diff mbox series

RISC-V: Add initial pipeline description for an out-of-order core.

Message ID 80530cd8-b0b6-43af-48b1-6e6cccfe5d6d@gmail.com
State New
Headers show
Series RISC-V: Add initial pipeline description for an out-of-order core. | expand

Commit Message

Robin Dapp Aug. 23, 2023, 1:48 p.m. UTC
Hi,

this adds a pipeline description for a generic out-of-order core.
Latency and units are not based on any real processor but more or less
educated guesses what such a processor could look like.
For the lack of a better name, I called the -mtune parameter "generic-ooo".

In order to account for latency scaling by LMUL != 1, sched_adjust_cost
is implemented.  It will scale an instruction's latency by its LMUL
so an LMUL == 8 instruction will take 8 times the number of cycles
the same instruction with LMUL == 1 would take.
As this potentially causes very high latencies which, in turn, might
lead to scheduling anomalies and a higher number of vsetvls emitted,
this feature is only enabled when specifying -madjust-lmul-cost.

Additionally, in order to easily recognize pre-RA vsetvls this patch
introduces an insn type vsetvl_pre which is used in sched_adjust_cost.

As mentioned, the latency numbers are guesswork at best.  I assumed
6-wide issue as most public announcements point into that direction
and obviously everything else is similarly coarse.  Feel free to
correct in case I unnecessarily pessimized or underestimated something.

Regards
 Robin

gcc/ChangeLog:

	* config/riscv/riscv-cores.def (RISCV_TUNE): Add parameter.
	* config/riscv/riscv-opts.h (enum riscv_microarchitecture_type):
	Add generic_ooo.
	* config/riscv/riscv.cc (riscv_sched_adjust_cost): Implement
	scheduler hook.
	(TARGET_SCHED_ADJUST_COST): Define.
	* config/riscv/riscv.md (no,yes"): Include generic-ooo.md
	* config/riscv/riscv.opt: Add -madjust-lmul-cost.
	* config/riscv/generic-ooo.md: New file.
	* config/riscv/vector.md: Add vsetvl_pre.
---
 gcc/config/riscv/generic-ooo.md  | 284 +++++++++++++++++++++++++++++++
 gcc/config/riscv/riscv-cores.def |   1 +
 gcc/config/riscv/riscv-opts.h    |   3 +-
 gcc/config/riscv/riscv.cc        |  87 ++++++++++
 gcc/config/riscv/riscv.md        |   5 +-
 gcc/config/riscv/riscv.opt       |   3 +
 gcc/config/riscv/vector.md       |   4 +-
 7 files changed, 383 insertions(+), 4 deletions(-)
 create mode 100644 gcc/config/riscv/generic-ooo.md

Comments

钟居哲 Aug. 23, 2023, 2:29 p.m. UTC | #1
Does this patch fix these 2 following PR:
108271 – Missed RVV cost model (gnu.org)
108412 – RISC-V: Negative optimization of GCSE && LOOP INVARIANTS (gnu.org)

If yes, plz append these 2 cases into testsuite and indicate those 2 PR are fixed.
So that we can close them.



juzhe.zhong@rivai.ai
 
From: Robin Dapp
Date: 2023-08-23 21:48
To: gcc-patches; palmer; Kito Cheng; jeffreyalaw; juzhe.zhong@rivai.ai
CC: rdapp.gcc
Subject: [PATCH] RISC-V: Add initial pipeline description for an out-of-order core.
Hi,
 
this adds a pipeline description for a generic out-of-order core.
Latency and units are not based on any real processor but more or less
educated guesses what such a processor could look like.
For the lack of a better name, I called the -mtune parameter "generic-ooo".
 
In order to account for latency scaling by LMUL != 1, sched_adjust_cost
is implemented.  It will scale an instruction's latency by its LMUL
so an LMUL == 8 instruction will take 8 times the number of cycles
the same instruction with LMUL == 1 would take.
As this potentially causes very high latencies which, in turn, might
lead to scheduling anomalies and a higher number of vsetvls emitted,
this feature is only enabled when specifying -madjust-lmul-cost.
 
Additionally, in order to easily recognize pre-RA vsetvls this patch
introduces an insn type vsetvl_pre which is used in sched_adjust_cost.
 
As mentioned, the latency numbers are guesswork at best.  I assumed
6-wide issue as most public announcements point into that direction
and obviously everything else is similarly coarse.  Feel free to
correct in case I unnecessarily pessimized or underestimated something.
 
Regards
Robin
 
gcc/ChangeLog:
 
* config/riscv/riscv-cores.def (RISCV_TUNE): Add parameter.
* config/riscv/riscv-opts.h (enum riscv_microarchitecture_type):
Add generic_ooo.
* config/riscv/riscv.cc (riscv_sched_adjust_cost): Implement
scheduler hook.
(TARGET_SCHED_ADJUST_COST): Define.
* config/riscv/riscv.md (no,yes"): Include generic-ooo.md
* config/riscv/riscv.opt: Add -madjust-lmul-cost.
* config/riscv/generic-ooo.md: New file.
* config/riscv/vector.md: Add vsetvl_pre.
---
gcc/config/riscv/generic-ooo.md  | 284 +++++++++++++++++++++++++++++++
gcc/config/riscv/riscv-cores.def |   1 +
gcc/config/riscv/riscv-opts.h    |   3 +-
gcc/config/riscv/riscv.cc        |  87 ++++++++++
gcc/config/riscv/riscv.md        |   5 +-
gcc/config/riscv/riscv.opt       |   3 +
gcc/config/riscv/vector.md       |   4 +-
7 files changed, 383 insertions(+), 4 deletions(-)
create mode 100644 gcc/config/riscv/generic-ooo.md
 
