-#include <map>
+#include <vector>
+#include <algorithm>
#include "aco_ir.h"
#include "common/sid.h"
#include "ac_shader_util.h"
+#include "util/u_math.h"
namespace aco {
struct asm_context {
Program *program;
enum chip_class chip_class;
- std::map<int, SOPP_instruction*> branches;
+ std::vector<std::pair<int, SOPP_instruction*>> branches;
std::vector<unsigned> constaddrs;
const int16_t* opcode;
// TODO: keep track of branch instructions referring blocks
encoding |=
!instr->definitions.empty() && !(instr->definitions[0].physReg() == scc) ?
instr->definitions[0].physReg() << 16 :
- !instr->operands.empty() && !(instr->operands[0].physReg() == scc) ?
+ !instr->operands.empty() && instr->operands[0].physReg() <= 127 ?
instr->operands[0].physReg() << 16 : 0;
encoding |= sopk->imm;
out.push_back(encoding);
encoding |= opcode << 16;
encoding |= (uint16_t) sopp->imm;
if (sopp->block != -1)
- ctx.branches.insert({out.size(), sopp});
+ ctx.branches.emplace_back(out.size(), sopp);
out.push_back(encoding);
break;
}
}
if (is_load || instr->operands.size() >= 3) { /* SDATA */
- encoding |= (is_load ? instr->definitions[0].physReg().reg : instr->operands[2].physReg().reg) << 6;
+ encoding |= (is_load ? instr->definitions[0].physReg() : instr->operands[2].physReg()) << 6;
}
if (instr->operands.size() >= 1) { /* SBASE */
- encoding |= instr->operands[0].physReg().reg >> 1;
+ encoding |= instr->operands[0].physReg() >> 1;
}
out.push_back(encoding);
case Format::VOP2: {
uint32_t encoding = 0;
encoding |= opcode << 25;
- encoding |= (0xFF & instr->definitions[0].physReg().reg) << 17;
- encoding |= (0xFF & instr->operands[1].physReg().reg) << 9;
- encoding |= instr->operands[0].physReg().reg;
+ encoding |= (0xFF & instr->definitions[0].physReg()) << 17;
+ encoding |= (0xFF & instr->operands[1].physReg()) << 9;
+ encoding |= instr->operands[0].physReg();
out.push_back(encoding);
break;
}
case Format::VOP1: {
uint32_t encoding = (0b0111111 << 25);
- encoding |= (0xFF & instr->definitions[0].physReg().reg) << 17;
+ if (!instr->definitions.empty())
+ encoding |= (0xFF & instr->definitions[0].physReg()) << 17;
encoding |= opcode << 9;
- encoding |= instr->operands[0].physReg().reg;
+ if (!instr->operands.empty())
+ encoding |= instr->operands[0].physReg();
out.push_back(encoding);
break;
}
case Format::VOPC: {
uint32_t encoding = (0b0111110 << 25);
encoding |= opcode << 17;
- encoding |= (0xFF & instr->operands[1].physReg().reg) << 9;
- encoding |= instr->operands[0].physReg().reg;
+ encoding |= (0xFF & instr->operands[1].physReg()) << 9;
+ encoding |= instr->operands[0].physReg();
out.push_back(encoding);
break;
}
}
assert(encoding);
- encoding |= (0xFF & instr->definitions[0].physReg().reg) << 18;
+ encoding |= (0xFF & instr->definitions[0].physReg()) << 18;
encoding |= opcode << 16;
encoding |= interp->attribute << 10;
encoding |= interp->component << 8;
if (instr->opcode == aco_opcode::v_interp_mov_f32)
encoding |= (0x3 & instr->operands[0].constantValue());
else
- encoding |= (0xFF & instr->operands[0].physReg().reg);
+ encoding |= (0xFF & instr->operands[0].physReg());
out.push_back(encoding);
break;
}
encoding |= (0xFF & reg) << 16;
reg = instr->operands.size() >= 2 && !(instr->operands[1].physReg() == m0) ? instr->operands[1].physReg() : 0;
encoding |= (0xFF & reg) << 8;
- encoding |= (0xFF & instr->operands[0].physReg().reg);
+ encoding |= (0xFF & instr->operands[0].physReg());
out.push_back(encoding);
break;
}
encoding |= instr->operands[2].physReg() << 24;
encoding |= (mubuf->tfe ? 1 : 0) << 23;
encoding |= (instr->operands[1].physReg() >> 2) << 16;
- unsigned reg = instr->operands.size() > 3 ? instr->operands[3].physReg() : instr->definitions[0].physReg().reg;
