diff options
Diffstat (limited to 'gcc/config/i386')
| -rw-r--r-- | gcc/config/i386/i386-expand.cc | 50 | ||||
| -rw-r--r-- | gcc/config/i386/i386-options.cc | 4 | ||||
| -rw-r--r-- | gcc/config/i386/i386.cc | 215 | ||||
| -rw-r--r-- | gcc/config/i386/i386.h | 6 | ||||
| -rw-r--r-- | gcc/config/i386/i386.md | 4 | ||||
| -rw-r--r-- | gcc/config/i386/predicates.md | 14 | ||||
| -rw-r--r-- | gcc/config/i386/sse.md | 94 | ||||
| -rw-r--r-- | gcc/config/i386/x86-tune-costs.h | 123 | ||||
| -rw-r--r-- | gcc/config/i386/x86-tune-sched.cc | 15 |
9 files changed, 468 insertions, 57 deletions
diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc index cdfd94d..a314800 100644 --- a/gcc/config/i386/i386-expand.cc +++ b/gcc/config/i386/i386-expand.cc @@ -4138,6 +4138,10 @@ ix86_expand_sse_fp_minmax (rtx dest, enum rtx_code code, rtx cmp_op0, return false; mode = GET_MODE (dest); + if (immediate_operand (if_false, mode)) + if_false = force_reg (mode, if_false); + if (immediate_operand (if_true, mode)) + if_true = force_reg (mode, if_true); /* We want to check HONOR_NANS and HONOR_SIGNED_ZEROS here, but MODE may be a vector mode and thus not appropriate. */ @@ -4687,6 +4691,8 @@ ix86_expand_fp_movcc (rtx operands[]) compare_op = ix86_expand_compare (NE, tmp, const0_rtx); } + operands[2] = force_reg (mode, operands[2]); + operands[3] = force_reg (mode, operands[3]); emit_insn (gen_rtx_SET (operands[0], gen_rtx_IF_THEN_ELSE (mode, compare_op, operands[2], operands[3]))); @@ -19256,8 +19262,6 @@ ix86_emit_swdivsf (rtx res, rtx a, rtx b, machine_mode mode) e1 = gen_reg_rtx (mode); x1 = gen_reg_rtx (mode); - /* a / b = a * ((rcp(b) + rcp(b)) - (b * rcp(b) * rcp (b))) */ - b = force_reg (mode, b); /* x0 = rcp(b) estimate */ @@ -19270,20 +19274,42 @@ ix86_emit_swdivsf (rtx res, rtx a, rtx b, machine_mode mode) emit_insn (gen_rtx_SET (x0, gen_rtx_UNSPEC (mode, gen_rtvec (1, b), UNSPEC_RCP))); - /* e0 = x0 * b */ - emit_insn (gen_rtx_SET (e0, gen_rtx_MULT (mode, x0, b))); + unsigned vector_size = GET_MODE_SIZE (mode); + + /* (a - (rcp(b) * a * b)) * rcp(b) + rcp(b) * a + N-R step with 2 fma implementation. */ + if (TARGET_FMA + || (TARGET_AVX512F && vector_size == 64) + || (TARGET_AVX512VL && (vector_size == 32 || vector_size == 16))) + { + /* e0 = x0 * a */ + emit_insn (gen_rtx_SET (e0, gen_rtx_MULT (mode, x0, a))); + /* e1 = e0 * b - a */ + emit_insn (gen_rtx_SET (e1, gen_rtx_FMA (mode, e0, b, + gen_rtx_NEG (mode, a)))); + /* res = - e1 * x0 + e0 */ + emit_insn (gen_rtx_SET (res, gen_rtx_FMA (mode, + gen_rtx_NEG (mode, e1), + x0, e0))); + } + else + /* a / b = a * ((rcp(b) + rcp(b)) - (b * rcp(b) * rcp (b))) */ + { + /* e0 = x0 * b */ + emit_insn (gen_rtx_SET (e0, gen_rtx_MULT (mode, x0, b))); - /* e0 = x0 * e0 */ - emit_insn (gen_rtx_SET (e0, gen_rtx_MULT (mode, x0, e0))); + /* e1 = x0 + x0 */ + emit_insn (gen_rtx_SET (e1, gen_rtx_PLUS (mode, x0, x0))); - /* e1 = x0 + x0 */ - emit_insn (gen_rtx_SET (e1, gen_rtx_PLUS (mode, x0, x0))); + /* e0 = x0 * e0 */ + emit_insn (gen_rtx_SET (e0, gen_rtx_MULT (mode, x0, e0))); - /* x1 = e1 - e0 */ - emit_insn (gen_rtx_SET (x1, gen_rtx_MINUS (mode, e1, e0))); + /* x1 = e1 - e0 */ + emit_insn (gen_rtx_SET (x1, gen_rtx_MINUS (mode, e1, e0))); - /* res = a * x1 */ - emit_insn (gen_rtx_SET (res, gen_rtx_MULT (mode, a, x1))); + /* res = a * x1 */ + emit_insn (gen_rtx_SET (res, gen_rtx_MULT (mode, a, x1))); + } } /* Output code to perform a Newton-Rhapson approximation of a diff --git a/gcc/config/i386/i386-options.cc b/gcc/config/i386/i386-options.cc index a9fac01..964449f 100644 --- a/gcc/config/i386/i386-options.cc +++ b/gcc/config/i386/i386-options.cc @@ -2828,8 +2828,8 @@ ix86_option_override_internal (bool main_args_p, if (flag_nop_mcount) error ("%<-mnop-mcount%> is not compatible with this target"); #endif - if (flag_nop_mcount && flag_pic) - error ("%<-mnop-mcount%> is not implemented for %<-fPIC%>"); + if (flag_nop_mcount && flag_pic && !flag_plt) + error ("%<-mnop-mcount%> is not implemented for %<-fno-plt%>"); /* Accept -msseregparm only if at least SSE support is enabled. */ if (TARGET_SSEREGPARM_P (opts->x_target_flags) diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc index 4f8380c4..78df3d9 100644 --- a/gcc/config/i386/i386.cc +++ b/gcc/config/i386/i386.cc @@ -100,6 +100,7 @@ along with GCC; see the file COPYING3. If not see #include "i386-features.h" #include "function-abi.h" #include "rtl-error.h" +#include "gimple-pretty-print.h" /* This file should be included last. */ #include "target-def.h" @@ -458,6 +459,9 @@ int ix86_arch_specified; indirect thunk pushes the return address onto stack, destroying red-zone. + NB: Don't use red-zone for functions with no_caller_saved_registers + and 