nv50/ir: add support for indirect buffer loading

Signed-off-by: Ilia Mirkin <[email protected]>

2 files changed, 31 insertions, 10 deletions

diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp
index f5448293e54..735e2891cf2 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp
@@ -2253,7 +2253,10 @@ Converter::handleLOAD(Value *dst0[4])
             sym = makeSym(TGSI_FILE_BUFFER, r, -1, c, 4 * c);
          }
 
-         mkLoad(TYPE_U32, dst0[c], sym, off)->cache = tgsi.getCacheMode();
+         Instruction *ld = mkLoad(TYPE_U32, dst0[c], sym, off);
+         ld->cache = tgsi.getCacheMode();
+         if (tgsi.getSrc(0).isIndirect(0))
+            ld->setIndirect(0, 1, fetchSrc(tgsi.getSrc(0).getIndirect(0), 0, 0));
       }
       return;
    }
@@ -2350,8 +2353,10 @@ Converter::handleSTORE()
             sym = makeSym(TGSI_FILE_BUFFER, r, -1, c, 4 * c);
          }
 
-         mkStore(OP_STORE, TYPE_U32, sym, off, fetchSrc(1, c))
-            ->cache = tgsi.getCacheMode();
+         Instruction *st = mkStore(OP_STORE, TYPE_U32, sym, off, fetchSrc(1, c));
+         st->cache = tgsi.getCacheMode();
+         if (tgsi.getDst(0).isIndirect(0))
+            st->setIndirect(0, 1, fetchSrc(tgsi.getDst(0).getIndirect(0), 0, 0));
       }
       return;
    }
@@ -2432,6 +2437,8 @@ Converter::handleATOM(Value *dst0[4], DataType ty, uint16_t subOp)
          insn = mkOp2(OP_ATOM, ty, dst, sym, fetchSrc(2, c));
          if (tgsi.getSrc(1).getFile() != TGSI_FILE_IMMEDIATE)
             insn->setIndirect(0, 0, off);
+         if (tgsi.getSrc(0).isIndirect(0))
+            insn->setIndirect(0, 1, fetchSrc(tgsi.getSrc(0).getIndirect(0), 0, 0));
          insn->subOp = subOp;
          if (subOp == NV50_IR_SUBOP_ATOM_CAS)
             insn->setSrc(2, fetchSrc(3, 0));
@@ -3200,8 +3207,10 @@ Converter::handleInstruction(const struct tgsi_full_instruction *insn)
       handleATOM(dst0, dstTy, tgsi::opcodeToSubOp(tgsi.getOpcode()));
       break;
    case TGSI_OPCODE_RESQ:
-      mkOp1(OP_SUQ, TYPE_U32, dst0[0],
-            makeSym(TGSI_FILE_BUFFER, tgsi.getSrc(0).getIndex(0), -1, 0, 0));
+      geni = mkOp1(OP_SUQ, TYPE_U32, dst0[0],
+                   makeSym(TGSI_FILE_BUFFER, tgsi.getSrc(0).getIndex(0), -1, 0, 0));
+      if (tgsi.getSrc(0).isIndirect(0))
+         geni->setIndirect(0, 1, fetchSrc(tgsi.getSrc(0).getIndirect(0), 0, 0));
       break;
    case TGSI_OPCODE_IBFE:
    case TGSI_OPCODE_UBFE:
diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
index 7f65b13b55f..0a77dce85c2 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
@@ -1026,7 +1026,10 @@ bool
 NVC0LoweringPass::handleSUQ(Instruction *suq)
 {
    suq->op = OP_MOV;
-   suq->setSrc(0, loadResLength32(NULL, suq->getSrc(0)->reg.fileIndex * 16));
+   suq->setSrc(0, loadResLength32(suq->getIndirect(0, 1),
+                                  suq->getSrc(0)->reg.fileIndex * 16));
+   suq->setIndirect(0, 0, NULL);
+   suq->setIndirect(0, 1, NULL);
    return true;
 }
 
