Age | Commit message (Collapse) | Author | Files | Lines |
|
This patch adds support for conditional multiplication and division.
It's mostly mechanical, but a few notes:
* The *_optab name and the .md names are the same as the unconditional
forms, just with "cond_" added to the front. This means we still
have the awkward difference between sdiv and div, etc.
* It was easier to retain the difference between integer and FP
division in the function names, given that they map to different
tree codes (TRUNC_DIV_EXPR and RDIV_EXPR).
* SVE has no direct support for IFN_COND_MOD, but it seemed more
consistent to add it anyway.
* Adding IFN_COND_MUL enables an extra fully-masked reduction
in gcc.dg/vect/pr53773.c.
* In practice we don't actually use the integer division forms without
if-conversion support (added by a later patch).
2018-05-25 Richard Sandiford <richard.sandiford@linaro.org>
gcc/
* doc/sourcebuild.texi (vect_double_cond_arith): Include
multiplication and division.
* doc/md.texi (cond_mul@var{m}, cond_div@var{m}, cond_mod@var{m})
(cond_udiv@var{m}, cond_umod@var{m}): Document.
* optabs.def (cond_smul_optab, cond_sdiv_optab, cond_smod_optab)
(cond_udiv_optab, cond_umod_optab): New optabs.
* internal-fn.def (IFN_COND_MUL, IFN_COND_DIV, IFN_COND_MOD)
(IFN_COND_RDIV): New internal functions.
* internal-fn.c (get_conditional_internal_fn): Handle TRUNC_DIV_EXPR,
TRUNC_MOD_EXPR and RDIV_EXPR.
* match.pd (UNCOND_BINARY, COND_BINARY): Handle them.
* config/aarch64/iterators.md (UNSPEC_COND_MUL, UNSPEC_COND_DIV):
New unspecs.
(SVE_INT_BINARY): Include mult.
(SVE_COND_FP_BINARY): Include UNSPEC_MUL and UNSPEC_DIV.
(optab, sve_int_op): Handle mult.
(optab, sve_fp_op, commutative): Handle UNSPEC_COND_MUL and
UNSPEC_COND_DIV.
* config/aarch64/aarch64-sve.md (cond_<optab><mode>): New pattern
for SVE_INT_BINARY_SD.
gcc/testsuite/
* lib/target-supports.exp
(check_effective_target_vect_double_cond_arith): Include
multiplication and division.
* gcc.dg/vect/pr53773.c: Do not expect a scalar tail when using
fully-masked loops with a fixed vector length.
* gcc.dg/vect/vect-cond-arith-1.c: Add multiplication and division
tests.
* gcc.target/aarch64/sve/vcond_8.c: Likewise.
* gcc.target/aarch64/sve/vcond_9.c: Likewise.
* gcc.target/aarch64/sve/vcond_12.c: Add multiplication tests.
From-SVN: r260713
|
|
After the previous patch to prevent pessimisation of divisions
by constants, this patch adds support for the SVE integer division
instructions.
2018-05-25 Richard Sandiford <richard.sandiford@linaro.org>
gcc/
* config/aarch64/iterators.md (SVE_INT_BINARY_SD): New code iterator.
(optab, sve_int_op): Handle div and udiv.
* config/aarch64/aarch64-sve.md (<optab><mode>3): New expander
for SVE_INT_BINARY_SD.
(*<optab><mode>3): New insn for the same.
gcc/testsuite/
* gcc.target/aarch64/sve/div_1.c: New test.
* gcc.target/aarch64/sve/div_1_run.c: Likewise.
* gcc.target/aarch64/sve/mul_highpart_2.c: Likewise.
* gcc.target/aarch64/sve/mul_highpart_2_run.c: Likewise.
From-SVN: r260712
|
|
This patch adds the folds:
(vec_cond COND (foo A B) C) -> (IFN_COND_FOO COND A B C)
(vec_cond COND C (foo A B)) -> (IFN_COND_FOO (!COND) A B C)
with the usual implicit restriction that the target must support
the produced IFN_COND_FOO.
The results of these folds don't have identical semantics, since
the reverse transform would be invalid if (FOO A[i] B[i]) faults when
COND[i] is false. But this direction is OK since we're simply dropping
faults for operations whose results aren't needed.
The new gimple_resimplify4 doesn't try to do any constant folding
on the IFN_COND_*s. This is because a later patch will handle it
by folding the associated unconditional operation.
Doing this in gimple is better than doing it in .md patterns,
since the second form (with the inverted condition) is much more
common than the first, and it's better to fold away the inversion
in gimple and optimise the result before entering expand.
2018-05-24 Richard Sandiford <richard.sandiford@linaro.org>
gcc/
* doc/sourcebuild.texi (vect_double_cond_arith: Document.
* gimple-match.h (gimple_match_op::MAX_NUM_OPS): Bump to 4.
(gimple_match_op::gimple_match_op): Add an overload for 4 operands.
(gimple_match_op::set_op): Likewise.
