RISC-V: Add VLS BOOL mode vcond_mask[PR111751]

Richard patch resolve PR111751: https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=7c76c876e917a1f20a788f602cc78fff7d0a2a65

which cause ICE in RISC-V regression:

FAIL: gcc.dg/torture/pr53144.c   -O2  (internal compiler error: in gimple_expand_vec_cond_expr, at gimple-isel.cc:328)
FAIL: gcc.dg/torture/pr53144.c   -O2  (test for excess errors)
FAIL: gcc.dg/torture/pr53144.c   -O2 -flto -fno-use-linker-plugin -flto-partition=none  (internal compiler error: in gimple_expand_vec_cond_expr, at gimple-isel.cc:328)
FAIL: gcc.dg/torture/pr53144.c   -O2 -flto -fno-use-linker-plugin -flto-partition=none  (test for excess errors)
FAIL: gcc.dg/torture/pr53144.c   -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions  (internal compiler error: in gimple_expand_vec_cond_expr, at gimple-isel.cc:328)
FAIL: gcc.dg/torture/pr53144.c   -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions  (test for excess errors)
FAIL: gcc.dg/torture/pr53144.c   -O3 -g  (internal compiler error: in gimple_expand_vec_cond_expr, at gimple-isel.cc:328)
FAIL: gcc.dg/torture/pr53144.c   -O3 -g  (test for excess errors)

VLS BOOL modes vcond_mask is needed to fix this regression ICE.

More details: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111751

Tested and Committed.

PR target/111751

gcc/ChangeLog:

* config/riscv/autovec.md: Add VLS BOOL modes.

tree-optimization/111751 - support 1024 bit vector constant reinterpretation

The following ups the limit in fold_view_convert_expr to handle
1024bit vectors as used by GCN and RVV.  It also robustifies
the handling in visit_reference_op_load to properly give up when
constants cannot be re-interpreted.

PR tree-optimization/111751
* fold-const.cc (fold_view_convert_expr): Up the buffer size
to 128 bytes.
* tree-ssa-sccvn.cc (visit_reference_op_load): Special case
constants, giving up when re-interpretation to the target type
fails.

ada: Fix internal error on too large representation clause for small component

This is a small bug present on strict-alignment platforms for questionable
representation clauses.

gcc/ada/

* gcc-interface/decl.cc (inline_status_for_subprog): Minor tweak.
(gnat_to_gnu_field): Try harder to get a packable form of the type
for a bitfield.

ada: Tweak internal subprogram in Ada.Directories

The purpose of this patch is to work around false-positive warnings
emitted by GNAT SAS (also known as CodePeer). It does not change
the behavior of the modified subprogram.

gcc/ada/

* libgnat/a-direct.adb (Start_Search_Internal): Tweak subprogram
body.

ada: Remove superfluous setter procedure

It is only called once.

gcc/ada/

* sem_util.ads (Set_Scope_Is_Transient): Delete.
* sem_util.adb (Set_Scope_Is_Transient): Likewise.
* exp_ch7.adb (Create_Transient_Scope): Set Is_Transient directly.

ada: Fix bad finalization of limited aggregate in conditional expression

This happens when the conditional expression is immediately returned, for
example in an expression function.

gcc/ada/

* exp_aggr.adb (Is_Build_In_Place_Aggregate_Return): Return true
if the aggregate is a dependent expression of a conditional
expression being returned from a build-in-place function.

ada: Fix infinite loop with multiple limited with clauses

This occurs when one of the types has an incomplete declaration in addition
to its full declaration in its package. In this case AI05-129 says that the
incomplete type is not part of the limited view of the package, i.e. only
the full view is. Now, in the GNAT implementation, it's the opposite in the
regular view of the package, i.e. the incomplete type is the visible one.

That's why the implementation needs to also swap the types on the visibility
chain while it is swapping the views when the clauses are either installed
or removed. This works correctly for the installation, but does not for the
removal, so this change rewrites the code doing the latter.

gcc/ada/
PR ada/111434
* sem_ch10.adb (Replace): New procedure to replace an entity with
another on the homonym chain.
(Install_Limited_With_Clause): Rename Non_Lim_View to Typ for the
sake of consistency.  Call Replace to do the replacements and split
the code into the regular and the special cases.  Add debuggging
output controlled by -gnatdi.
(Install_With_Clause): Print the Parent_With and Implicit_With flags
in the debugging output controlled by -gnatdi.
(Remove_Limited_With_Unit.Restore_Chain_For_Shadow (Shadow)): Rewrite
using a direct replacement of E4 by E2.   Call Replace to do the
replacements.  Add debuggging output controlled by -gnatdi.

ada: Fix filesystem entry filtering

This patch fixes the behavior of Ada.Directories.Search when being
requested to filter out regular files or directories. One of the
configurations in which that behavior was incorrect was that when the
caller requested only the regular and special files but not the
directories, the directories would still be returned.

gcc/ada/

* libgnat/a-direct.adb: Fix filesystem entry filtering.

ada: Tweak documentation comments

The concept of extended nodes was retired at the same time Gen_IL
was introduced, but there was a reference to that concept left over
in a comment. This patch removes that reference.

Also, the description of the field Comes_From_Check_Or_Contract was
incorrectly placed in a section for fields present in all nodes in
sinfo.ads. This patch fixes this.

gcc/ada/

* atree.ads, nlists.ads, types.ads: Remove references to extended
nodes. Fix typo.
* sinfo.ads: Likewise and fix position of
Comes_From_Check_Or_Contract description.

ada: Crash processing pragmas Compile_Time_Error and Compile_Time_Warning

gcc/ada/

* sem_attr.adb (Analyze_Attribute): Protect the frontend against
replacing 'Size by its static value if 'Size is not known at
compile time and we are processing pragmas Compile_Time_Warning or
Compile_Time_Errors.

RISC-V: Add testcase for SCCVN optimization[PR111751]

Add testcase for PR111751 which has been fixed:
https://gcc.gnu.org/pipermail/gcc-patches/2023-October/632474.html

PR target/111751

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/autovec/pr111751.c: New test.

Fix missed CSE with a BLKmode entity

The following fixes fallout of r10-7145-g1dc00a8ec9aeba which made
us cautionous about CSEing a load to an object that has padding bits.
The added check also triggers for BLKmode entities like STRING_CSTs
but by definition a BLKmode entity does not have padding bits.

PR tree-optimization/111751
* tree-ssa-sccvn.cc (visit_reference_op_load): Exempt
BLKmode result from the padding bits check.

RISC-V Regression: Fix FAIL of bb-slp-pr65935.c for RVV

Here is the reference comparing dump IR between ARM SVE and RVV.

https://godbolt.org/z/zqess8Gss

We can see RVV has one more dump IR:
optimized: basic block part vectorized using 128 byte vectors
since RVV has 1024 bit vectors.

The codegen is reasonable good.

However, I saw GCN also has 1024 bit vector.
This patch may cause this case FAIL in GCN port ?

Hi, GCN folk, could you check this patch in GCN port for me ?

gcc/testsuite/ChangeLog:

* gcc.dg/vect/bb-slp-pr65935.c: Add vect1024 variant.
* lib/target-supports.exp: Ditto.

arc: Refurbish add.f combiner patterns

Refurbish add compare patterns: use 'r' constraint, fix identation,
and fix pattern to match 'if (a+b) { ... }' constructions.

gcc/

* config/arc/arc.cc (arc_select_cc_mode): Match NEG code with
the first operand.
* config/arc/arc.md (addsi_compare): Make pattern canonical.
(addsi_compare_2): Fix identation, constraint letters.
(addsi_compare_3): Likewise.

gcc/testsuite/

* gcc.target/arc/add_f-combine.c: New test.

Signed-off-by: Claudiu Zissulescu <claziss@gmail.com>

RISC-V: Add available vector size for RVV

For RVV, we have VLS modes enable according to TARGET_MIN_VLEN
from M1 to M8.

For example, when TARGET_MIN_VLEN = 128 bits, we enable
128/256/512/1024 bits VLS modes.

This patch fixes following FAIL:
FAIL: gcc.dg/vect/bb-slp-subgroups-2.c -flto -ffat-lto-objects  scan-tree-dump-times slp2 "optimized: basic block" 2
FAIL: gcc.dg/vect/bb-slp-subgroups-2.c scan-tree-dump-times slp2 "optimized: basic block" 2

gcc/testsuite/ChangeLog:

* lib/target-supports.exp: Add 256/512/1024

Daily bump.

Fixes for profile count/probability maintenance

Verifier checks have recently been strengthened to check that
all counts and probabilities are initialized. The checks fired
during autoprofiledbootstrap build and this patch fixes it.

Tested on x86_64-pc-linux-gnu.

gcc/ChangeLog:
* auto-profile.cc (afdo_calculate_branch_prob): Fix count comparisons
* tree-vect-loop-manip.cc (vect_do_peeling): Guard against zero count
when scaling loop profile

analyzer: fix build with gcc < 6

gcc/analyzer/ChangeLog:
* access-diagram.cc (boundaries::add): Explicitly state
"boundaries::" scope for "kind" enum.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

Ensure float equivalences include + and - zero.

A floating point equivalence may not properly reflect both signs of
zero, so be pessimsitic and ensure both signs are included.

PR tree-optimization/111694
gcc/
* gimple-range-cache.cc (ranger_cache::fill_block_cache): Adjust
equivalence range.
* value-relation.cc (adjust_equivalence_range): New.
* value-relation.h (adjust_equivalence_range): New prototype.

gcc/testsuite/
* gcc.dg/pr111694.c: New.

Remove unused get_identity_relation.

Turns out we didnt need this as there is no unordered relations
managed by the oracle.

* gimple-range-gori.cc (gori_compute::compute_operand1_range): Do
not call get_identity_relation.
(gori_compute::compute_operand2_range): Ditto.
* value-relation.cc (get_identity_relation): Remove.
* value-relation.h (get_identity_relation): Remove protyotype.

RISC-V Regression test: Fix slp-perm-4.c FAIL for RVV

RVV vectorize it with stride5 load_lanes.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/slp-perm-4.c: Adapt test for stride5 load_lanes.

RISC-V Regression tests: Fix FAIL of pr97832* for RVV

These cases are vectorized by vec_load_lanes with strided = 8 instead of SLP
with -fno-vect-cost-model.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/pr97832-2.c: Adapt dump check for target supports load_lanes with stride = 8.
* gcc.dg/vect/pr97832-3.c: Ditto.
* gcc.dg/vect/pr97832-4.c: Ditto.

RISC-V Regression test: Fix FAIL of slp-12a.c

This case is vectorized by stride8 load_lanes.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/slp-12a.c: Adapt for stride 8 load_lanes.

RISC-V Regression test: Fix FAIL of slp-reduc-4.c for RVV

RVV vectortizes this case with stride8 load_lanes.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/slp-reduc-4.c: Adapt test for stride8 load_lanes.

RISC-V Regression test: Adapt SLP tests like ARM SVE

Like ARM SVE, RVV is vectorizing these 2 cases in the same way.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/slp-23.c: Add RVV like ARM SVE.
* gcc.dg/vect/slp-perm-10.c: Ditto.

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

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 would look like.

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.

In the future we might also want a latency adjustment similar to lmul
for reductions, i.e. make the latency dependent on the type and its
number of units.

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.

RISC-V: Support movmisalign of RVV VLA modes

This patch fixed these following FAILs in regressions:
FAIL: gcc.dg/vect/slp-perm-11.c -flto -ffat-lto-objects  scan-tree-dump-times vect "vectorizing stmts using SLP" 1
FAIL: gcc.dg/vect/slp-perm-11.c scan-tree-dump-times vect "vectorizing stmts using SLP" 1
FAIL: gcc.dg/vect/vect-bitfield-read-2.c -flto -ffat-lto-objects  scan-tree-dump-not optimized "Invalid sum"
FAIL: gcc.dg/vect/vect-bitfield-read-2.c scan-tree-dump-not optimized "Invalid sum"
FAIL: gcc.dg/vect/vect-bitfield-read-4.c -flto -ffat-lto-objects  scan-tree-dump-not optimized "Invalid sum"
FAIL: gcc.dg/vect/vect-bitfield-read-4.c scan-tree-dump-not optimized "Invalid sum"
FAIL: gcc.dg/vect/vect-bitfield-write-2.c -flto -ffat-lto-objects  scan-tree-dump-not optimized "Invalid sum"
FAIL: gcc.dg/vect/vect-bitfield-write-2.c scan-tree-dump-not optimized "Invalid sum"
FAIL: gcc.dg/vect/vect-bitfield-write-3.c -flto -ffat-lto-objects  scan-tree-dump-not optimized "Invalid sum"
FAIL: gcc.dg/vect/vect-bitfield-write-3.c scan-tree-dump-not optimized "Invalid sum"

Previously, I removed the movmisalign pattern to fix the execution FAILs in this commit:
https://github.com/gcc-mirror/gcc/commit/f7bff24905a6959f85f866390db2fff1d6f95520

I was thinking that RVV doesn't allow misaligned at the beginning so I removed that pattern.
However, after deep investigation && reading RVV ISA again and experiment on SPIKE,
I realized I was wrong.

RVV ISA reference: https://github.com/riscv/riscv-v-spec/blob/master/v-spec.adoc#vector-memory-alignment-constraints

"If an element accessed by a vector memory instruction is not naturally aligned to the size of the element,
 either the element is transferred successfully or an address misaligned exception is raised on that element."

It's obvious that RVV ISA does allow misaligned vector load/store.

And experiment and confirm on SPIKE:

[jzzhong@rios-cad122:/work/home/jzzhong/work/toolchain/riscv/gcc/gcc/testsuite/gcc.dg/vect]$~/work/toolchain/riscv/build/dev-rv64gcv_zfh-lp64d-medany-newlib-spike-debug/install/bin/spike --isa=rv64gcv --varch=vlen:128,elen:64 ~/work/toolchain/riscv/build/dev-rv64gcv_zfh-lp64d-medany-newlib-spike-debug/install/riscv64-unknown-elf/bin/pk64  a.out
bbl loader
z  0000000000000000 ra 0000000000010158 sp 0000003ffffffb40 gp 0000000000012c48
tp 0000000000000000 t0 00000000000110da t1 000000000000000f t2 0000000000000000
s0 0000000000013460 s1 0000000000000000 a0 0000000000012ef5 a1 0000000000012018
a2 0000000000012a71 a3 000000000000000d a4 0000000000000004 a5 0000000000012a71
a6 0000000000012a71 a7 0000000000012018 s2 0000000000000000 s3 0000000000000000
s4 0000000000000000 s5 0000000000000000 s6 0000000000000000 s7 0000000000000000
s8 0000000000000000 s9 0000000000000000 sA 0000000000000000 sB 0000000000000000
t3 0000000000000000 t4 0000000000000000 t5 0000000000000000 t6 0000000000000000
pc 0000000000010258 va/inst 00000000020660a7 sr 8000000200006620
Store/AMO access fault!