diff --git a/gcc/config/riscv/generic-ooo.md b/gcc/config/riscv/generic-ooo.md
new file mode 100644
index 00000000000..78b9e48f935
--- /dev/null
+++ b/gcc/config/riscv/generic-ooo.md
@@ -0,0 +1,284 @@
+;; RISC-V generic out-of-order core scheduling model.
+;; Copyright (C) 2017-2023 Free Software Foundation, Inc.
+;;
+;; This file is part of GCC.
+;;
+;; GCC is free software; you can redistribute it and/or modify it
+;; under the terms of the GNU General Public License as published by
+;; the Free Software Foundation; either version 3, or (at your option)
+;; any later version.
+;;
+;; GCC is distributed in the hope that it will be useful, but
+;; WITHOUT ANY WARRANTY; without even the implied warranty of
+;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+;; General Public License for more details.
+;;
+;; You should have received a copy of the GNU General Public License
+;; along with GCC; see the file COPYING3.  If not see
+;; <http://www.gnu.org/licenses/>.
+
+(define_automaton "generic_ooo")
+
+;; Regarding functional units we assume a three-way split:
+;; - Integer ALU (IXU) - 4 symmetric units.
+;; - Floating-point (FXU) - 2 symmetric units.
+;; - Vector Unit (VXU) - 1 unit.
+
+;; We assume 6-wide issue:
+;; - 5-wide generic/integer issue.
+;; - 1-wide vector issue.
+
+;; For now, the only subunits are for non-pipelined integer division and
+;; vector div/mult/sqrt.
+;; No extra units for e.g. vector permutes, masking, everything is assumed to
+;; be on the same pipelined execution unit.
+
+;; Latency:
+;; - Regular integer operations take 1 cycle.
+;; - Multiplication/Division take multiple cycles.
+;; - Float operations take 4-6 cycles.
+;; - Regular vector operations take 2-6 cycles.
+;;   (This assumes LMUL = 1, latency for LMUL = 2, 4, 8 is scaled accordingly
+;;    by riscv_sched_adjust_cost when -madjust-lmul-cost is given)
+;; - Load/Store:
+;;   - To/From IXU: 4 cycles.
+;;   - To/From FXU: 6 cycles.
+;;   - To/From VXU: 6 cycles.
+
+;; Integer/float issue queues.
+(define_cpu_unit "issue0,issue1,issue2,issue3,issue4" "generic_ooo")
+
+;; Separate issue queue for vector instructions.
+(define_cpu_unit "generic_ooo_vxu_issue" "generic_ooo")
+
+;; Integer/float execution units.
+(define_cpu_unit "ixu0,ixu1,ixu2,ixu3" "generic_ooo")
+(define_cpu_unit "fxu0,fxu1" "generic_ooo")
+
+;; Integer subunit for division.
+(define_cpu_unit "generic_ooo_div" "generic_ooo")
+
+;; Vector execution unit.
+(define_cpu_unit "generic_ooo_vxu_alu" "generic_ooo")
+
+;; Vector subunit that does mult/div/sqrt.
+(define_cpu_unit "generic_ooo_vxu_multicycle" "generic_ooo")
+
+;; Shortcuts
+(define_reservation "generic_ooo_issue" "issue0|issue1|issue2|issue3|issue4")
+(define_reservation "generic_ooo_ixu_alu" "ixu0|ixu1|ixu2|ixu3")
+(define_reservation "generic_ooo_fxu" "fxu0|fxu1")
+
+
+;; Integer load/store
+(define_insn_reservation "generic_ooo_int_load" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "load"))
+  "generic_ooo_issue,generic_ooo_ixu_alu")
+
+(define_insn_reservation "generic_ooo_int_store" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "store"))
+  "generic_ooo_issue,generic_ooo_ixu_alu")
+
+;; Float load/store
+(define_insn_reservation "generic_ooo_float_load" 6
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "fpload"))
+  "generic_ooo_issue,generic_ooo_ixu_alu")
+
+(define_insn_reservation "generic_ooo_float_store" 6
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "fpstore"))
+  "generic_ooo_issue,generic_ooo_fxu")
+
+;; Vector load/store
+(define_insn_reservation "generic_ooo_vec_load" 6
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vlde,vldm,vlds,vldux,vldox,vldff,vldr"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+(define_insn_reservation "generic_ooo_vec_store" 6
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vste,vstm,vsts,vstux,vstox,vstr"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector segment loads/stores.
+(define_insn_reservation "generic_ooo_vec_loadstore_seg" 10
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vlsegde,vlsegds,vlsegdux,vlsegdox,vlsegdff,\
+ vssegte,vssegts,vssegtux,vssegtox"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+
+;; Generic integer instructions.
+(define_insn_reservation "generic_ooo_alu" 1
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "unknown,const,arith,shift,slt,multi,auipc,nop,logical,\
+ move,bitmanip,min,max,minu,maxu,clz,ctz"))
+  "generic_ooo_issue,generic_ooo_ixu_alu")
+
+
+;; Float move, convert and compare.
+(define_insn_reservation "generic_ooo_float_move" 3
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "fmove"))
+  "generic_ooo_issue,generic_ooo_fxu")
+
+(define_insn_reservation "generic_ooo_fcvt" 3
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "fcvt"))
+  "generic_ooo_issue,generic_ooo_fxu")
+
+(define_insn_reservation "generic_ooo_fcmp" 2
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "fcmp"))
+  "generic_ooo_issue,generic_ooo_fxu")
+
+;; Integer multiplication.
+(define_insn_reservation "generic_ooo_imul" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "imul"))
+  "generic_ooo_issue,generic_ooo_ixu_alu,generic_ooo_ixu_alu")
+
+;; Assume integer division is not pipelined.  Do not block the unit for more
+;; than three cycles so the DFA does not get too large.  Similar for other
+;; non-pipelined instructions.
+(define_insn_reservation "generic_ooo_idiv" 16
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "idiv"))
+  "generic_ooo_issue,generic_ooo_ixu_alu,generic_ooo_div,generic_ooo_div*3")
+
+;; Float addition and multiplication.
+(define_insn_reservation "generic_ooo_faddmul" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "fadd,fmul"))
+  "generic_ooo_issue,generic_ooo_fxu")
+
+;; Float FMA.
+(define_insn_reservation "generic_ooo_float_fma" 6
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "fmadd"))
+  "generic_ooo_issue,generic_ooo_fxu")
+
+;; Assume float division and sqrt are not pipelined.