(gimple_resimplify4): Declare.
* genmatch.c (get_operand_type): Handle CFN_COND_* functions.
(expr::gen_transform): Likewise.
(decision_tree::gen): Generate a simplification routine for 4 operands.
* gimple-match-head.c (gimple_simplify): Add an overload for
4 operands. In the top-level function, handle up to 4 call
arguments and call gimple_resimplify4.
(gimple_resimplify4): New function.
(build_call_internal): Pass a fourth operand.
(maybe_push_to_seq): Likewise.
* match.pd (UNCOND_BINARY, COND_BINARY): New operator lists.
Fold VEC_COND_EXPRs of an operation and a default value into
an IFN_COND_* function if possible.
* config/aarch64/iterators.md (UNSPEC_COND_MAX, UNSPEC_COND_MIN):
New unspecs.
(SVE_COND_FP_BINARY): Include them.
(optab, sve_fp_op): Handle them.
(SVE_INT_BINARY_REV): New code iterator.
(SVE_COND_FP_BINARY_REV): New int iterator.
(commutative): New int attribute.
* config/aarch64/aarch64-protos.h (aarch64_sve_prepare_conditional_op):
Declare.
* config/aarch64/aarch64.c (aarch64_sve_prepare_conditional_op): New
function.
* config/aarch64/aarch64-sve.md (cond_<optab><mode>): Use it.
(*cond_<optab><mode>): New patterns for reversed operands.
gcc/testsuite/
* lib/target-supports.exp
(check_effective_target_vect_double_cond_arith): New proc.
* gcc.dg/vect/vect-cond-arith-1.c: New test.
* gcc.target/aarch64/sve/vcond_8.c: Likewise.
* gcc.target/aarch64/sve/vcond_8_run.c: Likewise.
* gcc.target/aarch64/sve/vcond_9.c: Likewise.
* gcc.target/aarch64/sve/vcond_9_run.c: Likewise.
* gcc.target/aarch64/sve/vcond_12.c: Likewise.
* gcc.target/aarch64/sve/vcond_12_run.c: Likewise.
From-SVN: r260710
|
|
* configure.ac (gcc_cv_as_section_has_e): Move to common section.
Rename to...
(gcc_cv_as_section_exclude): ... this.
Try Solaris as #exclude syntax.
* configure: Regenerate.
* config.in: Regenerate.
* config/i386/i386.c (i386_solaris_elf_named_section): Handle
SECTION_EXCLUDE.
* config/sparc/sparc.c (sparc_solaris_elf_asm_named_section)
[HAVE_GAS_SECTION_EXCLUDE]: Handle SECTION_EXCLUDE.
* varasm.c (default_elf_asm_named_section): Don't check if
HAVE_GAS_SECTION_EXCLUDE is defined.
From-SVN: r260708
|
|
As suggested by Richard B, this patch changes the IFN_COND_*
functions so that they take the else value of the ?: operation
as a final argument, rather than always using argument 1.
All current callers will still use the equivalent of argument 1,
so this patch makes the SVE code assert that for now. Later patches
add the general case.
2018-05-25 Richard Sandiford <richard.sandiford@linaro.org>
gcc/
* doc/md.texi: Update the documentation of the cond_* optabs
to mention the new final operand. Fix GET_MODE_NUNITS call.
Describe the scalar case too.
* internal-fn.def (IFN_EXTRACT_LAST): Change type to fold_left.
* internal-fn.c (expand_cond_unary_optab_fn): Expect 3 operands
instead of 2.
(expand_cond_binary_optab_fn): Expect 4 operands instead of 3.
(get_conditional_internal_fn): Update comment.
* tree-vect-loop.c (vectorizable_reduction): Pass the original
accumulator value as a final argument to conditional functions.
* config/aarch64/aarch64-sve.md (cond_<optab><mode>): Turn into
a define_expand and add an "else" operand. Assert for now that
the else operand is equal to operand 2. Use SVE_INT_BINARY and
SVE_COND_FP_BINARY instead of SVE_COND_INT_OP and SVE_COND_FP_OP.
(*cond_<optab><mode>): New patterns.
* config/aarch64/iterators.md (UNSPEC_COND_SMAX, UNSPEC_COND_UMAX)
(UNSPEC_COND_SMIN, UNSPEC_COND_UMIN, UNSPEC_COND_AND, UNSPEC_COND_ORR)
(UNSPEC_COND_EOR): Delete.
(optab): Remove associated mappings.
(SVE_INT_BINARY): New code iterator.
(sve_int_op): Remove int attribute and add "minus" to the code
attribute.
(SVE_COND_INT_OP): Delete.
(SVE_COND_FP_OP): Rename to...
(SVE_COND_FP_BINARY): ...this.
From-SVN: r260707
|
|
mnemonic.
* config/i386/sse.md (cvtusi2<ssescalarmodesuffix>64<round_name>):
Add {q} suffix to insn mnemonic.
testsuite/Changelog:
* gcc.target/i386/avx512f-vcvtusi2sd64-1.c: Update scan string.