32 GPRs since 128-byte red-zone is too small for 31 GPRs. + TODO: If we can reserve the first 2 WORDs, for PUSH and, another for CALL, in red-zone, we can allow local indirect jumps with indirect thunk. */ @@ -467,6 +471,9 @@ ix86_using_red_zone (void) { return (TARGET_RED_ZONE && !TARGET_64BIT_MS_ABI + && (!TARGET_APX_EGPR + || (cfun->machine->call_saved_registers + != TYPE_NO_CALLER_SAVED_REGISTERS)) && (!cfun->machine->has_local_indirect_jump || cfun->machine->indirect_branch_type == indirect_branch_keep)); } @@ -21810,6 +21817,25 @@ ix86_insn_cost (rtx_insn *insn, bool speed) return insn_cost + pattern_cost (PATTERN (insn), speed); } +/* Return cost of SSE/AVX FP->FP conversion (extensions and truncates). */ + +static int +vec_fp_conversion_cost (const struct processor_costs *cost, int size) +{ + if (size < 128) + return cost->cvtss2sd; + else if (size < 256) + { + if (TARGET_SSE_SPLIT_REGS) + return cost->cvtss2sd * size / 64; + return cost->cvtss2sd; + } + if (size < 512) + return cost->vcvtps2pd256; + else + return cost->vcvtps2pd512; +} + /* Compute a (partial) cost for rtx X. Return true if the complete cost has been computed, and false if subexpressions should be scanned. In either case, *TOTAL contains the cost result. */ @@ -22473,17 +22499,18 @@ ix86_rtx_costs (rtx x, machine_mode mode, int outer_code_i, int opno, return false; case FLOAT_EXTEND: + /* x87 represents all values extended to 80bit. */ if (!SSE_FLOAT_MODE_SSEMATH_OR_HFBF_P (mode)) *total = 0; else - *total = ix86_vec_cost (mode, cost->addss); + *total = vec_fp_conversion_cost (cost, GET_MODE_BITSIZE (mode)); return false; case FLOAT_TRUNCATE: if (!SSE_FLOAT_MODE_SSEMATH_OR_HFBF_P (mode)) *total = cost->fadd; else - *total = ix86_vec_cost (mode, cost->addss); + *total = vec_fp_conversion_cost (cost, GET_MODE_BITSIZE (mode)); return false; case ABS: @@ -23158,6 +23185,12 @@ x86_print_call_or_nop (FILE *file, const char *target) if (flag_nop_mcount || !strcmp (target, "nop")) /* 5 byte nop: nopl 0(%[re]ax,%[re]ax,1) */ fprintf (file, "1:" ASM_BYTE "0x0f, 0x1f, 0x44, 0x00, 0x00\n"); + else if (!TARGET_PECOFF && flag_pic) + { + gcc_assert (flag_plt); + + fprintf (file, "1:\tcall\t%s@PLT\n", target); + } else fprintf (file, "1:\tcall\t%s\n", target); } @@ -23321,7 +23354,7 @@ x86_function_profiler (FILE *file, int labelno ATTRIBUTE_UNUSED) break; case CM_SMALL_PIC: case CM_MEDIUM_PIC: - if (!ix86_direct_extern_access) + if (!flag_plt) { if (ASSEMBLER_DIALECT == ASM_INTEL) fprintf (file, "1:\tcall\t[QWORD PTR %s@GOTPCREL[rip]]\n", @@ -23352,7 +23385,9 @@ x86_function_profiler (FILE *file, int labelno ATTRIBUTE_UNUSED) "\tleal\t%sP%d@GOTOFF(%%ebx), %%" PROFILE_COUNT_REGISTER "\n", LPREFIX, labelno); #endif - if (ASSEMBLER_DIALECT == ASM_INTEL) + if (flag_plt) + x86_print_call_or_nop (file, mcount_name); + else if (ASSEMBLER_DIALECT == ASM_INTEL) fprintf (file, "1:\tcall\t[DWORD PTR %s@GOT[ebx]]\n", mcount_name); else fprintf (file, "1:\tcall\t*%s@GOT(%%ebx)\n", mcount_name); @@ -24669,7 +24704,7 @@ ix86_builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost, switch (type_of_cost) { case scalar_stmt: - return fp ? ix86_cost->addss : COSTS_N_INSNS (1); + return fp ? ix86_cost->addss : COSTS_N_INSNS (1); case scalar_load: /* load/store costs are relative to register move which is 2. Recompute @@ -24740,7 +24775,11 @@ ix86_builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost, return ix86_cost->cond_not_taken_branch_cost; case vec_perm: + return ix86_vec_cost (mode, ix86_cost->sse_op); + case vec_promote_demote: + if (fp) + return vec_fp_conversion_cost (ix86_tune_cost, mode); return ix86_vec_cost (mode, ix86_cost->sse_op); case vec_construct: @@ -25261,7 +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 @@ -25316,7 +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); @@ -25325,27 +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 = 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; } } @@ -25369,6 +25533,29 @@ ix86_vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind, break; } + if (kind == vec_promote_demote + && fp && FLOAT_TYPE_P (TREE_TYPE (gimple_assign_rhs1 (stmt_info->stmt)))) + { + 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 (AGU and load ports). Try to account for this by scaling the diff --git a/gcc/config/i386/i386.h b/gcc/config/i386/i386.h index 8507243..18aa42d 100644 --- a/gcc/config/i386/i386.h +++ b/gcc/config/i386/i386.h @@ -207,6 +207,12 @@ struct processor_costs { const int divsd; /* cost of DIVSD instructions. */ const int sqrtss; /* cost of SQRTSS instructions. */ const int sqrtsd; /* cost of SQRTSD instructions. */ + const int cvtss2sd; /* cost SSE FP conversions, + such as CVTSS2SD. */ + const int vcvtps2pd256; /* cost 256bit packed FP conversions, + such as VCVTPD2PS with larger reg in ymm. */ + const int vcvtps2pd512; /* cost 512bit packed FP conversions, + such as VCVTPD2PS with larger reg in zmm. */ const int reassoc_int, reassoc_fp, reassoc_vec_int, reassoc_vec_fp; /* Specify reassociation width for integer, fp, vector integer and vector fp diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md index d6b2f29..e170da3 100644 --- a/gcc/config/i386/i386.md +++ b/gcc/config/i386/i386.md @@ -26592,8 +26592,8 @@ [(set (match_operand:X87MODEF 0 "register_operand") (if_then_else:X87MODEF (match_operand 1 "comparison_operator") - (match_operand:X87MODEF 2 "register_operand") - (match_operand:X87MODEF 3 "register_operand")))] + (match_operand:X87MODEF 2 "nonimm_or_0_or_1s_operand") + (match_operand:X87MODEF 3 "nonimm_or_0_operand")))] "(TARGET_80387 && TARGET_CMOVE) || (SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH)" "if (ix86_expand_fp_movcc (operands)) DONE; else FAIL;") diff --git a/gcc/config/i386/predicates.md b/gcc/config/i386/predicates.md index 3d3848c..4b23e18 100644 --- a/gcc/config/i386/predicates.md +++ b/gcc/config/i386/predicates.md @@ -1267,6 +1267,14 @@ (match_operand 0 "vector_memory_operand") (match_code "const_vector"))) +; Return true when OP is register_operand, vector_memory_operand, +; const_vector zero or const_vector all ones. +(define_predicate "vector_or_0_or_1s_operand" + (ior (match_operand 0 "register_operand") + (match_operand 0 "vector_memory_operand") + (match_operand 0 "const0_operand") + (match_operand 0 "int_float_vector_all_ones_operand"))) + (define_predicate "bcst_mem_operand" (and (match_code "vec_duplicate") (and (match_test "TARGET_AVX512F") @@ -1333,6 +1341,12 @@ (ior (match_operand 0 "nonimmediate_operand") (match_operand 0 "const0_operand"))) +; Return true when OP is a nonimmediate or zero or all ones. +(define_predicate "nonimm_or_0_or_1s_operand" + (ior (match_operand 0 "nonimmediate_operand") + (match_operand 0 "const0_operand") + (match_operand 0 "int_float_vector_all_ones_operand"))) + ;; Return true for RTX codes that force SImode address. (define_predicate "SImode_address_operand" (match_code "subreg,zero_extend,and")) diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md index b280676..2ed348c 100644 --- a/gcc/config/i386/sse.md +++ b/gcc/config/i386/sse.md @@ -527,6 +527,16 @@ (V16SF "TARGET_EVEX512") (V8SF "TARGET_AVX512VL") (V4SF "TARGET_AVX512VL") (V8DF "TARGET_EVEX512") (V4DF "TARGET_AVX512VL") (V2DF "TARGET_AVX512VL")]) +(define_mode_iterator V48_AVX512VL_4 + [(V4SF "TARGET_AVX512VL") (V4DF "TARGET_AVX512VL") + (V4SI "TARGET_AVX512VL") (V4DI "TARGET_AVX512VL")]) + +(define_mode_iterator VI48_AVX512VL_4 + [(V4SI "TARGET_AVX512VL") (V4DI "TARGET_AVX512VL")]) + +(define_mode_iterator V8_AVX512VL_2 + [(V2DF "TARGET_AVX512VL") (V2DI "TARGET_AVX512VL")]) + (define_mode_iterator VFH_AVX10_2 [(V32HF "TARGET_AVX10_2") V16HF V8HF (V16SF "TARGET_AVX10_2") V8SF V4SF @@ -4410,7 +4420,7 @@ (unspec:<V48H_AVX512VL:avx512fmaskmode> [(match_operand:V48H_AVX512VL 1 "nonimmediate_operand" "v") (match_operand:V48H_AVX512VL 2 "nonimmediate_operand" "vm") - (match_operand:SI 3 "const_0_to_7_operand" "n")] + (match_operand:SI 3 "<cmp_imm_predicate>" "n")] UNSPEC_PCMP)))] "TARGET_AVX512F && (!VALID_MASK_AVX512BW_MODE (<SWI248x:MODE>mode) || TARGET_AVX512BW) @@ -4428,7 +4438,7 @@ (unspec:<V48H_AVX512VL:avx512fmaskmode> [(match_operand:V48H_AVX512VL 1 "nonimmediate_operand") (match_operand:V48H_AVX512VL 2 "nonimmediate_operand") - (match_operand:SI 3 "const_0_to_7_operand")] + (match_operand:SI 3 "<cmp_imm_predicate>")] UNSPEC_PCMP))) (set (match_operand:<V48H_AVX512VL:avx512fmaskmode> 4 "register_operand") (unspec:<V48H_AVX512VL:avx512fmaskmode> @@ -4469,7 +4479,8 @@ (match_operand:V48H_AVX512VL 2 "nonimmediate_operand") (match_operand:SI 3 "<cmp_imm_predicate>" "n")] UNSPEC_PCMP)))] - "TARGET_AVX512F && ix86_pre_reload_split ()" + "TARGET_AVX512F && GET_MODE_NUNITS (<MODE>mode) >= 8 + && ix86_pre_reload_split ()" "#" "&& 1" [(set (match_dup 0) @@ -4480,6 +4491,70 @@ UNSPEC_PCMP))] "operands[4] = GEN_INT (INTVAL (operands[3]) ^ 4);") +(define_insn "*<avx512>_cmp<mode>3_and15" + [(set (match_operand:QI 0 "register_operand" "=k") + (and:QI + (unspec:QI + [(match_operand:V48_AVX512VL_4 1 "nonimmediate_operand" "v") + (match_operand:V48_AVX512VL_4 2 "nonimmediate_operand" "vm") + (match_operand:SI 3 "<cmp_imm_predicate>" "n")] + UNSPEC_PCMP) + (const_int 15)))] + "TARGET_AVX512F" + "v<ssecmpintprefix>cmp<ssemodesuffix>\t{%3, %2, %1, %0|%0, %1, %2, %3}" + [(set_attr "type" "ssecmp") + (set_attr "length_immediate" "1") + (set_attr "prefix" "evex") + (set_attr "mode" "<sseinsnmode>")]) + +(define_insn "*<avx512>_ucmp<mode>3_and15" + [(set (match_operand:QI 0 "register_operand" "=k") + (and:QI + (unspec:QI + [(match_operand:VI48_AVX512VL_4 1 "nonimmediate_operand" "v") + (match_operand:VI48_AVX512VL_4 2 "nonimmediate_operand" "vm") + (match_operand:SI 3 "const_0_to_7_operand" "n")] + UNSPEC_UNSIGNED_PCMP) + (const_int 15)))] + "TARGET_AVX512F" + "vpcmpu<ssemodesuffix>\t{%3, %2, %1, %0|%0, %1, %2, %3}" + [(set_attr "type" "ssecmp") + (set_attr "length_immediate" "1") + (set_attr "prefix" "evex") + (set_attr "mode" "<sseinsnmode>")]) + +(define_insn "*<avx512>_cmp<mode>3_and3" + [(set (match_operand:QI 0 "register_operand" "=k") + (and:QI + (unspec:QI + [(match_operand:V8_AVX512VL_2 1 "nonimmediate_operand" "v") + (match_operand:V8_AVX512VL_2 2 "nonimmediate_operand" "vm") + (match_operand:SI 3 "<cmp_imm_predicate>" "n")] + UNSPEC_PCMP) + (const_int 3)))] + "TARGET_AVX512F" + "v<ssecmpintprefix>cmp<ssemodesuffix>\t{%3, %2, %1, %0|%0, %1, %2, %3}" + [(set_attr "type" "ssecmp") + (set_attr "length_immediate" "1") + (set_attr "prefix" "evex") + (set_attr "mode" "<sseinsnmode>")]) + +(define_insn "*avx512vl_ucmpv2di3_and3" + [(set (match_operand:QI 0 "register_operand" "=k") + (and:QI + (unspec:QI + [(match_operand:V2DI 1 "nonimmediate_operand" "v") + (match_operand:V2DI 2 "nonimmediate_operand" "vm") + (match_operand:SI 3 "const_0_to_7_operand" "n")] + UNSPEC_UNSIGNED_PCMP) + (const_int 3)))] + "TARGET_AVX512F" + "vpcmpuq\t{%3, %2, %1, %0|%0, %1, %2, %3}" + [(set_attr "type" "ssecmp") + (set_attr "length_immediate" "1") + (set_attr "prefix" "evex") + (set_attr "mode" "TI")]) + (define_insn "<avx512>_cmp<mode>3<mask_scalar_merge_name>" [(set (match_operand:<avx512fmaskmode> 0 "register_operand" "=k") (unspec:<avx512fmaskmode> @@ -4762,7 +4837,8 @@ (match_operand:VI48_AVX512VL 2 "nonimmediate_operand") (match_operand:SI 3 "const_0_to_7_operand")] UNSPEC_UNSIGNED_PCMP)))] - "TARGET_AVX512F && ix86_pre_reload_split ()" + "TARGET_AVX512F && ix86_pre_reload_split () + && GET_MODE_NUNITS (<MODE>mode) >= 8" "#" "&& 1" [(set (match_dup 0) @@ -5142,7 +5218,7 @@ (define_expand "vcond_mask_<mode><sseintvecmodelower>" [(set (match_operand:VI_256_AVX2 0 "register_operand") (vec_merge:VI_256_AVX2 - (match_operand:VI_256_AVX2 1 "nonimmediate_operand") + (match_operand:VI_256_AVX2 1 "nonimm_or_0_or_1s_operand") (match_operand:VI_256_AVX2 2 "nonimm_or_0_operand") (match_operand:<sseintvecmode> 3 "register_operand")))] "TARGET_AVX" @@ -5155,7 +5231,7 @@ (define_expand "vcond_mask_<mode><sseintvecmodelower>" [(set (match_operand:VI_128 0 "register_operand") (vec_merge:VI_128 - (match_operand:VI_128 1 "vector_operand") + (match_operand:VI_128 1 "vector_or_0_or_1s_operand") (match_operand:VI_128 2 "nonimm_or_0_operand") (match_operand:<sseintvecmode> 3 "register_operand")))] "TARGET_SSE2" @@ -5168,7 +5244,7 @@ (define_expand "vcond_mask_v1tiv1ti" [(set (match_operand:V1TI 0 "register_operand") (vec_merge:V1TI - (match_operand:V1TI 1 "vector_operand") + (match_operand:V1TI 1 "vector_or_0_or_1s_operand") (match_operand:V1TI 2 "nonimm_or_0_operand") (match_operand:V1TI 3 "register_operand")))] "TARGET_SSE2" @@ -5181,7 +5257,7 @@ (define_expand "vcond_mask_<mode><sseintvecmodelower>" [(set (match_operand:VF_256 0 "register_operand") (vec_merge:VF_256 - (match_operand:VF_256 1 "nonimmediate_operand") + (match_operand:VF_256 1 "nonimm_or_0_or_1s_operand") (match_operand:VF_256 2 "nonimm_or_0_operand") (match_operand:<sseintvecmode> 3 "register_operand")))] "TARGET_AVX" @@ -5194,7 +5270,7 @@ (define_expand "vcond_mask_<mode><sseintvecmodelower>" [(set (match_operand:VF_128 0 "register_operand") (vec_merge:VF_128 - (match_operand:VF_128 1 "vector_operand") + (match_operand:VF_128 1 "vector_or_0_or_1s_operand") (match_operand:VF_128 2 "nonimm_or_0_operand") (match_operand:<sseintvecmode> 3 "register_operand")))] "TARGET_SSE" diff --git a/gcc/config/i386/x86-tune-costs.h b/gcc/config/i386/x86-tune-costs.h index 7c8cb73..cddcf61 100644 --- a/gcc/config/i386/x86-tune-costs.h +++ b/gcc/config/i386/x86-tune-costs.h @@ -121,16 +121,19 @@ struct processor_costs ix86_size_cost = {/* costs for tuning for size */ COSTS_N_BYTES (2), /* cost of FCHS instruction. */ COSTS_N_BYTES (2), /* cost of FSQRT instruction. */ - COSTS_N_BYTES (2), /* cost of cheap SSE instruction. */ - COSTS_N_BYTES (2), /* cost of ADDSS/SD SUBSS/SD insns. */ - COSTS_N_BYTES (2), /* cost of MULSS instruction. */ - COSTS_N_BYTES (2), /* cost of MULSD instruction. */ - COSTS_N_BYTES (2), /* cost of FMA SS instruction. */ - COSTS_N_BYTES (2), /* cost of FMA SD instruction. */ - COSTS_N_BYTES (2), /* cost of DIVSS instruction. */ - COSTS_N_BYTES (2), /* cost of