+(define_insn_reservation "generic_ooo_float_div_single" 12
+  (and (eq_attr "tune" "generic_ooo")
+       (and (eq_attr "type" "fdiv,fsqrt")
+     (eq_attr "mode" "SF")))
+  "generic_ooo_issue,generic_ooo_fxu,generic_ooo_div,generic_ooo_div*3")
+
+(define_insn_reservation "generic_ooo_float_div_double" 16
+  (and (eq_attr "tune" "generic_ooo")
+       (and (eq_attr "type" "fdiv,fsqrt")
+     (eq_attr "mode" "DF")))
+  "generic_ooo_issue,generic_ooo_fxu,generic_ooo_div,generic_ooo_div*3")
+
+;; Popcount and clmul.
+(define_insn_reservation "generic_ooo_popcount" 2
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "cpop,clmul"))
+  "generic_ooo_issue,generic_ooo_ixu_alu")
+
+;; Regular vector operations and integer comparisons.
+(define_insn_reservation "generic_ooo_vec_alu" 3
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vialu,viwalu,vext,vicalu,vshift,vnshift,viminmax,vicmp,\
+         vimov,vsalu,vaalu,vsshift,vnclip,vmov,vfmov"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector float comparison, conversion etc.
+(define_insn_reservation "generic_ooo_vec_fcmp" 3
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vfrecp,vfminmax,vfcmp,vfsgnj,vfclass,vfcvtitof,\
+ vfcvtftoi,vfwcvtitof,vfwcvtftoi,vfwcvtftof,vfncvtitof,\
+ vfncvtftoi,vfncvtftof"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector integer multiplication.
+(define_insn_reservation "generic_ooo_vec_imul" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vimul,viwmul,vimuladd,viwmuladd,vsmul"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector float addition.
+(define_insn_reservation "generic_ooo_vec_fadd" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vfalu,vfwalu"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector float multiplication and FMA.
+(define_insn_reservation "generic_ooo_vec_fmul" 6
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vfmul,vfwmul,vfmuladd,vfwmuladd"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector crypto, assumed to be a generic operation for now.
+(define_insn_reservation "generic_ooo_crypto" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "crypto"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector permute.
+(define_insn_reservation "generic_ooo_perm" 3
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vimerge,vfmerge,vslideup,vslidedown,vislide1up,\
+ vislide1down,vfslide1up,vfslide1down,vgather,vcompress"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector reduction.
+(define_insn_reservation "generic_ooo_vec_reduction" 8
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vired,viwred,vfredu,vfwredu"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_multicycle")
+
+;; Vector ordered reduction, assume the latency number is for
+;; a 128-bit vector.  It is scaled in riscv_sched_adjust_cost
+;; for larger vectors.
+(define_insn_reservation "generic_ooo_vec_ordered_reduction" 10
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vfredo,vfwredo"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_multicycle*3")
+
+;; Vector integer division, assume not pipelined.
+(define_insn_reservation "generic_ooo_vec_idiv" 16
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vidiv"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_multicycle*3")
+
+;; Vector float divisions and sqrt, assume not pipelined.
+(define_insn_reservation "generic_ooo_vec_float_divsqrt" 16
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vfdiv,vfsqrt"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_multicycle*3")
+
+;; Vector mask operations.
+(define_insn_reservation "generic_ooo_vec_mask" 2
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vmalu,vmpop,vmffs,vmsfs,vmiota,vmidx,vimovvx,vimovxv,\
+ vfmovvf,vfmovfv"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector vsetvl.
+(define_insn_reservation "generic_ooo_vec_vesetvl" 1
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vsetvl,vsetvl_pre"))
+  "generic_ooo_vxu_issue")
+
+;; Vector rounding mode setters, assume pipeline barrier.
+(define_insn_reservation "generic_ooo_vec_setrm" 20
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "wrvxrm,wrfrm"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_issue*3")
+
+;; Vector read vlen/vlenb.
+(define_insn_reservation "generic_ooo_vec_readlen" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "rdvlenb,rdvl"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_issue")
+
+;; Transfer from/to coprocessor.  Assume not pipelined.
+(define_insn_reservation "generic_ooo_xfer" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "mfc,mtc"))
+  "generic_ooo_issue,generic_ooo_ixu_alu,generic_ooo_ixu_alu*3")
diff --git a/gcc/config/riscv/riscv-cores.def b/gcc/config/riscv/riscv-cores.def
index 7d87ab7ce28..91deabb6079 100644
--- a/gcc/config/riscv/riscv-cores.def
+++ b/gcc/config/riscv/riscv-cores.def
@@ -38,6 +38,7 @@ RISCV_TUNE("sifive-3-series", generic, rocket_tune_info)
RISCV_TUNE("sifive-5-series", generic, rocket_tune_info)
RISCV_TUNE("sifive-7-series", sifive_7, sifive_7_tune_info)
RISCV_TUNE("thead-c906", generic, thead_c906_tune_info)
+RISCV_TUNE("generic-ooo", generic_ooo, generic_ooo_tune_info)
RISCV_TUNE("size", generic, optimize_size_tune_info)
#undef RISCV_TUNE
diff --git a/gcc/config/riscv/riscv-opts.h b/gcc/config/riscv/riscv-opts.h
index 378a17699cd..2f22fcaf9bd 100644
--- a/gcc/config/riscv/riscv-opts.h
+++ b/gcc/config/riscv/riscv-opts.h
@@ -52,7 +52,8 @@ extern enum riscv_isa_spec_class riscv_isa_spec;
/* Keep this list in sync with define_attr "tune" in riscv.md.  */
enum riscv_microarchitecture_type {
   generic,
-  sifive_7
+  sifive_7,
+  generic_ooo
};
extern enum riscv_microarchitecture_type riscv_microarchitecture;
diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc
index 480f3124496..0d7bc2ae9b6 100644
--- a/gcc/config/riscv/riscv.cc
+++ b/gcc/config/riscv/riscv.cc
@@ -349,6 +349,21 @@ static const struct riscv_tune_param thead_c906_tune_info = {
   false /* use_divmod_expansion */
};
