diff options
Diffstat (limited to 'gcc/config/i386/i386.cc')
| -rw-r--r-- | gcc/config/i386/i386.cc | 191 |
1 files changed, 153 insertions, 38 deletions
diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc index 28603c2..78df3d9 100644 --- a/gcc/config/i386/i386.cc +++ b/gcc/config/i386/i386.cc @@ -25257,32 +25257,6 @@ ix86_vectorize_create_costs (vec_info *vinfo, bool costing_for_scalar) return new ix86_vector_costs (vinfo, costing_for_scalar); } -/* Return cost of statement doing FP conversion. */ - -static unsigned -fp_conversion_stmt_cost (machine_mode mode, gimple *stmt, bool scalar_p) -{ - int outer_size - = tree_to_uhwi - (TYPE_SIZE - (TREE_TYPE (gimple_assign_lhs (stmt)))); - int inner_size - = tree_to_uhwi - (TYPE_SIZE - (TREE_TYPE (gimple_assign_rhs1 (stmt)))); - int stmt_cost = vec_fp_conversion_cost - (ix86_tune_cost, GET_MODE_BITSIZE (mode)); - /* VEC_PACK_TRUNC_EXPR: If inner size is greater than outer size we will end - up doing two conversions and packing them. */ - if (!scalar_p && inner_size > outer_size) - { - int n = inner_size / outer_size; - stmt_cost = stmt_cost * n - + (n - 1) * ix86_vec_cost (mode, ix86_cost->sse_op); - } - return stmt_cost; -} - unsigned ix86_vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind, stmt_vec_info stmt_info, slp_tree node, @@ -25326,7 +25300,7 @@ ix86_vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind, else if (X87_FLOAT_MODE_P (mode)) stmt_cost = ix86_cost->fadd; else - stmt_cost = ix86_cost->add; + stmt_cost = ix86_cost->add; } else stmt_cost = ix86_vec_cost (mode, fp ? ix86_cost->addss @@ -25381,7 +25355,7 @@ ix86_vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind, (subcode == RSHIFT_EXPR && !TYPE_UNSIGNED (TREE_TYPE (op1))) ? ASHIFTRT : LSHIFTRT, mode, - TREE_CODE (op2) == INTEGER_CST, + TREE_CODE (op2) == INTEGER_CST, cst_and_fits_in_hwi (op2) ? int_cst_value (op2) : -1, false, false, NULL, NULL); @@ -25390,30 +25364,152 @@ ix86_vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind, case NOP_EXPR: /* Only sign-conversions are free. */ if (tree_nop_conversion_p - (TREE_TYPE (gimple_assign_lhs (stmt_info->stmt)), + (TREE_TYPE (gimple_assign_lhs (stmt_info->stmt)), TREE_TYPE (gimple_assign_rhs1 (stmt_info->stmt)))) stmt_cost = 0; else if (fp) - stmt_cost = fp_conversion_stmt_cost (mode, stmt_info->stmt, - scalar_p); + stmt_cost = vec_fp_conversion_cost + (ix86_tune_cost, GET_MODE_BITSIZE (mode)); + break; + + case COND_EXPR: + { + /* SSE2 conditinal move sequence is: + pcmpgtd %xmm5, %xmm0 (accounted separately) + pand %xmm0, %xmm2 + pandn %xmm1, %xmm0 + por %xmm2, %xmm0 + while SSE4 uses cmp + blend + and AVX512 masked moves. + + The condition is accounted separately since we usually have + p = a < b + c = p ? x : y + and we will account first statement as setcc. Exception is when + p is loaded from memory as bool and then we will not acocunt + the compare, but there is no way to check for this. */ + + int ninsns = TARGET_SSE4_1 ? 1 : 3; + + /* If one of parameters is 0 or -1 the sequence will be simplified: + (if_true & mask) | (if_false & ~mask) -> if_true & mask */ + if (ninsns > 1 + && (zerop (gimple_assign_rhs2 (stmt_info->stmt)) + || zerop (gimple_assign_rhs3 (stmt_info->stmt)) + || integer_minus_onep + (gimple_assign_rhs2 (stmt_info->stmt)) + || integer_minus_onep + (gimple_assign_rhs3 (stmt_info->stmt)))) + ninsns = 1; + + if (SSE_FLOAT_MODE_SSEMATH_OR_HFBF_P (mode)) + stmt_cost = ninsns * ix86_cost->sse_op; + else if (X87_FLOAT_MODE_P (mode)) + /* x87 requires