+ unsigned reg = instr->operands.size() > 3 ? instr->operands[3].physReg() : instr->definitions[0].physReg();
encoding |= (0xFF & reg) << 8;
- encoding |= (0xFF & instr->operands[0].physReg().reg);
+ encoding |= (0xFF & instr->operands[0].physReg());
out.push_back(encoding);
break;
}
uint32_t img_format = ac_get_tbuffer_format(ctx.chip_class, mtbuf->dfmt, mtbuf->nfmt);
uint32_t encoding = (0b111010 << 26);
+ assert(img_format <= 0x7F);
assert(!mtbuf->dlc || ctx.chip_class >= GFX10);
encoding |= (mtbuf->dlc ? 1 : 0) << 15; /* DLC bit replaces one bit of the OPCODE on GFX10 */
encoding |= (mtbuf->glc ? 1 : 0) << 14;
out.push_back(encoding);
encoding = 0;
- encoding |= instr->operands[2].physReg().reg << 24;
+ encoding |= instr->operands[2].physReg() << 24;
encoding |= (mtbuf->tfe ? 1 : 0) << 23;
encoding |= (mtbuf->slc ? 1 : 0) << 22;
- encoding |= (instr->operands[1].physReg().reg >> 2) << 16;
- unsigned reg = instr->operands.size() > 3 ? instr->operands[3].physReg().reg : instr->definitions[0].physReg().reg;
+ encoding |= (instr->operands[1].physReg() >> 2) << 16;
+ unsigned reg = instr->operands.size() > 3 ? instr->operands[3].physReg() : instr->definitions[0].physReg();
encoding |= (0xFF & reg) << 8;
- encoding |= (0xFF & instr->operands[0].physReg().reg);
+ encoding |= (0xFF & instr->operands[0].physReg());
if (ctx.chip_class >= GFX10) {
encoding |= (((opcode & 0x08) >> 4) << 21); /* MSB of 4-bit OPCODE */
}
encoding |= (0xF & mimg->dmask) << 8;
out.push_back(encoding);
- encoding = (0xFF & instr->operands[0].physReg().reg); /* VADDR */
+ encoding = (0xFF & instr->operands[0].physReg()); /* VADDR */
if (!instr->definitions.empty()) {
- encoding |= (0xFF & instr->definitions[0].physReg().reg) << 8; /* VDATA */
+ encoding |= (0xFF & instr->definitions[0].physReg()) << 8; /* VDATA */
} else if (instr->operands.size() == 4) {
- encoding |= (0xFF & instr->operands[3].physReg().reg) << 8; /* VDATA */
+ encoding |= (0xFF & instr->operands[3].physReg()) << 8; /* VDATA */
}
encoding |= (0x1F & (instr->operands[1].physReg() >> 2)) << 16; /* T# (resource) */
if (instr->operands.size() > 2)
if (ctx.chip_class <= GFX9) {
assert(flat->offset <= 0x1fff);
encoding |= flat->offset & 0x1fff;
+ } else if (instr->format == Format::FLAT) {
+ /* GFX10 has a 12-bit immediate OFFSET field,
+ * but it has a hw bug: it ignores the offset, called FlatSegmentOffsetBug
+ */
+ assert(flat->offset == 0);
} else {
- assert(flat->offset <= 0x0fff);
- encoding |= flat->offset & 0x0fff;
+ assert(flat->offset <= 0xfff);
+ encoding |= flat->offset & 0xfff;
}
if (instr->format == Format::SCRATCH)
encoding |= 1 << 14;
encoding = (0xFF & instr->operands[0].physReg());
if (!instr->definitions.empty())
encoding |= (0xFF & instr->definitions[0].physReg()) << 24;
- else
+ if (instr->operands.size() >= 3)
encoding |= (0xFF & instr->operands[2].physReg()) << 8;
if (!instr->operands[1].isUndefined()) {
assert(ctx.chip_class >= GFX10 || instr->operands[1].physReg() != 0x7F);
encoding |= exp->dest << 4;
encoding |= exp->enabled_mask;
out.push_back(encoding);
- encoding = 0xFF & exp->operands[0].physReg().reg;
- encoding |= (0xFF & exp->operands[1].physReg().reg) << 8;
- encoding |= (0xFF & exp->operands[2].physReg().reg) << 16;
- encoding |= (0xFF & exp->operands[3].physReg().reg) << 24;
+ encoding = 0xFF & exp->operands[0].physReg();
+ encoding |= (0xFF & exp->operands[1].physReg()) << 8;
+ encoding |= (0xFF & exp->operands[2].physReg()) << 16;
+ encoding |= (0xFF & exp->operands[3].physReg()) << 24;
out.push_back(encoding);
break;
}
opcode = opcode + 0x270;
}
- // TODO: op_sel
uint32_t encoding;
