aco/gfx10: Fix VcmpxExecWARHazard mitigation.
authorTimur Kristóf <timur.kristof@gmail.com>
Fri, 24 Jan 2020 14:17:44 +0000 (15:17 +0100)
committerMarge Bot <eric+marge@anholt.net>
Fri, 24 Jan 2020 16:21:08 +0000 (16:21 +0000)
The SOPP instruction shouldn't have a definition, and its block
should be set to -1 in order to prevent it from being recognized
as a branch.
Also fix a typo in the readme.

Fixes: d6dfce02d074d615a3b88a3fccd8ee8c7e13c010
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3552>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3552>

src/amd/compiler/README.md
src/amd/compiler/aco_insert_NOPs.cpp

index 822ecd79cbc4a0b7fca4b8f1ef5f1862e55e87af..7c7e68f458eee53eb56238b59e9acdbb98de7720 100644 (file)
@@ -154,7 +154,6 @@ A VALU instruction or an `s_waitcnt vmcnt(0)` between the two instructions.
 
 Triggered by:
 An SMEM instruction reads an SGPR. Then, a VALU instruction writes that same SGPR.
-Despite LLVM
 
 Mitigated by:
 Any non-SOPP SALU instruction (except `s_setvskip`, `s_version`, and any non-lgkmcnt `s_waitcnt`).
index b9eaaed96db0697f85cca5208cd93d1124ef0741..689d5e25acc2a2cc3a14267edf9c4ea7d0deedea 100644 (file)
@@ -491,9 +491,9 @@ void handle_instruction_gfx10(Program *program, NOP_ctx_gfx10 &ctx, aco_ptr<Inst
          ctx.has_nonVALU_exec_read = false;
 
          /* Insert s_waitcnt_depctr instruction with magic imm to mitigate the problem */
-         aco_ptr<SOPP_instruction> depctr{create_instruction<SOPP_instruction>(aco_opcode::s_waitcnt_depctr, Format::SOPP, 0, 1)};
+         aco_ptr<SOPP_instruction> depctr{create_instruction<SOPP_instruction>(aco_opcode::s_waitcnt_depctr, Format::SOPP, 0, 0)};
          depctr->imm = 0xfffe;
-         depctr->definitions[0] = Definition(sgpr_null, s1);
+         depctr->block = -1;
          new_instructions.emplace_back(std::move(depctr));
       } else if (instr_writes_sgpr(instr)) {
          /* Any VALU instruction that writes an SGPR mitigates the problem */