This approach addresses issues with exceptions, as well as issues with
denormal/invalid values. An obvious exception to the rule is a division,
where the value != 0.0 should be loaded into the upper half of the
denominator to avoid division by zero exception.
The patch effectively tightens the solution from PR95046 by clearing upper
halves of all operand registers before every potentially trapping instruction.
The testcase:
The implementation emits V4SFmode operation, so we can remove all "emulated"
SSE2 V2SFmode trapping instructions and remove "emulated" SSE2 V2SFmode
alternatives from 3dNOW! insn patterns.
PR target/110762
gcc/ChangeLog:
* config/i386/i386.md (plusminusmult): New code iterator.
* config/i386/mmx.md (mmxdoublevecmode): New mode attribute.
(movq_<mode>_to_sse): New expander.
(<plusminusmult:insn>v2sf3): Macroize expander from addv2sf3,
subv2sf3 and mulv2sf3 using plusminusmult code iterator. Rewrite
as a wrapper around V4SFmode operation.
(mmx_addv2sf3): Change operand 1 and operand 2 predicates to
nonimmediate_operand.
(*mmx_addv2sf3): Remove SSE alternatives. Change operand 1 and
operand 2 predicates to nonimmediate_operand.
(mmx_subv2sf3): Change operand 2 predicate to nonimmediate_operand.
(mmx_subrv2sf3): Change operand 1 predicate to nonimmediate_operand.
(*mmx_subv2sf3): Remove SSE alternatives. Change operand 1 and
operand 2 predicates to nonimmediate_operand.
(mmx_mulv2sf3): Change operand 1 and operand 2 predicates to
nonimmediate_operand.
(*mmx_mulv2sf3): Remove SSE alternatives. Change operand 1 and
operand 2 predicates to nonimmediate_operand.
(divv2sf3): Rewrite as a wrapper around V4SFmode operation.
(<smaxmin:code>v2sf3): Ditto.
(mmx_<smaxmin:code>v2sf3): Change operand 1 and operand 2
predicates to nonimmediate_operand.
(*mmx_<smaxmin:code>v2sf3): Remove SSE alternatives. Change
operand 1 and operand 2 predicates to nonimmediate_operand.
(mmx_ieee_<ieee_maxmin>v2sf3): Ditto.
(sqrtv2sf2): Rewrite as a wrapper around V4SFmode operation.
(*mmx_haddv2sf3_low): Ditto.
(*mmx_hsubv2sf3_low): Ditto.
(vec_addsubv2sf3): Ditto.
(*mmx_maskcmpv2sf3_comm): Remove.
(*mmx_maskcmpv2sf3): Remove.
(vec_cmpv2sfv2si): Rewrite as a wrapper around V4SFmode operation.
(vcond<V2FI:mode>v2sf): Ditto.
(fmav2sf4): Ditto.
(fmsv2sf4): Ditto.
(fnmav2sf4): Ditto.
(fnmsv2sf4): Ditto.
(fix_truncv2sfv2si2): Ditto.
(fixuns_truncv2sfv2si2): Ditto.
(mmx_fix_truncv2sfv2si2): Remove SSE alternatives.
Change operand 1 predicate to nonimmediate_operand.
(floatv2siv2sf2): Rewrite as a wrapper around V4SFmode operation.
(floatunsv2siv2sf2): Ditto.
(mmx_floatv2siv2sf2): Remove SSE alternatives.
Change operand 1 predicate to nonimmediate_operand.
(nearbyintv2sf2): Rewrite as a wrapper around V4SFmode operation.
(rintv2sf2): Ditto.
(lrintv2sfv2si2): Ditto.
(ceilv2sf2): Ditto.
(lceilv2sfv2si2): Ditto.
(floorv2sf2): Ditto.
(lfloorv2sfv2si2): Ditto.
(btruncv2sf2): Ditto.
(roundv2sf2): Ditto.
(lroundv2sfv2si2): Ditto.
(*mmx_roundv2sf2): Remove.
Richard Biener [Wed, 26 Jul 2023 07:28:10 +0000 (09:28 +0200)]
tree-optimization/110799 - fix bug in code hoisting
Code hoisting part of GIMPLE PRE failed to adjust the TBAA behavior
of common loads in the case the alias set of the ref was the same
but the base alias set was not. It also failed to adjust the
base behavior, assuming it would match. The following plugs this
hole.
PR tree-optimization/110799
* tree-ssa-pre.cc (compute_avail): More thoroughly match
up TBAA behavior of redundant loads.
Jakub Jelinek [Wed, 26 Jul 2023 08:50:50 +0000 (10:50 +0200)]
range-op-float: Fix up -frounding-math frange_arithmetic +- handling [PR110755]
IEEE754 says that x + (-x) and x - x result in +0 in all rounding modes
but rounding towards negative infinity, in which case the result is -0
for all finite x. x + x and x - (-x) if it is zero retain sign of x.
Now, range_arithmetic implements the normal rounds to even rounding,
and as the addition or subtraction in those cases is exact, we don't do any
further rounding etc. and e.g. on the testcase below distilled from glibc
compute a range [+0, +INF], which is fine for -fno-rounding-math or
if we'd have a guarantee that those statements aren't executed with rounding
towards negative infinity.
I believe it is only +- which has this problematic behavior and I think
it is best to deal with it in frange_arithmetic; if we know -frounding-math
is on, it is x + (-x) or x - x and we are asked to round to negative
infinity (i.e. want low bound rather than high bound), change +0 result to
-0.
2023-07-26 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/110755
* range-op-float.cc (frange_arithmetic): Change +0 result to -0
for PLUS_EXPR or MINUS_EXPR if -frounding-math, inf is negative and
it is exact op1 + (-op1) or op1 - op1.
Kewen Lin [Wed, 26 Jul 2023 08:42:29 +0000 (03:42 -0500)]
rs6000: Correct vsx operands output for xxeval [PR110741]
PR110741 exposes one issue that we didn't use the correct
character for vsx operands in output operand substitution,
consequently it can map to the wrong registers which hold
some unexpected values.
PR target/110741
gcc/ChangeLog:
* config/rs6000/vsx.md (define_insn xxeval): Correct vsx
operands output with "x".
It looks like we missed out on bitmasks for unary operators because we
were using bit_value_binop exclusively. This patch hands off to
bit_value_unop when appropriate, thus allowing us to handle ABS and
BIT_NOT_EXPR, and others. Follow-up patches will add the tweaks for the
range-ops entries themselves.
c++: Track lifetimes in constant evaluation [PR70331,PR96630,PR98675]
This adds rudimentary lifetime tracking in C++ constexpr contexts,
allowing the compiler to report errors with using values after their
backing has gone out of scope. We don't yet handle other ways of
accessing values outside their lifetime (e.g. following explicit
destructor calls).
PR c++/96630
PR c++/98675
PR c++/70331
gcc/cp/ChangeLog:
* constexpr.cc (constexpr_global_ctx::is_outside_lifetime): New
function.
(constexpr_global_ctx::get_value): Don't return expired values.
(constexpr_global_ctx::get_value_ptr): Likewise.
(constexpr_global_ctx::remove_value): Mark value outside
lifetime.
(outside_lifetime_error): New function.
(cxx_eval_call_expression): No longer track save_exprs.
(cxx_eval_loop_expr): Likewise.
(cxx_eval_constant_expression): Add checks for outside lifetime
values. Remove local variables at end of bind exprs, and
temporaries after cleanup points.
gcc/testsuite/ChangeLog:
* g++.dg/cpp1y/constexpr-lifetime1.C: New test.
