<html>
  <head>
    <meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
  </head>
  <body text="#000000" bgcolor="#FFFFFF">
    <p><br>
    </p>
    <div class="moz-cite-prefix">On 4/23/19 5:16 PM, Jason Ekstrand
      wrote:<br>
    </div>
    <blockquote type="cite"
cite="mid:CAOFGe94RBM09CMhhYpe0cgTpdQdRDN3y2VmGwFExpXpWhuvW8w@mail.gmail.com">
      <meta http-equiv="content-type" content="text/html; charset=UTF-8">
      <div dir="ltr">
        <div class="gmail_quote">
          <div dir="ltr" class="gmail_attr">On Tue, Apr 23, 2019 at 7:46
            AM Samuel Pitoiset <<a
              href="mailto:samuel.pitoiset@gmail.com"
              moz-do-not-send="true">samuel.pitoiset@gmail.com</a>>
            wrote:<br>
          </div>
          <blockquote class="gmail_quote" style="margin:0px 0px 0px
            0.8ex;border-left:1px solid
            rgb(204,204,204);padding-left:1ex"><br>
            On 4/23/19 10:45 AM, Bas Nieuwenhuizen wrote:<br>
            > On Tue, Apr 23, 2019 at 9:35 AM Samuel Pitoiset<br>
            > <<a href="mailto:samuel.pitoiset@gmail.com"
              target="_blank" moz-do-not-send="true">samuel.pitoiset@gmail.com</a>>
            wrote:<br>
            >> Signed-off-by: Samuel Pitoiset <<a
              href="mailto:samuel.pitoiset@gmail.com" target="_blank"
              moz-do-not-send="true">samuel.pitoiset@gmail.com</a>><br>
            >> ---<br>
            >>   src/amd/vulkan/radv_shader.c                 |  2
            +-<br>
            >>   src/compiler/nir/nir.h                       |  3
            ++-<br>
            >>   src/compiler/nir/nir_opt_if.c                | 17
            ++++++++++-------<br>
            >>   src/freedreno/ir3/ir3_nir.c                  |  2
            +-<br>
            >>   src/gallium/auxiliary/nir/tgsi_to_nir.c      |  2
            +-<br>
            >>   src/gallium/drivers/freedreno/a2xx/ir2_nir.c |  2
            +-<br>
            >>   src/gallium/drivers/radeonsi/si_shader_nir.c |  2
            +-<br>
            >>   src/intel/compiler/brw_nir.c                 |  2
            +-<br>
            >>   src/mesa/state_tracker/st_glsl_to_nir.cpp    |  2
            +-<br>
            >>   9 files changed, 19 insertions(+), 15
            deletions(-)<br>
            >><br>
            >> diff --git a/src/amd/vulkan/radv_shader.c
            b/src/amd/vulkan/radv_shader.c<br>
            >> index 13f1f9aa9dc..54a4e732230 100644<br>
            >> --- a/src/amd/vulkan/radv_shader.c<br>
            >> +++ b/src/amd/vulkan/radv_shader.c<br>
            >> @@ -158,7 +158,7 @@ radv_optimize_nir(struct
            nir_shader *shader, bool optimize_conservatively,<br>
            >>                          NIR_PASS(progress, shader,
            nir_opt_remove_phis);<br>
            >>                           NIR_PASS(progress,
            shader, nir_opt_dce);<br>
            >>                   }<br>
            >> -                NIR_PASS(progress, shader,
            nir_opt_if, true);<br>
            >> +                NIR_PASS(progress, shader,
            nir_opt_if, true, false);<br>
            >>                   NIR_PASS(progress, shader,
            nir_opt_dead_cf);<br>
            >>                   NIR_PASS(progress, shader,
            nir_opt_cse);<br>
            >>                   NIR_PASS(progress, shader,
            nir_opt_peephole_select, 8, true, true);<br>
            >> diff --git a/src/compiler/nir/nir.h
            b/src/compiler/nir/nir.h<br>
            >> index 7d2062d3691..d7506d6ddd1 100644<br>
            >> --- a/src/compiler/nir/nir.h<br>
            >> +++ b/src/compiler/nir/nir.h<br>
            >> @@ -3474,7 +3474,8 @@ bool nir_opt_gcm(nir_shader
            *shader, bool value_number);<br>
            >><br>
            >>   bool nir_opt_idiv_const(nir_shader *shader,
            unsigned min_bit_size);<br>
            >><br>
            >> -bool nir_opt_if(nir_shader *shader, bool
            aggressive_last_continue);<br>
            >> +bool nir_opt_if(nir_shader *shader, bool
            aggressive_last_continue,<br>
            >> +                bool skip_alu_of_phi);<br>
            > Can we have a flag for this instead (e.g. something
            like<br>
            > nir_opt_if_skip_alu_of_phi)? I think have a function
            with a bunch of<br>
            > bools is less than ideal as you can't see at the
            calling site what is<br>
            > for what arg.<br>
            Yes, that seems better to me.<br>
          </blockquote>
          <div><br>
          </div>
          <div>This is the worst kind of hack all around.  We're making
            NIR more complicated and adding a flag to disable a useful
            and correct piece of an optimization, not because it causes
            a perf regression but because the back-end compiler is
            broken and this is easier than fixing it properly. 