* gcc.target/i386/avx512f-vcvtusi2ss64-1.c: Ditto.
From-SVN: r260691
|
|
* config/msp430/msp430.c (TARGET_WARN_FUNC_RETURN): Define.
(msp430_warn_func_return): New.
From-SVN: r260690
|
|
PR target/85903
* config/i386/sse.md (movdi_to_sse): Do not generate pseudo
when memory input operand is handled.
From-SVN: r260681
|
|
Switch from using generic address costs to using Falkor-specific ones, which
give Falkor better results overall.
gcc/ChangeLog:
2018-05-24 Luis Machado <luis.machado@linaro.org>
* config/aarch64/aarch64.c (qdf24xx_addrcost_table): New static
global.
(qdf24xx_tunings) <addr_costs>: Set to qdf24xx_addrcost_table.
From-SVN: r260675
|
|
The operand constraint for the memory address of store/load pair lanes was
enforcing strictly hardware registers be allowed as memory addresses. We want
to relax that such that these patterns can be used by combine. During register
allocation the register constraint will enforce the correct register is chosen.
gcc
2018-05-24 Andre Vieira <andre.simoesdiasvieira@arm.com>
PR target/83009
* config/aarch64/predicates.md (aarch64_mem_pair_lanes_operand): Make
address check not strict.
gcc/testsuite
2018-05-24 Andre Vieira <andre.simoesdiasvieira@arm.com>
PR target/83009
* gcc/target/aarch64/store_v2vec_lanes.c: Add extra tests.
From-SVN: r260635
|
|
The following patch adds an option to control software prefetching of memory
references with non-constant/unknown strides.
Currently we prefetch these references if the pass thinks there is benefit to
doing so. But, since this is all based on heuristics, it's not always the case
that we end up with better performance.
For Falkor there is also the problem of conflicts with the hardware prefetcher,
so we need to be more conservative in terms of what we issue software prefetch
hints for.
This also aligns GCC with what LLVM does for Falkor.
Similarly to the previous patch, the defaults guarantee no change in behavior
for other targets and architectures.
gcc/ChangeLog:
2018-05-23 Luis Machado <luis.machado@linaro.org>
* config/aarch64/aarch64-protos.h (cpu_prefetch_tune)
<prefetch_dynamic_strides>: New const bool field.
* config/aarch64/aarch64.c (generic_prefetch_tune): Update to include
prefetch_dynamic_strides.
(exynosm1_prefetch_tune): Likewise.
(thunderxt88_prefetch_tune): Likewise.
(thunderx_prefetch_tune): Likewise.
(thunderx2t99_prefetch_tune): Likewise.
(qdf24xx_prefetch_tune): Likewise. Set prefetch_dynamic_strides to
false.
(aarch64_override_options_internal): Update to set
PARAM_PREFETCH_DYNAMIC_STRIDES.
* doc/invoke.texi (prefetch-dynamic-strides): Document new option.
* params.def (PARAM_PREFETCH_DYNAMIC_STRIDES): New.
* params.h (PARAM_PREFETCH_DYNAMIC_STRIDES): Define.
* tree-ssa-loop-prefetch.c (should_issue_prefetch_p): Account for
prefetch-dynamic-strides setting.
From-SVN: r260618
|
|
This patch adds a new option to control the minimum stride, for a memory
reference, after which the loop prefetch pass may issue software prefetch
hints for. There are two motivations:
* Make the pass less aggressive, only issuing prefetch hints for bigger strides
that are more likely to benefit from prefetching. I've noticed a case in cpu2017
where we were issuing thousands of hints, for example.
* For processors that have a hardware prefetcher, like Falkor, it allows the
loop prefetch pass to defer prefetching of smaller (less than the threshold)
strides to the hardware prefetcher instead. This prevents conflicts between
the software prefetcher and the hardware prefetcher.
I've noticed considerable reduction in the number of prefetch hints and
slightly positive performance numbers. This aligns GCC and LLVM in terms of
prefetch behavior for Falkor.
The default settings should guarantee no changes for existing targets. Those
are free to tweak the settings as necessary.
gcc/ChangeLog:
2018-05-23 Luis Machado <luis.machado@linaro.org>
* config/aarch64/aarch64-protos.h (cpu_prefetch_tune)
<minimum_stride>: New const int field.
* config/aarch64/aarch64.c (generic_prefetch_tune): Update to include
minimum_stride field defaulting to -1.
(exynosm1_prefetch_tune): Likewise.
(thunderxt88_prefetch_tune): Likewise.
(thunderx_prefetch_tune): Likewise.
(thunderx2t99_prefetch_tune): Likewise.
(qdf24xx_prefetch_tune) <minimum_stride>: Set to 2048.
<default_opt_level>: Set to 3.
(aarch64_override_options_internal): Update to set
PARAM_PREFETCH_MINIMUM_STRIDE.