+/* Costs to use when optimizing for a generic ooo profile.  */
+static const struct riscv_tune_param generic_ooo_tune_info = {
+  {COSTS_N_INSNS (2), COSTS_N_INSNS (2)}, /* fp_add */
+  {COSTS_N_INSNS (5), COSTS_N_INSNS (6)}, /* fp_mul */
+  {COSTS_N_INSNS (7), COSTS_N_INSNS (8)}, /* fp_div */
+  {COSTS_N_INSNS (2), COSTS_N_INSNS (2)}, /* int_mul */
+  {COSTS_N_INSNS (6), COSTS_N_INSNS (6)}, /* int_div */
+  1, /* issue_rate */
+  3, /* branch_cost */
+  4, /* memory_cost */
+  4, /* fmv_cost */
+  false, /* slow_unaligned_access */
+  false, /* use_divmod_expansion */
+};
+
/* Costs to use when optimizing for size.  */
static const struct riscv_tune_param optimize_size_tune_info = {
   {COSTS_N_INSNS (1), COSTS_N_INSNS (1)}, /* fp_add */
@@ -6744,6 +6759,75 @@ riscv_sched_variable_issue (FILE *, int, rtx_insn *insn, int more)
   return more - 1;
}
+/* Adjust the cost/latency of instructions for scheduling.
+   For now this is just used to change the latency of vector instructions
+   according to their LMUL.  We assume that an insn with LMUL == 8 requires
+   eight times more execution cycles than the same insn with LMUL == 1.
+   As this may cause very high latencies which lead to scheduling artifacts
+   we currently only perform the adjustment when -madjust-lmul-cost is given.
+   */
+static int
+riscv_sched_adjust_cost (rtx_insn *, int, rtx_insn *insn, int cost,
+ unsigned int)
+{
+  /* Only do adjustments for the generic out-of-order scheduling model.  */
+  if (!TARGET_VECTOR || riscv_microarchitecture != generic_ooo)
+    return cost;
+
+  if (recog_memoized (insn) < 0)
+    return cost;
+
+  enum attr_type type = get_attr_type (insn);
+
+  if (type == TYPE_VFREDO || type == TYPE_VFWREDO)
+    {
+      /* TODO: For ordered reductions scale the base cost relative to the
+ number of units.  */
+      ;
+    }
+
+  /* Don't do any LMUL-based latency adjustment unless explicitly asked to.  */
+  if (!TARGET_ADJUST_LMUL_COST)
+    return cost;
+
+  /* vsetvl has a vlmul attribute but its latency does not depend on it.  */
+  if (type == TYPE_VSETVL || type == TYPE_VSETVL_PRE)
+    return cost;
+
+  enum riscv_vector::vlmul_type lmul =
+    (riscv_vector::vlmul_type)get_attr_vlmul (insn);
+
+  double factor = 1;
+  switch (lmul)
+    {
+    case riscv_vector::LMUL_2:
+      factor = 2;
+      break;
+    case riscv_vector::LMUL_4:
+      factor = 4;
+      break;
+    case riscv_vector::LMUL_8:
+      factor = 8;
+      break;
+    case riscv_vector::LMUL_F2:
+      factor = 0.5;
+      break;
+    case riscv_vector::LMUL_F4:
+      factor = 0.25;
+      break;
+    case riscv_vector::LMUL_F8:
+      factor = 0.125;
+      break;
+    default:
+      factor = 1;
+    }
+
+  /* If the latency was nonzero, keep it that way.  */
+  int new_cost = MAX (cost > 0 ? 1 : 0, cost * factor);
+
+  return new_cost;
+}
+
/* Auxiliary function to emit RISC-V ELF attribute. */
static void
riscv_emit_attribute ()
@@ -8469,6 +8553,9 @@ riscv_frame_pointer_required (void)
#undef  TARGET_SCHED_VARIABLE_ISSUE
#define TARGET_SCHED_VARIABLE_ISSUE riscv_sched_variable_issue
+#undef  TARGET_SCHED_ADJUST_COST
+#define TARGET_SCHED_ADJUST_COST riscv_sched_adjust_cost
+
#undef TARGET_FUNCTION_OK_FOR_SIBCALL
#define TARGET_FUNCTION_OK_FOR_SIBCALL riscv_function_ok_for_sibcall
diff --git a/gcc/config/riscv/riscv.md b/gcc/config/riscv/riscv.md
index b456fa6abb3..2a9bad0d8ee 100644
--- a/gcc/config/riscv/riscv.md
+++ b/gcc/config/riscv/riscv.md
@@ -406,7 +406,7 @@ (define_attr "type"
    mtc,mfc,const,arith,logical,shift,slt,imul,idiv,move,fmove,fadd,fmul,
    fmadd,fdiv,fcmp,fcvt,fsqrt,multi,auipc,sfb_alu,nop,ghost,bitmanip,rotate,
    clmul,min,max,minu,maxu,clz,ctz,cpop,
-   atomic,condmove,crypto,rdvlenb,rdvl,wrvxrm,wrfrm,rdfrm,vsetvl,
+   atomic,condmove,crypto,rdvlenb,rdvl,wrvxrm,wrfrm,rdfrm,vsetvl,vsetvl_pre,
    vlde,vste,vldm,vstm,vlds,vsts,
    vldux,vldox,vstux,vstox,vldff,vldr,vstr,
    vlsegde,vssegte,vlsegds,vssegts,vlsegdux,vlsegdox,vssegtux,vssegtox,vlsegdff,
@@ -534,7 +534,7 @@ (define_attr "cannot_copy" "no,yes" (const_string "no"))
;; Microarchitectures we know how to tune for.
;; Keep this in sync with enum riscv_microarchitecture.
(define_attr "tune"
-  "generic,sifive_7"
+  "generic,sifive_7,generic_ooo"
   (const (symbol_ref "((enum attr_tune) riscv_microarchitecture)")))
;; Describe a user's asm statement.
@@ -3328,5 +3328,6 @@ (define_expand "msubhisi4"
(include "generic.md")
(include "sifive-7.md")
(include "thead.md")
+(include "generic-ooo.md")
(include "vector.md")
(include "zicond.md")
diff --git a/gcc/config/riscv/riscv.opt b/gcc/config/riscv/riscv.opt
index a962ea8f9d4..724a5429e5d 100644
--- a/gcc/config/riscv/riscv.opt
+++ b/gcc/config/riscv/riscv.opt
@@ -311,3 +311,6 @@ Enum(riscv_autovec_lmul) String(m8) Value(RVV_M8)
-param=riscv-autovec-lmul=
Target RejectNegative Joined Enum(riscv_autovec_lmul) Var(riscv_autovec_lmul) Init(RVV_M1)
-param=riscv-autovec-lmul=<string> Set the RVV LMUL of auto-vectorization in the RISC-V port.
+
+madjust-lmul-cost
+Target Var(TARGET_ADJUST_LMUL_COST) Init(0)
diff --git a/gcc/config/riscv/vector.md b/gcc/config/riscv/vector.md
index 6ceae25dbed..9c61d0ead8a 100644
--- a/gcc/config/riscv/vector.md
+++ b/gcc/config/riscv/vector.md
@@ -935,7 +935,9 @@ (define_insn "@vlmax_avl<mode>"
   [(set (match_operand:P 0 "register_operand" "=r")
(unspec:P [(match_operand:P 1 "const_int_operand" "i")] UNSPEC_VLMAX))]
   "TARGET_VECTOR"
-  "")
+  ""
+  [(set_attr "type" "vsetvl_pre")]
+  )
;; Set VXRM
(define_insn "vxrmsi"
Robin Dapp Aug. 23, 2023, 2:56 p.m. UTC | #2
> Does this patch fix these 2 following PR:
> 108271 – Missed RVV cost model (gnu.org) <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108271>
> 108412 – RISC-V: Negative optimization of GCSE && LOOP INVARIANTS (gnu.org) <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108412>
> 
> If yes, plz append these 2 cases into testsuite and indicate those 2 PR are fixed.
> So that we can close them.