conditional branch. We don't have cost for + that. */ + ; + else if (VECTOR_MODE_P (mode)) + stmt_cost = ix86_vec_cost (mode, ninsns * ix86_cost->sse_op); + else + /* compare (accounted separately) + cmov. */ + stmt_cost = ix86_cost->add; + } break; - case BIT_IOR_EXPR: - case ABS_EXPR: - case ABSU_EXPR: case MIN_EXPR: case MAX_EXPR: + if (fp) + { + if (X87_FLOAT_MODE_P (mode)) + /* x87 requires conditional branch. We don't have cost for + that. */ + ; + else + /* minss */ + stmt_cost = ix86_vec_cost (mode, ix86_cost->sse_op); + } + else + { + if (VECTOR_MODE_P (mode)) + { + stmt_cost = ix86_vec_cost (mode, ix86_cost->sse_op); + /* vpmin was introduced in SSE3. + SSE2 needs pcmpgtd + pand + pandn + pxor. + If one of parameters is 0 or -1 the sequence is simplified + to pcmpgtd + pand. */ + if (!TARGET_SSSE3) + { + if (zerop (gimple_assign_rhs2 (stmt_info->stmt)) + || integer_minus_onep + (gimple_assign_rhs2 (stmt_info->stmt))) + stmt_cost *= 2; + else + stmt_cost *= 4; + } + } + else + /* cmp + cmov. */ + stmt_cost = ix86_cost->add * 2; + } + break; + + case ABS_EXPR: + case ABSU_EXPR: + if (fp) + { + if (X87_FLOAT_MODE_P (mode)) + /* fabs. */ + stmt_cost = ix86_cost->fabs; + else + /* andss of sign bit. */ + stmt_cost = ix86_vec_cost (mode, ix86_cost->sse_op); + } + else + { + if (VECTOR_MODE_P (mode)) + { + stmt_cost = ix86_vec_cost (mode, ix86_cost->sse_op); + /* vabs was introduced in SSE3. + SSE3 uses psrat + pxor + psub. */ + if (!TARGET_SSSE3) + stmt_cost *= 3; + } + else + /* neg + cmov. */ + stmt_cost = ix86_cost->add * 2; + } + break; + + case BIT_IOR_EXPR: case BIT_XOR_EXPR: case BIT_AND_EXPR: case BIT_NOT_EXPR: - if (SSE_FLOAT_MODE_SSEMATH_OR_HFBF_P (mode)) - stmt_cost = ix86_cost->sse_op; - else if (VECTOR_MODE_P (mode)) + gcc_assert (!SSE_FLOAT_MODE_SSEMATH_OR_HFBF_P (mode) + && !X87_FLOAT_MODE_P (mode)); + if (VECTOR_MODE_P (mode)) stmt_cost = ix86_vec_cost (mode, ix86_cost->sse_op); else stmt_cost = ix86_cost->add; break; + default: + if (truth_value_p (subcode)) + { + if (SSE_FLOAT_MODE_SSEMATH_OR_HFBF_P (mode)) + /* CMPccS? insructions are cheap, so use sse_op. While they + produce a mask which may need to be turned to 0/1 by and, + expect that this will be optimized away in a common case. */ + stmt_cost = ix86_cost->sse_op; + else if (X87_FLOAT_MODE_P (mode)) + /* fcmp + setcc. */ + stmt_cost = ix86_cost->fadd + ix86_cost->add; + else if (VECTOR_MODE_P (mode)) + stmt_cost = ix86_vec_cost (mode, ix86_cost->sse_op); + else + /* setcc. */ + stmt_cost = ix86_cost->add; + break; + } break; } } @@ -25439,7 +25535,26 @@ ix86_vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind, if (kind == vec_promote_demote && fp && FLOAT_TYPE_P (TREE_TYPE (gimple_assign_rhs1 (stmt_info->stmt)))) - stmt_cost = fp_conversion_stmt_cost (mode, stmt_info->stmt, scalar_p); + { + int outer_size + = tree_to_uhwi + (TYPE_SIZE + (TREE_TYPE (gimple_assign_lhs (stmt_info->stmt)))); + int inner_size + = tree_to_uhwi + (TYPE_SIZE + (TREE_TYPE (gimple_assign_rhs1 (stmt_info->stmt)))); + int stmt_cost = vec_fp_conversion_cost + (ix86_tune_cost, GET_MODE_BITSIZE (mode)); + /* VEC_PACK_TRUNC_EXPR: If inner size is greater than outer size we will end + up doing two conversions and packing them. */ + if (inner_size > outer_size) + { + int n = inner_size / outer_size; + stmt_cost = stmt_cost * n + + (n - 1) * ix86_vec_cost (mode, ix86_cost->sse_op); + } + } /* If we do elementwise loads into a vector then we are bound by latency and execution resources for the many scalar loads |