* g++.dg/cpp1y/constexpr-lifetime2.C: New test.
* g++.dg/cpp1y/constexpr-lifetime3.C: New test.
* g++.dg/cpp1y/constexpr-lifetime4.C: New test.
* g++.dg/cpp1y/constexpr-lifetime5.C: New test.
* g++.dg/cpp1y/constexpr-lifetime6.C: New test.
c++: Prevent dangling pointers from becoming nullptr in constexpr [PR110619]
Currently, when typeck discovers that a return statement will refer to a
local variable it rewrites to return a null pointer. This causes the
error messages for using the return value in a constant expression to be
unhelpful, especially for reference return values, and is also a visible
change to otherwise valid code (as in the linked PR).
The transformation is nonetheless important, however, both as a safety
guard against attackers being able to gain a handle to other data on the
stack, and to prevent duplicate warnings from later null-dereference
warning passes.
As such, this patch just delays the transformation until cp_genericize,
after constexpr function definitions have been generated.
PR c++/110619
gcc/cp/ChangeLog:
* cp-gimplify.cc (cp_genericize_r): Transform RETURN_EXPRs to
not return dangling pointers.
* cp-tree.h (RETURN_EXPR_LOCAL_ADDR_P): New flag.
(check_return_expr): Add a new parameter.
* semantics.cc (finish_return_stmt): Set flag on RETURN_EXPR
when referring to dangling pointer.
* typeck.cc (check_return_expr): Disable transformation of
dangling pointers, instead pass this information to caller.
c++: Improve location information in constant evaluation
This patch updates 'input_location' during constant evaluation to ensure
that errors in subexpressions that lack location information still
provide accurate diagnostics.
By itself this change causes some small regressions in diagnostic
quality for circumstances where errors used 'input_location' but the
location of the parent subexpression doesn't make sense, so this patch
also includes a small diagnostic improvement to fix the most egregious
case.
gcc/cp/ChangeLog:
* constexpr.cc (modifying_const_object_error): Find the source
location of the const object's declaration.
(cxx_eval_constant_expression): Update input_location to the
location of the currently evaluated expression, if possible.
David Faust [Tue, 25 Jul 2023 18:50:38 +0000 (11:50 -0700)]
bpf: don't print () in bpf_print_operand_address
Unfortunately, the pseudo-C dialect syntax used for some of the v3
atomic instructions clashes with unconditionally printing the
surrounding parentheses in bpf_print_operand_address.
Instead, place the parentheses in the output templates where needed.
gcc/
* config/bpf/bpf.cc (bpf_print_operand_address): Don't print
enclosing parentheses for pseudo-C dialect.
* config/bpf/bpf.md (zero_exdendhidi2): Add parentheses around
operands of pseudo-C dialect output templates where needed.
(zero_extendqidi2): Likewise.
(zero_extendsidi2): Likewise.
(*mov<MM:mode>): Likewise.
Andrew Pinski [Tue, 25 Jul 2023 21:50:33 +0000 (21:50 +0000)]
Fix 110803: use of plain char instead of signed char
So the problem here is that plain char can either be signed
or unsigned depending on the target (powerpc and aarch64 are
unsigned while most other targets are signed). So the testcase
gcc.c-torture/execute/pr109986.c was assuming plain char was signed
char which is wrong so it is better to just change the `char` to be
`signed char`.
Note gcc.c-torture/execute/pr109986.c includes gcc.dg/tree-ssa/pr109986.c
where the plain char was being used.
Committed as obvious after a quick test to make sure gcc.c-torture/execute/pr109986.c
now passes and gcc.dg/tree-ssa/pr109986.c still passes.
gcc/testsuite/ChangeLog:
PR testsuite/110803
* gcc.dg/tree-ssa/pr109986.c: Change plain char to be
`signed char`.
Marek Polacek [Tue, 25 Jul 2023 18:36:47 +0000 (14:36 -0400)]
c++: clear tf_partial et al in instantiate_template [PR108960]
In <https://gcc.gnu.org/pipermail/gcc-patches/2023-February/612929.html>
we concluded that we might clear all flags except tf_warning_or_error
when performing instantiate_template.
We still need #1 though. E.g., in constexpr-96241.C, we never
set ctx.ctor/object before calling cxx_eval_array_reference, so
we have to build a CONSTRUCTOR there. And in constexpr-101371-2.C
we have a ctx.ctor, but it has the wrong type, so we need a new one.
We can fix the problem by always clearing the object, and, as an
optimization, only create/free a new ctor when actually needed.
PR c++/110382
gcc/cp/ChangeLog:
* constexpr.cc (cxx_eval_array_reference): Create a new constructor
only when we don't already have a matching one. Clear the object
when the type is non-scalar.
OpenMP/Fortran: Reject declarations between target + teams
While commit r14-2754-g2e31fe431b08b0302e1fa8a1c18ee51adafd41df
detected executable statements, declarations do not show up as
executable statements. Hence, we now check whether the first
statement after TARGET is TEAMS - such that we can detect data
statements like type or variable declarations. Fortran semantics
ensures that only executable directives/statemens can come after
'!$omp end teams' such that those can be detected with the
previous check.
Note that statements returning ST_NONE such as 'omp nothing' or
'omp error at(compilation)' will still slip through.
PR fortran/110725
PR middle-end/71065
gcc/fortran/ChangeLog:
* gfortran.h (gfc_omp_clauses): Add target_first_st_is_teams.
* parse.cc (parse_omp_structured_block): Set it if the first
statement in the structured block of a TARGET is TEAMS or
a combined/composite starting with TEAMS.
* openmp.cc (resolve_omp_target): Also show an error for
contains_teams_construct without target_first_st_is_teams.
Marc Poulhiès [Tue, 25 Jul 2023 12:11:54 +0000 (14:11 +0200)]
Adjust one Ada test
Recent change modified how the loops are created, with the first
iteration being extracted out of the loops in the 2 test cases.
Adjust the text to match from the unroll dump.
When changing the dg-error for this PR, somehow the addition of a
dg-error in a second line was lost (the message uses (1) and (2) as
location, showing two lines, both need a dg-error with the same message).
rs6000: Implemented f[min/max]_optab by xs[min/max]dp
gcc/
PR target/103605
* config/rs6000/rs6000-builtin.cc (rs6000_gimple_fold_builtin): Gimple
fold RS6000_BIF_XSMINDP and RS6000_BIF_XSMAXDP when fast-math is set.
* config/rs6000/rs6000.md (FMINMAX): New int iterator.
(minmax_op): New int attribute.
(UNSPEC_FMAX, UNSPEC_FMIN): New unspecs.
(f<minmax_op><mode>3): New pattern by UNSPEC_FMAX and UNSPEC_FMIN.
* config/rs6000/rs6000-builtins.def (__builtin_vsx_xsmaxdp): Set
pattern to fmaxdf3.
(__builtin_vsx_xsmindp): Set pattern to fmindf3.
Gaius Mulley [Tue, 25 Jul 2023 02:21:12 +0000 (03:21 +0100)]
PR modula2/110174 Bugfixes to M2GenGCC.mod:CodeInline preventing an ICE
This patch calls skip_const_decl before chaining parameter values and
ensures that all strings passed to build_stmt (..., ASM_EXPR, ...) are
nul terminated. It also improves the accuracy of locations in
function calls and asm statements.
gcc/m2/
PR modula2/110174
* gm2-compiler/M2GCCDeclare.def (PromoteToCString): New procedure
function.