[jzzhong@rios-cad122:/work/home/jzzhong/work/toolchain/riscv/gcc/gcc/testsuite/gcc.dg/vect]$~/work/toolchain/riscv/build/dev-rv64gcv_zfh-lp64d-medany-newlib-spike-debug/install/bin/spike --misaligned --isa=rv64gcv --varch=vlen:128,elen:64 ~/work/toolchain/riscv/build/dev-rv64gcv_zfh-lp64d-medany-newlib-spike-debug/install/riscv64-unknown-elf/bin/pk64  a.out
bbl loader

We can see SPIKE can pass previous *FAILED* execution tests with specifying --misaligned to SPIKE.

So, to honor RVV ISA SPEC, we should add movmisalign pattern back base on the investigations I have done since
it can improve multiple vectorization tests and fix dumple FAILs.

This patch adds TARGET_VECTOR_MISALIGN_SUPPORTED to decide whether we support misalign pattern for VLA modes (By default it is enabled).

Consider this following case:

struct s {
    unsigned i : 31;
    char a : 4;
};

#define N 32
#define ELT0 {0x7FFFFFFFUL, 0}
#define ELT1 {0x7FFFFFFFUL, 1}
#define ELT2 {0x7FFFFFFFUL, 2}
#define ELT3 {0x7FFFFFFFUL, 3}
#define RES 48
struct s A[N]
  = { ELT0, ELT1, ELT2, ELT3, ELT0, ELT1, ELT2, ELT3,
      ELT0, ELT1, ELT2, ELT3, ELT0, ELT1, ELT2, ELT3,
      ELT0, ELT1, ELT2, ELT3, ELT0, ELT1, ELT2, ELT3,
      ELT0, ELT1, ELT2, ELT3, ELT0, ELT1, ELT2, ELT3};

int __attribute__ ((noipa))
f(struct s *ptr, unsigned n) {
    int res = 0;
    for (int i = 0; i < n; ++i)
      res += ptr[i].a;
    return res;
}

-O3 -S -fno-vect-cost-model (default strict-align):

f:
mv a4,a0
beq a1,zero,.L9
addiw a5,a1,-1
li a3,14
vsetivli zero,16,e64,m8,ta,ma
bleu a5,a3,.L3
andi a5,a0,127
bne a5,zero,.L3
srliw a3,a1,4
slli a3,a3,7
li a0,15
slli a0,a0,32
add a3,a3,a4
mv a5,a4
li a2,32
vmv.v.x v16,a0
vsetvli zero,zero,e32,m4,ta,ma
vmv.v.i v4,0
.L4:
vsetvli zero,zero,e64,m8,ta,ma
vle64.v v8,0(a5)
addi a5,a5,128
vand.vv v8,v8,v16
vsetvli zero,zero,e32,m4,ta,ma
vnsrl.wx v8,v8,a2
vadd.vv v4,v4,v8
bne a5,a3,.L4
li a3,0
andi a5,a1,15
vmv.s.x v1,a3
andi a3,a1,-16
vredsum.vs v1,v4,v1
vmv.x.s a0,v1
mv a2,a0
beq a5,zero,.L15
slli a5,a3,3
add a5,a4,a5
lw a0,4(a5)
andi a0,a0,15
addiw a4,a3,1
addw a0,a0,a2
bgeu a4,a1,.L15
lw a2,12(a5)
andi a2,a2,15
addiw a4,a3,2
addw a0,a2,a0
bgeu a4,a1,.L15
lw a2,20(a5)
andi a2,a2,15
addiw a4,a3,3
addw a0,a2,a0
bgeu a4,a1,.L15
lw a2,28(a5)
andi a2,a2,15
addiw a4,a3,4
addw a0,a2,a0
bgeu a4,a1,.L15
lw a2,36(a5)
andi a2,a2,15
addiw a4,a3,5
addw a0,a2,a0
bgeu a4,a1,.L15
lw a2,44(a5)
andi a2,a2,15
addiw a4,a3,6
addw a0,a2,a0
bgeu a4,a1,.L15
lw a2,52(a5)
andi a2,a2,15
addiw a4,a3,7
addw a0,a2,a0
bgeu a4,a1,.L15
lw a4,60(a5)
andi a4,a4,15
addw a4,a4,a0
addiw a2,a3,8
mv a0,a4
bgeu a2,a1,.L15
lw a0,68(a5)
andi a0,a0,15
addiw a2,a3,9
addw a0,a0,a4
bgeu a2,a1,.L15
lw a2,76(a5)
andi a2,a2,15
addiw a4,a3,10
addw a0,a2,a0
bgeu a4,a1,.L15
lw a2,84(a5)
andi a2,a2,15
addiw a4,a3,11
addw a0,a2,a0
bgeu a4,a1,.L15
lw a2,92(a5)
andi a2,a2,15
addiw a4,a3,12
addw a0,a2,a0
bgeu a4,a1,.L15
lw a2,100(a5)
andi a2,a2,15
addiw a4,a3,13
addw a0,a2,a0
bgeu a4,a1,.L15
lw a4,108(a5)
andi a4,a4,15
addiw a3,a3,14
addw a0,a4,a0
bgeu a3,a1,.L15
lw a5,116(a5)
andi a5,a5,15
addw a0,a5,a0
ret
.L9:
li a0,0
.L15:
ret
.L3:
mv a5,a4
slli a4,a1,32
srli a1,a4,29
add a1,a5,a1
li a0,0
.L7:
lw a4,4(a5)
andi a4,a4,15
addi a5,a5,8
addw a0,a4,a0
bne a5,a1,.L7
ret

-O3 -S -mno-strict-align -fno-vect-cost-model:

f:
beq a1,zero,.L4
slli a1,a1,32
li a5,15
vsetvli a4,zero,e64,m1,ta,ma
slli a5,a5,32
srli a1,a1,32
li a6,32
vmv.v.x v3,a5
vsetvli zero,zero,e32,mf2,ta,ma
vmv.v.i v2,0
.L3:
vsetvli a5,a1,e64,m1,ta,ma
vle64.v v1,0(a0)
vsetvli a3,zero,e64,m1,ta,ma
slli a2,a5,3
vand.vv v1,v1,v3
sub a1,a1,a5
vsetvli zero,zero,e32,mf2,ta,ma
add a0,a0,a2
vnsrl.wx v1,v1,a6
vsetvli zero,a5,e32,mf2,tu,ma
vadd.vv v2,v2,v1
bne a1,zero,.L3
li a5,0
vsetvli a3,zero,e32,mf2,ta,ma
vmv.s.x v1,a5
vredsum.vs v2,v2,v1
vmv.x.s a0,v2
ret
.L4:
li a0,0
ret

We can see it improves this case codegen a lot.

gcc/ChangeLog:

* config/riscv/riscv-opts.h (TARGET_VECTOR_MISALIGN_SUPPORTED): New macro.
* config/riscv/riscv.cc (riscv_support_vector_misalignment): Depend on movmisalign pattern.
* config/riscv/vector.md (movmisalign<mode>): New pattern.

THead: Fix missing CFI directives for th.sdd in prologue.

When generating CFI directives for the store-pair instruction,
if we add two parallel REG_FRAME_RELATED_EXPR expr_lists like
  (expr_list:REG_FRAME_RELATED_EXPR (set (mem/c:DI (plus:DI (reg/f:DI 2 sp)
    (const_int 8 [0x8])) [1  S8 A64])
    (reg:DI 1 ra))
  (expr_list:REG_FRAME_RELATED_EXPR (set (mem/c:DI (reg/f:DI 2 sp) [1  S8 A64])
    (reg:DI 8 s0))
only the first expr_list will be recognized by dwarf2out_frame_debug
funciton. So, here we generate a SEQUENCE expression of REG_FRAME_RELATED_EXPR,
which includes two sub-expressions of RTX_FRAME_RELATED_P. Then the
dwarf2out_frame_debug_expr function will iterate through all the sub-expressions
and generate the corresponding CFI directives.

gcc/
* config/riscv/thead.cc (th_mempair_save_regs): Fix missing CFI
directives for store-pair instruction.

gcc/testsuite/
* gcc.target/riscv/xtheadmempair-4.c: New test.

tree-optimization/111715 - improve TBAA for access paths with pun

The following improves basic TBAA for access paths formed by
C++ abstraction where we are able to combine a path from an
address-taking operation with a path based on that access using
a pun to avoid memory access semantics on the address-taking part.

The trick is to identify the point the semantic memory access path
starts which allows us to use the alias set of the outermost access
instead of only that of the base of this path.

PR tree-optimization/111715
* alias.cc (reference_alias_ptr_type_1): When we have
a type-punning ref at the base search for the access
path part that's still semantically valid.

* gcc.dg/tree-ssa/ssa-fre-102.c: New testcase.

RISC-V: Refine bswap16 auto vectorization code gen

Update in v2

* Remove emit helper functions.
* Take expand_binop instead.

Original log:

This patch would like to refine the code gen for the bswap16.

We will have VEC_PERM_EXPR after rtl expand when invoking
__builtin_bswap. It will generate about 9 instructions in
loop as below, no matter it is bswap16, bswap32 or bswap64.

  .L2:
1 vle16.v v4,0(a0)
2 vmv.v.x v2,a7
3 vand.vv v2,v6,v2
4 slli    a2,a5,1
5 vrgatherei16.vv v1,v4,v2
6 sub     a4,a4,a5
7 vse16.v v1,0(a3)
8 add     a0,a0,a2
9 add     a3,a3,a2
  bne     a4,zero,.L2

But for bswap16 we may have a even simple code gen, which
has only 7 instructions in loop as below.

  .L5
1 vle8.v  v2,0(a5)
2 addi    a5,a5,32
3 vsrl.vi v4,v2,8
4 vsll.vi v2,v2,8
5 vor.vv  v4,v4,v2
6 vse8.v  v4,0(a4)
7 addi    a4,a4,32
  bne     a5,a6,.L5

Unfortunately, this way will make the insn in loop will grow up to
13 and 24 for bswap32 and bswap64. Thus, we will refine the code
gen for the bswap16 only, and leave both the bswap32 and bswap64
as is.

gcc/ChangeLog:

* config/riscv/riscv-v.cc (shuffle_bswap_pattern): New func impl
for shuffle bswap.
(expand_vec_perm_const_1): Add handling for shuffle bswap pattern.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/autovec/vls/perm-4.c: Adjust checker.
* gcc.target/riscv/rvv/autovec/unop/bswap16-0.c: New test.
* gcc.target/riscv/rvv/autovec/unop/bswap16-run-0.c: New test.
* gcc.target/riscv/rvv/autovec/vls/bswap16-0.c: New test.

Signed-off-by: Pan Li <pan2.li@intel.com>

RISC-V Regression test: Fix FAIL of pr45752.c for RVV

RVV use load_lanes with stride = 5 vectorize this case with -fno-vect-cost-model
instead of SLP.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/pr45752.c: Adapt dump check for target supports load_lanes with stride = 5.

testsuite: Fix vect_cond_arith_* dump checks for RVV.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/vect-cond-arith-2.c: Also match COND_LEN.
* gcc.dg/vect/vect-cond-arith-4.c: Ditto.
* gcc.dg/vect/vect-cond-arith-5.c: Ditto.
* gcc.dg/vect/vect-cond-arith-6.c: Ditto.

RISC-V Regression test: Fix FAIL of fast-math-slp-38.c for RVV

Reference: https://godbolt.org/z/G9jzf5Grh

RVV is able to vectorize this case using SLP. However, with -fno-vect-cost-model,
RVV vectorize it by vec_load_lanes with stride 6.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/fast-math-slp-38.c: Add ! vect_strided6.

i386: Implement doubleword right shifts by 1 bit using s[ha]r+rcr.

This patch tweaks the i386 back-end's ix86_split_ashr and ix86_split_lshr
functions to implement doubleword right shifts by 1 bit, using a shift
of the highpart that sets the carry flag followed by a rotate-carry-right
(RCR) instruction on the lowpart.

Conceptually this is similar to the recent left shift patch, but with two
complicating factors.  The first is that although the RCR sequence is
shorter, and is a ~3x performance improvement on AMD, my microbenchmarking
shows it ~10% slower on Intel.  Hence this patch also introduces a new
X86_TUNE_USE_RCR tuning parameter.  The second is that I believe this is
the first time a "rotate-right-through-carry" and a right shift that sets
the carry flag from the least significant bit has been modelled in GCC RTL
(on a MODE_CC target).  For this I've used the i386 back-end's UNSPEC_CC_NE
which seems appropriate.  Finally rcrsi2 and rcrdi2 are separate
define_insns so that we can use their generator functions.

For the pair of functions:
unsigned __int128 foo(unsigned __int128 x) { return x >> 1; }
__int128 bar(__int128 x) { return x >> 1; }

with -O2 -march=znver4 we previously generated:

foo: movq    %rdi, %rax
        movq    %rsi, %rdx
        shrdq   $1, %rsi, %rax
        shrq    %rdx
        ret
bar: movq    %rdi, %rax
        movq    %rsi, %rdx
        shrdq   $1, %rsi, %rax
        sarq    %rdx
        ret

with this patch we now generate:

foo: movq    %rsi, %rdx
        movq    %rdi, %rax
        shrq    %rdx
        rcrq    %rax
        ret
bar: movq    %rsi, %rdx
        movq    %rdi, %rax
        sarq    %rdx
        rcrq    %rax
        ret

2023-10-09  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* config/i386/i386-expand.cc (ix86_split_ashr): Split shifts by
one into ashr[sd]i3_carry followed by rcr[sd]i2, if TARGET_USE_RCR
or -Oz.
(ix86_split_lshr): Likewise, split shifts by one bit into
lshr[sd]i3_carry followed by rcr[sd]i2, if TARGET_USE_RCR or -Oz.
* config/i386/i386.h (TARGET_USE_RCR): New backend macro.
* config/i386/i386.md (rcrsi2): New define_insn for rcrl.
(rcrdi2): New define_insn for rcrq.
(<anyshiftrt><mode>3_carry): New define_insn for right shifts that
set the carry flag from the least significant bit, modelled using
UNSPEC_CC_NE.
* config/i386/x86-tune.def (X86_TUNE_USE_RCR): New tuning parameter
controlling use of rcr 1 vs. shrd, which is significantly faster on
AMD processors.

gcc/testsuite/ChangeLog
* gcc.target/i386/rcr-1.c: New 64-bit test case.
* gcc.target/i386/rcr-2.c: New 32-bit test case.

Allow -mno-evex512 usage

gcc/ChangeLog:

* config/i386/i386.opt: Allow -mno-evex512.

gcc/testsuite/ChangeLog:

* gcc.target/i386/noevex512-1.c: New test.
* gcc.target/i386/noevex512-2.c: Ditto.
* gcc.target/i386/noevex512-3.c: Ditto.

