_EX2: 195957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller bld.mkOp1(OP_PREEX2, TYPE_F32, i->getDef(0), i->getSrc(0)); 196057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller i->setSrc(0, i->getDef(0)); 196157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller break; 196257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_POW: 196357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return handlePOW(i); 196457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_DIV: 196557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return handleDIV(i); 196657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_MOD: 196757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return handleMOD(i); 196857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_SQRT: 196957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return handleSQRT(i); 197057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_EXPORT: 19719d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin ret = handleEXPORT(i); 19729d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin break; 197357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_EMIT: 197457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_RESTART: 197557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return handleOUT(i); 197657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_RDSV: 197757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return handleRDSV(i); 197857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_WRSV: 197957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return handleWRSV(i); 1980c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin case OP_STORE: 198157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller case OP_LOAD: 19829362d4bc0a03860ec386156cf499e855a9c2d2a5Christoph Bumiller if (i->src(0).getFile() == FILE_SHADER_INPUT) { 1983d105b3df14283a4dd80cecc1e6cab58432368ef6Christoph Bumiller if (prog->getType() == Program::TYPE_COMPUTE) { 1984d105b3df14283a4dd80cecc1e6cab58432368ef6Christoph Bumiller i->getSrc(0)->reg.file = FILE_MEMORY_CONST; 1985d105b3df14283a4dd80cecc1e6cab58432368ef6Christoph Bumiller i->getSrc(0)->reg.fileIndex = 0; 1986b3f82e1a63e8a58f0e7ac297fc5e94ebe76c3339Bryan Cain } else 1987b3f82e1a63e8a58f0e7ac297fc5e94ebe76c3339Bryan Cain if (prog->getType() == Program::TYPE_GEOMETRY && 1988b3f82e1a63e8a58f0e7ac297fc5e94ebe76c3339Bryan Cain i->src(0).isIndirect(0)) { 1989b3f82e1a63e8a58f0e7ac297fc5e94ebe76c3339Bryan Cain // XXX: this assumes vec4 units 1990b3f82e1a63e8a58f0e7ac297fc5e94ebe76c3339Bryan Cain Value *ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(), 1991b3f82e1a63e8a58f0e7ac297fc5e94ebe76c3339Bryan Cain i->getIndirect(0, 0), bld.mkImm(4)); 1992b3f82e1a63e8a58f0e7ac297fc5e94ebe76c3339Bryan Cain i->setIndirect(0, 0, ptr); 1993217301843aea0299ab245e260b20af7ad250e9d8Ilia Mirkin i->op = OP_VFETCH; 1994d105b3df14283a4dd80cecc1e6cab58432368ef6Christoph Bumiller } else { 1995d105b3df14283a4dd80cecc1e6cab58432368ef6Christoph Bumiller i->op = OP_VFETCH; 1996d105b3df14283a4dd80cecc1e6cab58432368ef6Christoph Bumiller assert(prog->getType() != Program::TYPE_FRAGMENT); // INTERP 1997d105b3df14283a4dd80cecc1e6cab58432368ef6Christoph Bumiller } 19980ddc28b026688df79e54d3af1d7914ff04b12fedIlia Mirkin } else if (i->src(0).getFile() == FILE_MEMORY_CONST) { 19990ddc28b026688df79e54d3af1d7914ff04b12fedIlia Mirkin if (i->src(0).isIndirect(1)) { 20000ddc28b026688df79e54d3af1d7914ff04b12fedIlia Mirkin Value *ptr; 20010ddc28b026688df79e54d3af1d7914ff04b12fedIlia Mirkin if (i->src(0).isIndirect(0)) 20020ddc28b026688df79e54d3af1d7914ff04b12fedIlia Mirkin ptr = bld.mkOp3v(OP_INSBF, TYPE_U32, bld.getSSA(), 20030ddc28b026688df79e54d3af1d7914ff04b12fedIlia Mirkin i->getIndirect(0, 1), bld.mkImm(0x1010), 20040ddc28b026688df79e54d3af1d7914ff04b12fedIlia Mirkin i->getIndirect(0, 0)); 20050ddc28b026688df79e54d3af1d7914ff04b12fedIlia Mirkin else 20060ddc28b026688df79e54d3af1d7914ff04b12fedIlia Mirkin ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getSSA(), 20070ddc28b026688df79e54d3af1d7914ff04b12fedIlia Mirkin i->getIndirect(0, 1), bld.mkImm(16)); 20080ddc28b026688df79e54d3af1d7914ff04b12fedIlia Mirkin i->setIndirect(0, 1, NULL); 20090ddc28b026688df79e54d3af1d7914ff04b12fedIlia Mirkin i->setIndirect(0, 0, ptr); 20100ddc28b026688df79e54d3af1d7914ff04b12fedIlia Mirkin i->subOp = NV50_IR_SUBOP_LDC_IS; 20110ddc28b026688df79e54d3af1d7914ff04b12fedIlia Mirkin } 2012e3e2df01bf855f3b435e03224a762649081c6558Ilia Mirkin } else if (i->src(0).getFile() == FILE_SHADER_OUTPUT) { 2013e3e2df01bf855f3b435e03224a762649081c6558Ilia Mirkin assert(prog->getType() == Program::TYPE_TESSELLATION_CONTROL); 2014e3e2df01bf855f3b435e03224a762649081c6558Ilia Mirkin i->op = OP_VFETCH; 2015c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin } else if (i->src(0).getFile() == FILE_MEMORY_GLOBAL) { 20167b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin Value *ind = i->getIndirect(0, 1); 20177b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin Value *ptr = loadResInfo64(ind, i->getSrc(0)->reg.fileIndex * 16); 2018c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin // XXX come up with a way not to do this for EVERY little access but 2019c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin // rather to batch these up somehow. Unfortunately we've lost the 2020c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin // information about the field width by the time we get here. 2021c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin Value *offset = bld.loadImm(NULL, i->getSrc(0)->reg.data.offset + typeSizeof(i->sType)); 20227b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin Value *length = loadResLength32(ind, i->getSrc(0)->reg.fileIndex * 16); 2023c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin Value *pred = new_LValue(func, FILE_PREDICATE); 