aco: insert s_nop before VGPR deallocation

A hazard requires this, apparently.

https://reviews.llvm.org/D155681

fossil-db (gfx1100):
Totals from 23175 (17.36% of 133461) affected shaders:
Instrs: 41240100 -> 41263275 (+0.06%)
CodeSize: 211858524 -> 211951224 (+0.04%)
Latency: 469738570 -> 469738576 (+0.00%)
InvThroughput: 58013998 -> 58013999 (+0.00%)

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24320>

diff --git a/src/amd/compiler/aco_ir.cpp b/src/amd/compiler/aco_ir.cpp
index 179d282..ecf34ed 100644 (file)
--- a/src/amd/compiler/aco_ir.cpp
+++ b/src/amd/compiler/aco_ir.cpp
@@ -1351,6 +1351,8 @@ dealloc_vgprs(Program* program)
    Builder bld(program);
    if (!block.instructions.empty() && block.instructions.back()->opcode == aco_opcode::s_endpgm) {
       bld.reset(&block.instructions, block.instructions.begin() + (block.instructions.size() - 1));
+      /* Due to a hazard, an s_nop is needed before "s_sendmsg sendmsg_dealloc_vgprs". */
+      bld.sopp(aco_opcode::s_nop, -1, 0);
       bld.sopp(aco_opcode::s_sendmsg, -1, sendmsg_dealloc_vgprs);
    }