* gm2-compiler/M2GCCDeclare.mod (PromoteToCString): New procedure
function.
* gm2-compiler/M2GenGCC.mod (BuildTreeFromInterface): Call
skip_const_decl before chaining the parameter value.
Use PromoteToCString to ensure the string is nul terminated.
(CodeInline): Remove all parameters and replace with quad.
Use GetQuadOtok to get operand token numbers.
Remove call to DeclareConstant and replace it with PromoteToCString.
* gm2-compiler/M2Quads.def (BuildInline): Rename into ...
(BuildAsm): ... this.
* gm2-compiler/M2Quads.mod: (BuildInline): Rename into ...
(BuildAsm): ... this.
(BuildAsmElement): Add debugging.
* gm2-compiler/P1Build.bnf: Remove import of BuildInline.
* gm2-compiler/P2Build.bnf: Remove import of BuildInline.
* gm2-compiler/P3Build.bnf: Remove import of BuildInline and
import BuildAsm.
* gm2-compiler/PHBuild.bnf: Remove import of BuildInline.
* gm2-libs-iso/SysClock.mod (foo): Remove.
* gm2-libs/FIO.mod (BufferedRead): Rename parameter a to dest.
Rename variable t to src.
* m2pp.cc (pf): Correct block comment.
(pe): Correct block comment.
(m2pp_asm_expr): New function.
(m2pp_statement): Call m2pp_asm_expr.
gcc/testsuite/
PR modula2/110174
* gm2/pim/pass/program2.mod: Remove import of BuildInline.
* gm2/extensions/asm/fail/extensions-asm-fail.exp: New test.
* gm2/extensions/asm/fail/stressreturn.mod: New test.
* gm2/extensions/asm/pass/extensions-asm-pass.exp: New test.
* gm2/extensions/asm/pass/fooasm.mod: New test.
libstdc++: Add missing constexpr specifiers in <format>
A couple of virtual functions in the libstdc++ format header are marked
constexpr in the base class, but not in the derived class. This was
causing build failures when trying to compile latest gcc libstdc++ with
clang 16 using c++20. Adding the constexpr specifier resolves the issue.
OpenMP/Fortran: Reject not strictly nested target -> teams [PR110725, PR71065]
OpenMP requires: "If a teams region is nested inside a target region, the
corresponding target construct must not contain any statements, declarations
or directives outside of the corresponding teams construct."
This commit checks now for this restriction.
PR fortran/110725
PR middle-end/71065
gcc/fortran/ChangeLog:
* gfortran.h (gfc_omp_clauses): Add contains_teams_construct.
* openmp.cc (resolve_omp_target): New; check for teams nesting.
(gfc_resolve_omp_directive): Call it.
* parse.cc (decode_omp_directive): Set contains_teams_construct
on enclosing ST_OMP_TARGET.
gcc/testsuite/ChangeLog:
* gfortran.dg/gomp/pr99226.f90: Update dg-error.
* gfortran.dg/gomp/teams-5.f90: New test.
Jonathan Wakely [Mon, 24 Jul 2023 10:45:43 +0000 (11:45 +0100)]
libstdc++; Do not use strtold for hppa-hpux [PR110653]
When I switched std::stold to depend on HAVE_STRTOLD that enabled it for
hppa-hpux which defines HAVE_BROKEN_STRTOLD. Add a check for that macro
so that we don't use strtold, and fall through to the check for double
and long double having the same representation. That should mean we
define a conforming std::stold in terms of std::stod, instead of trying
to use the broken strtold.
Also fix a logic error in the fallback definition of std::stod, which
should not treat zero as a subnormal number.
libstdc++-v3/ChangeLog:
PR libstdc++/110653
* include/bits/basic_string.h [!HAVE_STOF] (stof): Do not
throw an exception for zero result.
[HAVE_BROKEN_STRTOLD] (stold): Do not use strtold.
Adds a simplification for (~X | Y) ^ X to be folded into ~(X & Y).
Also adds the macro bitwise_equal_p for generic and gimple which
returns true iff EXPR1 and EXPR2 have the same value. This helps
to reduce the number of nop_converts necessary to match the pattern.
PR middle-end/109986
gcc/ChangeLog:
* generic-match-head.cc (bitwise_equal_p): New macro.
* gimple-match-head.cc (bitwise_equal_p): New macro.
(gimple_nop_convert): Declare.
(gimple_bitwise_equal_p): Helper for bitwise_equal_p.
* match.pd ((~X | Y) ^ X -> ~(X & Y)): New simplification.
gcc/testsuite/ChangeLog:
* gcc.c-torture/execute/pr109986.c: New test.
* gcc.dg/tree-ssa/pr109986.c: New test.
Jose E. Marchesi [Mon, 24 Jul 2023 13:57:05 +0000 (15:57 +0200)]
bpf: sdiv/smod are now part of BPF V4
We used to support signed division and signed modulus instructions in
the XBPF GCC-specific extensions to BPF. However, BPF catched up by
adding these instructions in the V4 of the ISA.
This patch changes GCC in order to use sdiv/smod instructions when
-mcpu=v4 or higher. The testsuite and the manual have been updated
accordingly.
PR target/110783
* gcc.target/bpf/xbpf-sdiv-1.c: Renamed to sdiv-1.c
* gcc.target/bpf/xbpf-smod-1.c: Renamed to smod-1.c
* gcc.target/bpf/sdiv-1.c: Renamed from xbpf-sdiv-1.c, use -mcpu=v4.
* gcc.target/bpf/smod-1.c: Renamed from xbpf-smod-1.c, use -mcpu=v4.
* gcc.target/bpf/diag-sdiv.c: Use -mcpu=v3.
* gcc.target/bpf/diag-smod.c: Likewise.
Jeff Law [Mon, 24 Jul 2023 13:56:36 +0000 (07:56 -0600)]
[committed][RISC-V] Fix minor issues in diagnostic message
This fixes two minor issues with the recently adding warning about too large
VLEN in the RISC-V backend. These prevent the RISC-V port from bootstrapping
as both Andreas and I have found.
Specifically we'll get warnings for the use of '>' in the recently added
message as well as using "can not" vs "cannot". While these warnings may
seem annoying, they're in place to make it easier for the translators.
This patch fixes the message in the fairly obvious way. Spells out the
greater than and uses cannot. There's a similar issue in another recently
added diagnostic that I'll push momentarily.
gcc/
* config/riscv/riscv.cc (riscv_option_override): Spell out
greater than and use cannot in diagnostic string.
Robin Dapp [Thu, 13 Jul 2023 07:10:06 +0000 (09:10 +0200)]
vect: Handle demoting FLOAT and promoting FIX_TRUNC.
The recent changes that allowed multi-step conversions for
"non-packing/unpacking", i.e. modifier == NONE targets included
promoting to-float and demoting to-int variants. This patch
adds the missing demoting to-float and promoting to-int handling.
gcc/ChangeLog:
* tree-vect-stmts.cc (vectorizable_conversion): Handle
more demotion/promotion for modifier == NONE.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/conversions/vec-narrow-int64-float16.c: New test.
* gcc.target/riscv/rvv/autovec/conversions/vec-widen-float16-int64.c: New test.
Roger Sayle [Mon, 24 Jul 2023 11:34:23 +0000 (12:34 +0100)]
[Committed] PR target/110787: Revert QImode offsets in {zero,sign}_extract.