DIVSD instruction. */ - COSTS_N_BYTES (2), /* cost of SQRTSS instruction. */ - COSTS_N_BYTES (2), /* cost of SQRTSD instruction. */ + COSTS_N_BYTES (4), /* cost of cheap SSE instruction. */ + COSTS_N_BYTES (4), /* cost of ADDSS/SD SUBSS/SD insns. */ + COSTS_N_BYTES (4), /* cost of MULSS instruction. */ + COSTS_N_BYTES (4), /* cost of MULSD instruction. */ + COSTS_N_BYTES (4), /* cost of FMA SS instruction. */ + COSTS_N_BYTES (4), /* cost of FMA SD instruction. */ + COSTS_N_BYTES (4), /* cost of DIVSS instruction. */ + COSTS_N_BYTES (4), /* cost of DIVSD instruction. */ + COSTS_N_BYTES (4), /* cost of SQRTSS instruction. */ + COSTS_N_BYTES (4), /* cost of SQRTSD instruction. */ + COSTS_N_BYTES (4), /* cost of CVTSS2SD etc. */ + COSTS_N_BYTES (4), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_BYTES (6), /* cost of 512bit VCVTPS2PD etc. */ 1, 1, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ ix86_size_memcpy, ix86_size_memset, @@ -243,6 +246,9 @@ struct processor_costs i386_cost = { /* 386 specific costs */ COSTS_N_INSNS (88), /* cost of DIVSD instruction. */ COSTS_N_INSNS (122), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (122), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (27), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (54), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (108), /* cost of 512bit VCVTPS2PD etc. */ 1, 1, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ i386_memcpy, i386_memset, @@ -356,6 +362,9 @@ struct processor_costs i486_cost = { /* 486 specific costs */ COSTS_N_INSNS (74), /* cost of DIVSD instruction. */ COSTS_N_INSNS (83), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (83), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (8), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (16), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (32), /* cost of 512bit VCVTPS2PD etc. */ 1, 1, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ i486_memcpy, i486_memset, @@ -467,6 +476,9 @@ struct processor_costs pentium_cost = { COSTS_N_INSNS (39), /* cost of DIVSD instruction. */ COSTS_N_INSNS (70), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (70), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (3), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (6), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (12), /* cost of 512bit VCVTPS2PD etc. */ 1, 1, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ pentium_memcpy, pentium_memset, @@ -571,6 +583,9 @@ struct processor_costs lakemont_cost = { COSTS_N_INSNS (60), /* cost of DIVSD instruction. */ COSTS_N_INSNS (31), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (63), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (5), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (10), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (20), /* cost of 512bit VCVTPS2PD etc. */ 1, 1, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ pentium_memcpy, pentium_memset, @@ -690,6 +705,9 @@ struct processor_costs pentiumpro_cost = { COSTS_N_INSNS (18), /* cost of DIVSD instruction. */ COSTS_N_INSNS (31), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (31), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (3), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (6), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (12), /* cost of 512bit VCVTPS2PD etc. */ 1, 1, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ pentiumpro_memcpy, pentiumpro_memset, @@ -800,6 +818,9 @@ struct processor_costs geode_cost = { COSTS_N_INSNS (47), /* cost of DIVSD instruction. */ COSTS_N_INSNS (54), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (54), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (6), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (12), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (24), /* cost of 512bit VCVTPS2PD etc. */ 1, 1, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ geode_memcpy, geode_memset, @@ -913,6 +934,9 @@ struct processor_costs k6_cost = { COSTS_N_INSNS (56), /* cost of DIVSD instruction. */ COSTS_N_INSNS (56), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (56), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (2), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (4), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (8), /* cost of 512bit VCVTPS2PD etc. */ 1, 1, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ k6_memcpy, k6_memset, @@ -1027,6 +1051,9 @@ struct processor_costs athlon_cost = { COSTS_N_INSNS (24), /* cost of DIVSD instruction. */ COSTS_N_INSNS (19), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (19), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (4), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (8), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (16), /* cost