* doc/invoke.texi (prefetch-minimum-stride): Document new option.
* params.def (PARAM_PREFETCH_MINIMUM_STRIDE): New.
* params.h (PARAM_PREFETCH_MINIMUM_STRIDE): Define.
* tree-ssa-loop-prefetch.c (should_issue_prefetch_p): Return false if
stride is constant and is below the minimum stride threshold.
From-SVN: r260617
|
|
* config/arm/arm-cpus.in (mode26): Delete.
(armv4): Delete mode26 reference.
* config/arm/arm.c (arm_configure_build_target): Delete use of
isa_bit_mode26.
From-SVN: r260615
|
|
* config/i386/i386.md (*floatuns<SWI48:mode><MODEF:mode>2_avx512):
New insn pattern.
(floatunssi<mode>2): Also enable for AVX512F and TARGET_SSE_MATH.
Rewrite expander pattern. Emit gen_floatunssi<mode>2_i387_with_xmm
for non-SSE modes.
(floatunsdisf2): Rewrite expander pattern. Hanlde TARGET_AVX512F.
(floatunsdidf2): Ditto.
* config/i386/i386.md (fixuns_trunc<mode>di2): New insn pattern.
(fixuns_trunc<mode>si2_avx512f): Ditto.
(*fixuns_trunc<mode>si2_avx512f_zext): Ditto.
(fixuns_trunc<mode>si2): Also enable for AVX512F and TARGET_SSE_MATH.
Emit fixuns_trunc<mode>si2_avx512f for AVX512F targets.
testsuite/ChangeLog:
* gcc.target/i386/cvt-2.c: New test.
* gcc.target/i386/cvt-3.c: New test.
From-SVN: r260614
|
|
Simplify frame pointer logic. Add aarch64_needs_frame_chain to
decide when to emit the frame chain using clearer logic.
Introduce aarch64_use_frame_pointer which contains the value of
-fno-omit-frame-pointer (flag_omit_frame_pointer is set to a magic
value so that the mid-end won't force the frame pointer in all
cases, and leaf frame pointer emission can't be supported).
gcc/
* config/aarch64/aarch64.c (aarch64_use_frame_pointer):
Add new boolean.
(aarch64_needs_frame_chain): New function.
(aarch64_parse_override_string): Set aarch64_use_frame_pointer.
From-SVN: r260606
|
|
*** gcc/ChangeLog ***
2018-05-23 Sudakshina Das <sudi.das@arm.com>
PR target/84882
* common/config/aarch64/aarch64-common.c (aarch64_handle_option):
Check val before adding MASK_STRICT_ALIGN to opts->x_target_flags.
* config/aarch64/aarch64.opt (mstrict-align): Remove RejectNegative.
* config/aarch64/aarch64.c (aarch64_attributes): Mark allow_neg
as true for strict-align.
(aarch64_can_inline_p): Perform checks even when callee has no
attributes to check for strict alignment.
* doc/extend.texi (AArch64 Function Attributes): Document
no-strict-align.
* doc/invoke.texi: (AArch64 Options): Likewise.
*** gcc/testsuite/ChangeLog ***
2018-05-23 Sudakshina Das <sudi.das@arm.com>
PR target/84882
* gcc.target/aarch64/pr84882.c: New test.
* gcc.target/aarch64/target_attr_18.c: Likewise.
From-SVN: r260604
|
|
A customer reported the following missed opportunities to combine a couple
instructions into a sbfiz.
int sbfiz32 (int x)
{
return x << 29 >> 10;
}
long sbfiz64 (long x)
{
return x << 58 >> 20;
}
This gets converted to the following pattern:
(set (reg:SI 98)
(ashift:SI (sign_extend:SI (reg:HI 0 x0 [ xD.3334 ]))
(const_int 6 [0x6])))
Currently, gcc generates the following:
sbfiz32:
lsl x0, x0, 29
asr x0, x0, 10
ret
sbfiz64:
lsl x0, x0, 58
asr x0, x0, 20
ret
It could generate this instead:
sbfiz32:
sbfiz w0, w0, 19, 3
ret
sbfiz64::
sbfiz x0, x0, 38, 6
ret
The unsigned versions already generate ubfiz for the same code, so the lack of
a sbfiz pattern may have been an oversight.
This particular sbfiz pattern shows up in both CPU2006 (~ 80 hits) and
CPU2017 (~ 280 hits). It's not a lot, but seems beneficial in any case. No
significant performance differences, probably due to the small number of
occurrences or cases outside hot areas.
gcc/ChangeLog:
2018-05-22 Luis Machado <luis.machado@linaro.org>
gcc/
* config/aarch64/aarch64.md (*ashift<mode>_extv_bfiz): New pattern.
gcc/testsuite/ChangeLog:
2018-05-22 Luis Machado <luis.machado@linaro.org>
gcc/testsuite/
* gcc.target/aarch64/lsl_asr_sbfiz.c: New test.