My recent patch to use QImode for bit offsets in ZERO_EXTRACTs and
SIGN_EXTRACTs in the i386 backend shouldn't have resulted in any change
behaviour, but as reported by Rainer it produces a bootstrap failure in
gm2. This reverts the problematic patch whilst we investigate the
underlying cause.
Committed as obvious.
2023-07-23 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
PR target/110787
PR target/110790
Revert patch.
* config/i386/i386.md (extv<mode>): Use QImode for offsets.
(extzv<mode>): Likewise.
(insv<mode>): Likewise.
(*testqi_ext_3): Likewise.
(*btr<mode>_2): Likewise.
(define_split): Likewise.
(*btsq_imm): Likewise.
(*btrq_imm): Likewise.
(*btcq_imm): Likewise.
(define_peephole2 x3): Likewise.
(*bt<mode>): Likewise
(*bt<mode>_mask): New define_insn_and_split.
(*jcc_bt<mode>): Use QImode for offsets.
(*jcc_bt<mode>_1): Delete obsolete pattern.
(*jcc_bt<mode>_mask): Use QImode offsets.
(*jcc_bt<mode>_mask_1): Likewise.
(define_split): Likewise.
(*bt<mode>_setcqi): Likewise.
(*bt<mode>_setncqi): Likewise.
(*bt<mode>_setnc<mode>): Likewise.
(*bt<mode>_setncqi_2): Likewise.
(*bt<mode>_setc<mode>_mask): New define_insn_and_split.
(bmi2_bzhi_<mode>3): Use QImode offsets.
(*bmi2_bzhi_<mode>3): Likewise.
(*bmi2_bzhi_<mode>3_1): Likewise.
(*bmi2_bzhi_<mode>3_1_ccz): Likewise.
(@tbm_bextri_<mode>): Likewise.
Jose E. Marchesi [Mon, 24 Jul 2023 09:52:30 +0000 (11:52 +0200)]
bpf: remove -mkernel option and BPF_KERNEL_VERSION_CODE
Having the ability of specifying a target kernel version when building
a BPF program is one of these things that sound pretty good in theory,
but simply don't work in practice: kernels in practice contain
backports, etc. Also, the addition of CO-RE to BPF has made this
uneccessary.
This patch removes the -mkernel command line option and also the
associated BPF_KERNEL_VERSION_CODE pre-processor constant.
Tested in bpf-unknown-none.
gcc/ChangeLog
* config/bpf/bpf-opts.h (enum bpf_kernel_version): Remove enum.
* config/bpf/bpf.opt (mkernel): Remove option.
* config/bpf/bpf.cc (bpf_target_macros): Do not define
BPF_KERNEL_VERSION_CODE.
Jose E. Marchesi [Mon, 24 Jul 2023 08:56:27 +0000 (10:56 +0200)]
bpf: make use of the bswap{16,32,64} V4 BPF instruction
This patch makes the BPF backend to use the new V4 bswap{16,32,64}
instructions in order to implement the __builtin_bswap{16,32,64}
built-ins. It also adds support for -mcpu=v4 and -m[no]bswap
command-line options. Tests and doc updates are includes.
Tested in bpf-unknown-none.
gcc/ChangeLog
PR target/110786
* config/bpf/bpf.opt (mcpu): Add ISA_V4 and make it the default.
(mbswap): New option.
* config/bpf/bpf-opts.h (enum bpf_isa_version): New value ISA_V4.
* config/bpf/bpf.cc (bpf_option_override): Set bpf_has_bswap.
* config/bpf/bpf.md: Use bswap instructions if available for
bswap* insn, and fix constraint.
* doc/invoke.texi (eBPF Options): Document -mcpu=v4 and -mbswap.
gcc/testsuite/ChangeLog
PR target/110786
* gcc.target/bpf/bswap-1.c: Pass -mcpu=v3 to build test.
* gcc.target/bpf/bswap-2.c: New test.
This patch is depending on:
https://gcc.gnu.org/pipermail/gcc-patches/2023-July/624995.html
Consider this following case:
float foo (float *__restrict a, int n)
{
float result = 1.0;
for (int i = 0; i < n; i++)
result += a[i];
return result;
}
Compile with **NO** -ffast-math:
Before this patch:
<source>:4:21: missed: couldn't vectorize loop
<source>:1:7: missed: not vectorized: relevant phi not supported: result_14 = PHI <result_11(6), 1.0e+0(5)>
After this patch:
foo:
lui a5,%hi(.LC0)
flw fa0,%lo(.LC0)(a5)
ble a1,zero,.L4
.L3:
vsetvli a5,a1,e32,m1,ta,ma
vle32.v v1,0(a0)
slli a4,a5,2
sub a1,a1,a5
vfmv.s.f v2,fa0
add a0,a0,a4
vfredosum.vs v1,v1,v2 ----------> FOLD_LEFT_PLUS
vfmv.f.s fa0,v1
bne a1,zero,.L3
ret
.L4:
ret
gcc/ChangeLog:
* config/riscv/autovec.md (fold_left_plus_<mode>): New pattern.
(mask_len_fold_left_plus_<mode>): Ditto.
* config/riscv/riscv-protos.h (enum insn_type): New enum.
(enum reduction_type): Ditto.
(expand_reduction): Add in-order reduction.
* config/riscv/riscv-v.cc (emit_nonvlmax_fp_reduction_insn): New function.
(expand_reduction): Add in-order reduction.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/reduc/reduc_strict-1.c: New test.
* gcc.target/riscv/rvv/autovec/reduc/reduc_strict-2.c: New test.
* gcc.target/riscv/rvv/autovec/reduc/reduc_strict-3.c: New test.
* gcc.target/riscv/rvv/autovec/reduc/reduc_strict-4.c: New test.
* gcc.target/riscv/rvv/autovec/reduc/reduc_strict-5.c: New test.
* gcc.target/riscv/rvv/autovec/reduc/reduc_strict-6.c: New test.
* gcc.target/riscv/rvv/autovec/reduc/reduc_strict-7.c: New test.
* gcc.target/riscv/rvv/autovec/reduc/reduc_strict_run-1.c: New test.
* gcc.target/riscv/rvv/autovec/reduc/reduc_strict_run-2.c: New test.
Kewen Lin [Mon, 24 Jul 2023 06:20:30 +0000 (01:20 -0500)]
vect: Don't vectorize a single scalar iteration loop [PR110740]
The function vect_update_epilogue_niters which has been
removed by r14-2281 has some code taking care of that if
there is only one scalar iteration left for epilogue then
we won't try to vectorize it any more.
Although costing should be able to care about it eventually,
I think we still want this special casing without costing
enabled, so this patch is to add it back in function
vect_analyze_loop_costing, and make it more general for
both main and epilogue loops as Richi suggested, it can fix
some exposed failures on Power10:
Andrew Pinski [Sat, 22 Jul 2023 15:52:42 +0000 (08:52 -0700)]
Fix PR 110066: crash with -pg -static on riscv
The problem -fasynchronous-unwind-tables is on by default for riscv linux
We need turn it off for crt*.o because it would make __EH_FRAME_BEGIN__ point
to .eh_frame data from crtbeginT.o instead of the user-defined object
during static linking.
This turns it off.
OK?
libgcc/ChangeLog:
* config.host (riscv*-*-linux*): Add t-crtstuff to tmake_file.
