fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2025-12-21 15:50:11 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

aco: add v_nop inbetween exec write and VMEM/DS/FLAT

Browse source 
LLVM and the proprietary compiler seem to do this

Fixes: b01847bd9 ("aco/gfx10: Fix mitigation of VMEMtoScalarWriteHazard.")
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
This commit is contained in:
Rhys Perry 2019-11-27 17:20:15 +00:00
parent 54742e157d
commit a9fc81b098
1 changed files with 8 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

13
src/amd/compiler/aco_insert_NOPs.cpp
View file

@ -385,7 +385,7 @@ void insert_NOPs_gfx8_9(Program* program)
} }
} }
void handle_instruction_gfx10(NOP_ctx_gfx10 &ctx, aco_ptr<Instruction>& instr, void handle_instruction_gfx10(Program *program, NOP_ctx_gfx10 &ctx, aco_ptr<Instruction>& instr,
std::vector<aco_ptr<Instruction>>& old_instructions, std::vector<aco_ptr<Instruction>>& old_instructions,
std::vector<aco_ptr<Instruction>>& new_instructions) std::vector<aco_ptr<Instruction>>& new_instructions)
{ {
@ -396,6 +396,9 @@ void handle_instruction_gfx10(NOP_ctx_gfx10 &ctx, aco_ptr<Instruction>& instr,
instr->format == Format::SCRATCH || instr->format == Format::DS) { instr->format == Format::SCRATCH || instr->format == Format::DS) {
/* Remember all SGPRs that are read by the VMEM instruction */ /* Remember all SGPRs that are read by the VMEM instruction */
mark_read_regs(instr, ctx.sgprs_read_by_VMEM); mark_read_regs(instr, ctx.sgprs_read_by_VMEM);
ctx.sgprs_read_by_VMEM.set(exec);
if (program->wave_size == 64)
ctx.sgprs_read_by_VMEM.set(exec_hi);
} else if (instr->isSALU() || instr->format == Format::SMEM) { } else if (instr->isSALU() || instr->format == Format::SMEM) {
/* Check if SALU writes an SGPR that was previously read by the VALU */ /* Check if SALU writes an SGPR that was previously read by the VALU */
if (check_written_regs(instr, ctx.sgprs_read_by_VMEM)) { if (check_written_regs(instr, ctx.sgprs_read_by_VMEM)) {
@ -528,7 +531,7 @@ void handle_instruction_gfx10(NOP_ctx_gfx10 &ctx, aco_ptr<Instruction>& instr,
} }
} }
void handle_block_gfx10(NOP_ctx_gfx10& ctx, Block& block) void handle_block_gfx10(Program *program, NOP_ctx_gfx10& ctx, Block& block)
{ {
if (block.instructions.empty()) if (block.instructions.empty())
return; return;
@ -537,7 +540,7 @@ void handle_block_gfx10(NOP_ctx_gfx10& ctx, Block& block)
instructions.reserve(block.instructions.size()); instructions.reserve(block.instructions.size());
for (aco_ptr<Instruction>& instr : block.instructions) { for (aco_ptr<Instruction>& instr : block.instructions) {
handle_instruction_gfx10(ctx, instr, block.instructions, instructions); handle_instruction_gfx10(program, ctx, instr, block.instructions, instructions);
instructions.emplace_back(std::move(instr)); instructions.emplace_back(std::move(instr));
} }
@ -562,7 +565,7 @@ void mitigate_hazards_gfx10(Program *program)
for (unsigned b : program->blocks[idx].linear_preds) for (unsigned b : program->blocks[idx].linear_preds)
loop_block_ctx.join(all_ctx[b]); loop_block_ctx.join(all_ctx[b]);
handle_block_gfx10(loop_block_ctx, program->blocks[idx]); handle_block_gfx10(program, loop_block_ctx, program->blocks[idx]);
/* We only need to continue if the loop header context changed */ /* We only need to continue if the loop header context changed */
if (idx == loop_header_indices.top() && loop_block_ctx == all_ctx[idx]) if (idx == loop_header_indices.top() && loop_block_ctx == all_ctx[idx])
@ -577,7 +580,7 @@ void mitigate_hazards_gfx10(Program *program)
for (unsigned b : block.linear_preds) for (unsigned b : block.linear_preds)
ctx.join(all_ctx[b]); ctx.join(all_ctx[b]);
handle_block_gfx10(ctx, block); handle_block_gfx10(program, ctx, block);
} }
} }