of 512bit VCVTPS2PD etc. */ 1, 1, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ athlon_memcpy, athlon_memset, @@ -1150,6 +1177,9 @@ struct processor_costs k8_cost = { COSTS_N_INSNS (20), /* cost of DIVSD instruction. */ COSTS_N_INSNS (19), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (27), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (4), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (8), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (16), /* cost of 512bit VCVTPS2PD etc. */ 1, 1, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ k8_memcpy, k8_memset, @@ -1281,6 +1311,9 @@ struct processor_costs amdfam10_cost = { COSTS_N_INSNS (20), /* cost of DIVSD instruction. */ COSTS_N_INSNS (19), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (27), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (4), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (8), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (16), /* cost of 512bit VCVTPS2PD etc. */ 1, 1, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ amdfam10_memcpy, amdfam10_memset, @@ -1405,6 +1438,9 @@ const struct processor_costs bdver_cost = { COSTS_N_INSNS (27), /* cost of DIVSD instruction. */ COSTS_N_INSNS (15), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (26), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (4), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (7), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (14), /* cost of 512bit VCVTPS2PD etc. */ 1, 2, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ bdver_memcpy, bdver_memset, @@ -1553,6 +1589,10 @@ struct processor_costs znver1_cost = { COSTS_N_INSNS (13), /* cost of DIVSD instruction. */ COSTS_N_INSNS (10), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (15), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (3), /* cost of CVTSS2SD etc. */ + /* Real latency is 4, but for split regs multiply cost of half op by 2. */ + COSTS_N_INSNS (6), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (12), /* cost of 512bit VCVTPS2PD etc. */ /* Zen can execute 4 integer operations per cycle. FP operations take 3 cycles and it can execute 2 integer additions and 2 multiplications thus reassociation may make sense up to with of 6. SPEC2k6 bencharks suggests @@ -1712,6 +1752,9 @@ struct processor_costs znver2_cost = { COSTS_N_INSNS (13), /* cost of DIVSD instruction. */ COSTS_N_INSNS (10), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (15), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (3), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (5), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (10), /* cost of 512bit VCVTPS2PD etc. */ /* Zen can execute 4 integer operations per cycle. FP operations take 3 cycles and it can execute 2 integer additions and 2 multiplications thus reassociation may make sense up to with of 6. @@ -1847,6 +1890,9 @@ struct processor_costs znver3_cost = { COSTS_N_INSNS (13), /* cost of DIVSD instruction. */ COSTS_N_INSNS (10), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (15), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (3), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (5), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (10), /* cost of 512bit VCVTPS2PD etc. */ /* Zen can execute 4 integer operations per cycle. FP operations take 3 cycles and it can execute 2 integer additions and 2 multiplications thus reassociation may make sense up to with of 6. @@ -1984,6 +2030,10 @@ struct processor_costs znver4_cost = { COSTS_N_INSNS (13), /* cost of DIVSD instruction. */ COSTS_N_INSNS (15), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (21), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (3), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (5), /* cost of 256bit VCVTPS2PD etc. */ + /* Real latency is 6, but for split regs multiply cost of half op by 2. */ + COSTS_N_INSNS (10), /* cost of 512bit VCVTPS2PD etc. */ /* Zen can execute 4 integer operations per cycle. FP operations take 3 cycles and it can execute 2 integer additions and 2 multiplications thus reassociation may make sense up to with of 6. @@ -2120,7 +2170,7 @@ struct processor_costs znver5_cost = { COSTS_N_INSNS (1), /* cost of cheap SSE instruction. */ /* ADDSS has throughput 2 and latency 2 (in some cases when source is another addition). */ - COSTS_N_INSNS (2), /* cost of ADDSS/SD SUBSS/SD insns. */ + COSTS_N_INSNS (3), /* cost of ADDSS/SD SUBSS/SD insns. */ /* MULSS has throughput 2 and latency 3. */ COSTS_N_INSNS (3), /* cost of MULSS instruction. */ COSTS_N_INSNS (3), /* cost of MULSD instruction. */ @@ -2135,6 +2185,9 @@ struct processor_costs