From-SVN: r260546
|
|
This patch removes a lot of duplicated code in aarch64-ldpstp.md.
The patterns that did not previously generate a base register now
do not check for aarch64_mem_pair_operand in the pattern. This has
been extracted to a check in aarch64_operands_ok_for_ldpstp.
All patterns in the file used to have explicit switching code to
swap loads and stores that were in the wrong order.
This has been extracted into aarch64_operands_ok_for_ldpstp
as a final operation after all the checks have been performed.
2018-05-22 Jackson Woodruff <jackson.woodruff@arm.com>
Kyrylo Tkachov <kyrylo.tkachov@arm.com>
* config/aarch64/aarch64-ldpstp.md: Replace uses of
aarch64_mem_pair_operand with memory_operand and delete operand swapping
code.
* config/aarch64/aarch64.c (aarch64_operands_ok_for_ldpstp):
Add check for legitimate_address.
(aarch64_gen_adjusted_ldpstp): Swap operands where appropriate.
(aarch64_swap_ldrstr_operands): New.
* config/aarch64/aarch64-protos.h (aarch64_swap_ldrstr_operands):
Define prototype.
Co-Authored-By: Kyrylo Tkachov <kyrylo.tkachov@arm.com>
From-SVN: r260539
|
|
This patch merges loads and stores from D-registers that are of different modes.
Code like this:
typedef int __attribute__((vector_size(8))) vec;
struct pair
{
vec v;
double d;
}
Now generates a store pair instruction:
void
assign (struct pair *p, vec v)
{
p->v = v;
p->d = 1.0;
}
Whereas previously it generated two `str` instructions.
This patch also merges storing of double zero values with
long integer values:
struct pair
{
long long l;
double d;
}
void
foo (struct pair *p)
{
p->l = 10;
p->d = 0.0;
}
Now generates a single store pair instruction rather than two `str` instructions.
The patch basically generalises the mode iterators on the patterns in aarch64.md
and the peepholes in aarch64-ldpstp.md to take all combinations of pairs of modes
so, while it may be a large-ish patch, it does fairly mechanical stuff.
2018-05-22 Jackson Woodruff <jackson.woodruff@arm.com>
Kyrylo Tkachov <kyrylo.tkachov@arm.com>
* config/aarch64/aarch64.md: New patterns to generate stp
and ldp.
(store_pair_sw, store_pair_dw): New patterns to generate stp for
single words and double words.
(load_pair_sw, load_pair_dw): Likewise.
(store_pair_sf, store_pair_df, store_pair_si, store_pair_di):
Delete.
(load_pair_sf, load_pair_df, load_pair_si, load_pair_di):
Delete.
* config/aarch64/aarch64-ldpstp.md: Modify peephole
for different mode ldpstp and add peephole for merged zero stores.
Likewise for loads.
* config/aarch64/aarch64.c (aarch64_operands_ok_for_ldpstp):
Add size check.
(aarch64_gen_store_pair): Rename calls to match new patterns.
(aarch64_gen_load_pair): Rename calls to match new patterns.
* config/aarch64/aarch64-simd.md (load_pair<mode>): Rename to...
(load_pair<DREG:mode><DREG2:mode>): ... This.
(store_pair<mode>): Rename to...
(vec_store_pair<DREG:mode><DREG2:mode>): ... This.
* config/aarch64/iterators.md (DREG, DREG2, DX2, SX, SX2, DSX):
New mode iterators.
(V_INT_EQUIV): Handle SImode.
* config/aarch64/predicates.md (aarch64_reg_zero_or_fp_zero):
New predicate.
* gcc.target/aarch64/ldp_stp_6.c: New.
* gcc.target/aarch64/ldp_stp_7.c: New.
* gcc.target/aarch64/ldp_stp_8.c: New.
Co-Authored-By: Kyrylo Tkachov <kyrylo.tkachov@arm.com>
From-SVN: r260538
|
|
the IBM double-double format)
[gcc]
2018-05-21 Michael Meissner <meissner@linux.ibm.com>
PR target/85657
* config/rs6000/rs6000-c.c (rs6000_cpu_cpp_builtins): Do not
define __ibm128 as long double.
* config/rs6000/rs6000.c (rs6000_init_builtins): Always create
__ibm128 as a distinct type.
(init_float128_ieee): Fix up conversions between IFmode and IEEE
128-bit types to use the correct functions.
(rs6000_expand_float128_convert): Use explicit FLOAT_EXTEND to
convert between 128-bit floating point types that have different
modes but the same representation, instead of using gen_lowpart to
makean alias.
* config/rs6000/rs6000.md (IFKF): New iterator for IFmode and
KFmode.
(IFKF_reg): New attributes to give the register constraints for
IFmode and KFmode.
(extend<mode>tf2_internal): New insns to mark an explicit
conversion between 128-bit floating point types that have a
different mode but share the same representation.