(riscv*-*-freebsd*): Likewise.
* config/riscv/t-crtstuff: New file.
RISC-V: optim const DF +0.0 store to mem [PR/110748]
Fixes: ef85d150b5963 ("RISC-V: Enable TARGET_SUPPORTS_WIDE_INT")
DF +0.0 is bitwise all zeros so int x0 store to mem can be used to optimize it.
void zd(double *) { *d = 0.0; }
currently:
| fmv.d.x fa5,zero
| fsd fa5,0(a0)
| ret
With patch
| sd zero,0(a0)
| ret
The fix updates predicate const_0_operand() so reg_or_0_operand () now
includes const_double, enabling movdf expander -> riscv_legitimize_move ()
to generate below vs. an intermediate set (reg:DF) const_double:DF
This change also enables such insns to be recog() by later passes.
The md pattern "*movdf_hardfloat_rv64" despite already supporting the
needed constraints {"m","G"} mem/const 0.0 was failing to match because
the additional condition check reg_or_0_operand() was failing due to
missing const_double.
This failure to recog() was triggering an ICE when testing the in-flight
f-m-o patches and is how all of this started, but then was deemed to be
an independent optimization of it's own [1].
Its worthwhile to note all the set peices were already there and working
up until my own commit mentioned at top regressed the whole thing.
Ran thru full multilib testsuite and no surprises. There was 1 false
failure due to random string "lw" appearing in lto build assembler output,
which is also fixed here.
gcc/ChangeLog:
PR target/110748
* config/riscv/predicates.md (const_0_operand): Add back
const_double.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/pr110748-1.c: New Test.
* gcc.target/riscv/xtheadfmv-fmv.c: Add '\t' around test
patterns to avoid random string matches.
Roger Sayle [Sat, 22 Jul 2023 20:52:55 +0000 (21:52 +0100)]
i386: Don't use insvti_{high,low}part with -O0 (for compile-time).
This patch attempts to help with PR rtl-optimization/110587, a regression
of -O0 compile time for the pathological pr28071.c. My recent patch helps
a bit, but hasn't returned -O0 compile-time to where it was before my
ix86_expand_move changes. The obvious solution/workaround is to guard
these new TImode parameter passing optimizations with "&& optimize", so
they don't trigger when compiling with -O0. The very minor complication
is that "&& optimize" alone leads to the regression of pr110533.c, where
our improved TImode parameter passing fixes a wrong-code issue with naked
functions, importantly, when compiling with -O0. This should explain
the one line fix below "&& (optimize || ix86_function_naked (cfun))".
I've an additional fix/tweak or two for this compile-time issue, but
this change eliminates the part of the regression that I've caused.
2023-07-22 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* config/i386/i386-expand.cc (ix86_expand_move): Disable the
64-bit insertions into TImode optimizations with -O0, unless
the function has the "naked" attribute (for PR target/110533).
Andrew Pinski [Sat, 22 Jul 2023 20:34:41 +0000 (20:34 +0000)]
Fix alpha building
The problem is after r14-2587-gd8105b10fff951, the definition of
extended_count now takes a bool as its last argument but we only
have a declaration for the version which takes an int as the last
argument. This fixes the problem by changing the declaration to be
a bool too.
Committed as obvious after building a cross to alpha-linux-gnu.
gcc/ChangeLog:
PR target/110778
* rtl.h (extended_count): Change last argument type
to bool.
Roger Sayle [Sat, 22 Jul 2023 20:50:06 +0000 (21:50 +0100)]
i386: Use QImode for offsets in zero_extract/sign_extract in i386.md
As suggested by Uros, this patch changes the ZERO_EXTRACTs and SIGN_EXTRACTs
in i386.md to consistently use QImode for bit offsets (i.e. third and fourth
operands), matching the use of QImode for bit counts in shifts and rotates.
There's no change in functionality, and the new patterns simply ensure that
we continue to generate the same code (match revised patterns) as before.
2023-07-22 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* config/i386/i386.md (extv<mode>): Use QImode for offsets.
(extzv<mode>): Likewise.
(insv<mode>): Likewise.
(*testqi_ext_3): Likewise.
(*btr<mode>_2): Likewise.
(define_split): Likewise.
(*btsq_imm): Likewise.
(*btrq_imm): Likewise.
(*btcq_imm): Likewise.
(define_peephole2 x3): Likewise.
(*bt<mode>): Likewise
(*bt<mode>_mask): New define_insn_and_split.
(*jcc_bt<mode>): Use QImode for offsets.
(*jcc_bt<mode>_1): Delete obsolete pattern.
(*jcc_bt<mode>_mask): Use QImode offsets.
(*jcc_bt<mode>_mask_1): Likewise.
(define_split): Likewise.
(*bt<mode>_setcqi): Likewise.
(*bt<mode>_setncqi): Likewise.
(*bt<mode>_setnc<mode>): Likewise.
(*bt<mode>_setncqi_2): Likewise.
(*bt<mode>_setc<mode>_mask): New define_insn_and_split.
(bmi2_bzhi_<mode>3): Use QImode offsets.
(*bmi2_bzhi_<mode>3): Likewise.
(*bmi2_bzhi_<mode>3_1): Likewise.
(*bmi2_bzhi_<mode>3_1_ccz): Likewise.
(@tbm_bextri_<mode>): Likewise.
Jeff Law [Sat, 22 Jul 2023 15:47:21 +0000 (09:47 -0600)]
[committed] Fix length computation bug in bfin port
The tester seemed to occasionally ping-pong a compilation failure on the
builtin-bitops-1.c test. I long suspected it was something like length
computations.
I finally got a few minutes to dig into it, and sure enough the blackfin
port was claiming the "ones" operation was 2 bytes when it is in fact 4 bytes.
This fixes the compilation failure for the builtin-bitops-1.c test. Sadly,
it doesn't fix any of the other failures on the bfin port.
Gaius Mulley [Sat, 22 Jul 2023 09:01:02 +0000 (10:01 +0100)]
PR modula2/110631 Bugfix to FIO WriteCardinal
FIO.WriteCardinal fails to write binary data. This patch fixes two
bugs in FIO.mod and provides a testcase which writes and reads binary
cardinals. There was an off by one error when using HIGH (a) to
determine the number of bytes and the dest/src pointers were switched
when calling memcpy.
gcc/m2/ChangeLog:
PR modula2/110631
* gm2-libs/FIO.def (ReadAny): Correct comment as
HIGH (a) + 1 is number of bytes.
(WriteAny): Correct comment as HIGH (a) + 1 is number of
bytes.
* gm2-libs/FIO.mod (ReadAny): Correct comment as
HIGH (a) + 1 is number of bytes. Also pass HIGH (a) + 1
to BufferedRead.
(WriteAny): Correct comment as HIGH (a) + 1 is number of
bytes. Also pass HIGH (a) + 1 to BufferedWrite.
(BufferedWrite): Rename parameter a to src, rename variable
t to dest. Correct parameter order to memcpy.
gcc/testsuite/ChangeLog:
PR modula2/110631
* gm2/pimlib/run/pass/testfiobinary.mod: New test.
[LRA]: Fix sparc bootstrap after recent patch for fp elimination for avr LRA port
The recent patch for fp elimination for avr LRA port modified an assert
which can be wrong for targets using hard frame pointer different from
frame pointer. Also for such ports spilling pseudos assigned to fp
was wrong too in the new code. Although this code is not used for any target
currently using LRA except for avr. Given patch fixes the issues.
gcc/ChangeLog:
* lra-eliminations.cc (update_reg_eliminate): Fix the assert.