znver5_cost = { COSTS_N_INSNS (14), /* cost of SQRTSS instruction. */ /* DIVSD has throughtput 0.13 and latency 20. */ COSTS_N_INSNS (20), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (3), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (5), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (5), /* cost of 512bit VCVTPS2PD etc. */ /* Zen5 can execute: - integer ops: 6 per cycle, at most 3 multiplications. latency 1 for additions, 3 for multiplications (pipelined) @@ -2274,6 +2327,9 @@ struct processor_costs skylake_cost = { COSTS_N_INSNS (14), /* cost of DIVSD instruction. */ COSTS_N_INSNS (12), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (18), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (2), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (2), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (4), /* cost of 512bit VCVTPS2PD etc. */ 1, 4, 2, 2, /* reassoc int, fp, vec_int, vec_fp. */ skylake_memcpy, skylake_memset, @@ -2403,6 +2459,9 @@ struct processor_costs icelake_cost = { COSTS_N_INSNS (14), /* cost of DIVSD instruction. */ COSTS_N_INSNS (12), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (18), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (2), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (2), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (2), /* cost of 512bit VCVTPS2PD etc. */ 1, 4, 2, 2, /* reassoc int, fp, vec_int, vec_fp. */ icelake_memcpy, icelake_memset, @@ -2526,6 +2585,9 @@ struct processor_costs alderlake_cost = { COSTS_N_INSNS (17), /* cost of DIVSD instruction. */ COSTS_N_INSNS (14), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (18), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (2), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (2), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (2), /* cost of 512bit VCVTPS2PD etc. */ 1, 4, 3, 3, /* reassoc int, fp, vec_int, vec_fp. */ alderlake_memcpy, alderlake_memset, @@ -2642,6 +2704,9 @@ const struct processor_costs btver1_cost = { COSTS_N_INSNS (17), /* cost of DIVSD instruction. */ COSTS_N_INSNS (14), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (48), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (4), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (7), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (14), /* cost of 512bit VCVTPS2PD etc. */ 1, 1, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ btver1_memcpy, btver1_memset, @@ -2755,6 +2820,9 @@ const struct processor_costs btver2_cost = { COSTS_N_INSNS (19), /* cost of DIVSD instruction. */ COSTS_N_INSNS (16), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (21), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (4), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (7), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (14), /* cost of 512bit VCVTPS2PD etc. */ 1, 1, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ btver2_memcpy, btver2_memset, @@ -2867,6 +2935,9 @@ struct processor_costs pentium4_cost = { COSTS_N_INSNS (38), /* cost of DIVSD instruction. */ COSTS_N_INSNS (23), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (38), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (10), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (20), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (40), /* cost of 512bit VCVTPS2PD etc. */ 1, 1, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ pentium4_memcpy, pentium4_memset, @@ -2982,6 +3053,9 @@ struct processor_costs nocona_cost = { COSTS_N_INSNS (40), /* cost of DIVSD instruction. */ COSTS_N_INSNS (32), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (41), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (10), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (20), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (40), /* cost of 512bit VCVTPS2PD etc. */ 1, 1, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ nocona_memcpy, nocona_memset, @@ -3095,6 +3169,9 @@ struct processor_costs atom_cost = { COSTS_N_INSNS (60), /* cost of DIVSD instruction. */ COSTS_N_INSNS (31), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (63), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (6), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (12), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (24), /* cost of 512bit VCVTPS2PD etc. */ 2, 2, 2, 2, /* reassoc int, fp, vec_int, vec_fp. */ atom_memcpy, atom_memset, @@ -3208,6 +3285,9 @@ struct processor_costs slm_cost = { COSTS_N_INSNS (69), /* cost of DIVSD instruction. */ COSTS_N_INSNS (20), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (35), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (3), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (6), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (12), /* cost