[gcc/testsuite]
2018-05-21 Michael Meissner <meissner@linux.ibm.com>
PR target/85657
* gcc.target/powerpc/pr85657-1.c: New test for converting between
__float128, __ibm128, and long double.
* gcc.target/powerpc/pr85657-2.c: Likewise.
* gcc.target/powerpc/pr85657-3.c: Likewise.
* g++.dg/pr85667.C: New test to make sure __ibm128 is
implementated as a separate type internally, and is not just an
alias for long double.
From-SVN: r260489
|
|
This patch implements the usadv16qi and ssadv16qi standard names.
See the thread at on gcc@gcc.gnu.org [1] for background.
The V16QImode variant is important to get right as it is the most commonly used pattern:
reducing vectors of bytes into an int.
The midend expects the optab to compute the absolute differences of operands 1 and 2 and
reduce them while widening along the way up to SImode. So the inputs are V16QImode and
the output is V4SImode.
I've tried out a few different strategies for that, the one I settled with is to emit:
UABDL2 tmp.8h, op1.16b, op2.16b
UABAL tmp.8h, op1.16b, op2.16b
UADALP op3.4s, tmp.8h
To work through the semantics let's say operands 1 and 2 are:
op1 { a0, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12, a13, a14, a15 }
op2 { b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, b10, b11, b12, b13, b14, b15 }
op3 { c0, c1, c2, c3 }
The UABDL2 takes the upper V8QI elements, computes their absolute differences, widens them and stores them into the V8HImode tmp:
tmp { ABS(a[8]-b[8]), ABS(a[9]-b[9]), ABS(a[10]-b[10]), ABS(a[11]-b[11]), ABS(a[12]-b[12]), ABS(a[13]-b[13]), ABS(a[14]-b[14]), ABS(a[15]-b[15]) }
The UABAL after that takes the lower V8QI elements, computes their absolute differences, widens them and accumulates them into the V8HImode tmp from the previous step:
tmp { ABS(a[8]-b[8])+ABS (a[0]-b[0]), ABS(a[9]-b[9])+ABS(a[1]-b[1]), ABS(a[10]-b[10])+ABS(a[2]-b[2]), ABS(a[11]-b[11])+ABS(a[3]-b[3]), ABS(a[12]-b[12])+ABS(a[4]-b[4]), ABS(a[13]-b[13])+ABS(a[5]-b[5]), ABS(a[14]-b[14])+ABS(a[6]-b[6]), ABS(a[15]-b[15])+ABS(a[7]-b[7]) }
Finally the UADALP does a pairwise widening reduction and accumulation into the V4SImode op3:
op3 { c0+ABS(a[8]-b[8])+ABS(a[0]-b[0])+ABS(a[9]-b[9])+ABS(a[1]-b[1]), c1+ABS(a[10]-b[10])+ABS(a[2]-b[2])+ABS(a[11]-b[11])+ABS(a[3]-b[3]), c2+ABS(a[12]-b[12])+ABS(a[4]-b[4])+ABS(a[13]-b[13])+ABS(a[5]-b[5]), c3+ABS(a[14]-b[14])+ABS(a[6]-b[6])+ABS(a[15]-b[15])+ABS(a[7]-b[7]) }
(sorry for the text dump)
Remember, according to [1] the exact reduction sequence doesn't matter (for integer arithmetic at least).
I've considered other sequences as well (thanks Wilco), for example
* UABD + UADDLP + UADALP
* UABLD2 + UABDL + UADALP + UADALP
I ended up settling in the sequence in this patch as it's short (3 instructions) and in the future we can potentially
look to optimise multiple occurrences of these into something even faster (for example accumulating into H registers for longer
before doing a single UADALP in the end to accumulate into the final S register).
If your microarchitecture has some some strong preferences for a particular sequence, please let me know or, even better, propose a patch
to parametrise the generation sequence by code (or the appropriate RTX cost).
This expansion allows the vectoriser to avoid unpacking the bytes in two steps and performing V4SI arithmetic on them.
So, for the code:
unsigned char pix1[N], pix2[N];
int foo (void)
{
int i_sum = 0;
int i;
for (i = 0; i < 16; i++)
i_sum += __builtin_abs (pix1[i] - pix2[i]);
return i_sum;
}
we now generate on aarch64:
foo:
adrp x1, pix1
add x1, x1, :lo12:pix1
movi v0.4s, 0
adrp x0, pix2
add x0, x0, :lo12:pix2
ldr q2, [x1]
ldr q3, [x0]
uabdl2 v1.8h, v2.16b, v3.16b
uabal v1.8h, v2.8b, v3.8b
uadalp v0.4s, v1.8h
addv s0, v0.4s
umov w0, v0.s[0]
ret
instead of:
foo:
adrp x1, pix1
adrp x0, pix2
add x1, x1, :lo12:pix1
add x0, x0, :lo12:pix2
ldr q0, [x1]
ldr q4, [x0]
ushll v1.8h, v0.8b, 0
ushll2 v0.8h, v0.16b, 0
ushll v2.8h, v4.8b, 0
ushll2 v4.8h, v4.16b, 0
usubl v3.4s, v1.4h, v2.4h
usubl2 v1.4s, v1.8h, v2.8h
usubl v2.4s, v0.4h, v4.4h
usubl2 v0.4s, v0.8h, v4.8h
abs v3.4s, v3.4s
abs v1.4s, v1.4s
abs v2.4s, v2.4s
abs v0.4s, v0.4s
add v1.4s, v3.4s, v1.4s
add v1.4s, v2.4s, v1.4s
add v0.4s, v0.4s, v1.4s
addv s0, v0.4s
umov w0, v0.s[0]
ret
So I expect this new expansion to be better than the status quo in any case.
