intel/fs: Constant fold OR and AND

The path taken in fs_visitor::swizzle_nir_scratch_addr for DG2 generates
some AND and OR instructions before the SHL. This commit folds those so
the whold calculation becomes a constant (like on older platforms).

v2: Fix return type of src_as_uint. Noticed by Marcin.

shader-db results:

DG2
total instructions in shared programs: 23190475 -> 23179540 (-0.05%)
instructions in affected programs: 36026 -> 25091 (-30.35%)
helped: 7 / HURT: 0

total cycles in shared programs: 841196807 -> 841142563 (<.01%)
cycles in affected programs: 1660670 -> 1606426 (-3.27%)
helped: 7 / HURT: 0

No shader-db changes on any older Intel platforms.

fossil-db results:

DG2
Totals:
Instrs: 197780372 -> 197773966 (-0.00%)
Cycles: 14066410782 -> 14066399378 (-0.00%); split: -0.00%, +0.00%
Subgroup size: 8438104 -> 8438112 (+0.00%)
Send messages: 8049445 -> 8049446 (+0.00%)
Scratch Memory Size: 14263296 -> 14264320 (+0.01%)

Totals from 9 (0.00% of 668055) affected shaders:
Instrs: 24547 -> 18141 (-26.10%)
Cycles: 1984791 -> 1973387 (-0.57%); split: -0.98%, +0.40%
Subgroup size: 88 -> 96 (+9.09%)
Send messages: 867 -> 868 (+0.12%)
Scratch Memory Size: 69632 -> 70656 (+1.47%)

No fossil-db changes on any older Intel platforms.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23884>

diff --git a/src/intel/compiler/brw_fs.cpp b/src/intel/compiler/brw_fs.cpp
index 3aaac3c..6e10036 100644 (file)
--- a/src/intel/compiler/brw_fs.cpp
+++ b/src/intel/compiler/brw_fs.cpp
@@ -2568,6 +2568,62 @@ fs_visitor::lower_constant_loads()
    invalidate_analysis(DEPENDENCY_INSTRUCTIONS);
 }
 
+static uint64_t
+src_as_uint(const fs_reg &src)
+{
+   assert(src.file == IMM);
+
+   switch (src.type) {
+   case BRW_REGISTER_TYPE_W:
+      return (uint64_t)(int16_t)(src.ud & 0xffff);
+
+   case BRW_REGISTER_TYPE_UW:
+      return (uint64_t)(uint16_t)(src.ud & 0xffff);
+
+   case BRW_REGISTER_TYPE_D:
+      return (uint64_t)src.d;
+
+   case BRW_REGISTER_TYPE_UD:
+      return (uint64_t)src.ud;
+
+   case BRW_REGISTER_TYPE_Q:
+      return src.d64;
+
+   case BRW_REGISTER_TYPE_UQ:
+      return src.u64;
+
+   default:
+      unreachable("Invalid integer type.");
+   }
+}
+
+static fs_reg
+brw_imm_for_type(uint64_t value, enum brw_reg_type type)
+{
+   switch (type) {
+   case BRW_REGISTER_TYPE_W:
+      return brw_imm_w(value);
+
+   case BRW_REGISTER_TYPE_UW:
+      return brw_imm_uw(value);
+
+   case BRW_REGISTER_TYPE_D:
+      return brw_imm_d(value);
+
+   case BRW_REGISTER_TYPE_UD:
+      return brw_imm_ud(value);
+
+   case BRW_REGISTER_TYPE_Q:
+      return brw_imm_d(value);
+
+   case BRW_REGISTER_TYPE_UQ:
+      return brw_imm_uq(value);
+
+   default:
+      unreachable("Invalid integer type.");
+   }
+}
+
 bool
 fs_visitor::opt_algebraic()
 {
@@ -2701,7 +2757,35 @@ fs_visitor::opt_algebraic()
             break;
          }
          break;
+
+      case BRW_OPCODE_AND:
+         if (inst->src[0].file == IMM && inst->src[1].file == IMM) {
+            const uint64_t src0 = src_as_uint(inst->src[0]);
+            const uint64_t src1 = src_as_uint(inst->src[1]);
+
+            inst->opcode = BRW_OPCODE_MOV;
+            inst->sources = 1;
+            inst->src[0] = brw_imm_for_type(src0 & src1, inst->dst.type);
+            inst->src[1] = reg_undef;
+            progress = true;
+            break;
+         }
+
+         break;
+
       case BRW_OPCODE_OR:
+         if (inst->src[0].file == IMM && inst->src[1].file == IMM) {
+            const uint64_t src0 = src_as_uint(inst->src[0]);
+            const uint64_t src1 = src_as_uint(inst->src[1]);
+
+            inst->opcode = BRW_OPCODE_MOV;
+            inst->sources = 1;
+            inst->src[0] = brw_imm_for_type(src0 | src1, inst->dst.type);
+            inst->src[1] = reg_undef;
+            progress = true;
+            break;
+         }
+
          if (inst->src[0].equals(inst->src[1]) ||
              inst->src[1].is_zero()) {
             /* On Gfx8+, the OR instruction can have a source modifier that

diff --git a/src/intel/compiler/brw_fs_copy_propagation.cpp b/src/intel/compiler/brw_fs_copy_propagation.cpp
index d16077c..ba0ebc5 100644 (file)
--- a/src/intel/compiler/brw_fs_copy_propagation.cpp
+++ b/src/intel/compiler/brw_fs_copy_propagation.cpp
@@ -935,8 +935,6 @@ fs_visitor::try_constant_propagate(fs_inst *inst, acp_entry *entry)
       case BRW_OPCODE_MUL:
       case SHADER_OPCODE_MULH:
       case BRW_OPCODE_ADD:
-      case BRW_OPCODE_OR:
-      case BRW_OPCODE_AND:
       case BRW_OPCODE_XOR:
       case BRW_OPCODE_ADDC:
          if (i == 1) {
@@ -1072,6 +1070,8 @@ fs_visitor::try_constant_propagate(fs_inst *inst, acp_entry *entry)
          }
          break;
 
+      case BRW_OPCODE_AND:
+      case BRW_OPCODE_OR:
       case SHADER_OPCODE_TEX_LOGICAL:
       case SHADER_OPCODE_TXD_LOGICAL:
       case SHADER_OPCODE_TXF_LOGICAL:
@@ -1120,6 +1120,17 @@ fs_visitor::try_constant_propagate(fs_inst *inst, acp_entry *entry)
       }
    }
 
+   /* If only one of the sources of a 2-source, commutative instruction (e.g.,
+    * AND) is immediate, it must be src1. If both are immediate, opt_algebraic
+    * should fold it away.
+    */
+   if (progress && inst->sources == 2 && inst->is_commutative() &&
+       inst->src[0].file == IMM && inst->src[1].file != IMM) {
+      const auto src1 = inst->src[1];
+      inst->src[1] = inst->src[0];
+      inst->src[0] = src1;
+   }
+
    return progress;
 }