            Seriously?  Can't we just fix the LLVM back-end?  Or, if
            this optimization is actually doing something wrong, fix
            it?  Or maybe actually figure out what pattern is causing
            LLVM to fall over and have a hack in your NIR -> LLVM
            pass?  On the list of "good ways to fix this problem", this
            seems to be pretty far down if it hasn't fallen off the
            bottom.<br>
          </div>
        </div>
      </div>
    </blockquote>
    <p>Best hack of the month? :-)</p>
    <p>As discussed over IRC, this is definitely not the best solution,
      I don't like it either as I said.</p>
    <p>I will work on a different solution maybe in our NIR->LLVM
      pass.<br>
    </p>
    <p>The aggressive_last_continue option should also be removed (make
      it default?).<br>
    </p>
    <blockquote type="cite"
cite="mid:CAOFGe94RBM09CMhhYpe0cgTpdQdRDN3y2VmGwFExpXpWhuvW8w@mail.gmail.com">
      <div dir="ltr">
        <div class="gmail_quote">
          <div><br>
          </div>
          <div>--Jason<br>
          </div>
          <div> </div>
          <blockquote class="gmail_quote" style="margin:0px 0px 0px
            0.8ex;border-left:1px solid
            rgb(204,204,204);padding-left:1ex">
            >>   bool nir_opt_intrinsics(nir_shader *shader);<br>
            >><br>
            >> diff --git a/src/compiler/nir/nir_opt_if.c
            b/src/compiler/nir/nir_opt_if.c<br>
            >> index f674185f1e2..149b3bd1659 100644<br>
            >> --- a/src/compiler/nir/nir_opt_if.c<br>
            >> +++ b/src/compiler/nir/nir_opt_if.c<br>
            >> @@ -1385,7 +1385,8 @@ opt_if_cf_list(nir_builder
            *b, struct exec_list *cf_list,<br>
            >>    * not do anything to cause the metadata to
            become invalid.<br>
            >>    */<br>
            >>   static bool<br>
            >> -opt_if_safe_cf_list(nir_builder *b, struct
            exec_list *cf_list)<br>
            >> +opt_if_safe_cf_list(nir_builder *b, struct
            exec_list *cf_list,<br>
            >> +                    bool skip_alu_of_phi)<br>
            >>   {<br>
            >>      bool progress = false;<br>
            >>      foreach_list_typed(nir_cf_node, cf_node, node,
            cf_list) {<br>
            >> @@ -1395,16 +1396,17 @@
            opt_if_safe_cf_list(nir_builder *b, struct exec_list
            *cf_list)<br>
            >><br>
            >>         case nir_cf_node_if: {<br>
            >>            nir_if *nif =
            nir_cf_node_as_if(cf_node);<br>
            >> -         progress |= opt_if_safe_cf_list(b,
            &nif->then_list);<br>
            >> -         progress |= opt_if_safe_cf_list(b,
            &nif->else_list);<br>
            >> +         progress |= opt_if_safe_cf_list(b,
            &nif->then_list, skip_alu_of_phi);<br>
            >> +         progress |= opt_if_safe_cf_list(b,
            &nif->else_list, skip_alu_of_phi);<br>
            >>            progress |=
            opt_if_evaluate_condition_use(b, nif);<br>
            >>            break;<br>
            >>         }<br>
            >><br>
            >>         case nir_cf_node_loop: {<br>
            >>            nir_loop *loop =
            nir_cf_node_as_loop(cf_node);<br>
            >> -         progress |= opt_if_safe_cf_list(b,
            &loop->body);<br>
            >> -         progress |= opt_split_alu_of_phi(b,
            loop);<br>
            >> +         progress |= opt_if_safe_cf_list(b,
            &loop->body, skip_alu_of_phi);<br>
            >> +         if (!skip_alu_of_phi)<br>
            >> +            progress |= opt_split_alu_of_phi(b,
            loop);<br>
            >>            break;<br>
            >>         }<br>
            >><br>
            >> @@ -1417,7 +1419,8 @@
            opt_if_safe_cf_list(nir_builder *b, struct exec_list
            *cf_list)<br>
            >>   }<br>
            >><br>