(lra_update_fp2sp_elimination): Use HARD_FRAME_POINTER_REGNUM
instead of FRAME_POINTER_REGNUM to spill pseudos.
bpf: fixed template for neg (added second operand)
This patch fixes define_insn for "neg" to support 2 operands.
Initial implementation assumed the format "neg %0" while the instruction
allows both a destination and source operands. The second operand can
either be a register or an immediate value.
gcc/ChangeLog:
* config/bpf/bpf.md: fixed template for neg instruction.
Marek Polacek [Tue, 18 Jul 2023 20:02:21 +0000 (16:02 -0400)]
c++: fix ICE with is_really_empty_class [PR110106]
is_really_empty_class is liable to crash when it gets an incomplete
or dependent type. Since r11-557, we pass the yet-uninstantiated
class type S<0> of the PARM_DECL s to is_really_empty_class -- because
of the potential_rvalue_constant_expression -> is_rvalue_constant_expression
change in cp_parser_constant_expression. Here we're not parsing
a template so we did not check COMPLETE_TYPE_P as we should.
It should work to complete the type before checking COMPLETE_TYPE_P.
PR c++/110106
gcc/cp/ChangeLog:
* constexpr.cc (potential_constant_expression_1): Try to complete the
type when !processing_template_decl.
Jan Hubicka [Fri, 21 Jul 2023 17:38:26 +0000 (19:38 +0200)]
Avoid scaling flat loop profiles of vectorized loops
As discussed, when vectorizing loop with static profile, it is not always good idea
to divide the header frequency by vectorization factor because the profile may
not realistically represent the expected number of iterations. Since in such cases
we default to relatively low iteration counts (based on average for spec2k17), this
will make vectorized loop body look cold.
This patch makes vectorizer to look for flat profiles and only possibly reduce the
profile by known upper bound on iteration counts.
gcc/ChangeLog:
PR target/110727
* tree-vect-loop.cc (scale_profile_for_vect_loop): Avoid scaling flat
profiles by vectorization factor.
(vect_transform_loop): Check for flat profiles.
Jan Hubicka [Fri, 21 Jul 2023 15:34:31 +0000 (17:34 +0200)]
Implement flat loop profile detection
This patch adds maybe_flat_loop_profile which can be used in loop profile udpate
to detect situation where the profile may be unrealistically flat and should
not be dwonscalled after vectorizing, unrolling and other transforms that
assume that loop has high iteration count even if the CFG profile says
otherwise.
Profile is flat if it was statically detected and at that time we had
no idea about actual number of iterations or we artificially capped them.
So the function considers flat all profiles that have guessed or lower
reliability in their count and there is no nb_iteration_bounds/estimate
which would prove that the profile iteration count is high enough.
gcc/ChangeLog:
* cfgloop.h (maybe_flat_loop_profile): Declare
* cfgloopanal.cc (maybe_flat_loop_profile): New function.
* tree-cfg.cc (print_loop_info): Print info about flat profiles.
Jan Hubicka [Fri, 21 Jul 2023 15:31:34 +0000 (17:31 +0200)]
Fix gcc.dg/tree-ssa/copy-headers-9.c and gcc.dg/tree-ssa/dce-1.c failures
This patch fixes template in the two testcases so it matches the output
correctly. I did not re-test after last changes in the previous patch,
sorry for that.
gcc/testsuite/ChangeLog:
* gcc.dg/tree-ssa/copy-headers-9.c: Fix template for tree-ssa-loop-ch.cc changes.
* gcc.dg/tree-ssa/dce-1.c: Likewise.
Jan Hubicka [Fri, 21 Jul 2023 14:44:01 +0000 (16:44 +0200)]
Fix sreal::to_int and implement sreal::to_nearest_int
while exploring new loop estimate dumps, I noticed that loop iterating 1.8
times by profile is etimated as iterating once instead of 2 by nb_estimate.
While nb_estimate should really be a sreal and I will convert it incrementally,
I found problem is in previous patch doing:
+ *nit = (snit + 0.5).to_int ();
this does not work for sreal because it has only constructor from integer, so
first 0.5 is rounded to 0 and then added to snit.
Some code uses sreal(1, -1) which produces 0.5, but it reuqires unnecessary
addition, so I decided to add to_nearest_int. Testing it I noticed that to_int
is buggy:
(sreal(3)/2).to_int () == 1
while
(sreal(-3)/2).to_int () == -2
Fix is easy, we need to correctly shift in positive values. This patch fixes
it and adds the to_nearest_int alternative.
Jan Hubicka [Fri, 21 Jul 2023 12:54:23 +0000 (14:54 +0200)]
loop-ch improvements, part 5
Currently loop-ch skips all do-while loops. But when loop is not do-while
in addition to original goal of turining it to do-while it can do additional
things:
1) move out loop invariant computations
2) duplicate loop invariant conditionals and eliminate them in loop body.
3) prove that some exits are always true in first iteration
and can be skipped
Most of time 1 can be done by lim (exception is when the invariant computation
is conditional). For 2 we however don't really have other place doing it except
for loop unswitching that is more expensive (it will duplicate the loop and
then optimize out one path to non-loop).
3 can be done by loop peeling but it is also more expensive by duplicating full
loop body.
This patch improves heuristics by not giving up on do-while loops and trying
to find sequence of BBs to duplicate to obtain one of goals:
- turn loop to do-while
- eliminate invariant conditional in loop body
- do partial "peeling" as long as code optimizes enough so this does not
increase code size.
Bootstrapped/regtested x86_64-linux, OK?
gcc/ChangeLog:
* tree-ssa-loop-ch.cc (enum ch_decision): New enum.
(should_duplicate_loop_header_p): Return info on profitability.
(do_while_loop_p): Watch for constant conditionals.
(update_profile_after_ch): Do not sanity check that all
static exits are taken.
(ch_base::copy_headers): Run on all loops.
(pass_ch::process_loop_p): Improve heuristics by handling also
do_while loop and duplicating shortest sequence containing all
winning blocks.
gcc/testsuite/ChangeLog:
* gcc.dg/loop-unswitch-17.c: Disable ch.
* gcc.dg/pr103079.c: Disable ch.
* gcc.dg/tree-ssa/copy-headers-7.c: Update so ch behaves
as expected.
* gcc.dg/tree-ssa/copy-headers.c: Update template.
* gcc.dg/tree-ssa/copy-headers-9.c: New test.
gcc.dg/tree-ssa/forwprop-12.c looks for reconstruction of an
ARRAY_REF from pointer arithmetic and dereference. That's not
safe because ARRAY_REFs carry special semantics we later exploit
during data dependence analysis.
The following removes the testcase, closing the bug as WONTFIX.
Jan Hubicka [Fri, 21 Jul 2023 11:57:34 +0000 (13:57 +0200)]
finite_loop_p tweak
We have finite_p flag in loop structure. finite_loop_p already know to
use it, but we also may set the flag when we prove loop to be finite by
SCEV analysis to avoid duplicated work.
Bootstrapped/regtested x86_64-linux, OK?
gcc/ChangeLog:
* tree-ssa-loop-niter.cc (finite_loop_p): Reorder to do cheap
tests first; update finite_p flag.
Jan Hubicka [Fri, 21 Jul 2023 11:38:29 +0000 (13:38 +0200)]
improfe loop dumps
we have flow_loop_dump and print_loop. While print_loop was extended to dump
stuff from loop structure we added over years (loop info), flow_loop_dump was not.