Support -mevex512 for AVX512FP16 intrins

gcc/ChangeLog:

* config/i386/sse.md (V48H_AVX512VL): Add TARGET_EVEX512.
(VFH): Ditto.
(VF2H): Ditto.
(VFH_AVX512VL): Ditto.
(VHFBF): Ditto.
(VHF_AVX512VL): Ditto.
(VI2H_AVX512VL): Ditto.
(VI2F_256_512): Ditto.
(VF48_I1248): Remove unused iterator.
(VF48H_AVX512VL): Add TARGET_EVEX512.
(VF_AVX512): Remove unused iterator.
(REDUC_PLUS_MODE): Add TARGET_EVEX512.
(REDUC_SMINMAX_MODE): Ditto.
(FMAMODEM): Ditto.
(VFH_SF_AVX512VL): Ditto.
(VEC_PERM_AVX2): Ditto.

Co-authored-by: Hu, Lin1 <lin1.hu@intel.com>

Support -mevex512 for AVX512{IFMA,VBMI,VNNI,BF16,VPOPCNTDQ,VBMI2,BITALG,VP2INTERSECT},VAES,GFNI,VPCLMULQDQ intrins

gcc/ChangeLog:

* config/i386/sse.md (VI1_AVX512VL): Add TARGET_EVEX512.
(VI8_FVL): Ditto.
(VI1_AVX512F): Ditto.
(VI1_AVX512VNNI): Ditto.
(VI1_AVX512VL_F): Ditto.
(VI12_VI48F_AVX512VL): Ditto.
(*avx512f_permvar_truncv32hiv32qi_1): Ditto.
(sdot_prod<mode>): Ditto.
(VEC_PERM_AVX2): Ditto.
(VPERMI2): Ditto.
(VPERMI2I): Ditto.
(vpmadd52<vpmadd52type>v8di): Ditto.
(usdot_prod<mode>): Ditto.
(vpdpbusd_v16si): Ditto.
(vpdpbusds_v16si): Ditto.
(vpdpwssd_v16si): Ditto.
(vpdpwssds_v16si): Ditto.
(VI48_AVX512VP2VL): Ditto.
(avx512vp2intersect_2intersectv16si): Ditto.
(VF_AVX512BF16VL): Ditto.
(VF1_AVX512_256): Ditto.

gcc/testsuite/ChangeLog:

* gcc.target/i386/pr90096.c: Adjust error message.

Co-authored-by: Hu, Lin1 <lin1.hu@intel.com>

Support -mevex512 for AVX512BW intrins

gcc/Changelog:

* config/i386/i386-expand.cc (ix86_expand_vector_init_duplicate):
Make sure there is EVEX512 enabled.
(ix86_expand_vecop_qihi2): Refuse V32QI->V32HI when no EVEX512.
* config/i386/i386.cc (ix86_hard_regno_mode_ok): Disable 64 bit mask
when !TARGET_EVEX512.
* config/i386/i386.md (avx512bw_512): New.
(SWI1248_AVX512BWDQ_64): Add TARGET_EVEX512.
(*zero_extendsidi2): Change isa to avx512bw_512.
(kmov_isa): Ditto.
(*anddi_1): Ditto.
(*andn<mode>_1): Change isa to kmov_isa.
(*<code><mode>_1): Ditto.
(*notxor<mode>_1): Ditto.
(*one_cmpl<mode>2_1): Ditto.
(*one_cmplsi2_1_zext): Change isa to avx512bw_512.
(*ashl<mode>3_1): Change isa to kmov_isa.
(*lshr<mode>3_1): Ditto.
* config/i386/sse.md (VI12HFBF_AVX512VL): Add TARGET_EVEX512.
(VI1248_AVX512VLBW): Ditto.
(VHFBF_AVX512VL): Ditto.
(VI): Ditto.
(VIHFBF): Ditto.
(VI_AVX2): Ditto.
(VI1_AVX512): Ditto.
(VI12_256_512_AVX512VL): Ditto.
(VI2_AVX2_AVX512BW): Ditto.
(VI2_AVX512VNNIBW): Ditto.
(VI2_AVX512VL): Ditto.
(VI2HFBF_AVX512VL): Ditto.
(VI8_AVX2_AVX512BW): Ditto.
(VIMAX_AVX2_AVX512BW): Ditto.
(VIMAX_AVX512VL): Ditto.
(VI12_AVX2_AVX512BW): Ditto.
(VI124_AVX2_24_AVX512F_1_AVX512BW): Ditto.
(VI248_AVX512VL): Ditto.
(VI248_AVX512VLBW): Ditto.
(VI248_AVX2_8_AVX512F_24_AVX512BW): Ditto.
(VI248_AVX512BW): Ditto.
(VI248_AVX512BW_AVX512VL): Ditto.
(VI248_512): Ditto.
(VI124_256_AVX512F_AVX512BW): Ditto.
(VI_AVX512BW): Ditto.
(VIHFBF_AVX512BW): Ditto.
(SWI1248_AVX512BWDQ): Ditto.
(SWI1248_AVX512BW): Ditto.
(SWI1248_AVX512BWDQ2): Ditto.
(*knotsi_1_zext): Ditto.
(define_split for zero_extend + not): Ditto.
(kunpckdi): Ditto.
(REDUC_SMINMAX_MODE): Ditto.
(VEC_EXTRACT_MODE): Ditto.
(*avx512bw_permvar_truncv16siv16hi_1): Ditto.
(*avx512bw_permvar_truncv16siv16hi_1_hf): Ditto.
(truncv32hiv32qi2): Ditto.
(avx512bw_<code>v32hiv32qi2): Ditto.
(avx512bw_<code>v32hiv32qi2_mask): Ditto.
(avx512bw_<code>v32hiv32qi2_mask_store): Ditto.
(usadv64qi): Ditto.
(VEC_PERM_AVX2): Ditto.
(AVX512ZEXTMASK): Ditto.
(SWI24_MASK): New.
(vec_pack_trunc_<mode>): Change iterator to SWI24_MASK.
(avx512bw_packsswb<mask_name>): Add TARGET_EVEX512.
(avx512bw_packssdw<mask_name>): Ditto.
(avx512bw_interleave_highv64qi<mask_name>): Ditto.
(avx512bw_interleave_lowv64qi<mask_name>): Ditto.
(<mask_codefor>avx512bw_pshuflwv32hi<mask_name>): Ditto.
(<mask_codefor>avx512bw_pshufhwv32hi<mask_name>): Ditto.
(vec_unpacks_lo_di): Ditto.
(SWI48x_MASK): New.
(vec_unpacks_hi_<mode>): Change iterator to SWI48x_MASK.
(avx512bw_umulhrswv32hi3<mask_name>): Add TARGET_EVEX512.
(VI1248_AVX512VL_AVX512BW): Ditto.
(avx512bw_<code>v32qiv32hi2<mask_name>): Ditto.
(*avx512bw_zero_extendv32qiv32hi2_1): Ditto.
(*avx512bw_zero_extendv32qiv32hi2_2): Ditto.
(<insn>v32qiv32hi2): Ditto.
(pbroadcast_evex_isa): Change isa attribute to avx512bw_512.
(VPERMI2): Add TARGET_EVEX512.
(VPERMI2I): Ditto.

Support -mevex512 for AVX512DQ intrins

gcc/ChangeLog:

* config/i386/i386-expand.cc (ix86_expand_sse2_mulvxdi3):
Add TARGET_EVEX512 for 512 bit usage.
* config/i386/i386.cc (standard_sse_constant_opcode): Ditto.
* config/i386/sse.md (VF1_VF2_AVX512DQ): Ditto.
(VF1_128_256VL): Ditto.
(VF2_AVX512VL): Ditto.
(VI8_256_512): Ditto.
(<mask_codefor>fixuns_trunc<mode><sseintvecmodelower>2<mask_name>):
Ditto.
(AVX512_VEC): Ditto.
(AVX512_VEC_2): Ditto.
(VI4F_BRCST32x2): Ditto.
(VI8F_BRCST64x2): Ditto.

Support -mevex512 for AVX512F intrins

gcc/ChangeLog:

* config/i386/i386-builtins.cc
(ix86_vectorize_builtin_gather): Disable 512 bit gather
when !TARGET_EVEX512.
* config/i386/i386-expand.cc (ix86_valid_mask_cmp_mode):
Add TARGET_EVEX512.
(ix86_expand_int_sse_cmp): Ditto.
(ix86_expand_vector_init_one_nonzero): Disable subroutine
when !TARGET_EVEX512.
(ix86_emit_swsqrtsf): Add TARGET_EVEX512.
(ix86_vectorize_vec_perm_const): Disable subroutine when
!TARGET_EVEX512.
* config/i386/i386.cc
(standard_sse_constant_p): Add TARGET_EVEX512.
(standard_sse_constant_opcode): Ditto.
(ix86_get_ssemov): Ditto.
(ix86_legitimate_constant_p): Ditto.
(ix86_vectorize_builtin_scatter): Diable 512 bit scatter
when !TARGET_EVEX512.
* config/i386/i386.md (avx512f_512): New.
(movxi): Add TARGET_EVEX512.
(*movxi_internal_avx512f): Ditto.
(*movdi_internal): Change alternative 12 to ?Yv. Adjust mode
for alternative 13.
(*movsi_internal): Change alternative 8 to ?Yv. Adjust mode for
alternative 9.
(*movhi_internal): Change alternative 11 to *Yv.
(*movdf_internal): Change alternative 12 to Yv.
(*movsf_internal): Change alternative 5 to Yv. Adjust mode for
alternative 5 and 6.
(*mov<mode>_internal): Change alternative 4 to Yv.
(define_split for convert SF to DF): Add TARGET_EVEX512.
(extendbfsf2_1): Ditto.
* config/i386/predicates.md (bcst_mem_operand): Disable predicate
for 512 bit when !TARGET_EVEX512.
* config/i386/sse.md (VMOVE): Add TARGET_EVEX512.
(V48_AVX512VL): Ditto.
(V48_256_512_AVX512VL): Ditto.
(V48H_AVX512VL): Ditto.
(VI12_AVX512VL): Ditto.
(V): Ditto.
(V_512): Ditto.
(V_256_512): Ditto.
(VF): Ditto.
(VF1_VF2_AVX512DQ): Ditto.
(VFH): Ditto.
(VFB): Ditto.
(VF1): Ditto.
(VF1_AVX2): Ditto.
(VF2): Ditto.
(VF2H): Ditto.
(VF2_512_256): Ditto.
(VF2_512_256VL): Ditto.
(VF_512): Ditto.
(VFB_512): Ditto.
(VI48_AVX512VL): Ditto.
(VI1248_AVX512VLBW): Ditto.
(VF_AVX512VL): Ditto.
(VFH_AVX512VL): Ditto.
(VF1_AVX512VL): Ditto.
(VI): Ditto.
(VIHFBF): Ditto.
(VI_AVX2): Ditto.
(VI8): Ditto.
(VI8_AVX512VL): Ditto.
(VI2_AVX512F): Ditto.
(VI4_AVX512F): Ditto.
(VI4_AVX512VL): Ditto.
(VI48_AVX512F_AVX512VL): Ditto.
(VI8_AVX2_AVX512F): Ditto.
(VI8_AVX_AVX512F): Ditto.
(V8FI): Ditto.
(V16FI): Ditto.
(VI124_AVX2_24_AVX512F_1_AVX512BW): Ditto.
(VI248_AVX512VLBW): Ditto.
(VI248_AVX2_8_AVX512F_24_AVX512BW): Ditto.
(VI248_AVX512BW): Ditto.
(VI248_AVX512BW_AVX512VL): Ditto.
(VI48_AVX512F): Ditto.
(VI48_AVX_AVX512F): Ditto.
(VI12_AVX_AVX512F): Ditto.
(VI148_512): Ditto.
(VI124_256_AVX512F_AVX512BW): Ditto.
(VI48_512): Ditto.
(VI_AVX512BW): Ditto.
(VIHFBF_AVX512BW): Ditto.
(VI4F_256_512): Ditto.
(VI48F_256_512): Ditto.
(VI48F): Ditto.
(VI12_VI48F_AVX512VL): Ditto.
(V32_512): Ditto.
(AVX512MODE2P): Ditto.
(STORENT_MODE): Ditto.
(REDUC_PLUS_MODE): Ditto.
(REDUC_SMINMAX_MODE): Ditto.
(*andnot<mode>3): Change isa attribute to avx512f_512.
(*andnot<mode>3): Ditto.
(<code><mode>3): Ditto.
(<code>tf3): Ditto.
(FMAMODEM): Add TARGET_EVEX512.
(FMAMODE_AVX512): Ditto.
(VFH_SF_AVX512VL): Ditto.
(avx512f_fix_notruncv16sfv16si<mask_name><round_name>): Ditto.
(fix<fixunssuffix>_truncv16sfv16si2<mask_name><round_saeonly_name>):
Ditto.
(avx512f_cvtdq2pd512_2): Ditto.
(avx512f_cvtpd2dq512<mask_name><round_name>): Ditto.
(fix<fixunssuffix>_truncv8dfv8si2<mask_name><round_saeonly_name>):
Ditto.
(<mask_codefor>avx512f_cvtpd2ps512<mask_name><round_name>): Ditto.
(vec_unpacks_lo_v16sf): Ditto.
(vec_unpacks_hi_v16sf): Ditto.
(vec_unpacks_float_hi_v16si): Ditto.
(vec_unpacks_float_lo_v16si): Ditto.
(vec_unpacku_float_hi_v16si): Ditto.
(vec_unpacku_float_lo_v16si): Ditto.
(vec_pack_sfix_trunc_v8df): Ditto.
(avx512f_vec_pack_sfix_v8df): Ditto.
(<mask_codefor>avx512f_unpckhps512<mask_name>): Ditto.
(<mask_codefor>avx512f_unpcklps512<mask_name>): Ditto.
(<mask_codefor>avx512f_movshdup512<mask_name>): Ditto.
(<mask_codefor>avx512f_movsldup512<mask_name>): Ditto.
(AVX512_VEC): Ditto.
(AVX512_VEC_2): Ditto.
(vec_extract_lo_v64qi): Ditto.
(vec_extract_hi_v64qi): Ditto.
(VEC_EXTRACT_MODE): Ditto.
(<mask_codefor>avx512f_unpckhpd512<mask_name>): Ditto.
(avx512f_movddup512<mask_name>): Ditto.
(avx512f_unpcklpd512<mask_name>): Ditto.
(*<avx512>_vternlog<mode>_all): Ditto.
(*<avx512>_vpternlog<mode>_1): Ditto.
(*<avx512>_vpternlog<mode>_2): Ditto.
(*<avx512>_vpternlog<mode>_3): Ditto.
(avx512f_shufps512_mask): Ditto.
(avx512f_shufps512_1<mask_name>): Ditto.
(avx512f_shufpd512_mask): Ditto.
(avx512f_shufpd512_1<mask_name>): Ditto.
(<mask_codefor>avx512f_interleave_highv8di<mask_name>): Ditto.
(<mask_codefor>avx512f_interleave_lowv8di<mask_name>): Ditto.
(vec_dupv2df<mask_name>): Ditto.
(trunc<pmov_src_lower><mode>2): Ditto.
(*avx512f_<code><pmov_src_lower><mode>2): Ditto.
(*avx512f_vpermvar_truncv8div8si_1): Ditto.
(avx512f_<code><pmov_src_lower><mode>2_mask): Ditto.
(avx512f_<code><pmov_src_lower><mode>2_mask_store): Ditto.
(truncv8div8qi2): Ditto.
(avx512f_<code>v8div16qi2): Ditto.
(*avx512f_<code>v8div16qi2_store_1): Ditto.
(*avx512f_<code>v8div16qi2_store_2): Ditto.
(avx512f_<code>v8div16qi2_mask): Ditto.
(*avx512f_<code>v8div16qi2_mask_1): Ditto.
(*avx512f_<code>v8div16qi2_mask_store_1): Ditto.
(avx512f_<code>v8div16qi2_mask_store_2): Ditto.
(vec_widen_umult_even_v16si<mask_name>): Ditto.
(*vec_widen_umult_even_v16si<mask_name>): Ditto.
(vec_widen_smult_even_v16si<mask_name>): Ditto.
(*vec_widen_smult_even_v16si<mask_name>): Ditto.
(VEC_PERM_AVX2): Ditto.
(one_cmpl<mode>2): Ditto.
(<mask_codefor>one_cmpl<mode>2<mask_name>): Ditto.
(*one_cmpl<mode>2_pternlog_false_dep): Ditto.
(define_split to xor): Ditto.
(*andnot<mode>3): Ditto.
(define_split for ior): Ditto.
(*iornot<mode>3): Ditto.
(*xnor<mode>3): Ditto.
(*<nlogic><mode>3): Ditto.
(<mask_codefor>avx512f_interleave_highv16si<mask_name>): Ditto.
(<mask_codefor>avx512f_interleave_lowv16si<mask_name>): Ditto.
(avx512f_pshufdv3_mask): Ditto.
(avx512f_pshufd_1<mask_name>): Ditto.
(*vec_extractv4ti): Ditto.
(VEXTRACTI128_MODE): Ditto.
(define_split to vec_extract): Ditto.
(VI1248_AVX512VL_AVX512BW): Ditto.
(<mask_codefor>avx512f_<code>v16qiv16si2<mask_name>): Ditto.
(<insn>v16qiv16si2): Ditto.
(avx512f_<code>v16hiv16si2<mask_name>): Ditto.
(<insn>v16hiv16si2): Ditto.
(avx512f_zero_extendv16hiv16si2_1): Ditto.
(avx512f_<code>v8qiv8di2<mask_name>): Ditto.
(*avx512f_<code>v8qiv8di2<mask_name>_1): Ditto.
(*avx512f_<code>v8qiv8di2<mask_name>_2): Ditto.
(<insn>v8qiv8di2): Ditto.
(avx512f_<code>v8hiv8di2<mask_name>): Ditto.
(<insn>v8hiv8di2): Ditto.
(avx512f_<code>v8siv8di2<mask_name>): Ditto.
(*avx512f_zero_extendv8siv8di2_1): Ditto.
(*avx512f_zero_extendv8siv8di2_2): Ditto.
(<insn>v8siv8di2): Ditto.
(avx512f_roundps512_sfix): Ditto.
(vashrv8di3): Ditto.
(vashrv16si3): Ditto.
(pbroadcast_evex_isa): Change isa attribute to avx512f_512.
(vec_dupv4sf): Add TARGET_EVEX512.
(*vec_dupv4si): Ditto.
(*vec_dupv2di): Ditto.
(vec_dup<mode>): Change isa attribute to avx512f_512.
(VPERMI2): Add TARGET_EVEX512.
(VPERMI2I): Ditto.
(VEC_INIT_MODE): Ditto.
(VEC_INIT_HALF_MODE): Ditto.
(<mask_codefor>avx512f_vcvtph2ps512<mask_name><round_saeonly_name>):
Ditto.
(avx512f_vcvtps2ph512_mask_sae): Ditto.
(<mask_codefor>avx512f_vcvtps2ph512<mask_name><round_saeonly_name>):
Ditto.
(*avx512f_vcvtps2ph512<merge_mask_name>): Ditto.
(INT_BROADCAST_MODE): Ditto.