if (ctx.chip_class <= GFX9) {
encoding = (0b110100 << 26);
encoding |= (vop3->clamp ? 1 : 0) << 15;
for (unsigned i = 0; i < 3; i++)
encoding |= vop3->abs[i] << (8+i);
+ for (unsigned i = 0; i < 4; i++)
+ encoding |= vop3->opsel[i] << (11+i);
if (instr->definitions.size() == 2)
encoding |= instr->definitions[1].physReg() << 8;
- encoding |= (0xFF & instr->definitions[0].physReg().reg);
+ encoding |= (0xFF & instr->definitions[0].physReg());
out.push_back(encoding);
encoding = 0;
if (instr->opcode == aco_opcode::v_interp_mov_f32) {
for (unsigned i = 0; i < 3; i++)
encoding |= vop3->neg[i] << (29+i);
out.push_back(encoding);
- return;
} else if (instr->isDPP()){
/* first emit the instruction without the DPP operand */
encoding |= dpp->neg[0] << 20;
encoding |= dpp->bound_ctrl << 19;
encoding |= dpp->dpp_ctrl << 8;
- encoding |= (0xFF) & dpp_op.physReg().reg;
+ encoding |= (0xFF) & dpp_op.physReg();
out.push_back(encoding);
return;
} else {
unreachable("unimplemented instruction format");
}
+ break;
}
/* append literal dword */
}
}
+static void fix_branches_gfx10(asm_context& ctx, std::vector<uint32_t>& out)
+{
+ /* Branches with an offset of 0x3f are buggy on GFX10, we workaround by inserting NOPs if needed. */
+ bool gfx10_3f_bug = false;
+
+ do {
+ auto buggy_branch_it = std::find_if(ctx.branches.begin(), ctx.branches.end(), [&ctx](const auto &branch) -> bool {
+ return ((int)ctx.program->blocks[branch.second->block].offset - branch.first - 1) == 0x3f;
+ });
+
+ gfx10_3f_bug = buggy_branch_it != ctx.branches.end();
+
+ if (gfx10_3f_bug) {
+ /* Insert an s_nop after the branch */
+ constexpr uint32_t s_nop_0 = 0xbf800000u;
+ int s_nop_pos = buggy_branch_it->first + 1;
+ auto out_pos = std::next(out.begin(), s_nop_pos);
+ out.insert(out_pos, s_nop_0);
+
+ /* Update the offset of each affected block */
+ for (Block& block : ctx.program->blocks) {
+ if (block.offset > (unsigned)buggy_branch_it->first)
+ block.offset++;
+ }
+
+ /* Update the branches following the current one */
+ for (auto branch_it = std::next(buggy_branch_it); branch_it != ctx.branches.end(); ++branch_it)
+ branch_it->first++;
+
+ /* Find first constant address after the inserted instruction */
+ auto caddr_it = std::find_if(ctx.constaddrs.begin(), ctx.constaddrs.end(), [s_nop_pos](const int &caddr_pos) -> bool {
+ return caddr_pos >= s_nop_pos;
+ });
+
+ /* Update the locations of constant addresses */
+ for (; caddr_it != ctx.constaddrs.end(); ++caddr_it)
+ (*caddr_it)++;
+
+ }
+ } while (gfx10_3f_bug);
+}
+
void fix_branches(asm_context& ctx, std::vector<uint32_t>& out)
{
- for (std::pair<int, SOPP_instruction*> branch : ctx.branches)
- {
+ if (ctx.chip_class >= GFX10)
+ fix_branches_gfx10(ctx, out);
+
+ for (std::pair<int, SOPP_instruction*> &branch : ctx.branches) {
int offset = (int)ctx.program->blocks[branch.second->block].offset - branch.first - 1;
out[branch.first] |= (uint16_t) offset;
}
}
fix_branches(ctx, code);
+
+ unsigned exec_size = code.size() * sizeof(uint32_t);
+
+ if (program->chip_class >= GFX10) {
+ /* Pad output with s_code_end so instruction prefetching doesn't cause
+ * page faults */
+ unsigned final_size = align(code.size() + 3 * 16, 16);
+ while (code.size() < final_size)
+ code.push_back(0xbf9f0000u);
+ }
+
fix_constaddrs(ctx, code);
- unsigned constant_data_offset = code.size() * sizeof(uint32_t);
while (program->constant_data.size() % 4u)
program->constant_data.push_back(0);
/* Copy constant data */
code.insert(code.end(), (uint32_t*)program->constant_data.data(),
(uint32_t*)(program->constant_data.data() + program->constant_data.size()));
- return constant_data_offset;
+ return exec_size;
}
}