aco: don't create sendmsg(dealloc_vgprs) if scratch is used

LLVM does something similar: https://reviews.llvm.org/D153295

fossil-db (gfx1100):
Totals from 21 (0.02% of 133461) affected shaders:
Instrs: 147428 -> 147396 (-0.02%)
CodeSize: 797188 -> 797060 (-0.02%)

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Fixes: 2930317cea53 ("aco/gfx11: deallocate VGPRs at the end of the shader")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24669>

diff --git a/src/amd/compiler/aco_ir.cpp b/src/amd/compiler/aco_ir.cpp
index ecf34ed..9bc20b7 100644 (file)
--- a/src/amd/compiler/aco_ir.cpp
+++ b/src/amd/compiler/aco_ir.cpp
@@ -1345,6 +1345,11 @@ dealloc_vgprs(Program* program)
    if (program->max_reg_demand.vgpr <= get_addr_vgpr_from_waves(program, max_waves))
       return false;
 
+   /* sendmsg(dealloc_vgprs) releases scratch, so this isn't safe if there is a in-progress scratch
+    * store. */
+   if (uses_scratch(program))
+      return false;
+
    Block& block = program->blocks.back();
 
    /* don't bother checking if there is a pending VMEM store or export: there almost always is */

diff --git a/src/amd/compiler/aco_ir.h b/src/amd/compiler/aco_ir.h
index b1f0280..716d019 100644 (file)
--- a/src/amd/compiler/aco_ir.h
+++ b/src/amd/compiler/aco_ir.h
@@ -2334,6 +2334,8 @@ uint16_t get_vgpr_alloc(Program* program, uint16_t addressable_vgprs);
 uint16_t get_addr_sgpr_from_waves(Program* program, uint16_t max_waves);
 uint16_t get_addr_vgpr_from_waves(Program* program, uint16_t max_waves);
 
+bool uses_scratch(Program* program);
+
 typedef struct {
    const int16_t opcode_gfx7[static_cast<int>(aco_opcode::num_opcodes)];
    const int16_t opcode_gfx9[static_cast<int>(aco_opcode::num_opcodes)];

diff --git a/src/amd/compiler/aco_live_var_analysis.cpp b/src/amd/compiler/aco_live_var_analysis.cpp
index 989e26b..ee1455f 100644 (file)
--- a/src/amd/compiler/aco_live_var_analysis.cpp
+++ b/src/amd/compiler/aco_live_var_analysis.cpp
@@ -325,15 +325,18 @@ calc_waves_per_workgroup(Program* program)
 }
 } /* end namespace */
 
+bool
+uses_scratch(Program* program)
+{
+   /* RT uses scratch but we don't yet know how much. */
+   return program->config->scratch_bytes_per_wave || program->stage == raytracing_cs;
+}
+
 uint16_t
 get_extra_sgprs(Program* program)
 {
-   /* We don't use this register on GFX6-8 and it's removed on GFX10+. RT uses scratch but we don't
-    * yet know how much.
-    */
-   bool needs_flat_scr =
-      (program->config->scratch_bytes_per_wave || program->stage == raytracing_cs) &&
-      program->gfx_level == GFX9;
+   /* We don't use this register on GFX6-8 and it's removed on GFX10+. */
+   bool needs_flat_scr = uses_scratch(program) && program->gfx_level == GFX9;
 
    if (program->gfx_level >= GFX10) {
       assert(!program->dev.xnack_enabled);