Disable zmm register and 512 bit libmvec call when !TARGET_EVEX512

gcc/ChangeLog:

* config/i386/i386-expand.cc (ix86_broadcast_from_constant):
Disable zmm broadcast for !TARGET_EVEX512.
* config/i386/i386-options.cc (ix86_option_override_internal):
Do not use PVW_512 when no-evex512.
(ix86_simd_clone_adjust): Add evex512 target into string.
* config/i386/i386.cc (type_natural_mode): Report ABI warning
when using zmm register w/o evex512.
(ix86_return_in_memory): Do not allow zmm when !TARGET_EVEX512.
(ix86_hard_regno_mode_ok): Ditto.
(ix86_set_reg_reg_cost): Ditto.
(ix86_rtx_costs): Ditto.
(ix86_vector_mode_supported_p): Ditto.
(ix86_preferred_simd_mode): Ditto.
(ix86_get_mask_mode): Ditto.
(ix86_simd_clone_compute_vecsize_and_simdlen): Disable 512 bit
libmvec call when !TARGET_EVEX512.
(ix86_simd_clone_usable): Ditto.
* config/i386/i386.h (BIGGEST_ALIGNMENT): Disable 512 alignment
when !TARGET_EVEX512
(MOVE_MAX): Do not use PVW_512 when !TARGET_EVEX512.
(STORE_MAX_PIECES): Ditto.

[PATCH 5/5] Add OPTION_MASK_ISA2_EVEX512 for 512 bit builtins

gcc/ChangeLog:

* config/i386/i386-builtin.def (BDESC): Add
OPTION_MASK_ISA2_EVEX512.

[PATCH 4/5] Add OPTION_MASK_ISA2_EVEX512 for 512 bit builtins

gcc/ChangeLog:

* config/i386/i386-builtin.def (BDESC): Add
OPTION_MASK_ISA2_EVEX512.

[PATCH 3/5] Add OPTION_MASK_ISA2_EVEX512 for 512 bit builtins

gcc/ChangeLog:

* config/i386/i386-builtin.def (BDESC): Add
OPTION_MASK_ISA2_EVEX512.

[PATCH 2/5] Add OPTION_MASK_ISA2_EVEX512 for 512 bit builtins

gcc/ChangeLog:

* config/i386/i386-builtin.def (BDESC): Add
OPTION_MASK_ISA2_EVEX512.

[PATCH 1/5] Add OPTION_MASK_ISA2_EVEX512 for 512 bit builtins

gcc/ChangeLog:

* config/i386/i386-builtin.def (BDESC): Add
OPTION_MASK_ISA2_EVEX512.
* config/i386/i386-builtins.cc
(ix86_init_mmx_sse_builtins): Ditto.

[PATCH 5/5] Push evex512 target for 512 bit intrins

gcc/Changelog:

* config/i386/avx512fp16intrin.h: Add evex512 target for 512 bit
intrins.

Co-authored-by: Hu, Lin1 <lin1.hu@intel.com>

[PATCH 4/5] Push evex512 target for 512 bit intrins

gcc/ChangeLog:

* config.gcc: Add avx512bitalgvlintrin.h.
* config/i386/avx5124fmapsintrin.h: Add evex512 target for 512 bit
intrins.
* config/i386/avx5124vnniwintrin.h: Ditto.
* config/i386/avx512bf16intrin.h: Ditto.
* config/i386/avx512bitalgintrin.h: Add evex512 target for 512 bit
intrins. Split 128/256 bit intrins to avx512bitalgvlintrin.h.
* config/i386/avx512erintrin.h: Add evex512 target for 512 bit
intrins
* config/i386/avx512ifmaintrin.h: Ditto
* config/i386/avx512pfintrin.h: Ditto
* config/i386/avx512vbmi2intrin.h: Ditto.
* config/i386/avx512vbmiintrin.h: Ditto.
* config/i386/avx512vnniintrin.h: Ditto.
* config/i386/avx512vp2intersectintrin.h: Ditto.
* config/i386/avx512vpopcntdqintrin.h: Ditto.
* config/i386/gfniintrin.h: Ditto.
* config/i386/immintrin.h: Add avx512bitalgvlintrin.h.
* config/i386/vaesintrin.h: Add evex512 target for 512 bit intrins.
* config/i386/vpclmulqdqintrin.h: Ditto.
* config/i386/avx512bitalgvlintrin.h: New.

[PATCH 4/5] Push evex512 target for 512 bit intrins

gcc/ChangeLog:

* config/i386/avx512bwintrin.h: Add evex512 target for 512 bit
intrins.

[PATCH 2/5] Push evex512 target for 512 bit intrins

gcc/ChangeLog:

* config/i386/avx512dqintrin.h: Add evex512 target for 512 bit
intrins.

[PATCH 1/5] Push evex512 target for 512 bit intrins

gcc/ChangeLog:

* config/i386/avx512fintrin.h: Add evex512 target for 512 bit intrins.

Initial support for -mevex512

gcc/ChangeLog:

* common/config/i386/i386-common.cc
(OPTION_MASK_ISA2_EVEX512_SET): New.
(OPTION_MASK_ISA2_EVEX512_UNSET): Ditto.
(ix86_handle_option): Handle EVEX512.
* config/i386/i386-c.cc
(ix86_target_macros_internal): Handle EVEX512. Add __EVEX256__
when AVX512VL is set.
* config/i386/i386-options.cc: (isa2_opts): Handle EVEX512.
(ix86_valid_target_attribute_inner_p): Ditto.
(ix86_option_override_internal): Set EVEX512 target if it is not
explicitly set when AVX512 is enabled. Disable
AVX512{PF,ER,4VNNIW,4FAMPS} for -mno-evex512.
* config/i386/i386.opt: Add mevex512. Temporaily RejectNegative.

TEST: Fix dump FAIL for RVV (RISCV-V vector)

As this showed: https://godbolt.org/z/3K9oK7fx3

ARM SVE 2 times for FOLD_EXTRACT_LAST wheras RVV 4 times.

This is because RISC-V doesn't enable vec_pack_trunc so we will failed conversion and fold_extract_last at the first time analysis.
Then we succeed at the second time.

So RVV has 4 times of showing "FOLD_EXTRACT_LAST:.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/vect-cond-reduc-4.c: Add vect_pack_trunc variant.

rs6000: support 32bit inline lrint

gcc/
PR target/88558
* config/rs6000/rs6000.md (lrint<mode>di2): Remove TARGET_FPRND
from insn condition.
(lrint<mode>si2): New insn pattern for 32bit lrint.

gcc/testsuite/
PR target/106769
* gcc.target/powerpc/pr88558.h: New.
* gcc.target/powerpc/pr88558-p7.c: New.
* gcc.target/powerpc/pr88558-p8.c: New.

rs6000: enable SImode in FP register on P7

gcc/
PR target/88558
* config/rs6000/rs6000.cc (rs6000_hard_regno_mode_ok_uncached):
Enable SImode on FP registers for P7.
* config/rs6000/rs6000.md (*movsi_internal1): Add fmr for SImode
move between FP registers.  Set attribute isa of stfiwx to "*"
and attribute of stxsiwx to "p7".

s390: Make use of new copysign RTL

gcc/ChangeLog:

* config/s390/s390.md: Make use of new copysign RTL.

[i386] APX EGPR: fix missing patterns that prohibit egpr

For some pattern m/Bm constraint in alternative 0 and 1 could result in
egpr allocated on memory operand under -mapxf. Should use jm/ja instead.

gcc/ChangeLog:

* config/i386/sse.md (vec_concatv2di): Replace constraint "m"
with "jm" for alternative 0 and 1 of operand 2.
(sse4_1_<code><mode>3<mask_name>): Replace constraint "Bm" with
"ja" for alternative 0 and 1 of operand2.

Daily bump.

libcpp: eliminate LINEMAPS_{ORDINARY,MACRO}_MAPS

libcpp/ChangeLog:
* include/line-map.h (LINEMAPS_ORDINARY_MAPS): Delete.
(LINEMAPS_MACRO_MAPS): Delete.
* line-map.cc (linemap_tracks_macro_expansion_locs_p): Update for
deletion of LINEMAPS_MACRO_MAPS.
(linemap_get_statistics): Likewise.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

libcpp: eliminate LINEMAPS_{,ORDINARY_,MACRO_}CACHE

It's simpler to use field access than to go through these inline
functions that look as if they are macros.

No functional change intended.

libcpp/ChangeLog:
* include/line-map.h (maps_info_ordinary::cache): Rename to...
(maps_info_ordinary::m_cache): ...this.
(maps_info_macro::cache): Rename to...
(maps_info_macro::m_cache): ...this.
(LINEMAPS_CACHE): Delete.
(LINEMAPS_ORDINARY_CACHE): Delete.
(LINEMAPS_MACRO_CACHE): Delete.
* init.cc (read_original_filename): Update for adding "m_" prefix.
* line-map.cc (linemap_add): Eliminate LINEMAPS_ORDINARY_CACHE in
favor of a simple field access.
(linemap_enter_macro): Likewise for LINEMAPS_MACRO_CACHE.
(linemap_ordinary_map_lookup): Likewise for
LINEMAPS_ORDINARY_CACHE, twice.
(linemap_lookup_macro_index): Likewise for LINEMAPS_MACRO_CACHE.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

libcpp: eliminate LINEMAPS_LAST_ALLOCATED{,_ORDINARY,_MACRO}_MAP

Nothing uses these; delete them.

libcpp/ChangeLog:
* include/line-map.h (LINEMAPS_LAST_ALLOCATED_MAP): Delete.
(LINEMAPS_LAST_ALLOCATED_ORDINARY_MAP): Delete.
(LINEMAPS_LAST_ALLOCATED_MACRO_MAP): Delete.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

analyzer: improvements to out-of-bounds diagrams [PR111155]

Update out-of-bounds diagrams to show existing string values,
and the initial write index within a string buffer.