-fdump-tree-all files contains flow_loop_dump which makes it hard to see what
metadata we have attached to loop.
This patch unifies dumping of these fields from both functions. For example for:
int a[100];
main()
{
for (int i = 0; i < 10; i++)
a[i]=i;
}
we now print:
;; Loop 0
;; header 0, latch 1
;; depth 0, outer -1
;; nodes: 0 1 2 3 4 5
;;
;; Loop 1
;; header 4, latch 3
;; depth 1, outer 0, finite_p
;; upper_bound 10
;; likely_upper_bound 10
;; estimate 10
;; iterations by profile: 10.001101 (unreliable)
finite_p, upper_boud, likely_upper_bound estimate and iterations by profile is new.
Bootstrap/regtest on x86_64 in progress. OK if it passes?
Honza
gcc/ChangeLog:
* cfgloop.cc (flow_loop_dump): Use print_loop_info.
* cfgloop.h (print_loop_info): Declare.
* tree-cfg.cc (print_loop_info): Break out from ...; add
printing of missing fields and profile
(print_loop): ... here.
gcc/testsuite/ChangeLog:
* gcc.dg/tree-ssa/dce-1.c: Update for new loop dumps.
cleanup: make all cond_len_* and mask_len_* consistent on the order of mask and len
This patch is depending on:
https://gcc.gnu.org/pipermail/gcc-patches/2023-July/625121.html
Hi, Richard and Richi.
This patch is to align the order of mask and len.
Currently, According to this piece code:
if (final_len && final_mask)
call = gimple_build_call_internal (
IFN_LEN_MASK_GATHER_LOAD, 7, dataref_ptr,
vec_offset, scale, zero, final_mask, final_len,
bias);
You can see the order of mask and len, is {mask,len,bias}.
"mask" comes before "len". The reason of this order is that we want to
reuse the current codes of MASK_GATHER_LOAD/MASK_SCATTER_STORE.
Same situation for COND_LEN_*, we want to reuse the codes of COND_*.
Reusing codes from the existing MASK_* or COND_* can allow us not to
change the codes too much and make the codes elegant and easy to maintain && read.
To avoid any confusions of auto-vectorization patterns that includes both mask and len,
this patch align the order of mask and len for both Gimple IR and RTL pattern into
{mask, len, bias} to make everything cleaner and more elegant.
Since start from LEN_MASK_GATHER_LOAD/LEN_MASK_SCATTER_STORE, COND_LEN_* patterns,
the order of len and mask is {mask,len,bias}.
The reason we make "mask" argument comes before "len" is because we want to keep
the "mask" location same as mask_* or cond_* patterns to make use of current codes flow
of mask_* and cond_*. Otherwise, we will need to change codes much more and make codes
hard to maintain.
Now, we already have COND_LEN_*, it's naturally that we should rename "LEN_MASK" into "MASK_LEN"
to keep name scheme consistent.
This patch only changes the name "LEN_MASK" into "MASK_LEN".
No codes functionality change.
Richard Biener [Thu, 13 Jul 2023 06:58:58 +0000 (08:58 +0200)]
tree-optimization/88540 - FP x > y ? x : y if-conversion without -ffast-math
The following makes sure that FP x > y ? x : y style max/min operations
are if-converted at the GIMPLE level. While we can neither match
it to MAX_EXPR nor .FMAX as both have different semantics with IEEE
than the ternary ?: operation we can make sure to maintain this form
as a COND_EXPR so backends have the chance to match this to instructions
their ISA offers.
The patch does this in phiopt where we recognize min/max and instead
of giving up when we have to honor NaNs we alter the generated code
to a COND_EXPR.
This resolves PR88540 and we can then SLP vectorize the min operation
for its testcase. It also resolves part of the regressions observed
with the change matching bit-inserts of bit-field-refs to vec_perm.
Expansion from a COND_EXPR rather than from compare-and-branch
gcc.target/i386/pr54855-9.c by producing extra moves while the
corresponding min/max operations are now already synthesized by
RTL expansion, register selection isn't optimal. This can be also
provoked without this change by altering the operand order in the source.
I have XFAILed that part of the test.
PR tree-optimization/88540
* tree-ssa-phiopt.cc (minmax_replacement): Do not give up
with NaNs but handle the simple case by if-converting to a
COND_EXPR.
This adds a simple match pattern to simplify
`max<max<a,b>,a>` to `max<a,b>`. Reassociation handles
this already (r0-77700-ge969dbde29bfd396259357) but
seems like we should be able to handle this even before
reassociation.
This fixes part of PR tree-optimization/80574 but more
work is needed fix it the rest of the way. The original
testcase there is fixed but the RTL level is what fixes
it the rest of the way.
OK? Bootstrapped and tested on x86_64-linux-gnu.
gcc/ChangeLog:
* match.pd (minmax<minmax<a,b>,a>->minmax<a,b>): New
transformation.
gcc/testsuite/ChangeLog:
* gcc.dg/tree-ssa/reassoc-12.c: Disable all of
the passes that enables match-and-simplify.
* gcc.dg/tree-ssa/minmax-23.c: New test.
Richard Biener [Thu, 20 Jul 2023 11:09:17 +0000 (13:09 +0200)]
tree-optimization/110742 - fix latent issue with permuting existing vectors
When we materialize a layout we push edge permutes to constant/external
defs without checking we can actually do so. For externals defined
by vector stmts rather than scalar components we can't.
PR tree-optimization/110742
* tree-vect-slp.cc (vect_optimize_slp_pass::get_result_with_layout):
Do not materialize an edge permutation in an external node with
vector defs.
(vect_slp_analyze_node_operations_1): Guard purely internal
nodes better.
Jan Hubicka [Fri, 21 Jul 2023 06:52:00 +0000 (08:52 +0200)]
Cleanup expected_loop_iterations
this patch cleanups API for determining expected loop iteraitons from profile.
We started with having expected_loop_iterations and only source was the integer
represented BB counts. It did some work on guessing number of iteration if
profile was absent or bogus. Later we introduced loop_info and added
get_estimated_loop_iterations which made expected_loop_iterations useful mostly
when doing profile updates and not for loop optimization heuristics. The
naming is bit ambiguous so this difference is not clear. Even later we
introduced precision tracking to profile and exended the API to return
reliablity of result but did not update all uses to do reasonable stuff with
it. There is also some cofusion about +-1s concering latch execution counts
versus header execution counts.
This patch aims to obsolette expected_loop_iterations and
expected_loop_iterations_unbounded (and "suceeds" modulo 1 use of each of two).
It adds expected_loop_iterations_by_profile which computes sreal and does
correct precision/presence tracking.
Unlike old code, it is based on CFG profile only and does not attempt to
provide fake answer when info is missing and does not check sanity with
loop_info.
We now define iterations consistently as lath execution in loop_info so I use
that here too.
I converted almost all calls to new API: dumps, code produing loop_info from
CFG profile and profile updating. Remaining uses are in loop unrolling and
prefetching that needs more TLC I will do incrementally.
There are some improvements possible which I can play with incrementally.
- for simple loops with one exit dominating latch we can use exit
probability for easier to preserve info in loop itraionts.
THis is probably not too critical since all esitmates should be recorded
in loop_info and would help mostly if new loop is constructed or old
loop is lost and redicovered.