            >>   bool<br>
            >> -nir_opt_if(nir_shader *shader, bool
            aggressive_last_continue)<br>
            >> +nir_opt_if(nir_shader *shader, bool
            aggressive_last_continue,<br>
            >> +           bool skip_alu_of_phi)<br>
            >>   {<br>
            >>      bool progress = false;<br>
            >><br>
            >> @@ -1430,7 +1433,7 @@ nir_opt_if(nir_shader
            *shader, bool aggressive_last_continue)<br>
            >><br>
            >>         nir_metadata_require(function->impl,
            nir_metadata_block_index |<br>
            >>                             
            nir_metadata_dominance);<br>
            >> -      progress = opt_if_safe_cf_list(&b,
            &function->impl->body);<br>
            >> +      progress = opt_if_safe_cf_list(&b,
            &function->impl->body, skip_alu_of_phi);<br>
            >>         nir_metadata_preserve(function->impl,
            nir_metadata_block_index |<br>
            >>                             
             nir_metadata_dominance);<br>
            >><br>
            >> diff --git a/src/freedreno/ir3/ir3_nir.c
            b/src/freedreno/ir3/ir3_nir.c<br>
            >> index 76230e3be50..1bec3c030a9 100644<br>
            >> --- a/src/freedreno/ir3/ir3_nir.c<br>
            >> +++ b/src/freedreno/ir3/ir3_nir.c<br>
            >> @@ -147,7 +147,7 @@ ir3_optimize_loop(nir_shader
            *s)<br>
            >>                          OPT(s, nir_copy_prop);<br>
            >>                          OPT(s, nir_opt_dce);<br>
            >>                  }<br>
            >> -               progress |= OPT(s, nir_opt_if,
            false);<br>
            >> +               progress |= OPT(s, nir_opt_if,
            false, false);<br>
            >>                  progress |= OPT(s,
            nir_opt_remove_phis);<br>
            >>                  progress |= OPT(s, nir_opt_undef);<br>
            >><br>
            >> diff --git
            a/src/gallium/auxiliary/nir/tgsi_to_nir.c
            b/src/gallium/auxiliary/nir/tgsi_to_nir.c<br>
            >> index c55e8b84a41..6b40bff1f73 100644<br>
            >> --- a/src/gallium/auxiliary/nir/tgsi_to_nir.c<br>
            >> +++ b/src/gallium/auxiliary/nir/tgsi_to_nir.c<br>
            >> @@ -2066,7 +2066,7 @@ ttn_optimize_nir(nir_shader
            *nir, bool scalar)<br>
            >>            NIR_PASS(progress, nir, nir_opt_dce);<br>
            >>         }<br>
            >><br>
            >> -      NIR_PASS(progress, nir, nir_opt_if, false);<br>
            >> +      NIR_PASS(progress, nir, nir_opt_if, false,
            false);<br>
            >>         NIR_PASS(progress, nir, nir_opt_dead_cf);<br>
            >>         NIR_PASS(progress, nir, nir_opt_cse);<br>
            >>         NIR_PASS(progress, nir,
            nir_opt_peephole_select, 8, true, true);<br>
            >> diff --git
            a/src/gallium/drivers/freedreno/a2xx/ir2_nir.c
            b/src/gallium/drivers/freedreno/a2xx/ir2_nir.c<br>
            >> index ee348ca6a93..3522971e435 100644<br>
            >> --- a/src/gallium/drivers/freedreno/a2xx/ir2_nir.c<br>
            >> +++ b/src/gallium/drivers/freedreno/a2xx/ir2_nir.c<br>
            >> @@ -94,7 +94,7 @@ ir2_optimize_loop(nir_shader *s)<br>
            >>                          OPT(s, nir_opt_dce);<br>
            >>                  }<br>
            >>                  progress |= OPT(s,
            nir_opt_loop_unroll, nir_var_all);<br>
            >> -               progress |= OPT(s, nir_opt_if,
            false);<br>
            >> +               progress |= OPT(s, nir_opt_if,
            false, false);<br>
            >>                  progress |= OPT(s,
            nir_opt_remove_phis);<br>
            >>                  progress |= OPT(s, nir_opt_undef);<br>
            >><br>
            >> diff --git
            a/src/gallium/drivers/radeonsi/si_shader_nir.c