For example, given the out-of-bounds write in strcat in:

void test (void)
{
  char buf[10];
  strcpy (buf, "hello");
  strcat (buf, " world!");
}

the diagram improves from:

                           ┌─────┬─────┬────┬────┬────┐┌─────┬─────┬─────┐
                           │ [0] │ [1] │[2] │[3] │[4] ││ [5] │ [6] │ [7] │
                           ├─────┼─────┼────┼────┼────┤├─────┼─────┼─────┤
                           │ ' ' │ 'w' │'o' │'r' │'l' ││ 'd' │ '!' │ NUL │
                           ├─────┴─────┴────┴────┴────┴┴─────┴─────┴─────┤
                           │      string literal (type: 'char[8]')       │
                           └─────────────────────────────────────────────┘
                              │     │    │    │    │      │     │     │
                              │     │    │    │    │      │     │     │
                              v     v    v    v    v      v     v     v
  ┌─────┬────────────────────────────────────────┬────┐┌─────────────────┐
  │ [0] │                  ...                   │[9] ││                 │
  ├─────┴────────────────────────────────────────┴────┤│after valid range│
  │             'buf' (type: 'char[10]')              ││                 │
  └───────────────────────────────────────────────────┘└─────────────────┘
  ├─────────────────────────┬─────────────────────────┤├────────┬────────┤
                            │                                   │
                  ╭─────────┴────────╮                ╭─────────┴─────────╮
                  │capacity: 10 bytes│                │overflow of 3 bytes│
                  ╰──────────────────╯                ╰───────────────────╯

to:

                             ┌────┬────┬────┬────┬────┐┌─────┬─────┬─────┐
                             │[0] │[1] │[2] │[3] │[4] ││ [5] │ [6] │ [7] │
                             ├────┼────┼────┼────┼────┤├─────┼─────┼─────┤
                             │' ' │'w' │'o' │'r' │'l' ││ 'd' │ '!' │ NUL │
                             ├────┴────┴────┴────┴────┴┴─────┴─────┴─────┤
                             │     string literal (type: 'char[8]')      │
                             └───────────────────────────────────────────┘
                               │    │    │    │    │      │     │     │
                               │    │    │    │    │      │     │     │
                               v    v    v    v    v      v     v     v
  ┌─────┬────────────────────┬────┬──────────────┬────┐┌─────────────────┐
  │ [0] │        ...         │[5] │     ...      │[9] ││                 │
  ├─────┼────┬────┬────┬────┬┼────┼──────────────┴────┘│                 │
  │ 'h' │'e' │'l' │'l' │'o' ││NUL │                    │after valid range│
  ├─────┴────┴────┴────┴────┴┴────┴───────────────────┐│                 │
  │             'buf' (type: 'char[10]')              ││                 │
  └───────────────────────────────────────────────────┘└─────────────────┘
  ├─────────────────────────┬─────────────────────────┤├────────┬────────┤
                            │                                   │
                  ╭─────────┴────────╮                ╭─────────┴─────────╮
                  │capacity: 10 bytes│                │overflow of 3 bytes│
                  ╰──────────────────╯                ╰───────────────────╯

gcc/analyzer/ChangeLog:
PR analyzer/111155
* access-diagram.cc (boundaries::boundaries): Add logger param
(boundaries::add): Add logging.
(boundaries::get_hard_boundaries_in_range): New.
(boundaries::m_logger): New field.
(boundaries::get_table_x_for_offset): Make public.
(class svalue_spatial_item): New.
(class compound_svalue_spatial_item): New.
(add_ellipsis_to_gaps): New.
(valid_region_spatial_item::valid_region_spatial_item): Add theme
param.  Initialize m_boundaries, m_existing_sval, and
m_existing_sval_spatial_item.
(valid_region_spatial_item::add_boundaries): Set m_boundaries.
Add boundaries for any m_existing_sval_spatial_item.
(valid_region_spatial_item::add_array_elements_to_table): Rewrite
creation of min/max index in terms of
maybe_add_array_index_to_table.  Rewrite ellipsis code using
add_ellipsis_to_gaps. Add index values for any hard boundaries
within the valid region.
(valid_region_spatial_item::maybe_add_array_index_to_table): New,
based on code formerly in add_array_elements_to_table.
(valid_region_spatial_item::make_table): Make use of
m_existing_sval_spatial_item, if any.
(valid_region_spatial_item::m_boundaries): New field.
(valid_region_spatial_item::m_existing_sval): New field.
(valid_region_spatial_item::m_existing_sval_spatial_item): New
field.
(class svalue_spatial_item): Rename to...
(class written_svalue_spatial_item): ...this.
(class string_region_spatial_item): Rename to..
(class string_literal_spatial_item): ...this.  Add "kind".
(string_literal_spatial_item::add_boundaries): Use m_kind to
determine kind of boundary.  Update for renaming of m_actual_bits
to m_bits.
(string_literal_spatial_item::make_table): Likewise.  Support not
displaying a row for byte indexes, and not displaying a row for
the type.
(string_literal_spatial_item::add_column_for_byte): Make byte index
row optional.
(svalue_spatial_item::make): Convert to...
(make_written_svalue_spatial_item): ...this.
(make_existing_svalue_spatial_item): New.
(access_diagram_impl::access_diagram_impl): Pass theme to
m_valid_region_spatial_item ctor.  Update for renaming of
m_svalue_spatial_item.
(access_diagram_impl::find_boundaries): Pass logger to boundaries.
Update for renaming of...
(access_diagram_impl::m_svalue_spatial_item): Rename to...
(access_diagram_impl::m_written_svalue_spatial_item): ...this.

gcc/testsuite/ChangeLog:
PR analyzer/111155
* c-c++-common/analyzer/out-of-bounds-diagram-strcat-2.c: New test.
* c-c++-common/analyzer/out-of-bounds-diagram-strcat.c: New test.
* gcc.dg/analyzer/out-of-bounds-diagram-17.c: Update expected
result to show the existing content of "buf" and the index at
which the write starts.
* gcc.dg/analyzer/out-of-bounds-diagram-18.c: Likewise.
* gcc.dg/analyzer/out-of-bounds-diagram-19.c: Likewise.
* gcc.dg/analyzer/out-of-bounds-diagram-6.c: Update expected
output.

gcc/ChangeLog:
PR analyzer/111155
* text-art/table.cc (table::maybe_set_cell_span): New.
(table::add_other_table): New.
* text-art/table.h (class table::cell_placement): Add class table
as a friend.
(table::add_rows): New.
(table::add_row): Reimplement in terms of add_rows.
(table::maybe_set_cell_span): New decl.
(table::add_other_table): New decl.
* text-art/types.h (operator+): New operator for rect + coord.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

libcpp: eliminate COMBINE_LOCATION_DATA

This patch eliminates the function "COMBINE_LOCATION_DATA" (which hasn't
been a macro since r6-739-g0501dbd932a7e9) and the function
"get_combined_adhoc_loc" in favor of a new
line_maps::get_or_create_combined_loc member function.

No functional change intended.

gcc/cp/ChangeLog:
* module.cc (module_state::read_location): Update for renaming of
get_combined_adhoc_loc.

gcc/ChangeLog:
* genmatch.cc (main): Update for "m_" prefix of some fields of
line_maps.
* input.cc (make_location): Update for removal of
COMBINE_LOCATION_DATA.
(dump_line_table_statistics): Update for "m_" prefix of some
fields of line_maps.
(location_with_discriminator): Update for removal of
COMBINE_LOCATION_DATA.
(line_table_test::line_table_test): Update for "m_" prefix of some
fields of line_maps.
* toplev.cc (general_init): Likewise.
* tree.cc (set_block): Update for removal of
COMBINE_LOCATION_DATA.
(set_source_range): Likewise.

libcpp/ChangeLog:
* include/line-map.h (line_maps::reallocator): Rename to...
(line_maps::m_reallocator): ...this.
(line_maps::round_alloc_size): Rename to...
(line_maps::m_round_alloc_size): ...this.
(line_maps::location_adhoc_data_map): Rename to...
(line_maps::m_location_adhoc_data_map): ...this.
(line_maps::num_optimized_ranges): Rename to...
(line_maps::m_num_optimized_ranges): ..this.
(line_maps::num_unoptimized_ranges): Rename to...
(line_maps::m_num_unoptimized_ranges): ...this.
(get_combined_adhoc_loc): Delete decl.
(COMBINE_LOCATION_DATA): Delete.
* lex.cc (get_location_for_byte_range_in_cur_line): Update for
removal of COMBINE_LOCATION_DATA.
(warn_about_normalization): Likewise.
(_cpp_lex_direct): Likewise.
* line-map.cc (line_maps::~line_maps): Update for "m_" prefix of
some fields of line_maps.
(rebuild_location_adhoc_htab): Likewise.
(can_be_stored_compactly_p): Convert to...
(line_maps::can_be_stored_compactly_p): ...this private member
function.
(get_combined_adhoc_loc): Convert to...
(line_maps::get_or_create_combined_loc): ...this public member
function.
(line_maps::make_location): Update for removal of
COMBINE_LOCATION_DATA.
(get_data_from_adhoc_loc): Update for "m_" prefix of some fields
of line_maps.
(get_discriminator_from_adhoc_loc): Likewise.
(get_location_from_adhoc_loc): Likewise.
(get_range_from_adhoc_loc): Convert to...
(line_maps::get_range_from_adhoc_loc): ...this private member
function.
(line_maps::get_range_from_loc): Update for conversion of
get_range_from_adhoc_loc to a member function.
(linemap_init): Update for "m_" prefix of some fields of
line_maps.
(line_map_new_raw): Likewise.
(linemap_enter_macro): Likewise.
(linemap_get_statistics): Likewise.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

libcpp: "const" and other cleanups

No functional change intended.

gcc/ChangeLog:
* input.cc (make_location): Move implementation to
line_maps::make_location.

libcpp/ChangeLog:
* include/line-map.h (line_maps::pure_location_p): New decl.
(line_maps::get_pure_location): New decl.
(line_maps::get_range_from_loc): New decl.
(line_maps::get_start): New.
(line_maps::get_finish): New.
(line_maps::make_location): New decl.
(get_range_from_loc): Make line_maps param const.
(get_discriminator_from_loc): Likewise.
(pure_location_p): Likewise.
(get_pure_location): Likewise.
(linemap_check_files_exited): Likewise.
(linemap_tracks_macro_expansion_locs_p): Likewise.
(linemap_location_in_system_header_p): Likewise.
(linemap_location_from_macro_definition_p): Likewise.
(linemap_macro_map_loc_unwind_toward_spelling): Likewise.
(linemap_included_from_linemap): Likewise.
(first_map_in_common): Likewise.
(linemap_compare_locations): Likewise.
(linemap_location_before_p): Likewise.
(linemap_resolve_location): Likewise.
(linemap_unwind_toward_expansion): Likewise.
(linemap_unwind_to_first_non_reserved_loc): Likewise.
(linemap_expand_location): Likewise.
(linemap_get_file_highest_location): Likewise.
(linemap_get_statistics): Likewise.
(linemap_dump_location): Likewise.
(linemap_dump): Likewise.
(line_table_dump): Likewise.
* internal.h (linemap_get_expansion_line): Likewise.
(linemap_get_expansion_filename): Likewise.
* line-map.cc (can_be_stored_compactly_p): Likewise.
(get_data_from_adhoc_loc): Drop redundant "class".
(get_discriminator_from_adhoc_loc): Likewise.
(get_location_from_adhoc_loc): Likewise.
(get_range_from_adhoc_loc): Likewise.
(get_range_from_loc): Make const and move implementation to...
(line_maps::get_range_from_loc): ...this new function.
(get_discriminator_from_loc): Make line_maps param const.
(pure_location_p): Make const and move implementation to...
(line_maps::pure_location_p): ...this new function.
(get_pure_location): Make const and move implementation to...
(line_maps::get_pure_location): ...this new function.
(linemap_included_from_linemap): Make line_maps param const.
(linemap_check_files_exited): Likewise.
(linemap_tracks_macro_expansion_locs_p): Likewise.
(linemap_macro_map_loc_unwind_toward_spelling): Likewise.
(linemap_get_expansion_line): Likewise.
(linemap_get_expansion_filename): Likewise.
(linemap_location_in_system_header_p): Likewise.
(first_map_in_common_1): Likewise.
(linemap_compare_locations): Likewise.
(linemap_macro_loc_to_spelling_point): Likewise.
(linemap_macro_loc_to_def_point): Likewise.
(linemap_macro_loc_to_exp_point): Likewise.
(linemap_resolve_location): Likewise.
(linemap_location_from_macro_definition_p): Likewise.
(linemap_unwind_toward_expansion): Likewise.
(linemap_unwind_to_first_non_reserved_loc): Likewise.
(linemap_expand_location): Likewise.
(linemap_dump): Likewise.
(linemap_dump_location): Likewise.
(linemap_get_file_highest_location): Likewise.
(linemap_get_statistics): Likewise.
(line_table_dump): Likewise.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

diagnostics: fix ICE on sarif output when source file is unreadable [PR111700]

gcc/ChangeLog:
PR driver/111700
* input.cc (file_cache::add_file): Update leading comment to
clarify that it can fail.
(file_cache::lookup_or_add_file): Likewise.
(file_cache::get_source_file_content): Gracefully handle
lookup_or_add_file failing.

gcc/testsuite/ChangeLog:
PR driver/111700
* c-c++-common/diagnostic-format-sarif-file-pr111700.c: New test.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

Support signbit/xorsign/copysign/abs/neg/and/xor/ior/andn for V2HF/V4HF.

gcc/ChangeLog:

* config/i386/i386.cc (ix86_build_const_vector): Handle V2HF
and V4HFmode.
(ix86_build_signbit_mask): Ditto.
* config/i386/mmx.md (mmxintvecmode): Ditto.
(<code><mode>2): New define_expand.
(*mmx_<code><mode>): New define_insn_and_split.
(*mmx_nabs<mode>2): Ditto.
(*mmx_andnot<mode>3): New define_insn.
(<code><mode>3): Ditto.
(copysign<mode>3): New define_expand.
(xorsign<mode>3): Ditto.
(signbit<mode>2): Ditto.

gcc/testsuite/ChangeLog:

* gcc.target/i386/part-vect-absneghf.c: New test.
* gcc.target/i386/part-vect-copysignhf.c: New test.
* gcc.target/i386/part-vect-xorsignhf.c: New test.

Support smin/smax for V2HF/V4HF

gcc/ChangeLog:

* config/i386/mmx.md (VHF_32_64): New mode iterator.
(<insn><mode>3): New define_expand, merged from ..
(<insn>v4hf3): .. this and
(<insn>v2hf3): .. this.
(movd_v2hf_to_sse_reg): New define_expand, splitted from ..
(movd_v2hf_to_sse): .. this.
(<code><mode>3): New define_expand.

gcc/testsuite/ChangeLog:

* gcc.target/i386/part-vect-vminmaxph-1.c: New test.
* gcc.target/i386/avx512fp16-64-32-vecop-1.c: Scan-assembler
only for { target { ! ia32 } }.