- We may want to avoid trusting the profile if it is obviously inconsistent
on header.
gcc/ChangeLog:
* cfgloop.cc: Include sreal.h.
(flow_loop_dump): Dump sreal iteration exsitmate.
(get_estimated_loop_iterations): Update.
* cfgloop.h (expected_loop_iterations_by_profile): Declare.
* cfgloopanal.cc (expected_loop_iterations_by_profile): New function.
(expected_loop_iterations_unbounded): Use new API.
* cfgloopmanip.cc (scale_loop_profile): Use
expected_loop_iterations_by_profile
* predict.cc (pass_profile::execute): Likewise.
* profile.cc (branch_prob): Likewise.
* tree-ssa-loop-niter.cc: Include sreal.h.
(estimate_numbers_of_iterations): Likewise
Andrew Pinski [Fri, 21 Jul 2023 02:26:09 +0000 (02:26 +0000)]
libfortran: Fix build for targets that don't have 10byte or 16 byte floating point
So the problem here is EXPAND_INTER_MACRO_16 expands to nothing if 16 byte FP does not
exist but we still add a comma after it and that causes a build failure.
The same is true for EXPAND_INTER_MACRO_10 too.
Committed as obvious after a bootstrap and test on x86_64-linux-gnu and aarch64-linux-gnu.
libgfortran/ChangeLog:
PR libfortran/110759
* ieee/ieee_arithmetic.F90
(COMP_INTERFACE): Remove the comma after EXPAND_INTER_MACRO_16
and EXPAND_INTER_MACRO_10.
(EXPAND_INTER_MACRO_16): Add comma here if 16 byte fp exist.
(EXPAND_INTER_MACRO_10): Likewise.
Kewen Lin [Fri, 21 Jul 2023 05:18:19 +0000 (00:18 -0500)]
sccvn: Correct the index of bias for IFN_LEN_STORE [PR110744]
Commit r14-2267-gb8806f6ffbe72e adjusts the arguments order
of LEN_STORE from {len,vector,bias} to {len,bias,vector},
in order to make them consistent with LEN_MASK_STORE and
MASK_STORE. But it missed to update the related handlings
in tree-ssa-sccvn.cc, it caused the failure shown in PR
110744. This patch is to fix the related handlings with
the correct index.
PR tree-optimization/110744
gcc/ChangeLog:
* tree-ssa-sccvn.cc (vn_reference_lookup_3): Correct the index of bias
operand for ifn IFN_LEN_STORE.
Kewen Lin [Fri, 21 Jul 2023 05:16:29 +0000 (00:16 -0500)]
testsuite: Add a test case for PR110729 [PR110729]
As PR110729 reported, there was one issue for .section
__patchable_function_entries with -ffunction-sections, that
is we put the same symbol as link_to section symbol for all
functions wrongly. The commit r13-4294 for PR99889 has
fixed this with the corresponding label LPFE* which sits in
the function_section.
As Fangrui suggested [1], this patch is to add a bit more
test coverage. I didn't find a good way to check all
linked_to symbols are different, so I checked for LPFE[012].
liuhongt [Fri, 12 May 2023 07:15:08 +0000 (15:15 +0800)]
Provide -fcf-protection=branch,return.
Use EnumSet instead of EnumBitSet since CF_FULL is not power of 2.
It is a bit tricky for sets classification, cf_branch and cf_return
should be in different sets, but they both "conflicts" cf_full,
cf_none. And current EnumSet don't handle this well.
So in the current implementation, only cf_full,cf_none are exclusive
to each other, but they can be combined with any cf_branch, cf_return,
cf_check. It's not perfect, but still an improvement than original
one.
gcc/ChangeLog:
PR target/89701
* common.opt: (fcf-protection=): Add EnumSet attribute to
support combination of params.
gcc/testsuite/ChangeLog:
* c-c++-common/fcf-protection-10.c: New test.
* c-c++-common/fcf-protection-11.c: New test.
* c-c++-common/fcf-protection-12.c: New test.
* c-c++-common/fcf-protection-8.c: New test.
* c-c++-common/fcf-protection-9.c: New test.
* gcc.target/i386/pr89701-1.c: New test.
* gcc.target/i386/pr89701-2.c: New test.
* gcc.target/i386/pr89701-3.c: New test.
> I see some regressions most likely with this change on i686-linux,
> in particular:
> +FAIL: gcc.dg/pr107547.c (test for excess errors)
> +FAIL: gcc.dg/torture/floatn-convert.c -O0 (test for excess errors)
> +UNRESOLVED: gcc.dg/torture/floatn-convert.c -O0 compilation failed to produce executable
> +FAIL: gcc.dg/torture/floatn-convert.c -O1 (test for excess errors)
> +UNRESOLVED: gcc.dg/torture/floatn-convert.c -O1 compilation failed to produce executable
> +FAIL: gcc.dg/torture/floatn-convert.c -O2 (test for excess errors)
> +UNRESOLVED: gcc.dg/torture/floatn-convert.c -O2 compilation failed to produce executable
> +FAIL: gcc.dg/torture/floatn-convert.c -O2 -flto (test for excess errors)
> +UNRESOLVED: gcc.dg/torture/floatn-convert.c -O2 -flto compilation failed to produce executable
> +FAIL: gcc.dg/torture/floatn-convert.c -O2 -flto -flto-partition=none (test for excess errors)
> +UNRESOLVED: gcc.dg/torture/floatn-convert.c -O2 -flto -flto-partition=none compilation failed to produce executable
> +FAIL: gcc.dg/torture/floatn-convert.c -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions (test for excess errors)
> +UNRESOLVED: gcc.dg/torture/floatn-convert.c -O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer -finline-functions compilation failed to produce executable
> +FAIL: gcc.dg/torture/floatn-convert.c -O3 -g (test for excess errors)
> +UNRESOLVED: gcc.dg/torture/floatn-convert.c -O3 -g compilation failed to produce executable
> +FAIL: gcc.dg/torture/floatn-convert.c -Os (test for excess errors)
> +UNRESOLVED: gcc.dg/torture/floatn-convert.c -Os compilation failed to produce executable
> +FAIL: gcc.target/i386/float16-7.c (test for errors, line 7)
>
> Perhaps we need to tweak
> gcc/testsuite/lib/target-supports.exp (add_options_for_float16)
> so that it adds -msse2 for i?86-*-* x86_64-*-* (that would likely
> fix up floatn-convert) and for the others perhaps
> /* { dg-add-options float16 } */
> ?
David Malcolm [Fri, 21 Jul 2023 00:24:10 +0000 (20:24 -0400)]
analyzer: avoid usage of TYPE_PRECISION on vector types [PR110455]
gcc/analyzer/ChangeLog:
PR analyzer/110455
* region-model.cc (region_model::get_gassign_result): Only check
for bad shift counts when dealing with an integral type.
gcc/testsuite/ChangeLog:
PR analyzer/110455
* gcc.dg/analyzer/pr110455.c: New test.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>
David Malcolm [Fri, 21 Jul 2023 00:24:01 +0000 (20:24 -0400)]
analyzer: fix ICE on certain pointer subtractions [PR110387]
gcc/analyzer/ChangeLog:
PR analyzer/110387
* region.h (struct cast_region::key_t): Support "m_type" being
null by using "m_original_region" for empty/deleted slots.
gcc/testsuite/ChangeLog:
PR analyzer/110387
* gcc.dg/analyzer/out-of-bounds-pr110387.c: New test.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>