glsl: Add new expressions for INTEL_shader_integer_functions2

v2: Re-write iadd64_saturate and isub64_saturate to avoid undefined
overflow behavior.  Also fix copy-and-paste bug in isub64_saturate.
Suggested by Caio.

v3: Avoid signed integer overflow for abs_sub(0, INT_MIN).  Noticed by
Caio.

v4: Alternate fix for signed integer overflow for abs_sub(0, INT_MIN).
I tried the previous methon in a small test program with -ftrapv, and it
failed.

Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/767>

src/compiler/glsl/ir.cpp

@@ -257,6 +257,7 @@ ir_expression::ir_expression(int op, ir_rvalue *op0)
    case ir_unop_dFdy_fine:
    case ir_unop_bitfield_reverse:
    case ir_unop_interpolate_at_centroid:
+   case ir_unop_clz:
    case ir_unop_saturate:
    case ir_unop_atan:
       this->type = op0->type;
@@ -500,6 +501,7 @@ ir_expression::ir_expression(int op, ir_rvalue *op0, ir_rvalue *op1)
       break;
 
    case ir_binop_imul_high:
+   case ir_binop_mul_32x16:
    case ir_binop_carry:
    case ir_binop_borrow:
    case ir_binop_lshift:
@@ -510,6 +512,44 @@ ir_expression::ir_expression(int op, ir_rvalue *op0, ir_rvalue *op1)
       this->type = op0->type;
       break;
 
+   case ir_binop_add_sat:
+   case ir_binop_sub_sat:
+   case ir_binop_avg:
+   case ir_binop_avg_round:
+      assert(op0->type == op1->type);
+      this->type = op0->type;
+      break;
+
+   case ir_binop_abs_sub: {
+      enum glsl_base_type base;
+
+      assert(op0->type == op1->type);
+
+      switch (op0->type->base_type) {
+      case GLSL_TYPE_UINT:
+      case GLSL_TYPE_INT:
+         base = GLSL_TYPE_UINT;
+         break;
+      case GLSL_TYPE_UINT8:
+      case GLSL_TYPE_INT8:
+         base = GLSL_TYPE_UINT8;
+         break;
+      case GLSL_TYPE_UINT16:
+      case GLSL_TYPE_INT16:
+         base = GLSL_TYPE_UINT16;
+         break;
+      case GLSL_TYPE_UINT64:
+      case GLSL_TYPE_INT64:
+         base = GLSL_TYPE_UINT64;
+         break;
+      default:
+         unreachable(!"Invalid base type.");
+      }
+
+      this->type = glsl_type::get_instance(base, op0->type->vector_elements, 1);
+      break;
+   }
+
    case ir_binop_vector_extract:
       this->type = op0->type->get_scalar_type();
       break;

src/compiler/glsl/ir_constant_expression.cpp

@@ -416,6 +416,42 @@ unpack_half_1x16(uint16_t u)
    return _mesa_half_to_float(u);
 }
 
+static int32_t
+iadd_saturate(int32_t a, int32_t b)
+{
+   return CLAMP(int64_t(a) + int64_t(b), INT32_MIN, INT32_MAX);
+}
+
+static int64_t
+iadd64_saturate(int64_t a, int64_t b)
+{
+   if (a < 0 && b < INT64_MIN - a)
+      return INT64_MIN;
+
+   if (a > 0 && b > INT64_MAX - a)
+      return INT64_MAX;
+
+   return a + b;
+}
+
+static int32_t
+isub_saturate(int32_t a, int32_t b)
+{
+   return CLAMP(int64_t(a) - int64_t(b), INT32_MIN, INT32_MAX);
+}
+
+static int64_t
+isub64_saturate(int64_t a, int64_t b)
+{
+   if (b > 0 && a < INT64_MIN + b)
+      return INT64_MIN;
+
+   if (b < 0 && a > INT64_MAX + b)
+      return INT64_MAX;
+
+   return a - b;
+}
+
 /**
  * Get the constant that is ultimately referenced by an r-value, in a constant
  * expression evaluation context.

src/compiler/glsl/ir_expression_operation.py

@@ -539,6 +539,7 @@ ir_expression_operation = [
    operation("bit_count", 1, source_types=(uint_type, int_type), dest_type=int_type, c_expression="util_bitcount({src0})"),
    operation("find_msb", 1, source_types=(uint_type, int_type), dest_type=int_type, c_expression={'u': "find_msb_uint({src0})", 'i': "find_msb_int({src0})"}),
    operation("find_lsb", 1, source_types=(uint_type, int_type), dest_type=int_type, c_expression="find_msb_uint({src0} & -{src0})"),
+   operation("clz", 1, source_types=(uint_type,), dest_type=uint_type, c_expression="(unsigned)(31 - find_msb_uint({src0}))"),
 
    operation("saturate", 1, printable_name="sat", source_types=(float_type,), c_expression="CLAMP({src0}, 0.0f, 1.0f)"),
 