The second one is fixed on my local branch, the first not yet because there
is more to it still.  The second one is more due to pressure-aware scheduling
and I'm going to add it to the commit as well as the PR to the commit once this
is verified.

Regards
 Robin
Jeff Law Aug. 23, 2023, 11:21 p.m. UTC | #3
On 8/23/23 08:56, Robin Dapp wrote:
>> Does this patch fix these 2 following PR:
>> 108271 – Missed RVV cost model (gnu.org) <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108271>
>> 108412 – RISC-V: Negative optimization of GCSE && LOOP INVARIANTS (gnu.org) <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108412>
>>
>> If yes, plz append these 2 cases into testsuite and indicate those 2 PR are fixed.
>> So that we can close them.
> 
> The second one is fixed on my local branch, the first not yet because there
> is more to it still.  The second one is more due to pressure-aware scheduling
> and I'm going to add it to the commit as well as the PR to the commit once this
> is verified.
Basically there's a little conditional in IRA which basically says don't 
rematerialize constants at their use points if register pressure 
sensitive scheduling has been enabled.    So turning on -fsched-pressure 
helps, it's a bit of a cop-out, but probably good enough for now.

Jeff
钟居哲 Aug. 28, 2023, 2:25 a.m. UTC | #4
Ok.

I am not familiar with scheduling stuff but I hope you can fix those 2 issues.

I have no objection with this patch and I prefer Jeff or kito make the decision for this patch.

Thanks.


juzhe.zhong@rivai.ai
 
From: Robin Dapp
Date: 2023-08-23 22:56
To: 钟居哲; gcc-patches; palmer; kito.cheng; Jeff Law
CC: rdapp.gcc
Subject: Re: [PATCH] RISC-V: Add initial pipeline description for an out-of-order core.
> Does this patch fix these 2 following PR:
> 108271 – Missed RVV cost model (gnu.org) <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108271>
> 108412 – RISC-V: Negative optimization of GCSE && LOOP INVARIANTS (gnu.org) <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108412>
> 
> If yes, plz append these 2 cases into testsuite and indicate those 2 PR are fixed.
> So that we can close them.
 
The second one is fixed on my local branch, the first not yet because there
is more to it still.  The second one is more due to pressure-aware scheduling
and I'm going to add it to the commit as well as the PR to the commit once this
is verified.
 
