Mesa (main): aco: don't create DPP instructions with SGPR operands

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: 6afba805349db4e4664065ad71ba6b99e62ff481
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=6afba805349db4e4664065ad71ba6b99e62ff481

Author: Rhys Perry <pendingchaos02 at gmail.com>
Date:   Mon Nov 29 16:34:15 2021 +0000

aco: don't create DPP instructions with SGPR operands

Signed-off-by: Rhys Perry <pendingchaos02 at gmail.com>
Reviewed-by: Daniel Schürmann <daniel at schuermann.dev>
Fixes: 2e6834d4f6c ("aco: combine DPP into VALU before RA")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13976>

---

 src/amd/compiler/aco_ir.cpp               |  2 ++
 src/amd/compiler/tests/test_optimizer.cpp | 20 ++++++++++++++++++--
 2 files changed, 20 insertions(+), 2 deletions(-)

diff --git a/src/amd/compiler/aco_ir.cpp b/src/amd/compiler/aco_ir.cpp
index 5923fea9589..156227a345c 100644
--- a/src/amd/compiler/aco_ir.cpp
+++ b/src/amd/compiler/aco_ir.cpp
@@ -318,6 +318,8 @@ can_use_DPP(const aco_ptr<Instruction>& instr, bool pre_ra)
          return false;
       if (instr->format == Format::VOP3)
          return false;
+      if (instr->operands.size() > 1 && !instr->operands[1].isOfType(RegType::vgpr))
+         return false;
    }
 
    /* there are more cases but those all take 64-bit inputs */
diff --git a/src/amd/compiler/tests/test_optimizer.cpp b/src/amd/compiler/tests/test_optimizer.cpp
index 31a229f99e9..f33de4caf76 100644
--- a/src/amd/compiler/tests/test_optimizer.cpp
+++ b/src/amd/compiler/tests/test_optimizer.cpp
@@ -954,13 +954,14 @@ BEGIN_TEST(optimize.denorm_propagation)
 END_TEST
 
 BEGIN_TEST(optimizer.dpp)
-   //>> v1: %a, v1: %b, s2: %c = p_startpgm
-   if (!setup_cs("v1 v1 s2", GFX10_3))
+   //>> v1: %a, v1: %b, s2: %c, s1: %d = p_startpgm
+   if (!setup_cs("v1 v1 s2 s1", GFX10_3))
       return;
 
    Operand a(inputs[0]);
    Operand b(inputs[1]);
    Operand c(inputs[2]);
+   Operand d(inputs[3]);
 
    /* basic optimization */
    //! v1: %res0 = v_add_f32 %a, %b row_mirror bound_ctrl:1
@@ -1028,6 +1029,21 @@ BEGIN_TEST(optimizer.dpp)
    Temp res8 = bld.vop2(aco_opcode::v_cndmask_b32, bld.def(v1), tmp8, b, c);
    writeout(8, res8);
 
+   /* sgprs */
+   //! v1: %tmp9 = v_mov_b32 %a row_mirror bound_ctrl:1
+   //! v1: %res9 = v_add_f32 %tmp9, %d
+   //! p_unit_test 9, %res9
+   Temp tmp9 = bld.vop1_dpp(aco_opcode::v_mov_b32, bld.def(v1), a, dpp_row_mirror);
+   Temp res9 = bld.vop2_e64(aco_opcode::v_add_f32, bld.def(v1), tmp9, d);
+   writeout(9, res9);
+
+   //! v1: %tmp10 = v_mov_b32 %a row_mirror bound_ctrl:1
+   //! v1: %res10 = v_add_f32 %d, %tmp10
+   //! p_unit_test 10, %res10
+   Temp tmp10 = bld.vop1_dpp(aco_opcode::v_mov_b32, bld.def(v1), a, dpp_row_mirror);
+   Temp res10 = bld.vop2(aco_opcode::v_add_f32, bld.def(v1), d, tmp10);
+   writeout(10, res10);
+
    finish_opt_test();
 END_TEST