i965/fs: Skip emitting MACH/MOV for small integers.

The vector backend already implemented this optimization, but
surprisingly, we never bothered to implement it in the scalar backend.

In addition to saving two instructions, this eliminates a use of the
accumulator as an explicit source, which is unsupported in SIMD16 mode
on Gen7+, which could help us gain SIMD16 programs.

Cuts 19.23% of the instructions in dolphin/efb2ram.shader_test.

v2: Rebase on is_16bit_integer_constant -> is_uint16_constant rename.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
This commit is contained in:
Kenneth Graunke
2014-04-06 18:38:46 -07:00
parent 7540be22d1
commit 4311f9878d
+19 -10
View File
@@ -458,18 +458,27 @@ fs_visitor::visit(ir_expression *ir)
* of one of the operands (src0 on gen6, src1 on gen7). The
* MACH accumulates in the contribution of the upper 16 bits
* of that operand.
*
* FINISHME: Emit just the MUL if we know an operand is small
* enough.
*/
if (brw->gen >= 7)
no16("SIMD16 explicit accumulator operands unsupported\n");
*/
if (ir->operands[0]->is_uint16_constant()) {
if (brw->gen < 7)
emit(MUL(this->result, op[0], op[1]));
else
emit(MUL(this->result, op[1], op[0]));
} else if (ir->operands[1]->is_uint16_constant()) {
if (brw->gen < 7)
emit(MUL(this->result, op[1], op[0]));
else
emit(MUL(this->result, op[0], op[1]));
} else {
if (brw->gen >= 7)
no16("SIMD16 explicit accumulator operands unsupported\n");
struct brw_reg acc = retype(brw_acc_reg(), this->result.type);
struct brw_reg acc = retype(brw_acc_reg(), this->result.type);
emit(MUL(acc, op[0], op[1]));
emit(MACH(reg_null_d, op[0], op[1]));
emit(MOV(this->result, fs_reg(acc)));
emit(MUL(acc, op[0], op[1]));
emit(MACH(reg_null_d, op[0], op[1]));
emit(MOV(this->result, fs_reg(acc)));
}
} else {
emit(MUL(this->result, op[0], op[1]));
}