Regards
Robin
diff mbox series

Patch

diff --git a/gcc/config/riscv/generic-ooo.md b/gcc/config/riscv/generic-ooo.md
new file mode 100644
index 00000000000..78b9e48f935
--- /dev/null
+++ b/gcc/config/riscv/generic-ooo.md
@@ -0,0 +1,284 @@ 
+;; RISC-V generic out-of-order core scheduling model.
+;; Copyright (C) 2017-2023 Free Software Foundation, Inc.
+;;
+;; This file is part of GCC.
+;;
+;; GCC is free software; you can redistribute it and/or modify it
+;; under the terms of the GNU General Public License as published by
+;; the Free Software Foundation; either version 3, or (at your option)
+;; any later version.
+;;
+;; GCC is distributed in the hope that it will be useful, but
+;; WITHOUT ANY WARRANTY; without even the implied warranty of
+;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+;; General Public License for more details.
+;;
+;; You should have received a copy of the GNU General Public License
+;; along with GCC; see the file COPYING3.  If not see
+;; <http://www.gnu.org/licenses/>.
+
+(define_automaton "generic_ooo")
+
+;; Regarding functional units we assume a three-way split:
+;; - Integer ALU (IXU) - 4 symmetric units.
+;; - Floating-point (FXU) - 2 symmetric units.
+;; - Vector Unit (VXU) - 1 unit.
+
+;; We assume 6-wide issue:
+;; - 5-wide generic/integer issue.
+;; - 1-wide vector issue.
+
+;; For now, the only subunits are for non-pipelined integer division and
+;; vector div/mult/sqrt.
+;; No extra units for e.g. vector permutes, masking, everything is assumed to
+;; be on the same pipelined execution unit.
+
+;; Latency:
+;; - Regular integer operations take 1 cycle.
+;; - Multiplication/Division take multiple cycles.
+;; - Float operations take 4-6 cycles.
+;; - Regular vector operations take 2-6 cycles.
+;;   (This assumes LMUL = 1, latency for LMUL = 2, 4, 8 is scaled accordingly
+;;    by riscv_sched_adjust_cost when -madjust-lmul-cost is given)
+;; - Load/Store:
+;;   - To/From IXU: 4 cycles.
+;;   - To/From FXU: 6 cycles.
+;;   - To/From VXU: 6 cycles.
+
+;; Integer/float issue queues.
+(define_cpu_unit "issue0,issue1,issue2,issue3,issue4" "generic_ooo")
+
+;; Separate issue queue for vector instructions.
+(define_cpu_unit "generic_ooo_vxu_issue" "generic_ooo")
+
+;; Integer/float execution units.
+(define_cpu_unit "ixu0,ixu1,ixu2,ixu3" "generic_ooo")
+(define_cpu_unit "fxu0,fxu1" "generic_ooo")
+
+;; Integer subunit for division.
+(define_cpu_unit "generic_ooo_div" "generic_ooo")
+
+;; Vector execution unit.
+(define_cpu_unit "generic_ooo_vxu_alu" "generic_ooo")
+
+;; Vector subunit that does mult/div/sqrt.
+(define_cpu_unit "generic_ooo_vxu_multicycle" "generic_ooo")
+
+;; Shortcuts
+(define_reservation "generic_ooo_issue" "issue0|issue1|issue2|issue3|issue4")
+(define_reservation "generic_ooo_ixu_alu" "ixu0|ixu1|ixu2|ixu3")
+(define_reservation "generic_ooo_fxu" "fxu0|fxu1")
+
+
+;; Integer load/store
+(define_insn_reservation "generic_ooo_int_load" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "load"))
+  "generic_ooo_issue,generic_ooo_ixu_alu")
+
+(define_insn_reservation "generic_ooo_int_store" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "store"))
+  "generic_ooo_issue,generic_ooo_ixu_alu")
+
+;; Float load/store
+(define_insn_reservation "generic_ooo_float_load" 6
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "fpload"))
+  "generic_ooo_issue,generic_ooo_ixu_alu")
+
+(define_insn_reservation "generic_ooo_float_store" 6
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "fpstore"))
+  "generic_ooo_issue,generic_ooo_fxu")
+
+;; Vector load/store
+(define_insn_reservation "generic_ooo_vec_load" 6
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vlde,vldm,vlds,vldux,vldox,vldff,vldr"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+(define_insn_reservation "generic_ooo_vec_store" 6
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vste,vstm,vsts,vstux,vstox,vstr"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector segment loads/stores.
+(define_insn_reservation "generic_ooo_vec_loadstore_seg" 10
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vlsegde,vlsegds,vlsegdux,vlsegdox,vlsegdff,\
+			vssegte,vssegts,vssegtux,vssegtox"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+
+;; Generic integer instructions.
+(define_insn_reservation "generic_ooo_alu" 1
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "unknown,const,arith,shift,slt,multi,auipc,nop,logical,\
+			move,bitmanip,min,max,minu,maxu,clz,ctz"))
+  "generic_ooo_issue,generic_ooo_ixu_alu")
+
+
+;; Float move, convert and compare.
+(define_insn_reservation "generic_ooo_float_move" 3
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "fmove"))
+  "generic_ooo_issue,generic_ooo_fxu")
+
+(define_insn_reservation "generic_ooo_fcvt" 3
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "fcvt"))
+  "generic_ooo_issue,generic_ooo_fxu")
+
+(define_insn_reservation "generic_ooo_fcmp" 2
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "fcmp"))
+  "generic_ooo_issue,generic_ooo_fxu")
+
+;; Integer multiplication.
+(define_insn_reservation "generic_ooo_imul" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "imul"))
+  "generic_ooo_issue,generic_ooo_ixu_alu,generic_ooo_ixu_alu")
+
+;; Assume integer division is not pipelined.  Do not block the unit for more
+;; than three cycles so the DFA does not get too large.  Similar for other
+;; non-pipelined instructions.
+(define_insn_reservation "generic_ooo_idiv" 16
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "idiv"))
+  "generic_ooo_issue,generic_ooo_ixu_alu,generic_ooo_div,generic_ooo_div*3")
+
+;; Float addition and multiplication.
+(define_insn_reservation "generic_ooo_faddmul" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "fadd,fmul"))
+  "generic_ooo_issue,generic_ooo_fxu")
+
+;; Float FMA.
+(define_insn_reservation "generic_ooo_float_fma" 6
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "fmadd"))
+  "generic_ooo_issue,generic_ooo_fxu")
+
+;; Assume float division and sqrt are not pipelined.
+(define_insn_reservation "generic_ooo_float_div_single" 12
+  (and (eq_attr "tune" "generic_ooo")
+       (and (eq_attr "type" "fdiv,fsqrt")
+	    (eq_attr "mode" "SF")))
+  "generic_ooo_issue,generic_ooo_fxu,generic_ooo_div,generic_ooo_div*3")
+
+(define_insn_reservation "generic_ooo_float_div_double" 16
+  (and (eq_attr "tune" "generic_ooo")
+       (and (eq_attr "type" "fdiv,fsqrt")
+	    (eq_attr "mode" "DF")))
+  "generic_ooo_issue,generic_ooo_fxu,generic_ooo_div,generic_ooo_div*3")
+
+;; Popcount and clmul.
+(define_insn_reservation "generic_ooo_popcount" 2
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "cpop,clmul"))
+  "generic_ooo_issue,generic_ooo_ixu_alu")
+
+;; Regular vector operations and integer comparisons.
+(define_insn_reservation "generic_ooo_vec_alu" 3
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vialu,viwalu,vext,vicalu,vshift,vnshift,viminmax,vicmp,\
+		        vimov,vsalu,vaalu,vsshift,vnclip,vmov,vfmov"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector float comparison, conversion etc.
+(define_insn_reservation "generic_ooo_vec_fcmp" 3
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vfrecp,vfminmax,vfcmp,vfsgnj,vfclass,vfcvtitof,\
+			vfcvtftoi,vfwcvtitof,vfwcvtftoi,vfwcvtftof,vfncvtitof,\
+			vfncvtftoi,vfncvtftof"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector integer multiplication.
+(define_insn_reservation "generic_ooo_vec_imul" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vimul,viwmul,vimuladd,viwmuladd,vsmul"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector float addition.
+(define_insn_reservation "generic_ooo_vec_fadd" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vfalu,vfwalu"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector float multiplication and FMA.
+(define_insn_reservation "generic_ooo_vec_fmul" 6
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vfmul,vfwmul,vfmuladd,vfwmuladd"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector crypto, assumed to be a generic operation for now.
+(define_insn_reservation "generic_ooo_crypto" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "crypto"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector permute.
+(define_insn_reservation "generic_ooo_perm" 3
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vimerge,vfmerge,vslideup,vslidedown,vislide1up,\
+			vislide1down,vfslide1up,vfslide1down,vgather,vcompress"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector reduction.
+(define_insn_reservation "generic_ooo_vec_reduction" 8
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vired,viwred,vfredu,vfwredu"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_multicycle")
+
+;; Vector ordered reduction, assume the latency number is for
+;; a 128-bit vector.  It is scaled in riscv_sched_adjust_cost
+;; for larger vectors.
+(define_insn_reservation "generic_ooo_vec_ordered_reduction" 10
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vfredo,vfwredo"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_multicycle*3")
+
+;; Vector integer division, assume not pipelined.
+(define_insn_reservation "generic_ooo_vec_idiv" 16
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vidiv"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_multicycle*3")
+
+;; Vector float divisions and sqrt, assume not pipelined.
+(define_insn_reservation "generic_ooo_vec_float_divsqrt" 16
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vfdiv,vfsqrt"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_multicycle*3")
+
+;; Vector mask operations.
+(define_insn_reservation "generic_ooo_vec_mask" 2
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vmalu,vmpop,vmffs,vmsfs,vmiota,vmidx,vimovvx,vimovxv,\
+			vfmovvf,vfmovfv"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_alu")
+
+;; Vector vsetvl.
+(define_insn_reservation "generic_ooo_vec_vesetvl" 1
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "vsetvl,vsetvl_pre"))
+  "generic_ooo_vxu_issue")
+
+;; Vector rounding mode setters, assume pipeline barrier.
+(define_insn_reservation "generic_ooo_vec_setrm" 20
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "wrvxrm,wrfrm"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_issue*3")
+
+;; Vector read vlen/vlenb.
+(define_insn_reservation "generic_ooo_vec_readlen" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "rdvlenb,rdvl"))
+  "generic_ooo_vxu_issue,generic_ooo_vxu_issue")
+
+;; Transfer from/to coprocessor.  Assume not pipelined.
+(define_insn_reservation "generic_ooo_xfer" 4
+  (and (eq_attr "tune" "generic_ooo")
+       (eq_attr "type" "mfc,mtc"))
+  "generic_ooo_issue,generic_ooo_ixu_alu,generic_ooo_ixu_alu*3")
diff --git a/gcc/config/riscv/riscv-cores.def b/gcc/config/riscv/riscv-cores.def
index 7d87ab7ce28..91deabb6079 100644
--- a/gcc/config/riscv/riscv-cores.def
+++ b/gcc/config/riscv/riscv-cores.def
@@ -38,6 +38,7 @@  RISCV_TUNE("sifive-3-series", generic, rocket_tune_info)
 RISCV_TUNE("sifive-5-series", generic, rocket_tune_info)
 RISCV_TUNE("sifive-7-series", sifive_7, sifive_7_tune_info)
 RISCV_TUNE("thead-c906", generic, thead_c906_tune_info)
+RISCV_TUNE("generic-ooo", generic_ooo, generic_ooo_tune_info)
 RISCV_TUNE("size", generic, optimize_size_tune_info)
 