@@ -584,8 +585,33 @@ ir_expression_operation = [
 
    operation("add", 2, printable_name="+", source_types=numeric_types, c_expression="{src0} + {src1}", flags=vector_scalar_operation),
    operation("sub", 2, printable_name="-", source_types=numeric_types, c_expression="{src0} - {src1}", flags=vector_scalar_operation),
+   operation("add_sat", 2, printable_name="add_sat", source_types=integer_types, c_expression={
+      'u': "({src0} + {src1}) < {src0} ? UINT32_MAX : ({src0} + {src1})",
+      'i': "iadd_saturate({src0}, {src1})",
+      'u64': "({src0} + {src1}) < {src0} ? UINT64_MAX : ({src0} + {src1})",
+      'i64': "iadd64_saturate({src0}, {src1})"
+   }),
+   operation("sub_sat", 2, printable_name="sub_sat", source_types=integer_types, c_expression={
+      'u': "({src1} > {src0}) ? 0 : {src0} - {src1}",
+      'i': "isub_saturate({src0}, {src1})",
+      'u64': "({src1} > {src0}) ? 0 : {src0} - {src1}",
+      'i64': "isub64_saturate({src0}, {src1})"
+   }),
+   operation("abs_sub", 2, printable_name="abs_sub", source_types=integer_types, c_expression={
+      'u': "({src1} > {src0}) ? {src1} - {src0} : {src0} - {src1}",
+      'i': "({src1} > {src0}) ? (unsigned){src1} - (unsigned){src0} : (unsigned){src0} - (unsigned){src1}",
+      'u64': "({src1} > {src0}) ? {src1} - {src0} : {src0} - {src1}",
+      'i64': "({src1} > {src0}) ? (uint64_t){src1} - (uint64_t){src0} : (uint64_t){src0} - (uint64_t){src1}",
+   }),
+   operation("avg", 2, printable_name="average", source_types=integer_types, c_expression="({src0} >> 1) + ({src1} >> 1) + (({src0} & {src1}) & 1)"),
+   operation("avg_round", 2, printable_name="average_rounded", source_types=integer_types, c_expression="({src0} >> 1) + ({src1} >> 1) + (({src0} | {src1}) & 1)"),
+
    # "Floating-point or low 32-bit integer multiply."
    operation("mul", 2, printable_name="*", source_types=numeric_types, c_expression="{src0} * {src1}"),
+   operation("mul_32x16", 2, printable_name="*", source_types=(uint_type, int_type), c_expression={
+      'u': "{src0} * (uint16_t){src1}",
+      'i': "{src0} * (int16_t){src0}"
+   }),
    operation("imul_high", 2),       # Calculates the high 32-bits of a 64-bit multiply.
    operation("div", 2, printable_name="/", source_types=numeric_types, c_expression={'u': "{src1} == 0 ? 0 : {src0} / {src1}", 'i': "{src1} == 0 ? 0 : {src0} / {src1}", 'u64': "{src1} == 0 ? 0 : {src0} / {src1}", 'i64': "{src1} == 0 ? 0 : {src0} / {src1}", 'default': "{src0} / {src1}"}, flags=vector_scalar_operation),
 

src/compiler/glsl/ir_validate.cpp

@@ -546,6 +546,11 @@ ir_validate::visit_leave(ir_expression *ir)
       assert(ir->type->base_type == GLSL_TYPE_INT);
       break;
 
+   case ir_unop_clz:
+      assert(ir->operands[0]->type == ir->type);
+      assert(ir->type->base_type == GLSL_TYPE_UINT);
+      break;
+
    case ir_unop_noise:
       /* XXX what can we assert here? */
       break;
@@ -649,6 +654,25 @@ ir_validate::visit_leave(ir_expression *ir)
       }
       break;
 
+   case ir_binop_abs_sub:
+      assert(ir->operands[0]->type == ir->operands[1]->type);
+      assert(ir->operands[0]->type->is_integer_32_64());
+      assert(ir->operands[0]->type->vector_elements ==
+             ir->type->vector_elements);
+      assert(ir->type->base_type == GLSL_TYPE_UINT ||
+             ir->type->base_type == GLSL_TYPE_UINT64);
+      break;
+
+   case ir_binop_add_sat:
+   case ir_binop_sub_sat:
+   case ir_binop_avg:
+   case ir_binop_avg_round:
+      assert(ir->type == ir->operands[0]->type);
+      assert(ir->type == ir->operands[1]->type);
+      assert(ir->type->is_integer_32_64());
+      break;
+
+   case ir_binop_mul_32x16:
    case ir_binop_imul_high:
       assert(ir->type == ir->operands[0]->type);
       assert(ir->type == ir->operands[1]->type);

src/mesa/program/ir_to_mesa.cpp

@@ -1296,6 +1296,12 @@ ir_to_mesa_visitor::visit(ir_expression *ir)
    case ir_binop_ldexp:
    case ir_binop_carry:
    case ir_binop_borrow:
+   case ir_binop_abs_sub:
+   case ir_binop_add_sat:
+   case ir_binop_sub_sat:
+   case ir_binop_avg:
+   case ir_binop_avg_round:
+   case ir_binop_mul_32x16:
    case ir_binop_imul_high:
    case ir_unop_interpolate_at_centroid:
    case ir_binop_interpolate_at_offset:
@@ -1340,6 +1346,7 @@ ir_to_mesa_visitor::visit(ir_expression *ir)
    case ir_unop_unpack_image_2x32:
    case ir_unop_atan:
    case ir_binop_atan2:
+   case ir_unop_clz:
       assert(!"not supported");
       break;