i965/fs: Unrestrict constant propagation into integer multiply.

Gen8+'s MUL instruction doesn't ignore the high 16-bits of one source
like on earlier platforms, so we can constant propagate into it without
worry. Integer multiplies (not into the accumulator, which is done for
imul_high) are lowered in lower_integer_multiplication(), so it's safe
there as well.

On Broadwell, fragment shaders only:
total instructions in shared programs: 4377769 -> 4377451 (-0.01%)
instructions in affected programs:     48064 -> 47746 (-0.66%)
helped:                                156

On Broadwell, vertex shaders only:
total instructions in shared programs: 2858885 -> 2856313 (-0.09%)
instructions in affected programs:     26380 -> 23808 (-9.75%)
helped:                                134

On Broadwell, vertex shaders only (with INTEL_USE_NIR=1):
total instructions in shared programs: 2911688 -> 2865984 (-1.57%)
instructions in affected programs:     1421715 -> 1376011 (-3.21%)
helped:                                6186

Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
This commit is contained in:
Matt Turner 2015-05-14 22:23:22 -07:00
parent 1e4e17fbd9
commit 81deefc45b
1 changed files with 9 additions and 1 deletions

View File

@ -541,8 +541,16 @@ fs_visitor::try_constant_propagate(fs_inst *inst, acp_entry *entry)
/* Fit this constant in by commuting the operands.
* Exception: we can't do this for 32-bit integer MUL/MACH
* because it's asymmetric.
*
* The BSpec says for Broadwell that
*
* "When multiplying DW x DW, the dst cannot be accumulator."
*
* Integer MUL with a non-accumulator destination will be lowered
* by lower_integer_multiplication(), so don't restrict it.
*/
if ((inst->opcode == BRW_OPCODE_MUL ||
if (((inst->opcode == BRW_OPCODE_MUL &&
inst->dst.is_accumulator()) ||
inst->opcode == BRW_OPCODE_MACH) &&
(inst->src[1].type == BRW_REGISTER_TYPE_D ||
inst->src[1].type == BRW_REGISTER_TYPE_UD))