of 512bit VCVTPS2PD etc. */ 1, 2, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ slm_memcpy, slm_memset, @@ -3335,6 +3415,9 @@ struct processor_costs tremont_cost = { COSTS_N_INSNS (17), /* cost of DIVSD instruction. */ COSTS_N_INSNS (14), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (18), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (3), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (6), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (12), /* cost of 512bit VCVTPS2PD etc. */ 1, 4, 3, 3, /* reassoc int, fp, vec_int, vec_fp. */ tremont_memcpy, tremont_memset, @@ -3448,6 +3531,9 @@ struct processor_costs intel_cost = { COSTS_N_INSNS (20), /* cost of DIVSD instruction. */ COSTS_N_INSNS (40), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (40), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (8), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (16), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (32), /* cost of 512bit VCVTPS2PD etc. */ 1, 4, 1, 1, /* reassoc int, fp, vec_int, vec_fp. */ intel_memcpy, intel_memset, @@ -3566,6 +3652,9 @@ struct processor_costs lujiazui_cost = { COSTS_N_INSNS (17), /* cost of DIVSD instruction. */ COSTS_N_INSNS (32), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (60), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (3), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (6), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (12), /* cost of 512bit VCVTPS2PD etc. */ 1, 4, 3, 3, /* reassoc int, fp, vec_int, vec_fp. */ lujiazui_memcpy, lujiazui_memset, @@ -3682,6 +3771,9 @@ struct processor_costs yongfeng_cost = { COSTS_N_INSNS (14), /* cost of DIVSD instruction. */ COSTS_N_INSNS (20), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (35), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (3), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (6), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (12), /* cost of 512bit VCVTPS2PD etc. */ 4, 4, 4, 4, /* reassoc int, fp, vec_int, vec_fp. */ yongfeng_memcpy, yongfeng_memset, @@ -3798,6 +3890,9 @@ struct processor_costs shijidadao_cost = { COSTS_N_INSNS (14), /* cost of DIVSD instruction. */ COSTS_N_INSNS (11), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (18), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (3), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (6), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (12), /* cost of 512bit VCVTPS2PD etc. */ 4, 4, 4, 4, /* reassoc int, fp, vec_int, vec_fp. */ shijidadao_memcpy, shijidadao_memset, @@ -3922,6 +4017,9 @@ struct processor_costs generic_cost = { COSTS_N_INSNS (17), /* cost of DIVSD instruction. */ COSTS_N_INSNS (14), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (18), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (3), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (4), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (5), /* cost of 512bit VCVTPS2PD etc. */ 1, 4, 3, 3, /* reassoc int, fp, vec_int, vec_fp. */ generic_memcpy, generic_memset, @@ -4051,6 +4149,9 @@ struct processor_costs core_cost = { COSTS_N_INSNS (32), /* cost of DIVSD instruction. */ COSTS_N_INSNS (30), /* cost of SQRTSS instruction. */ COSTS_N_INSNS (58), /* cost of SQRTSD instruction. */ + COSTS_N_INSNS (2), /* cost of CVTSS2SD etc. */ + COSTS_N_INSNS (2), /* cost of 256bit VCVTPS2PD etc. */ + COSTS_N_INSNS (2), /* cost of 512bit VCVTPS2PD etc. */ 1, 4, 2, 2, /* reassoc int, fp, vec_int, vec_fp. */ core_memcpy, core_memset, diff --git a/gcc/config/i386/x86-tune-sched.cc b/gcc/config/i386/x86-tune-sched.cc index 685a83c..15d3d91 100644 --- a/gcc/config/i386/x86-tune-sched.cc +++ b/gcc/config/i386/x86-tune-sched.cc @@ -81,6 +81,14 @@ ix86_issue_rate (void) case PROCESSOR_YONGFENG: case PROCESSOR_SHIJIDADAO: case PROCESSOR_GENERIC: + /* For znver5 decoder can handle 4 or 8 instructions per cycle, + op cache 12 instruction/cycle, dispatch 8 instructions + integer rename 8 instructions and Fp 6 instructions. + + The scheduler, without understanding out of order nature of the CPU + is not going to be able to use more than 4 instructions since that + is limits of the decoders. */ + case PROCESSOR_ZNVER5: return 4; case PROCESSOR_ICELAKE_CLIENT: @@ -91,13 +99,6 @@ ix86_issue_rate (void) return 5; case PROCESSOR_SAPPHIRERAPIDS: - /* For znver5 decoder can handle 4 or 8 instructions per cycle, - op cache 12 instruction/cycle, dispatch 8 instructions - integer rename 8 instructions and Fp 6 instructions. - - The scheduler, without understanding out of order nature of the CPU - is unlikely going to be able to fill all of these. */ - case PROCESSOR_ZNVER5: return 6; default: |