 #undef RISCV_TUNE
diff --git a/gcc/config/riscv/riscv-opts.h b/gcc/config/riscv/riscv-opts.h
index 378a17699cd..2f22fcaf9bd 100644
--- a/gcc/config/riscv/riscv-opts.h
+++ b/gcc/config/riscv/riscv-opts.h
@@ -52,7 +52,8 @@  extern enum riscv_isa_spec_class riscv_isa_spec;
 /* Keep this list in sync with define_attr "tune" in riscv.md.  */
 enum riscv_microarchitecture_type {
   generic,
-  sifive_7
+  sifive_7,
+  generic_ooo
 };
 extern enum riscv_microarchitecture_type riscv_microarchitecture;
 
diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc
index 480f3124496..0d7bc2ae9b6 100644
--- a/gcc/config/riscv/riscv.cc
+++ b/gcc/config/riscv/riscv.cc
@@ -349,6 +349,21 @@  static const struct riscv_tune_param thead_c906_tune_info = {
   false		/* use_divmod_expansion */
 };
 
+/* Costs to use when optimizing for a generic ooo profile.  */
+static const struct riscv_tune_param generic_ooo_tune_info = {
+  {COSTS_N_INSNS (2), COSTS_N_INSNS (2)},	/* fp_add */
+  {COSTS_N_INSNS (5), COSTS_N_INSNS (6)},	/* fp_mul */
+  {COSTS_N_INSNS (7), COSTS_N_INSNS (8)},	/* fp_div */
+  {COSTS_N_INSNS (2), COSTS_N_INSNS (2)},	/* int_mul */
+  {COSTS_N_INSNS (6), COSTS_N_INSNS (6)},	/* int_div */
+  1,						/* issue_rate */
+  3,						/* branch_cost */
+  4,						/* memory_cost */
+  4,						/* fmv_cost */
+  false,					/* slow_unaligned_access */
+  false,					/* use_divmod_expansion */
+};
+
 /* Costs to use when optimizing for size.  */
 static const struct riscv_tune_param optimize_size_tune_info = {
   {COSTS_N_INSNS (1), COSTS_N_INSNS (1)},	/* fp_add */
@@ -6744,6 +6759,75 @@  riscv_sched_variable_issue (FILE *, int, rtx_insn *insn, int more)
   return more - 1;
 }
 
