aco: Don't use opsel for p_insert.

This doesn't make sense, opsel preserves the not selected half of the register,
p_insert zeros it.

No Foz-DB changes.

Signed-off-by: Georg Lehmann <dadschoorse@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Fixes: 54292e99c78 ("aco: optimize 32-bit extracts and inserts using SDWA")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19253>

diff --git a/src/amd/compiler/aco_optimizer.cpp b/src/amd/compiler/aco_optimizer.cpp
index 48f429f..bc1fbf2 100644 (file)
--- a/src/amd/compiler/aco_optimizer.cpp
+++ b/src/amd/compiler/aco_optimizer.cpp
@@ -3201,21 +3201,13 @@ apply_insert(opt_ctx& ctx, aco_ptr<Instruction>& instr)
    SubdwordSel sel = parse_insert(def_info.instr);
    assert(sel);
 
-   if (instr->isVOP3() && sel.size() == 2 && !sel.sign_extend() &&
-       can_use_opsel(ctx.program->gfx_level, instr->opcode, -1)) {
-      if (instr->vop3().opsel & (1 << 3))
-         return false;
-      if (sel.offset())
-         instr->vop3().opsel |= 1 << 3;
-   } else {
-      if (!can_use_SDWA(ctx.program->gfx_level, instr, true))
-         return false;
+   if (!can_use_SDWA(ctx.program->gfx_level, instr, true))
+      return false;
 
-      to_SDWA(ctx, instr);
-      if (instr->sdwa().dst_sel.size() != 4)
-         return false;
-      static_cast<SDWA_instruction*>(instr.get())->dst_sel = sel;
-   }
+   to_SDWA(ctx, instr);
+   if (instr->sdwa().dst_sel.size() != 4)
+      return false;
+   static_cast<SDWA_instruction*>(instr.get())->dst_sel = sel;
 
    instr->definitions[0].swapTemp(def_info.instr->definitions[0]);
    ctx.info[instr->definitions[0].tempId()].label = 0;

diff --git a/src/amd/compiler/tests/test_sdwa.cpp b/src/amd/compiler/tests/test_sdwa.cpp
index 8d784ef..9df87eb 100644 (file)
--- a/src/amd/compiler/tests/test_sdwa.cpp
+++ b/src/amd/compiler/tests/test_sdwa.cpp
@@ -496,17 +496,15 @@ BEGIN_TEST(optimize.sdwa.insert)
       bld.pseudo(ins, bld.def(v1), val, Operand::c32(1u), Operand::c32(16u));
       writeout(10, val);
 
-      //~gfx8! v1: %tmp11 = v_sub_i16 %a, %b
-      //~gfx8! v1: %res11 = p_insert %tmp11, 0, 16
-      //~gfx(9|10)! v1: %res11 = v_sub_i16 %a, %b
-      //~gfx(8|9|10)! p_unit_test 11, %res11
+      //~gfx[^7]! v1: %tmp11 = v_sub_i16 %a, %b
+      //~gfx[^7]! v1: %res11 = p_insert %tmp11, 0, 16
+      //~gfx[^7]! p_unit_test 11, %res11
       val = bld.vop3(aco_opcode::v_sub_i16, bld.def(v1), inputs[0], inputs[1]);
       writeout(11, bld.pseudo(ins, bld.def(v1), val, Operand::zero(), Operand::c32(16u)));
 
-      //~gfx8! v1: %tmp12 = v_sub_i16 %a, %b
-      //~gfx8! v1: %res12 = p_insert %tmp12, 1, 16
-      //~gfx(9|10)! v1: %res12 = v_sub_i16 %a, %b opsel_hi
-      //~gfx(8|9|10)! p_unit_test 12, %res12
+      //~gfx[^7]! v1: %tmp12 = v_sub_i16 %a, %b
+      //~gfx[^7]! v1: %res12 = p_insert %tmp12, 1, 16
+      //~gfx[^7]! p_unit_test 12, %res12
       val = bld.vop3(aco_opcode::v_sub_i16, bld.def(v1), inputs[0], inputs[1]);
       writeout(12, bld.pseudo(ins, bld.def(v1), val, Operand::c32(1u), Operand::c32(16u)));