aco: use scratch_* for VGPR spill/reload on GFX9+

fossil-db (navi21):
Totals from 12 (0.01% of 162293) affected shaders:
Instrs: 122808 -> 122782 (-0.02%); split: -0.11%, +0.09%
CodeSize: 711248 -> 710788 (-0.06%); split: -0.16%, +0.10%
SpillSGPRs: 928 -> 831 (-10.45%)
SpillVGPRs: 1626 -> 1624 (-0.12%)
Latency: 4960285 -> 4932547 (-0.56%)
InvThroughput: 2574083 -> 2559953 (-0.55%)
VClause: 3404 -> 3402 (-0.06%)
Copies: 36992 -> 37181 (+0.51%); split: -0.05%, +0.56%
Branches: 3582 -> 3585 (+0.08%)
PreVGPRs: 3055 -> 3057 (+0.07%)

fossil-db (vega10):
Totals from 12 (0.01% of 161355) affected shaders:
Instrs: 124817 -> 124383 (-0.35%); split: -0.46%, +0.12%
CodeSize: 705116 -> 703664 (-0.21%); split: -0.44%, +0.23%
SpillSGPRs: 1012 -> 898 (-11.26%)
SpillVGPRs: 1632 -> 1624 (-0.49%)
Scratch: 201728 -> 200704 (-0.51%)
Latency: 6160115 -> 6266025 (+1.72%); split: -0.34%, +2.06%
InvThroughput: 6440203 -> 6544595 (+1.62%); split: -0.35%, +1.97%
VClause: 3409 -> 3423 (+0.41%)
Copies: 37929 -> 37748 (-0.48%); split: -1.16%, +0.69%
Branches: 3851 -> 3855 (+0.10%); split: -0.13%, +0.23%
PreVGPRs: 3053 -> 3055 (+0.07%)

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/17079>

1 files changed, 56 insertions, 20 deletions

diff --git a/src/amd/compiler/aco_spill.cpp b/src/amd/compiler/aco_spill.cpp
index 848df3dfe95..4b1099e8b28 100644
--- a/src/amd/compiler/aco_spill.cpp
+++ b/src/amd/compiler/aco_spill.cpp
@@ -1408,6 +1408,10 @@ load_scratch_resource(spill_ctx& ctx, Temp& scratch_offset, Block& block,
       }
    }
 
+   /* GFX9+ uses scratch_* instructions, which don't use a resource. Return a SADDR instead. */
+   if (ctx.program->gfx_level >= GFX9)
+      return bld.copy(bld.def(s1), Operand::c32(offset));
+
    Temp private_segment_buffer = ctx.program->private_segment_buffer;
    if (ctx.program->stage.hw != HWStage::CS)
       private_segment_buffer =
@@ -1445,17 +1449,29 @@ setup_vgpr_spill_reload(spill_ctx& ctx, Block& block,
    Temp scratch_offset = ctx.program->scratch_offset;
 
    *offset = spill_slot * 4;
-
-   bool add_offset_to_sgpr = ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size +
-                                ctx.vgpr_spill_slots * 4 >
-                             4096;
-   if (!add_offset_to_sgpr)
-      *offset += ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size;
-
-   if (ctx.scratch_rsrc == Temp()) {
-      unsigned rsrc_offset = add_offset_to_sgpr ? ctx.program->config->scratch_bytes_per_wave : 0;
-      ctx.scratch_rsrc =
-         load_scratch_resource(ctx, scratch_offset, block, instructions, rsrc_offset);
+   if (ctx.program->gfx_level >= GFX9) {
+      *offset += ctx.program->dev.scratch_global_offset_min;
+
+      if (ctx.scratch_rsrc == Temp()) {
+         int32_t saddr = ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size -
+                         ctx.program->dev.scratch_global_offset_min;
+         ctx.scratch_rsrc =
+            load_scratch_resource(ctx, scratch_offset, block, instructions, saddr);
+      }
+   } else {
+      bool add_offset_to_sgpr =
+         ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size +
+            ctx.vgpr_spill_slots * 4 >
+         4096;
+      if (!add_offset_to_sgpr)
+         *offset += ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size;
+
+      if (ctx.scratch_rsrc == Temp()) {
+         unsigned rsrc_offset =
+            add_offset_to_sgpr ? ctx.program->config->scratch_bytes_per_wave : 0;
+         ctx.scratch_rsrc =
+            load_scratch_resource(ctx, scratch_offset, block, instructions, rsrc_offset);
+      }
    }
 }
 