Bootstrapped and tested on aarch64-none-linux-gnu.
This gives about 8% on 525.x264_r from SPEC2017 on a Cortex-A72.
* config/aarch64/aarch64.md ("unspec"): Define UNSPEC_SABAL,
UNSPEC_SABDL2, UNSPEC_SADALP, UNSPEC_UABAL, UNSPEC_UABDL2,
UNSPEC_UADALP values.
* config/aarch64/iterators.md (ABAL): New int iterator.
(ABDL2): Likewise.
(ADALP): Likewise.
(sur): Add mappings for the above.
* config/aarch64/aarch64-simd.md (aarch64_<sur>abdl2<mode>_3):
New define_insn.
(aarch64_<sur>abal<mode>_4): Likewise.
(aarch64_<sur>adalp<mode>_3): Likewise.
(<sur>sadv16qi): New define_expand.
* gcc.c-torture/execute/ssad-run.c: New test.
* gcc.c-torture/execute/usad-run.c: Likewise.
* gcc.target/aarch64/ssadv16qi.c: Likewise.
* gcc.target/aarch64/usadv16qi.c: Likewise.
From-SVN: r260437
|
|
2018-05-21 Alexander Nesterovskiy <alexander.nesterovskiy@intel.com>
gcc/
* config/i386/i386.md (*movsf_internal): AVX falsedep fix.
(*movdf_internal): Ditto.
(*rcpsf2_sse): Ditto.
(*rsqrtsf2_sse): Ditto.
(*sqrt<mode>2_sse): Ditto.
From-SVN: r260436
|
|
This patch adds the missing neon intrinsics for all 128 bit vector Integer modes for the
three-way XOR and negate and xor instructions for Arm8.2-a to Armv8.4-a.
gcc/
2018-05-21 Tamar Christina <tamar.christina@arm.com>
* config/aarch64/aarch64-simd.md (aarch64_eor3qv8hi): Change to
eor3q<mode>4.
(aarch64_bcaxqv8hi): Change to bcaxq<mode>4.
* config/aarch64/aarch64-simd-builtins.def (veor3q_u8, veor3q_u32,
veor3q_u64, veor3q_s8, veor3q_s16, veor3q_s32, veor3q_s64, vbcaxq_u8,
vbcaxq_u32, vbcaxq_u64, vbcaxq_s8, vbcaxq_s16, vbcaxq_s32,
vbcaxq_s64): New.
* config/aarch64/arm_neon.h: Likewise.
* config/aarch64/iterators.md (VQ_I): New.
gcc/testsuite/
2018-05-21 Tamar Christina <tamar.christina@arm.com>
* gcc.target/gcc.target/aarch64/sha3.h (veor3q_u8, veor3q_u32,
veor3q_u64, veor3q_s8, veor3q_s16, veor3q_s32, veor3q_s64, vbcaxq_u8,
vbcaxq_u32, vbcaxq_u64, vbcaxq_s8, vbcaxq_s16, vbcaxq_s32,
vbcaxq_s64): New.
* gcc.target/gcc.target/aarch64/sha3_1.c: Likewise.
* gcc.target/gcc.target/aarch64/sha3_1.c: Likewise.
* gcc.target/gcc.target/aarch64/sha3_1.c: Likewise.
From-SVN: r260435
|
|
We used to build baremetal (AKA Elf32) multilibbed toolchains for years
now but never made that for Linux targets since there were problems with
uClibc n multilib setup. Now with help of Crosstool-NG it is finally
possible to create uClibc-based multilibbed toolchains and so we add
relevant CPUs for multilib in case of configuration for "arc*-*-linux*".
This will be essentially useful for glibc-based multilibbbed toolchains
in the future.
gcc/
2018-05-16 Alexey Brodkin <abrodkin@synopsys.com>
* config.gcc: Add arc/t-multilib-linux to tmake_file for
arc*-*-linux*.
* config/arc/t-multilib-linux: Specify MULTILIB_OPTIONS and
MULTILIB_DIRNAMES
From-SVN: r260434
|
|
gcc/
* config/nds32/constraints.md (S): New constraint.
* config/nds32/nds32.md (call_internal): Use constraint S.
(call_value_internal): Likewise.
(sibcall_internal): Likewise.
(sibcall_value_internal): Likewise.