Fortran/OpenMP: Fix handling of strictly structured blocks

For strictly structured blocks, a BLOCK was created but the code
was placed after the block the outer structured block. Additionally,
labelled blocks were mishandled. As the code is now properly in a
BLOCK, it solves additional issues.

gcc/fortran/ChangeLog:

* parse.cc (parse_omp_structured_block): Make the user code end
up inside of BLOCK construct for strictly structured blocks;
fix fallout for 'section' and 'teams'.
* openmp.cc (resolve_omp_target): Fix changed BLOCK handling
for teams in target checking.

libgomp/ChangeLog:

* testsuite/libgomp.fortran/strictly-structured-block-1.f90: New test.

gcc/testsuite/ChangeLog:

* gfortran.dg/block_17.f90: New test.
* gfortran.dg/gomp/strictly-structured-block-5.f90: New test.

rs6000: build constant via li/lis;rldic

This patch checks if a constant is possible to be built by "li;rldic".
Only need to take care of "negative li", other forms do not need to check.
For example, "negative lis" is just a "negative li" with an additional shift.

gcc/ChangeLog:

* config/rs6000/rs6000.cc (can_be_built_by_li_and_rldic): New function.
(rs6000_emit_set_long_const): Call can_be_built_by_li_and_rldic.

gcc/testsuite/ChangeLog:

* gcc.target/powerpc/const-build.c: Add more tests.

rs6000: build constant via li/lis;rldicl/rldicr

If a constant is possible left/right cleaned on a rotated value from
a negative value of "li/lis".  Then, using "li/lis ; rldicl/rldicr"
to build the constant.

gcc/ChangeLog:

* config/rs6000/rs6000.cc (can_be_built_by_li_lis_and_rldicl): New
function.
(can_be_built_by_li_lis_and_rldicr): New function.
(rs6000_emit_set_long_const): Call can_be_built_by_li_lis_and_rldicr and
can_be_built_by_li_lis_and_rldicl.

gcc/testsuite/ChangeLog:

* gcc.target/powerpc/const-build.c: Add more tests.

rs6000: build constant via lis;rotldi

If a constant is possible to be rotated to/from a negative value from
"lis", then using "lis;rotldi" to build the constant.

The positive value of "lis" does not need to be analyzed.  Because if a
constant can be rotated from the positive value of "lis", it also can be
rotated from a positive value of "li".

gcc/ChangeLog:

* config/rs6000/rs6000.cc (can_be_rotated_to_negative_lis): New
function.
(can_be_built_by_li_and_rotldi): Rename to ...
(can_be_built_by_li_lis_and_rotldi): ... this function.
(rs6000_emit_set_long_const): Call can_be_built_by_li_lis_and_rotldi.

gcc/testsuite/ChangeLog:

* gcc.target/powerpc/const-build.c: Add more tests.

rs6000: build constant via li;rotldi

If a constant is possible to be rotated to/from a positive or negative
value which "li" can generated, then "li;rotldi" can be used to build
the constant.

gcc/ChangeLog:

* config/rs6000/rs6000.cc (can_be_built_by_li_and_rotldi): New function.
(rs6000_emit_set_long_const): Call can_be_built_by_li_and_rotldi.

gcc/testsuite/ChangeLog:

* gcc.target/powerpc/const-build.c: New test.

[i386] Fix apx test fails on 32bit target

Since -mapxf works similar as -muintr that will emit error for 32bit
target, add !ia32 target guard for apx related tests.

gcc/testsuite/ChangeLog:

* gcc.target/i386/apx-egprs-names.c: Compile for non-ia32.
* gcc.target/i386/apx-inline-gpr-norex2.c: Likewise.
* gcc.target/i386/apx-interrupt-1.c: Likewise.
* gcc.target/i386/apx-legacy-insn-check-norex2-asm.c: Likewise.
* gcc.target/i386/apx-legacy-insn-check-norex2.c: Likewise.

RISC-V: add static-pie support

We only need to pass options to the linker when static-pie is passed.
There's another patch to enable static-pie in glibc. And we need to
enable in GCC first.

gcc/ChangeLog:

* config/riscv/linux.h: Pass the static-pie specific options to
the linker.

Signed-off-by: Yanzhang Wang <yanzhang.wang@intel.com>

TEST: Fix XPASS of TSVC testsuites for RVV

Fix these following XPASS FAILs of TSVC for RVV:

XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1115.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1115.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s114.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s114.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1161.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1161.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1232.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1232.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s124.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s124.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1279.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1279.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s161.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s161.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s253.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s253.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s257.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s257.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s271.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s271.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s2711.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s2711.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s2712.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s2712.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s272.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s272.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s273.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s273.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s274.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s274.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s276.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s276.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s278.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s278.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s279.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s279.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s3111.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s3111.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s353.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s353.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s441.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s441.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s443.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s443.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-vif.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-vif.c scan-tree-dump vect "vectorized 1 loops"

gcc/testsuite/ChangeLog:

* gcc.dg/vect/tsvc/vect-tsvc-s1115.c: Fix TSVC XPASS.
* gcc.dg/vect/tsvc/vect-tsvc-s114.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s1161.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s1232.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s124.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s1279.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s161.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s253.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s257.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s271.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s2711.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s2712.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s272.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s273.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s274.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s276.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s278.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s279.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s3111.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s353.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s441.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-s443.c: Ditto.
* gcc.dg/vect/tsvc/vect-tsvc-vif.c: Ditto.

RISC-V: Enable more tests of "vect" for RVV

This patch enables almost full coverage vectorization tests for RVV, except these
following tests (not enabled yet):

1. Will enable soon:

check_effective_target_vect_call_lrint
check_effective_target_vect_call_btrunc
check_effective_target_vect_call_btruncf
check_effective_target_vect_call_ceil
check_effective_target_vect_call_ceilf
check_effective_target_vect_call_floor
check_effective_target_vect_call_floorf
check_effective_target_vect_call_lceil
check_effective_target_vect_call_lfloor
check_effective_target_vect_call_nearbyint
check_effective_target_vect_call_nearbyintf
check_effective_target_vect_call_round
check_effective_target_vect_call_roundf

2. Not sure we will need to enable or not:

check_effective_target_vect_complex_*
check_effective_target_vect_simd_clones
check_effective_target_vect_bswap
check_effective_target_vect_widen_shift
check_effective_target_vect_widen_mult_*
check_effective_target_vect_widen_sum_*
check_effective_target_vect_unpack
check_effective_target_vect_interleave
check_effective_target_vect_extract_even_odd
check_effective_target_vect_pack_trunc
check_effective_target_vect_check_ptrs
check_effective_target_vect_sdiv_pow2_si
check_effective_target_vect_usad_*
check_effective_target_vect_udot_*
check_effective_target_vect_sdot_*
check_effective_target_vect_gather_load_ifn

After this patch, we will have these following additional FAILs:
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1115.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1115.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s114.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s114.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1161.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1161.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1232.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1232.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s124.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s124.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1279.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s1279.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s161.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s161.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s253.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s253.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s257.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s257.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s271.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s271.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s2711.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s2711.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s2712.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s2712.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s272.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s272.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s273.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s273.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s274.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s274.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s276.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s276.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s278.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s278.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s279.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s279.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s3111.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s3111.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s353.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s353.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s441.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s441.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s443.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-s443.c scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-vif.c -flto -ffat-lto-objects  scan-tree-dump vect "vectorized 1 loops"
XPASS: gcc.dg/vect/tsvc/vect-tsvc-vif.c scan-tree-dump vect "vectorized 1 loops"
FAIL: gcc.dg/vect/vect-114.c -flto -ffat-lto-objects  scan-tree-dump-times vect "vectorized 0 loops" 1
FAIL: gcc.dg/vect/vect-114.c scan-tree-dump-times vect "vectorized 0 loops" 1

FAIL: gcc.dg/vect/vect-live-2.c -flto -ffat-lto-objects  scan-tree-dump-times vect "vec_stmt_relevant_p: stmt live but not relevant" 1
FAIL: gcc.dg/vect/vect-live-2.c scan-tree-dump-times vect "vec_stmt_relevant_p: stmt live but not relevant" 1
FAIL: gcc.dg/vect/vect-reduc-or_1.c -flto -ffat-lto-objects  scan-tree-dump vect "Reduce using vector shifts"
FAIL: gcc.dg/vect/vect-reduc-or_1.c scan-tree-dump vect "Reduce using vector shifts"
FAIL: gcc.dg/vect/vect-reduc-or_2.c -flto -ffat-lto-objects  scan-tree-dump vect "Reduce using vector shifts"
FAIL: gcc.dg/vect/vect-reduc-or_2.c scan-tree-dump vect "Reduce using vector shifts"

FAIL: gcc.dg/vect/vect-cond-arith-5.c -flto -ffat-lto-objects  scan-tree-dump optimized " = \\.COND_ADD"
FAIL: gcc.dg/vect/vect-cond-arith-5.c -flto -ffat-lto-objects  scan-tree-dump optimized " = \\.COND_SUB"
FAIL: gcc.dg/vect/vect-cond-arith-5.c -flto -ffat-lto-objects  scan-tree-dump optimized " = \\.COND_MUL"
FAIL: gcc.dg/vect/vect-cond-arith-5.c -flto -ffat-lto-objects  scan-tree-dump optimized " = \\.COND_RDIV"

They are all dump FAILs (No more ICE and execution FAILs).

Fixing those FAILs will be another separate patch.

But I think we should commit this patch first.

Ok for trunk ?

gcc/testsuite/ChangeLog:

* lib/target-supports.exp: Enable more vect tests for RVV.

Daily bump.

aarch64: Enable Cortex-X4 CPU

This patch adds support for the Cortex-X4 CPU to GCC.

gcc/ChangeLog:
* config/aarch64/aarch64-cores.def (AARCH64_CORE): Add support for
cortex-x4 core.
* config/aarch64/aarch64-tune.md: Regenerated.
* doc/invoke.texi: Add command-line option for cortex-x4 core.

Revert "RISC-V: Add more run test for FP rounding autovec"

Revert since other fails are introduced

This reverts commit 7866984ba427dc56a12ee1b8d99feb4927b834b1.

[APX EGPR] Handle vex insns that only support GPR16 (5/5)

These vex insn may have legacy counterpart that could support EGPR,
but they do not have evex counterpart. Split out its vex part from
patterns and set the vex part to non-EGPR supported by adjusting
constraints and attr_gpr32.

insn list:
1. vmovmskpd/vmovmskps
2. vpmovmskb
3. vrsqrtss/vrsqrtps
4. vrcpss/vrcpps
5. vhaddpd/vhaddps, vhsubpd/vhsubps
6. vldmxcsr/vstmxcsr
7. vaddsubpd/vaddsubps
8. vlddqu
9. vtestps/vtestpd
10. vmaskmovps/vmaskmovpd, vpmaskmovd/vpmaskmovq
11. vperm2f128/vperm2i128
12. vinserti128/vinsertf128
13. vbroadcasti128/vbroadcastf128
14. vcmppd/vcmpps, vcmpss/vcmpsd
15. vgatherdps/vgatherqps, vgatherdpd/vgatherqpd

gcc/ChangeLog:

* config/i386/constraints.md (jb): New constraint for vsib memory
that does not allow gpr32.
* config/i386/i386.md: (setcc_<mode>_sse): Replace m to jm for avx
alternative and set attr_gpr32 to 0.
(movmsk_df): Split avx/noavx alternatives and  replace "r" to "jr" for
avx alternative.
(<sse>_rcp<mode>2): Split avx/noavx alternatives and replace
"m/Bm" to "jm/ja" for avx alternative, set its gpr32 attr to 0.
(*rsqrtsf2_sse): Likewise.
* config/i386/mmx.md (mmx_pmovmskb): Split alternative 1 to
avx/noavx and assign jr/r constraint to dest.
* config/i386/sse.md (<sse>_movmsk<ssemodesuffix><avxsizesuffix>):
Split avx/noavx alternatives and replace "r" to "jr" for avx alternative.
(*<sse>_movmsk<ssemodesuffix><avxsizesuffix>_<u>ext): Likewise.
(*<sse>_movmsk<ssemodesuffix><avxsizesuffix>_lt): Likewise.
(*<sse>_movmsk<ssemodesuffix><avxsizesuffix>_<u>ext_lt): Likewise.
(*<sse>_movmsk<ssemodesuffix><avxsizesuffix>_shift): Likewise.
(*<sse>_movmsk<ssemodesuffix><avxsizesuffix>_<u>ext_shift): Likewise.
(<sse2_avx2>_pmovmskb): Likewise.
(*<sse2_avx2>_pmovmskb_zext): Likewise.
(*sse2_pmovmskb_ext): Likewise.
(*<sse2_avx2>_pmovmskb_lt): Likewise.
(*<sse2_avx2>_pmovmskb_zext_lt): Likewise.
(*sse2_pmovmskb_ext_lt): Likewise.
(<sse>_rcp<mode>2): Split avx/noavx alternatives and replace
"m/Bm" to "jm/ja" for avx alternative, set its attr_gpr32 to 0.
(sse_vmrcpv4sf2): Likewise.
(*sse_vmrcpv4sf2): Likewise.
(rsqrt<mode>2): Likewise.
(sse_vmrsqrtv4sf2): Likewise.
(*sse_vmrsqrtv4sf2): Likewise.
(avx_h<insn>v4df3): Likewise.
(sse3_hsubv2df3): Likewise.
(avx_h<insn>v8sf3): Likewise.
(sse3_h<insn>v4sf3): Likewise.
(<sse3>_lddqu<avxsizesuffix>): Likewise.
(avx_cmp<mode>3): Likewise.
(avx_vmcmp<mode>3): Likewise.
(*sse2_gt<mode>3): Likewise.
(sse_ldmxcsr): Likewise.
(sse_stmxcsr): Likewise.
(avx_vtest<ssemodesuffix><avxsizesuffix>): Replace m to jm for
avx alternative and set attr_gpr32 to 0.
(avx2_permv2ti): Likewise.
(*avx_vperm2f128<mode>_full): Likewise.
(*avx_vperm2f128<mode>_nozero): Likewise.
(vec_set_lo_v32qi): Likewise.
(<avx_avx2>_maskload<ssemodesuffix><avxsizesuffix>): Likewise.
(<avx_avx2>_maskstore<ssemodesuffix><avxsi)zesuffix>: Likewise.
(avx_cmp<mode>3): Likewise.
(avx_vmcmp<mode>3): Likewise.
(*<sse>_maskcmp<mode>3_comm): Likewise.
(*avx2_gathersi<VEC_GATHER_MODE:mode>): Replace Tv to jb and set
attr_gpr32 to 0.
(*avx2_gathersi<VEC_GATHER_MODE:mode>_2): Likewise.
(*avx2_gatherdi<VEC_GATHER_MODE:mode>): Likewise.
(*avx2_gatherdi<VEC_GATHER_MODE:mode>_2): Likewise.
(*avx2_gatherdi<VI4F_256:mode>_3): Likewise.
(*avx2_gatherdi<VI4F_256:mode>_4): Likewise.
(avx_vbroadcastf128_<mode>): Restrict non-egpr alternative to
noavx512vl, set its constraint to jm and set attr_gpr32 to 0.
(vec_set_lo_<mode><mask_name>): Likewise.
(vec_set_lo_<mode><mask_name>): Likewise for SF/SI modes.
(vec_set_hi_<mode><mask_name>): Likewise.
(vec_set_hi_<mode><mask_name>): Likewise for SF/SI modes.
(vec_set_hi_<mode>): Likewise.
(vec_set_lo_<mode>): Likewise.
(avx2_set_hi_v32qi): Likewise.