@@ -1485,11 +1501,19 @@ spill_vgpr(spill_ctx& ctx, Block& block, std::vector<aco_ptr<Instruction>>& inst
       bld.insert(split);
       for (unsigned i = 0; i < temp.size(); i++, offset += 4) {
          Temp elem = split->definitions[i].getTemp();
-         Instruction* instr =
-            bld.mubuf(aco_opcode::buffer_store_dword, ctx.scratch_rsrc, Operand(v1),
-                      ctx.program->scratch_offset, elem, offset, false, true);
-         instr->mubuf().sync = memory_sync_info(storage_vgpr_spill, semantic_private);
+         if (ctx.program->gfx_level >= GFX9) {
+            bld.scratch(aco_opcode::scratch_store_dword, Operand(v1), ctx.scratch_rsrc, elem,
+                        offset, memory_sync_info(storage_vgpr_spill, semantic_private));
+         } else {
+            Instruction* instr =
+               bld.mubuf(aco_opcode::buffer_store_dword, ctx.scratch_rsrc, Operand(v1),
+                         ctx.program->scratch_offset, elem, offset, false, true);
+            instr->mubuf().sync = memory_sync_info(storage_vgpr_spill, semantic_private);
+         }
       }
+   } else if (ctx.program->gfx_level >= GFX9) {
+      bld.scratch(aco_opcode::scratch_store_dword, Operand(v1), ctx.scratch_rsrc, temp, offset,
+                  memory_sync_info(storage_vgpr_spill, semantic_private));
    } else {
       Instruction* instr = bld.mubuf(aco_opcode::buffer_store_dword, ctx.scratch_rsrc, Operand(v1),
                                      ctx.program->scratch_offset, temp, offset, false, true);
@@ -1517,12 +1541,21 @@ reload_vgpr(spill_ctx& ctx, Block& block, std::vector<aco_ptr<Instruction>>& ins
       for (unsigned i = 0; i < def.size(); i++, offset += 4) {
          Temp tmp = bld.tmp(v1);
          vec->operands[i] = Operand(tmp);
-         Instruction* instr =
-            bld.mubuf(aco_opcode::buffer_load_dword, Definition(tmp), ctx.scratch_rsrc, Operand(v1),
-                      ctx.program->scratch_offset, offset, false, true);
-         instr->mubuf().sync = memory_sync_info(storage_vgpr_spill, semantic_private);
+         if (ctx.program->gfx_level >= GFX9) {
+            bld.scratch(aco_opcode::scratch_load_dword, Definition(tmp), Operand(v1),
+                        ctx.scratch_rsrc, offset,
+                        memory_sync_info(storage_vgpr_spill, semantic_private));
+         } else {
+            Instruction* instr =
+               bld.mubuf(aco_opcode::buffer_load_dword, Definition(tmp), ctx.scratch_rsrc,
+                         Operand(v1), ctx.program->scratch_offset, offset, false, true);
+            instr->mubuf().sync = memory_sync_info(storage_vgpr_spill, semantic_private);
+         }
       }
       bld.insert(vec);
+   } else if (ctx.program->gfx_level >= GFX9) {
+      bld.scratch(aco_opcode::scratch_load_dword, def, Operand(v1), ctx.scratch_rsrc, offset,
+                  memory_sync_info(storage_vgpr_spill, semantic_private));
    } else {
       Instruction* instr = bld.mubuf(aco_opcode::buffer_load_dword, def, ctx.scratch_rsrc,
                                      Operand(v1), ctx.program->scratch_offset, offset, false, true);
@@ -1907,7 +1940,10 @@ spill(Program* program, live& live_vars)
    }
    /* add extra SGPRs required for spilling VGPRs */
    if (demand.vgpr + extra_vgprs > vgpr_limit) {
-      extra_sgprs = 5; /* scratch_resource (s4) + scratch_offset (s1) */
+      if (program->gfx_level >= GFX9)
+         extra_sgprs = 1; /* SADDR */
+      else
+         extra_sgprs = 5; /* scratch_resource (s4) + scratch_offset (s1) */
       if (demand.sgpr + extra_sgprs > sgpr_limit) {
          /* re-calculate in case something has changed */
          unsigned sgpr_spills = demand.sgpr + extra_sgprs - sgpr_limit;