ir3: Use the new NIR lowering pass for integer multiplication

Shader-db stats courtesy of Eric Anholt:

total instructions in shared programs: 6480215 -> 6475457 (-0.07%)
instructions in affected programs: 662105 -> 657347 (-0.72%)
helped: 1209
HURT: 13
total constlen in shared programs: 1432704 -> 1427769 (-0.34%)
constlen in affected programs: 100063 -> 95128 (-4.93%)
helped: 512
HURT: 0
total max_sun in shared programs: 875561 -> 873387 (-0.25%)
max_sun in affected programs: 46179 -> 44005 (-4.71%)
helped: 1087
HURT: 0

Reviewed-by: Eric Anholt <eric@anholt.net>

diff --git a/src/freedreno/ir3/ir3_compiler_nir.c b/src/freedreno/ir3/ir3_compiler_nir.c
index 1a75181a96723077605e7112fb415301b2098129..e453f33fb91ffe968cc96345b13fc9d9b65904a0 100644 (file)
--- a/src/freedreno/ir3/ir3_compiler_nir.c
+++ b/src/freedreno/ir3/ir3_compiler_nir.c
@@ -542,23 +542,6 @@ emit_alu(struct ir3_context *ctx, nir_alu_instr *alu)
        case nir_op_umin:
                dst[0] = ir3_MIN_U(b, src[0], 0, src[1], 0);
                break;
-       case nir_op_imul:
-               if (bs[0] > 16 || bs[1] > 16) {
-                       /*
-                        * dst = (al * bl) + (ah * bl << 16) + (al * bh << 16)
-                        *   mull.u tmp0, a, b           ; mul low, i.e. al * bl
-                        *   madsh.m16 tmp1, a, b, tmp0  ; mul-add shift high mix,
-                        *                               ; i.e. ah * bl << 16
-                        *   madsh.m16 dst, b, a, tmp1   ; i.e. al * bh << 16
-                        */
-                       dst[0] = ir3_MADSH_M16(b, src[1], 0, src[0], 0,
-                                                                  ir3_MADSH_M16(b, src[0], 0, src[1], 0,
-                                                                                                ir3_MULL_U(b, src[0], 0,
-                                                                                                                       src[1], 0), 0), 0);
-               } else {
-                       dst[0] = ir3_MUL_S(b, src[0], 0, src[1], 0);
-               }
-               break;
        case nir_op_umul_low:
                dst[0] = ir3_MULL_U(b, src[0], 0, src[1], 0);
                break;

diff --git a/src/freedreno/ir3/ir3_context.c b/src/freedreno/ir3/ir3_context.c
index 99997427ec30db90399962919397ad6d8626bbb5..e572f33ef24f11869115f77f2ff9573726237bc4 100644 (file)
--- a/src/freedreno/ir3/ir3_context.c
+++ b/src/freedreno/ir3/ir3_context.c
@@ -84,6 +84,22 @@ ir3_context_init(struct ir3_compiler *compiler,
         */
        NIR_PASS_V(ctx->s, nir_lower_bool_to_int32);
        NIR_PASS_V(ctx->s, nir_lower_locals_to_regs);
+
+       /* We want to lower nir_op_imul as late as possible, to catch also
+        * those generated by earlier passes (e.g, nir_lower_locals_to_regs).
+        * However, we want a final swing of a few passes to have a chance
+        * at optimizing the result.
+        */
+       bool progress;
+       NIR_PASS(progress, ctx->s, ir3_nir_lower_imul);
+       if (progress) {
+               NIR_PASS_V(ctx->s, nir_opt_algebraic);
+               NIR_PASS_V(ctx->s, nir_opt_copy_prop_vars);
+               NIR_PASS_V(ctx->s, nir_opt_dead_write_vars);
+               NIR_PASS_V(ctx->s, nir_opt_dce);
+               NIR_PASS_V(ctx->s, nir_opt_constant_folding);
+       }
+
        NIR_PASS_V(ctx->s, nir_convert_from_ssa, true);
 
        if (ir3_shader_debug & IR3_DBG_DISASM) {