aco: limit hard clauses to 63 instructions

See https://reviews.llvm.org/D127391

fossil-db (gfx1100):
Totals from 4 (0.00% of 161689) affected shaders:
Latency: 24545 -> 24539 (-0.02%)
InvThroughput: 102867 -> 102835 (-0.03%)

fossil-db (navi10):
Totals from 4 (0.00% of 161220) affected shaders:
Latency: 25969 -> 25959 (-0.04%)
InvThroughput: 112917 -> 112869 (-0.04%)

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17710>

diff --git a/src/amd/compiler/aco_form_hard_clauses.cpp b/src/amd/compiler/aco_form_hard_clauses.cpp
index c32a813..32c0b56 100644 (file)
--- a/src/amd/compiler/aco_form_hard_clauses.cpp
+++ b/src/amd/compiler/aco_form_hard_clauses.cpp
@@ -248,7 +248,7 @@ form_hard_clauses(Program* program)
 {
    for (Block& block : program->blocks) {
       unsigned num_instrs = 0;
-      aco_ptr<Instruction> current_instrs[64];
+      aco_ptr<Instruction> current_instrs[63];
       clause_type current_type = clause_other;
 
       std::vector<aco_ptr<Instruction>> new_instructions;
@@ -259,7 +259,7 @@ form_hard_clauses(Program* program)
          aco_ptr<Instruction>& instr = block.instructions[i];
 
          clause_type type = get_type(program, instr);
-         if (type != current_type || num_instrs == 64 ||
+         if (type != current_type || num_instrs == 63 ||
              (num_instrs && !should_form_clause(current_instrs[0].get(), instr.get()))) {
             emit_clause(bld, num_instrs, current_instrs);
             num_instrs = 0;

diff --git a/src/amd/compiler/tests/test_hard_clause.cpp b/src/amd/compiler/tests/test_hard_clause.cpp
index 5274c24..d6b5a6b 100644 (file)
--- a/src/amd/compiler/tests/test_hard_clause.cpp
+++ b/src/amd/compiler/tests/test_hard_clause.cpp
@@ -173,30 +173,30 @@ BEGIN_TEST(form_hard_clauses.size)
    create_smem();
 
    //>> p_unit_test 1
-   //! s_clause imm:63
-   //; for i in range(64):
+   //! s_clause imm:62
+   //; for i in range(63):
    //;    search_re('s_load_dword')
    bld.pseudo(aco_opcode::p_unit_test, Operand::c32(1u));
-   for (unsigned i = 0; i < 64; i++)
+   for (unsigned i = 0; i < 63; i++)
       create_smem();
 
    //>> p_unit_test 2
-   //! s_clause imm:63
-   //; for i in range(65):
+   //! s_clause imm:62
+   //; for i in range(64):
    //;    search_re('s_load_dword')
    bld.pseudo(aco_opcode::p_unit_test, Operand::c32(2u));
-   for (unsigned i = 0; i < 65; i++)
+   for (unsigned i = 0; i < 64; i++)
       create_smem();
 
    //>> p_unit_test 3
-   //! s_clause imm:63
-   //; for i in range(64):
+   //! s_clause imm:62
+   //; for i in range(63):
    //;    search_re('s_load_dword')
    //! s_clause imm:1
    //; search_re('s_load_dword')
    //; search_re('s_load_dword')
    bld.pseudo(aco_opcode::p_unit_test, Operand::c32(3u));
-   for (unsigned i = 0; i < 66; i++)
+   for (unsigned i = 0; i < 65; i++)
       create_smem();
 
    finish_form_hard_clause_test();
@@ -370,22 +370,22 @@ BEGIN_TEST(form_hard_clauses.stores)
    /* Unimportant pass limitations */
    //>> p_unit_test 4
    //; search_re('buffer_store_dword')
-   //! s_clause imm:62
-   //; for i in range(63):
+   //! s_clause imm:61
+   //; for i in range(62):
    //;    search_re('buffer_load_dword')
    //; search_re('buffer_load_dword')
    bld.pseudo(aco_opcode::p_unit_test, Operand::c32(4u));
    create_mubuf_store();
-   for (unsigned i = 0; i < 64; i++)
+   for (unsigned i = 0; i < 63; i++)
       create_mubuf();
 
    //>> p_unit_test 5
-   //! s_clause imm:63
-   //; for i in range(64):
+   //! s_clause imm:62
+   //; for i in range(63):
    //;    search_re('buffer_load_dword')
    //; search_re('buffer_store_dword')
    bld.pseudo(aco_opcode::p_unit_test, Operand::c32(5u));
-   for (unsigned i = 0; i < 64; i++)
+   for (unsigned i = 0; i < 63; i++)
       create_mubuf();
    create_mubuf_store();