aco: try harder to not create v_mul_lo_u32

fossil-db (Vega):
Totals from 4 (0.00% of 137413) affected shaders:
CodeSize: 13708 -> 13716 (+0.06%)
Instrs: 2742 -> 2744 (+0.07%)
Cycles: 24348 -> 24236 (-0.46%)

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5390>

src/amd/compiler/aco_builder_h.py

@@ -371,6 +371,10 @@ public:
    Result v_mul_imm(Definition dst, Temp tmp, uint32_t imm, bool bits24=false)
    {
       assert(tmp.type() == RegType::vgpr);
+      bool has_lshl_add = program->chip_class >= GFX9;
+      /* v_mul_lo_u32 has 1.6x the latency of most VALU on GFX10 (8 vs 5 cycles),
+       * compared to 4x the latency on <GFX10. */
+      unsigned mul_cost = program->chip_class >= GFX10 ? 1 : (4 + Operand(imm).isLiteral());
       if (imm == 0) {
          return copy(dst, Operand(0u));
       } else if (imm == 1) {
@@ -379,10 +383,40 @@ public:
          return vop2(aco_opcode::v_lshlrev_b32, dst, Operand((uint32_t)ffs(imm) - 1u), tmp);
       } else if (bits24) {
         return vop2(aco_opcode::v_mul_u32_u24, dst, Operand(imm), tmp);
-      } else {
+      } else if (util_is_power_of_two_nonzero(imm - 1u)) {
-        Temp imm_tmp = copy(def(s1), Operand(imm));
+         return vadd32(dst, vop2(aco_opcode::v_lshlrev_b32, def(v1), Operand((uint32_t)ffs(imm - 1u) - 1u), tmp), tmp);
-        return vop3(aco_opcode::v_mul_lo_u32, dst, imm_tmp, tmp);
+      } else if (mul_cost > 2 && util_is_power_of_two_nonzero(imm + 1u)) {
+         return vsub32(dst, vop2(aco_opcode::v_lshlrev_b32, def(v1), Operand((uint32_t)ffs(imm + 1u) - 1u), tmp), tmp);
       }
+
+      unsigned instrs_required = util_bitcount(imm);
+      if (!has_lshl_add) {
+         instrs_required = util_bitcount(imm) - (imm & 0x1); /* shifts */
+         instrs_required += util_bitcount(imm) - 1; /* additions */
+      }
+      if (instrs_required < mul_cost) {
+         Result res(NULL);
+         Temp cur;
+         while (imm) {
+            unsigned shift = u_bit_scan(&imm);
+            Definition tmp_dst = imm ? def(v1) : dst;
+
+            if (shift && cur.id())
+               res = vadd32(Definition(tmp_dst), vop2(aco_opcode::v_lshlrev_b32, def(v1), Operand(shift), tmp), cur);
+            else if (shift)
+               res = vop2(aco_opcode::v_lshlrev_b32, Definition(tmp_dst), Operand(shift), tmp);
+            else if (cur.id())
+               res = vadd32(Definition(tmp_dst), tmp, cur);
+            else
+               tmp_dst = Definition(tmp);
+
+            cur = tmp_dst.getTemp();
+         }
+         return res;
+      }
+
+      Temp imm_tmp = copy(def(s1), Operand(imm));
+      return vop3(aco_opcode::v_mul_lo_u32, dst, imm_tmp, tmp);
    }
 
    Result v_mul24_imm(Definition dst, Temp tmp, uint32_t imm)