Co-authored-by: Hongyu Wang <hongyu.wang@intel.com>
Co-authored-by: Hongtao Liu <hongtao.liu@intel.com>

[APX_EGPR] Handle legacy insns that only support GPR16 (4/5)

The APX enabled hardware should also be AVX10 enabled, thus for map2/3 insns
with evex counterpart, we assume auto promotion to EGPR under APX_F if the
insn uses GPR32. So for below insns, we disabled EGPR usage for their sse
mnenomics, while allowing egpr generation of their v prefixed mnemonics.

insn list:
1. pabsb/pabsw/pabsd
2. pextrb/pextrw/pextrd/pextrq
3. pinsrb/pinsrd/pinsrq
4. pshufb
5. extractps/insertps
6. pmaddubsw
7. pmulhrsw
8. packusdw
9. palignr
10. movntdqa
11. mpsadbw
12. pmuldq/pmulld
13. pmaxsb/pmaxsd, pminsb/pminsd
    pmaxud/pmaxuw, pminud/pminuw
14. (pmovsxbw/pmovsxbd/pmovsxbq,
     pmovsxwd/pmovsxwq, pmovsxdq
     pmovzxbw/pmovzxbd/pmovzxbq,
     pmovzxwd/pmovzxwq, pmovzxdq)
15. aesdec/aesdeclast, aesenc/aesenclast
16. pclmulqdq
17. gf2p8affineqb/gf2p8affineinvqb/gf2p8mulb

gcc/ChangeLog:

* config/i386/i386.md (*movhi_internal): Split out non-gpr
supported pextrw with mem constraint to avx/noavx alternatives,
set jm and attr gpr32 0 to the noavx alternative.
(*mov<mode>_internal): Likewise.
* config/i386/mmx.md (mmx_pshufbv8qi3): Change "r/m/Bm" to
"jr/jm/ja" and set_attr gpr32 0 for noavx alternative.
(mmx_pshufbv4qi3): Likewise.
(*mmx_pinsrd): Likewise.
(*mmx_pinsrb): Likewise.
(*pinsrb): Likewise.
(mmx_pshufbv8qi3): Likewise.
(mmx_pshufbv4qi3): Likewise.
(@sse4_1_insertps_<mode>): Likewise.
(*mmx_pextrw): Split altrenatives and map non-EGPR
constraints, attr_gpr32 and attr_isa to noavx mnemonics.
(*movv2qi_internal): Likewise.
(*pextrw): Likewise.
(*mmx_pextrb): Likewise.
(*mmx_pextrb_zext): Likewise.
(*pextrb): Likewise.
(*pextrb_zext): Likewise.
(vec_extractv2si_1): Likewise.
(vec_extractv2si_1_zext): Likewise.
* config/i386/sse.md: (vi128_h_r): New mode attr for
pinsr{bw}/pextr{bw} with reg operand.
(*abs<mode>2): Split altrenatives and %v in mnemonics, map
non-EGPR constraints, gpr32 and isa attrs to noavx mnemonics.
(*vec_extract<mode>): Likewise.
(*vec_extract<mode>): Likewise for HFBF pattern.
(*vec_extract<PEXTR_MODE12:mode>_zext): Likewise.
(*vec_extractv4si_1): Likewise.
(*vec_extractv4si_zext): Likewise.
(*vec_extractv2di_1): Likewise.
(*vec_concatv2si_sse4_1): Likewise.
(<sse2p4_1>_pinsr<ssemodesuffix>): Likewise.
(vec_concatv2di): Likewise.
(*sse4_1_<code>v2qiv2di2<mask_name>_1): Likewise.
(ssse3_avx2>_pshufb<mode>3<mask_name>): Change "r/m/Bm" to
"jr/jm/ja" and set_attr gpr32 0 for noavx alternative, split
%v for avx/noavx alternatives if necessary.
(*vec_concatv2sf_sse4_1): Likewise.
(*sse4_1_extractps): Likewise.
(vec_set<mode>_0): Likewise for VI4F_128.
(*vec_setv4sf_sse4_1): Likewise.
(@sse4_1_insertps<mode>): Likewise.
(ssse3_pmaddubsw128): Likewise.
(*<ssse3_avx2>_pmulhrsw<mode>3<mask_name>): Likewise.
(<sse4_1_avx2>_packusdw<mask_name>): Likewise.
(<ssse3_avx2>_palignr<mode>): Likewise.
(<vi8_sse4_1_avx2_avx512>_movntdqa): Likewise.
(<sse4_1_avx2>_mpsadbw): Likewise.
(*sse4_1_mulv2siv2di3<mask_name>): Likewise.
(*<sse4_1_avx2>_mul<mode>3<mask_name>): Likewise.
(*sse4_1_<code><mode>3<mask_name>): Likewise.
(*<code>v8hi3): Likewise.
(*<code>v16qi3): Likewise.
(*sse4_1_<code>v8qiv8hi2<mask_name>_1): Likewise.
(*sse4_1_zero_extendv8qiv8hi2_3): Likewise.
(*sse4_1_zero_extendv8qiv8hi2_4): Likewise.
(*sse4_1_<code>v4qiv4si2<mask_name>_1): Likewise.
(*sse4_1_<code>v4hiv4si2<mask_name>_1): Likewise.
(*sse4_1_zero_extendv4hiv4si2_3): Likewise.
(*sse4_1_zero_extendv4hiv4si2_4): Likewise.
(*sse4_1_<code>v2hiv2di2<mask_name>_1): Likewise.
(*sse4_1_<code>v2siv2di2<mask_name>_1): Likewise.
(*sse4_1_zero_extendv2siv2di2_3): Likewise.
(*sse4_1_zero_extendv2siv2di2_4): Likewise.
(aesdec): Likewise.
(aesdeclast): Likewise.
(aesenc): Likewise.
(aesenclast): Likewise.
(pclmulqdq): Likewise.
(vgf2p8affineinvqb_<mode><mask_name>): Likewise.
(vgf2p8affineqb_<mode><mask_name>): Likewise.
(vgf2p8mulb_<mode><mask_name>): Likewise.

Co-authored-by: Hongyu Wang <hongyu.wang@intel.com>
Co-authored-by: Hongtao Liu <hongtao.liu@intel.com>

[APX EGPR] Handle legacy insns that only support GPR16 (3/5)

Disable EGPR usage for below legacy insns in opcode map2/3 that have vex
but no evex counterpart.

insn list:
1. phminposuw/vphminposuw
2. ptest/vptest
3. roundps/vroundps, roundpd/vroundpd,
   roundss/vroundss, roundsd/vroundsd
4. pcmpestri/vpcmpestri, pcmpestrm/vpcmpestrm
5. pcmpistri/vpcmpistri, pcmpistrm/vpcmpistrm
6. aesimc/vaesimc, aeskeygenassist/vaeskeygenassist

gcc/ChangeLog:

* config/i386/i386-protos.h (x86_evex_reg_mentioned_p): New
prototype.
* config/i386/i386.cc (x86_evex_reg_mentioned_p): New
function.
* config/i386/i386.md (sse4_1_round<mode>2): Set attr gpr32 0
and constraint jm to all non-evex alternatives, adjust
alternative outputs if evex reg is mentioned.
* config/i386/sse.md (<sse4_1>_ptest<mode>): Set attr gpr32 0
and constraint jm/ja to all non-evex alternatives.
(ptesttf2): Likewise.
(<sse4_1>_round<ssemodesuffix><avxsizesuffix): Likewise.
(sse4_1_round<ssescalarmodesuffix>): Likewise.
(sse4_2_pcmpestri): Likewise.
(sse4_2_pcmpestrm): Likewise.
(sse4_2_pcmpestr_cconly): Likewise.
(sse4_2_pcmpistr): Likewise.
(sse4_2_pcmpistri): Likewise.
(sse4_2_pcmpistrm): Likewise.
(sse4_2_pcmpistr_cconly): Likewise.
(aesimc): Likewise.
(aeskeygenassist): Likewise.

gcc/testsuite/ChangeLog:

* gcc.target/i386/apx-legacy-insn-check-norex2.c: Add intrinsic
tests.

Co-authored-by: Hongyu Wang <hongyu.wang@intel.com>
Co-authored-by: Hongtao Liu <hongtao.liu@intel.com>

[APX EGPR] Handle legacy insns that only support GPR16 (2/5)

These legacy insns in opcode map2/3 have vex but no evex
counterpart, disable EGPR for them by adjusting alternatives and
attr_gpr32.

insn list:
1. phaddw/vphaddw, phaddd/vphaddd, phaddsw/vphaddsw
2. phsubw/vphsubw, phsubd/vphsubd, phsubsw/vphsubsw
3. psignb/vpsginb, psignw/vpsignw, psignd/vpsignd
4. blendps/vblendps, blendpd/vblendpd
5. blendvps/vblendvps, blendvpd/vblendvpd
6. pblendvb/vpblendvb, pblendw/vpblendw
7. mpsadbw/vmpsadbw
8. dpps/vddps, dppd/vdppd
9. pcmpeqq/vpcmpeqq, pcmpgtq/vpcmpgtq

gcc/ChangeLog:

* config/i386/sse.md (avx2_ph<plusminus_mnemonic>wv16hi3): Set
attr gpr32 0 and constraint jm/ja to all mem alternatives.
(ssse3_ph<plusminus_mnemonic>wv8hi3): Likewise.
(ssse3_ph<plusminus_mnemonic>wv4hi3): Likewise.
(avx2_ph<plusminus_mnemonic>dv8si3): Likewise.
(ssse3_ph<plusminus_mnemonic>dv4si3): Likewise.
(ssse3_ph<plusminus_mnemonic>dv2si3): Likewise.
(<ssse3_avx2>_psign<mode>3): Likewise.
(ssse3_psign<mode>3): Likewise.
(<sse4_1>_blend<ssemodesuffix><avxsizesuffix): Likewise.
(<sse4_1>_blendv<ssemodesuffix><avxsizesuffix): Likewise.
(*<sse4_1>_blendv<ssemodesuffix><avxsizesuffix>_lt): Likewise.
(*<sse4_1>_blendv<ssefltmodesuff)ix><avxsizesuffix>_not_ltint: Likewise.
(<sse4_1>_dp<ssemodesuffix><avxsizesuffix>): Likewise.
(<sse4_1_avx2>_mpsadbw): Likewise.
(<sse4_1_avx2>_pblendvb): Likewise.
(*<sse4_1_avx2>_pblendvb_lt): Likewise.
(sse4_1_pblend<ssemodesuffix>): Likewise.
(*avx2_pblend<ssemodesuffix>): Likewise.
(avx2_permv2ti): Likewise.
(*avx_vperm2f128<mode>_nozero): Likewise.
(*avx2_eq<mode>3): Likewise.
(*sse4_1_eqv2di3): Likewise.
(sse4_2_gtv2di3): Likewise.
(avx2_gt<mode>3): Likewise.

gcc/testsuite/ChangeLog:

* gcc.target/i386/apx-legacy-insn-check-norex2.c: Add
sse/vex intrinsic tests.

Co-authored-by: Hongyu Wang <hongyu.wang@intel.com>
Co-authored-by: Hongtao Liu <hongtao.liu@intel.com>

[APX EGPR] Handle legacy insn that only support GPR16 (1/5)

These legacy insn in opcode map0/1 only support GPR16,
and do not have vex/evex counterpart, directly adjust constraints and
add gpr32 attr to patterns.

insn list:
1. xsave/xsave64, xrstor/xrstor64
2. xsaves/xsaves64, xrstors/xrstors64
3. xsavec/xsavec64
4. xsaveopt/xsaveopt64
5. fxsave64/fxrstor64

gcc/ChangeLog:

* config/i386/i386.md (<xsave>): Set attr gpr32 0 and constraint
jm.
(<xsave>_rex64): Likewise.
(<xrstor>_rex64): Likewise.
(<xrstor>64): Likewise.
(fxsave64): Likewise.
(fxstore64): Likewise.

gcc/testsuite/ChangeLog:

* lib/target-supports.exp: Add apxf check.
* gcc.target/i386/apx-legacy-insn-check-norex2.c: New test.
* gcc.target/i386/apx-legacy-insn-check-norex2-asm.c: New assembler test.

Co-authored-by: Hongyu Wang <hongyu.wang@intel.com>
Co-authored-by: Hongtao Liu <hongtao.liu@intel.com>

[APX EGPR] Handle GPR16 only vector move insns

For vector move insns like vmovdqa/vmovdqu, their evex counterparts
requrire explicit suffix 64/32/16/8. The usage of these instruction
are prohibited under AVX10_1 or AVX512F, so for we select
vmovaps/vmovups for vector load/store insns that contains EGPR if
ther is no AVX512VL, and keep the original move insn selection
otherwise.

gcc/ChangeLog:

* config/i386/i386.cc (ix86_get_ssemov): Check if egpr is used,
adjust mnemonic for vmovduq/vmovdqa.
* config/i386/sse.md (*<extract_type>_vinsert<shuffletype><extract_suf>_0):
Check if egpr is used, adjust mnemonic for vmovdqu/vmovdqa.
(avx_vec_concat<mode>): Likewise, and separate alternative 0 to
avx_noavx512f.

Co-authored-by: Kong Lingling <lingling.kong@intel.com>
Co-authored-by: Hongtao Liu <hongtao.liu@intel.com>

[APX EGPR] Map reg/mem constraints in inline asm to non-EGPR constraint.

In inline asm, we do not know if the insn can use EGPR, so disable EGPR
usage by default via mapping the common reg/mem constraint to non-EGPR
constraints.

The full list of mapping goes like

  "g" -> "jrjmi"
  "r" -> "jr"
  "m" -> "jm"
  "<" -> "j<"
  ">" -> "j>"
  "o" -> "jo"
  "V" -> "jV"
  "p" -> "jp"
  "Bm" -> "ja

For memory constraints, we add an option -mapx-inline-asm-use-gpr32
to allow/disallow gpr32 usage in any memory related constraints, as
base_reg_class/index_reg_class cannot aware whether the asm insn
support gpr32 or not.

gcc/ChangeLog:

* config/i386/i386.cc (map_egpr_constraints): New funciton to
map common constraints to EGPR prohibited constraints.
(ix86_md_asm_adjust): Calls map_egpr_constraints.
* config/i386/i386.opt: Add option mapx-inline-asm-use-gpr32.

gcc/testsuite/ChangeLog:

* gcc.target/i386/apx-inline-gpr-norex2.c: New test.

Co-authored-by: Hongyu Wang <hongyu.wang@intel.com>
Co-authored-by: Hongtao Liu <hongtao.liu@intel.com>

[APX EGPR] Add backend hook for base_reg_class/index_reg_class.

Add backend helper functions to verify if a rtx_insn can adopt EGPR to
its base/index reg of memory operand. The verification rule goes like
  1. For asm insn, enable/disable EGPR by ix86_apx_inline_asm_use_gpr32.
  2. Disable EGPR for unrecognized insn.
  3. If which_alternative is not decided, loop through enabled alternatives
  and check its attr_gpr32. Only enable EGPR when all enabled
  alternatives has attr_gpr32 = 1.
  4. If which_alternative is decided, enable/disable EGPR by its corresponding
  attr_gpr32.

gcc/ChangeLog:

* config/i386/i386-protos.h (ix86_insn_base_reg_class): New
prototype.
(ix86_regno_ok_for_insn_base_p): Likewise.
(ix86_insn_index_reg_class): Likewise.
* config/i386/i386.cc (ix86_memory_address_use_extended_reg_class_p):
New helper function to scan the insn.
(ix86_insn_base_reg_class): New function to choose BASE_REG_CLASS.
(ix86_regno_ok_for_insn_base_p): Likewise for base regno.
(ix86_insn_index_reg_class): Likewise for INDEX_REG_CLASS.
* config/i386/i386.h (INSN_BASE_REG_CLASS): Define.
(REGNO_OK_FOR_INSN_BASE_P): Likewise.
(INSN_INDEX_REG_CLASS): Likewise.
(enum reg_class): Add INDEX_GPR16.
(GENERAL_GPR16_REGNO_P): Define.
* config/i386/i386.md (gpr32): New attribute.