+/* Adjust the cost/latency of instructions for scheduling.
+   For now this is just used to change the latency of vector instructions
+   according to their LMUL.  We assume that an insn with LMUL == 8 requires
+   eight times more execution cycles than the same insn with LMUL == 1.
+   As this may cause very high latencies which lead to scheduling artifacts
+   we currently only perform the adjustment when -madjust-lmul-cost is given.
+   */
+static int
+riscv_sched_adjust_cost (rtx_insn *, int, rtx_insn *insn, int cost,
+			 unsigned int)
+{
+  /* Only do adjustments for the generic out-of-order scheduling model.  */
+  if (!TARGET_VECTOR || riscv_microarchitecture != generic_ooo)
+    return cost;
+
+  if (recog_memoized (insn) < 0)
+    return cost;
+
+  enum attr_type type = get_attr_type (insn);
+
+  if (type == TYPE_VFREDO || type == TYPE_VFWREDO)
+    {
+      /* TODO: For ordered reductions scale the base cost relative to the
+	 number of units.  */
+      ;
+    }
+
+  /* Don't do any LMUL-based latency adjustment unless explicitly asked to.  */
+  if (!TARGET_ADJUST_LMUL_COST)
+    return cost;
+
+  /* vsetvl has a vlmul attribute but its latency does not depend on it.  */
+  if (type == TYPE_VSETVL || type == TYPE_VSETVL_PRE)
+    return cost;
+
+  enum riscv_vector::vlmul_type lmul =
+    (riscv_vector::vlmul_type)get_attr_vlmul (insn);
+
+  double factor = 1;
+  switch (lmul)
+    {
+    case riscv_vector::LMUL_2:
+      factor = 2;
+      break;
+    case riscv_vector::LMUL_4:
+      factor = 4;
+      break;
+    case riscv_vector::LMUL_8:
+      factor = 8;
+      break;
+    case riscv_vector::LMUL_F2:
+      factor = 0.5;
+      break;
+    case riscv_vector::LMUL_F4:
+      factor = 0.25;
+      break;
+    case riscv_vector::LMUL_F8:
+      factor = 0.125;
+      break;
+    default:
+      factor = 1;
+    }
+
+  /* If the latency was nonzero, keep it that way.  */
+  int new_cost = MAX (cost > 0 ? 1 : 0, cost * factor);
+
+  return new_cost;
+}
+
 /* Auxiliary function to emit RISC-V ELF attribute. */
 static void
 riscv_emit_attribute ()
@@ -8469,6 +8553,9 @@  riscv_frame_pointer_required (void)
 #undef  TARGET_SCHED_VARIABLE_ISSUE
 #define TARGET_SCHED_VARIABLE_ISSUE riscv_sched_variable_issue
 
+#undef  TARGET_SCHED_ADJUST_COST
+#define TARGET_SCHED_ADJUST_COST riscv_sched_adjust_cost
+
 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
 #define TARGET_FUNCTION_OK_FOR_SIBCALL riscv_function_ok_for_sibcall
 
diff --git a/gcc/config/riscv/riscv.md b/gcc/config/riscv/riscv.md
index b456fa6abb3..2a9bad0d8ee 100644
--- a/gcc/config/riscv/riscv.md
+++ b/gcc/config/riscv/riscv.md
@@ -406,7 +406,7 @@  (define_attr "type"
    mtc,mfc,const,arith,logical,shift,slt,imul,idiv,move,fmove,fadd,fmul,
    fmadd,fdiv,fcmp,fcvt,fsqrt,multi,auipc,sfb_alu,nop,ghost,bitmanip,rotate,
    clmul,min,max,minu,maxu,clz,ctz,cpop,
-   atomic,condmove,crypto,rdvlenb,rdvl,wrvxrm,wrfrm,rdfrm,vsetvl,
+   atomic,condmove,crypto,rdvlenb,rdvl,wrvxrm,wrfrm,rdfrm,vsetvl,vsetvl_pre,
    vlde,vste,vldm,vstm,vlds,vsts,
    vldux,vldox,vstux,vstox,vldff,vldr,vstr,
    vlsegde,vssegte,vlsegds,vssegts,vlsegdux,vlsegdox,vssegtux,vssegtox,vlsegdff,
@@ -534,7 +534,7 @@  (define_attr "cannot_copy" "no,yes" (const_string "no"))
 ;; Microarchitectures we know how to tune for.
 ;; Keep this in sync with enum riscv_microarchitecture.
 (define_attr "tune"
-  "generic,sifive_7"
+  "generic,sifive_7,generic_ooo"
   (const (symbol_ref "((enum attr_tune) riscv_microarchitecture)")))
 
 ;; Describe a user's asm statement.
@@ -3328,5 +3328,6 @@  (define_expand "msubhisi4"
 (include "generic.md")
 (include "sifive-7.md")
 (include "thead.md")
+(include "generic-ooo.md")
 (include "vector.md")
 (include "zicond.md")
diff --git a/gcc/config/riscv/riscv.opt b/gcc/config/riscv/riscv.opt
index a962ea8f9d4..724a5429e5d 100644
--- a/gcc/config/riscv/riscv.opt
+++ b/gcc/config/riscv/riscv.opt
@@ -311,3 +311,6 @@  Enum(riscv_autovec_lmul) String(m8) Value(RVV_M8)
 -param=riscv-autovec-lmul=
 Target RejectNegative Joined Enum(riscv_autovec_lmul) Var(riscv_autovec_lmul) Init(RVV_M1)
 -param=riscv-autovec-lmul=<string>	Set the RVV LMUL of auto-vectorization in the RISC-V port.
+
+madjust-lmul-cost
+Target Var(TARGET_ADJUST_LMUL_COST) Init(0)
diff --git a/gcc/config/riscv/vector.md b/gcc/config/riscv/vector.md
index 6ceae25dbed..9c61d0ead8a 100644
--- a/gcc/config/riscv/vector.md
+++ b/gcc/config/riscv/vector.md
@@ -935,7 +935,9 @@  (define_insn "@vlmax_avl<mode>"
   [(set (match_operand:P 0 "register_operand" "=r")
 	(unspec:P [(match_operand:P 1 "const_int_operand" "i")] UNSPEC_VLMAX))]
   "TARGET_VECTOR"
-  "")
+  ""
+  [(set_attr "type" "vsetvl_pre")]
+  )
 
 ;; Set VXRM
 (define_insn "vxrmsi"