            b/src/gallium/drivers/radeonsi/si_shader_nir.c<br>
            >> index 5a925f19e09..7f1fe4ba2e9 100644<br>
            >> --- a/src/gallium/drivers/radeonsi/si_shader_nir.c<br>
            >> +++ b/src/gallium/drivers/radeonsi/si_shader_nir.c<br>
            >> @@ -879,7 +879,7 @@ si_lower_nir(struct
            si_shader_selector* sel)<br>
            >>                          NIR_PASS(progress,
            sel->nir, nir_copy_prop);<br>
            >>                          NIR_PASS(progress,
            sel->nir, nir_opt_dce);<br>
            >>                  }<br>
            >> -               NIR_PASS(progress, sel->nir,
            nir_opt_if, true);<br>
            >> +               NIR_PASS(progress, sel->nir,
            nir_opt_if, true, false);<br>
            >>                  NIR_PASS(progress, sel->nir,
            nir_opt_dead_cf);<br>
            >>                  NIR_PASS(progress, sel->nir,
            nir_opt_cse);<br>
            >>                  NIR_PASS(progress, sel->nir,
            nir_opt_peephole_select, 8, true, true);<br>
            >> diff --git a/src/intel/compiler/brw_nir.c
            b/src/intel/compiler/brw_nir.c<br>
            >> index e0a393fc298..ba911049ce3 100644<br>
            >> --- a/src/intel/compiler/brw_nir.c<br>
            >> +++ b/src/intel/compiler/brw_nir.c<br>
            >> @@ -607,7 +607,7 @@ brw_nir_optimize(nir_shader
            *nir, const struct brw_compiler *compiler,<br>
            >>            OPT(nir_copy_prop);<br>
            >>            OPT(nir_opt_dce);<br>
            >>         }<br>
            >> -      OPT(nir_opt_if, false);<br>
            >> +      OPT(nir_opt_if, false, false);<br>
            >>         if
            (nir->options->max_unroll_iterations != 0) {<br>
            >>            OPT(nir_opt_loop_unroll, indirect_mask);<br>
            >>         }<br>
            >> diff --git
            a/src/mesa/state_tracker/st_glsl_to_nir.cpp
            b/src/mesa/state_tracker/st_glsl_to_nir.cpp<br>
            >> index 97b2831b880..3f1f78e875b 100644<br>
            >> --- a/src/mesa/state_tracker/st_glsl_to_nir.cpp<br>
            >> +++ b/src/mesa/state_tracker/st_glsl_to_nir.cpp<br>
            >> @@ -324,7 +324,7 @@ st_nir_opts(nir_shader *nir,
            bool scalar)<br>
            >>            NIR_PASS(progress, nir, nir_copy_prop);<br>
            >>            NIR_PASS(progress, nir, nir_opt_dce);<br>
            >>         }<br>
            >> -      NIR_PASS(progress, nir, nir_opt_if, false);<br>
            >> +      NIR_PASS(progress, nir, nir_opt_if, false,
            false);<br>
            >>         NIR_PASS(progress, nir, nir_opt_dead_cf);<br>
            >>         NIR_PASS(progress, nir, nir_opt_cse);<br>
            >>         NIR_PASS(progress, nir,
            nir_opt_peephole_select, 8, true, true);<br>
            >> --<br>
            >> 2.21.0<br>
            >><br>
            >> _______________________________________________<br>
            >> mesa-dev mailing list<br>
            >> <a href="mailto:mesa-dev@lists.freedesktop.org"
              target="_blank" moz-do-not-send="true">mesa-dev@lists.freedesktop.org</a><br>
            >> <a
              href="https://lists.freedesktop.org/mailman/listinfo/mesa-dev"
              rel="noreferrer" target="_blank" moz-do-not-send="true">https://lists.freedesktop.org/mailman/listinfo/mesa-dev</a><br>
            _______________________________________________<br>
            mesa-dev mailing list<br>
            <a href="mailto:mesa-dev@lists.freedesktop.org"
              target="_blank" moz-do-not-send="true">mesa-dev@lists.freedesktop.org</a><br>
            <a
              href="https://lists.freedesktop.org/mailman/listinfo/mesa-dev"
              rel="noreferrer" target="_blank" moz-do-not-send="true">https://lists.freedesktop.org/mailman/listinfo/mesa-dev</a></blockquote>
        </div>
      </div>
    </blockquote>
  </body>
</html>