2024c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin if (i->src(0).isIndirect(0)) { 2025c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0)); 2026c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin bld.mkOp2(OP_ADD, TYPE_U32, offset, offset, i->getIndirect(0, 0)); 2027c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin } 20287b9a77b905bda3003dc57efb99879499ebc4ba41Ilia Mirkin i->setIndirect(0, 1, NULL); 2029c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin i->setIndirect(0, 0, ptr); 2030c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin bld.mkCmp(OP_SET, CC_GT, TYPE_U32, pred, TYPE_U32, offset, length); 2031c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin i->setPredicate(CC_NOT_P, pred); 2032c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin if (i->defExists(0)) { 2033c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin bld.mkMov(i->getDef(0), bld.mkImm(0)); 2034c3083c70823d8f4bfdabcf38f98dfebeff0a2b2bIlia Mirkin } 203557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 203657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller break; 2037c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller case OP_ATOM: 203875f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller { 203975f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller const bool cctl = i->src(0).getFile() == FILE_MEMORY_GLOBAL; 2040c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller handleATOM(i); 204175f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller handleCasExch(i, cctl); 204275f1f852b00ad0d766684d01695322b93a2acd55Christoph Bumiller } 2043c0fc3463e9be19495ed31018fc093e726d5f4fbbChristoph Bumiller break; 20444506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case OP_SULDB: 20454506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case OP_SULDP: 20464506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case OP_SUSTB: 20474506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case OP_SUSTP: 20484506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case OP_SUREDB: 20494506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller case OP_SUREDP: 20504506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller if (targ->getChipset() >= NVISA_GK104_CHIPSET) 20514506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller handleSurfaceOpNVE4(i->asTex()); 20524506ed28de7f9d76bbc99c0758a7891b84528729Christoph Bumiller break; 20532c4eeb0b5cf17caa06cb3fa46d4f64e6a8005d23Ilia Mirkin case OP_SUQ: 20542c4eeb0b5cf17caa06cb3fa46d4f64e6a8005d23Ilia Mirkin handleSUQ(i); 20552c4eeb0b5cf17caa06cb3fa46d4f64e6a8005d23Ilia Mirkin break; 205657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller default: 205757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller break; 20586bca283ad5ebdd85e268c6757842b3c808c6b73dJohannes Obermayr } 20599d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin 20609d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin /* Kepler+ has a special opcode to compute a new base address to be used 20619d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin * for indirect loads. 20629d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin */ 20639d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin if (targ->getChipset() >= NVISA_GK104_CHIPSET && !i->perPatch && 20649d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin (i->op == OP_VFETCH || i->op == OP_EXPORT) && i->src(0).isIndirect(0)) { 20659d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin Instruction *afetch = bld.mkOp1(OP_AFETCH, TYPE_U32, bld.getSSA(), 20669d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin cloneShallow(func, i->getSrc(0))); 20679d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin afetch->setIndirect(0, 0, i->getIndirect(0, 0)); 20689d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin i->src(0).get()->reg.data.offset = 0; 20699d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin i->setIndirect(0, 0, afetch->getDef(0)); 20709d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin } 20719d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin 20729d60793a03e40e1d139b78fce0144cad57438741Ilia Mirkin return ret; 207357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 207457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 207557594065c30feec9376be9b2132659f7d87362eeChristoph Bumillerbool 207657594065c30feec9376be9b2132659f7d87362eeChristoph BumillerTargetNVC0::runLegalizePass(Program *prog, CGStage stage) const 207757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller{ 207857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (stage == CG_STAGE_PRE_SSA) { 207957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller NVC0LoweringPass pass(prog); 208057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return pass.run(prog, false, true); 208157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } else 208257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (stage == CG_STAGE_POST_RA) { 2083e44089b2f79aa2dcaacf348911433d1e21235c0cChristoph Bumiller NVC0LegalizePostRA pass(prog); 208457594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return pass.run(prog, false, true); 208557594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } else 208657594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller if (stage == CG_STAGE_SSA) { 208757594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller NVC0LegalizeSSA pass; 208857594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return pass.run(prog, false, true); 208957594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller } 209057594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller return false; 209157594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} 209257594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller 209357594065c30feec9376be9b2132659f7d87362eeChristoph Bumiller} // namespace nv50_ir 2094