Co-authored-by: Hongyu Wang <hongyu.wang@intel.com>
Co-authored-by: Hongtao Liu <hongtao.liu@intel.com>

[APX EGPR] Add register and memory constraints that disallow EGPR

For APX, as we extended the GENERAL_REG_CLASS, new constraints are
needed to restrict insns that cannot adopt EGPR either in its reg or
memory operands. We added a series of constraints for general/backend
ones that related to GPR usage. All of them are prefixed with "j" to
indicate the constraints does not allow EGPR.

gcc/ChangeLog:

* config/i386/constraints.md (jr): New register constraint
that prohibits EGPR.
(jR): Constraint that force usage of EGPR.
(jm): New memory constraint that prohibits EGPR.
(ja): Likewise for Bm constraint.
(jb): Likewise for Tv constraint.
(j<): New auto-dec memory constraint that prohibits EGPR.
(j>): Likewise for ">" constraint.
(jo): Likewise for "o" constraint.
(jv): Likewise for "V" constraint.
(jp): Likewise for "p" constraint.
* config/i386/i386.h (enum reg_class): Add new reg class
GENERAL_GPR16.

Co-authored-by: Hongyu Wang <hongyu.wang@intel.com>
Co-authored-by: Hongtao Liu <hongtao.liu@intel.com>

[APX EGPR] Add 16 new integer general purpose registers

Extend GENERAL_REGS with extra r16-r31 registers like REX registers,
named as REX2 registers. They will only be enabled under
TARGET_APX_EGPR.

gcc/ChangeLog:

* config/i386/i386-protos.h (x86_extended_rex2reg_mentioned_p):
New function prototype.
* config/i386/i386.cc (regclass_map): Add mapping for 16 new
general registers.
(debugger64_register_map): Likewise.
(ix86_conditional_register_usage): Clear REX2 register when APX
disabled.
(ix86_code_end): Add handling for REX2 reg.
(print_reg): Likewise.
(ix86_output_jmp_thunk_or_indirect): Likewise.
(ix86_output_indirect_branch_via_reg): Likewise.
(ix86_attr_length_vex_default): Likewise.
(ix86_emit_save_regs): Adjust to allow saving r31.
(ix86_register_priority): Set REX2 reg priority same as REX.
(x86_extended_reg_mentioned_p): Add check for REX2 regs.
(x86_extended_rex2reg_mentioned_p): New function.
* config/i386/i386.h (CALL_USED_REGISTERS): Add new extended
registers.
(REG_ALLOC_ORDER): Likewise.
(FIRST_REX2_INT_REG): Define.
(LAST_REX2_INT_REG): Ditto.
(GENERAL_REGS): Add 16 new registers.
(INT_SSE_REGS): Likewise.
(FLOAT_INT_REGS): Likewise.
(FLOAT_INT_SSE_REGS): Likewise.
(INT_MASK_REGS): Likewise.
(ALL_REGS):Likewise.
(REX2_INT_REG_P): Define.
(REX2_INT_REGNO_P): Ditto.
(GENERAL_REGNO_P): Add REX2_INT_REGNO_P.
(REGNO_OK_FOR_INDEX_P): Ditto.
(REG_OK_FOR_INDEX_NONSTRICT_P): Add new extended registers.
* config/i386/i386.md: Add 16 new integer general
registers.

gcc/testsuite/ChangeLog:

* gcc.target/i386/apx-egprs-names.c: New test.
* gcc.target/i386/apx-spill_to_egprs-1.c: Likewise.
* gcc.target/i386/apx-interrupt-1.c: Likewise.

Co-authored-by: Hongyu Wang <hongyu.wang@intel.com>
Co-authored-by: Hongtao Liu <hongtao.liu@intel.com>

[APX_EGPR] Initial support for APX_F

Add -mapx-features= enumeration to separate subfeatures of APX_F.
-mapxf is treated same as previous ISA flag, while it sets
-mapx-features=apx_all that enables all subfeatures.

gcc/ChangeLog:

* common/config/i386/cpuinfo.h (XSTATE_APX_F): New macro.
(XCR_APX_F_ENABLED_MASK): Likewise.
(get_available_features): Detect APX_F under
* common/config/i386/i386-common.cc (OPTION_MASK_ISA2_APX_F_SET): New.
(OPTION_MASK_ISA2_APX_F_UNSET): Likewise.
(ix86_handle_option): Handle -mapxf.
* common/config/i386/i386-cpuinfo.h (FEATURE_APX_F): New.
* common/config/i386/i386-isas.h: Add entry for APX_F.
* config/i386/cpuid.h (bit_APX_F): New.
* config/i386/i386.h (bit_APX_F): (TARGET_APX_EGPR,
TARGET_APX_PUSH2POP2, TARGET_APX_NDD): New define.
* config/i386/i386-opts.h (enum apx_features): New enum.
* config/i386/i386-isa.def (APX_F): New DEF_PTA.
* config/i386/i386-options.cc (ix86_function_specific_save):
Save ix86_apx_features.
(ix86_function_specific_restore): Restore it.
(ix86_valid_target_attribute_inner_p): Add mapxf.
(ix86_option_override_internal): Set ix86_apx_features for PTA
and TARGET_APX_F. Also reports error when APX_F is set but not
having TARGET_64BIT.
* config/i386/i386.opt: (-mapxf): New ISA flag option.
(-mapx=): New enumeration option.
(apx_features): New enum type.
(apx_none): New enum value.
(apx_egpr): Likewise.
(apx_push2pop2): Likewise.
(apx_ndd): Likewise.
(apx_all): Likewise.
* doc/invoke.texi: Document mapxf.

gcc/testsuite/ChangeLog:

* gcc.target/i386/apx-1.c: New test.

Co-authored-by: Hongyu Wang <hongyu.wang@intel.com>
Co-authored-by: Hongtao Liu <hongtao.liu@intel.com>

[APX EGPR] middle-end: Add index_reg_class with insn argument.

Like base_reg_class, INDEX_REG_CLASS also does not support backend insn.
Add index_reg_class with insn argument for lra/reload usage.

gcc/ChangeLog:

* addresses.h (index_reg_class): New wrapper function like
base_reg_class.
* doc/tm.texi: Document INSN_INDEX_REG_CLASS.
* doc/tm.texi.in: Ditto.
* lra-constraints.cc (index_part_to_reg): Pass index_class.
(process_address_1): Calls index_reg_class with curr_insn and
replace INDEX_REG_CLASS with its return value index_cl.
* reload.cc (find_reloads_address): Likewise.
(find_reloads_address_1): Likewise.

Co-authored-by: Kong Lingling <lingling.kong@intel.com>
Co-authored-by: Hongtao Liu <hongtao.liu@intel.com>

[APX EGPR] middle-end: Add insn argument to base_reg_class

Current reload infrastructure does not support selective base_reg_class
for backend insn. Add new macros with insn parameters to base_reg_class
for lra/reload usage.

gcc/ChangeLog:

* addresses.h (base_reg_class): Add insn argument and new macro
INSN_BASE_REG_CLASS.
(regno_ok_for_base_p_1): Add insn argument and new macro
REGNO_OK_FOR_INSN_BASE_P.
(regno_ok_for_base_p): Add insn argument and parse to ok_for_base_p_1.
* doc/tm.texi: Document INSN_BASE_REG_CLASS and
REGNO_OK_FOR_INSN_BASE_P.
* doc/tm.texi.in: Ditto.
* lra-constraints.cc (process_address_1): Pass insn to
base_reg_class.
(curr_insn_transform): Ditto.
* reload.cc (find_reloads): Ditto.
(find_reloads_address): Ditto.
(find_reloads_address_1): Ditto.
(find_reloads_subreg_address): Ditto.
* reload1.cc (maybe_fix_stack_asms): Ditto.

Co-authored-by: Hongyu Wang <hongyu.wang@intel.com>
Co-authored-by: Hongtao Liu <hongtao.liu@intel.com>

RISC-V: Add more run test for FP rounding autovec

For _Float16 types, add run test for:
* ceil
* floor
* nearbyint
* rint
* round
* roundeven
* trunc

For float and double, add run test for:
* roundeven

The zfa extension is required for these run test cases, the simulation
target_board may look like below for rv64.

target_board="riscv-sim/-march=rv64gcv_zfa_zfh/-mabi=lp64d/-mcmodel=medlow"

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/rvv.exp: Add zfa for building.
* gcc.target/riscv/rvv/autovec/unop/math-ceil-run-0.c: New test.
* gcc.target/riscv/rvv/autovec/unop/math-floor-run-0.c: New test.
* gcc.target/riscv/rvv/autovec/unop/math-nearbyint-run-0.c: New test.
* gcc.target/riscv/rvv/autovec/unop/math-rint-run-0.c: New test.
* gcc.target/riscv/rvv/autovec/unop/math-round-run-0.c: New test.
* gcc.target/riscv/rvv/autovec/unop/math-roundeven-run-0.c: New test.
* gcc.target/riscv/rvv/autovec/unop/math-roundeven-run-1.c: New test.
* gcc.target/riscv/rvv/autovec/unop/math-roundeven-run-2.c: New test.
* gcc.target/riscv/rvv/autovec/unop/math-trunc-run-0.c: New test.

Signed-off-by: Pan Li <pan2.li@intel.com>

rs6000: use mtvsrws to move sf from si p9

As mentioned in PR108338, on p9, we could use mtvsrws to implement
the bitcast from SI to SF (or lowpart DI to SF).

For example:
  *(long long*)buff = di;
  float f = *(float*)(buff);

"sldi 9,3,32 ; mtvsrd 1,9 ; xscvspdpn 1,1" is generated.
A better one would be "mtvsrws 1,3 ; xscvspdpn 1,1".

PR target/108338

gcc/ChangeLog:

* config/rs6000/rs6000.md (movsf_from_si): Update to generate mtvsrws
for P9.

gcc/testsuite/ChangeLog:

* gcc.target/powerpc/pr108338.c: Updated to check mtvsrws for p9.

rs6000: optimize moving to sf from highpart di

Currently, we have the pattern "movsf_from_si2" which was trying
to support moving high part DI to SF.

But current pattern only accepts "ashiftrt":
XX:SF=bitcast:SF(subreg(YY:DI>>32),0), but actually "lshiftrt" should
also be ok.
And current pattern only supports BE.

Here, updating the pattern to support BE and "lshiftrt".

PR target/108338

gcc/ChangeLog:

* config/rs6000/predicates.md (lowpart_subreg_operator): New
define_predicate.
* config/rs6000/rs6000.md (any_rshift): New code_iterator.
(movsf_from_si2): Rename to ...
(movsf_from_si2_<code>): ... this.

gcc/testsuite/ChangeLog:

* gcc.target/powerpc/pr108338.c: New test.

RISC-V: Bugfix for legitimize address PR/111634

Given we have RTL as below.

(plus:DI (mult:DI (reg:DI 138 [ g.4_6 ])
                  (const_int 8 [0x8]))
         (lo_sum:DI (reg:DI 167)
                    (symbol_ref:DI ("f") [flags 0x86] <var_decl 0x7fa96ea1cc60 f>)
))

When handling (plus (plus (mult (a) (mem_shadd_constant)) (fp)) (C)) case,
the fp will be the lo_sum operand as above. We have assumption that the fp
is reg but actually not here. It will have ICE when building with option
--enable-checking=rtl.

This patch would like to fix it by adding the REG_P to ensure the operand
is a register. The test case gcc/testsuite/gcc.dg/pr109417.c covered this
fix when build with --enable-checking=rtl.

PR target/111634

gcc/ChangeLog:

* config/riscv/riscv.cc (riscv_legitimize_address): Ensure
object is a REG before extracting its' REGNO.

Signed-off-by: Pan Li <pan2.li@intel.com>

RISC-V: Fix scan-assembler-times of RVV test case

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/vsetvl/vlmax_back_prop-25.c: Adjust assembler times.
* gcc.target/riscv/rvv/vsetvl/vlmax_back_prop-26.c: Ditto.

Daily bump.

i386: Implement doubleword shift left by 1 bit using add+adc.

This patch tweaks the i386 back-end's ix86_split_ashl to implement
doubleword left shifts by 1 bit, using an add followed by an add-with-carry
(i.e. a doubleword x+x) instead of using the x86's shld instruction.
The replacement sequence both requires fewer bytes and is faster on
both Intel and AMD architectures (from Agner Fog's latency tables and
confirmed by my own micro-benchmarking).

For the test case:
__int128 foo(__int128 x) { return x << 1; }

with -O2 we previously generated:

foo: movq    %rdi, %rax
        movq    %rsi, %rdx
        shldq   $1, %rdi, %rdx
        addq    %rdi, %rax
        ret

with this patch we now generate:

foo: movq    %rdi, %rax
        movq    %rsi, %rdx
        addq    %rdi, %rax
        adcq    %rsi, %rdx
        ret

This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
and make -k check, both with and without --target_board=unix{-m32}
with no new failures.  Ok for mainline?

2023-10-06  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* config/i386/i386-expand.cc (ix86_split_ashl): Split shifts by
one into add3_cc_overflow_1 followed by add3_carry.
* config/i386/i386.md (@add<mode>3_cc_overflow_1): Renamed from
"*add<mode>3_cc_overflow_1" to provide generator function.

gcc/testsuite/ChangeLog
* gcc.target/i386/ashldi3-2.c: New 32-bit test case.
* gcc.target/i386/ashlti3-3.c: New 64-bit test case.

Makefile.tpl: disable -Werror for feedback stage [PR111663]

Without the change profiled bootstrap fails for various warnings on
master branch as:

    $ ../gcc/configure
    $ make profiledbootstrap
    ...
    gcc/genmodes.cc: In function ‘int main(int, char**)’:
    gcc/genmodes.cc:2152:1: error: ‘gcc/build/genmodes.gcda’ profile count data file not found [-Werror=missing-profile]
    ...
    gcc/gengtype-parse.cc: In function ‘void parse_error(const char*, ...)’:
    gcc/gengtype-parse.cc:142:21: error: ‘%s’ directive argument is null [-Werror=format-overflow=]

The change removes -Werror just like autofeedback does today.

/

PR bootstrap/111663
* Makefile.tpl (STAGEfeedback_CONFIGURE_FLAGS): Disable -Werror.
* Makefile.in: Regenerate.