@@ -1034,7 +1037,7 @@ bool
 NVC0LoweringPass::handleATOM(Instruction *atom)
 {
    SVSemantic sv;
-   Value *ptr = atom->getIndirect(0, 0), *base;
+   Value *ptr = atom->getIndirect(0, 0), *ind = atom->getIndirect(0, 1), *base;
 
    switch (atom->src(0).getFile()) {
    case FILE_MEMORY_LOCAL:
@@ -1045,7 +1048,7 @@ NVC0LoweringPass::handleATOM(Instruction *atom)
       break;
    default:
       assert(atom->src(0).getFile() == FILE_MEMORY_GLOBAL);
-      base = loadResInfo64(NULL, atom->getSrc(0)->reg.fileIndex * 16);
+      base = loadResInfo64(ind, atom->getSrc(0)->reg.fileIndex * 16);
       assert(base->reg.size == 8);
       if (ptr)
          base = bld.mkOp2v(OP_ADD, TYPE_U64, base, base, ptr);
@@ -1060,6 +1063,7 @@ NVC0LoweringPass::handleATOM(Instruction *atom)
    atom->getSrc(0)->reg.file = FILE_MEMORY_GLOBAL;
    if (ptr)
       base = bld.mkOp2v(OP_ADD, TYPE_U32, base, base, ptr);
+   atom->setIndirect(0, 1, NULL);
    atom->setIndirect(0, 0, base);
 
    return true;
@@ -1112,6 +1116,9 @@ NVC0LoweringPass::loadResInfo64(Value *ptr, uint32_t off)
    uint8_t b = prog->driver->io.resInfoCBSlot;
    off += prog->driver->io.suInfoBase;
 
+   if (ptr)
+      ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getScratch(), ptr, bld.mkImm(4));
+
    return bld.
       mkLoadv(TYPE_U64, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U64, off), ptr);
 }
@@ -1122,6 +1129,9 @@ NVC0LoweringPass::loadResLength32(Value *ptr, uint32_t off)
    uint8_t b = prog->driver->io.resInfoCBSlot;
    off += prog->driver->io.suInfoBase;
 
+   if (ptr)
+      ptr = bld.mkOp2v(OP_SHL, TYPE_U32, bld.getScratch(), ptr, bld.mkImm(4));
+
    return bld.
       mkLoadv(TYPE_U32, bld.mkSymbol(FILE_MEMORY_CONST, b, TYPE_U64, off + 8), ptr);
 }
@@ -1855,17 +1865,19 @@ NVC0LoweringPass::visit(Instruction *i)
          assert(prog->getType() == Program::TYPE_TESSELLATION_CONTROL);
          i->op = OP_VFETCH;
       } else if (i->src(0).getFile() == FILE_MEMORY_GLOBAL) {
-         Value *ptr = loadResInfo64(NULL, i->getSrc(0)->reg.fileIndex * 16);
+         Value *ind = i->getIndirect(0, 1);
+         Value *ptr = loadResInfo64(ind, i->getSrc(0)->reg.fileIndex * 16);
          // XXX come up with a way not to do this for EVERY little access but
          // rather to batch these up somehow. Unfortunately we've lost the
          // information about the field width by the time we get here.
          Value *offset = bld.loadImm(NULL, i->getSrc(0)->reg.data.offset + typeSizeof(i->sType));
-         Value *length = loadResLength32(NULL, i->getSrc(0)->reg.fileIndex * 16);
+         Value *length = loadResLength32(ind, i->getSrc(0)->reg.fileIndex * 16);
          Value *pred = new_LValue(func, FILE_PREDICATE);
          if (i->src(0).isIndirect(0)) {
             bld.mkOp2(OP_ADD, TYPE_U64, ptr, ptr, i->getIndirect(0, 0));
             bld.mkOp2(OP_ADD, TYPE_U32, offset, offset, i->getIndirect(0, 0));
          }
+         i->setIndirect(0, 1, NULL);
          i->setIndirect(0, 0, ptr);
          bld.mkCmp(OP_SET, CC_GT, TYPE_U32, pred, TYPE_U32, offset, length);
          i->setPredicate(CC_NOT_P, pred);