From-SVN: r260422
|
|
gcc/
* config/nds32/nds32.c (nds32_register_move_cost): Take garywolf cpu
into consideration.
Co-Authored-By: Chung-Ju Wu <jasonwucj@gmail.com>
From-SVN: r260412
|
|
gcc/
* config/nds32/nds32-cost.c (rtx_cost_model_t): New structure.
(insn_size_16bit, insn_size_32bit): New variables for cost evaluation.
(nds32_rtx_costs_impl): Simplify.
(nds32_address_cost_impl): Simplify.
(nds32_init_rtx_costs): New function.
(nds32_rtx_costs_speed_prefer): Likewise.
(nds32_rtx_costs_size_prefer): Likewise.
(nds32_address_cost_speed_prefer): Likewise.
(nds32_address_cost_speed_fwprop): Likewise.
(nds32_address_cost_size_prefer): Likewise.
* config/nds32/nds32-protos.h (nds32_init_rtx_costs): Declare.
* config/nds32/nds32.c (nds32_option_override): Use
nds32_init_rtx_costs function.
Co-Authored-By: Chung-Ju Wu <jasonwucj@gmail.com>
From-SVN: r260411
|
|
gcc/
* config/nds32/nds32.c (nds32_asm_file_start): Output pipeline model.
* config/nds32/nds32.h (TARGET_PIPELINE_N7): Define.
(TARGET_PIPELINE_N8): Likewise.
(TARGET_PIPELINE_N10): Likewise.
(TARGET_PIPELINE_N13): Likewise.
(TARGET_PIPELINE_GRAYWOLF): Likewise.
From-SVN: r260409
|
|
gcc/
* config/nds32/nds32-fpu.md: Update copyright year.
From-SVN: r260402
|
|
gcc/
* config/nds32/nds32.h (ASM_SPEC): Adjust spec rule.
From-SVN: r260401
|
|
gcc/
* config/nds32/nds32.c
(nds32_md_asm_adjust): Consider flag_inline_asm_r15 variable.
* config/nds32/nds32.opt (minline-asm-r15): New option.
From-SVN: r260400
|
|
gcc/
* common/config/nds32/nds32-common.c (TARGET_DEFAULT_TARGET_FLAGS): Add
MASK_HW_ABS.
* config/nds32/nds32.md (abssi2): New pattern.
From-SVN: r260398
|
|
* config/i386/i386.md (rex64namesuffix): New mode attribute.
* config/i386/sse.md (sse_cvtsi2ss<rex64namesuffix><round_name>):
Merge insn pattern from sse_cvtsi2ss<round_name> and
sse_cvtsi2ssq<round_name> using SWI48 mode iterator.
(sse_cvtss2si<rex64namesuffix><round_name>): Merge insn pattern
from sse_cvtss2si<round_name> and sse_cvtss2siq<round_name>
using SWI48 mode iterator.
(sse_cvtss2si<rex64namesuffix>_2): Merge insn pattern from
sse_cvtss2si_2 and sse_cvtss2siq_2 using SWI48 mode iterator.
(sse_cvttss2si<rex64namesuffix><round_saeonly_name>): Merge insn
pattern from sse_cvttss2si<round_saeonly_name>
and sse_cvttss2siq<round_saeonly_name> using SWI48 mode iterator.
(avx512f_vcvtss2usi<rex64namesuffix><round_name>): Merge insn pattern
from avx512f_vcvtss2usi<round_name> and avx512f_vcvtss2usiq<round_name>
using SWI48 mode iterator.
(avx512f_vcvttss2usi<rex64namesuffix><round_saeonly_name>): Merge
insn pattern from avx512f_vcvttss2usi<round_saeonly_name> and
avx512f_vcvttss2usiq<round_saeonly_name> using SWI48 mode iterator.
(avx512f_vcvtsd2usi<rex64namesuffix><round_name>): Merge insn pattern
from avx512f_vcvtsd2usi<round_name> and avx512f_vcvtsd2usiq<round_name>
using SWI48 mode iterator.
(avx512f_vcvttsd2usi<rex64namesuffix><round_saeonly_name>): Merge
insn pattern from avx512f_vcvttsd2usi<round_saeonly_name> and
avx512f_vcvttsd2usiq<round_saeonly_name> using SWI48 mode iterator.
(sse2_cvtsd2si<rex64namesuffix><round_name>): Merge insn pattern from
sse2_cvtsd2si<round_name> and sse2_cvtsd2siq<round_name> using
SWI48 mode iterator.
(sse2_cvtsd2si<rex64namesuffix>_2): Merge insn pattern from
sse2_cvtsd2si_2 and sse2_cvtsd2siq_2 using SWI48 mode iterator.
(sse_cvttsd2si<rex64namesuffix><round_saeonly_name>): Merge insn
pattern from sse_cvttsd2si<round_saeonly_name>
and sse_cvttsd2siq<round_saeonly_name> using SWI48 mode iterator.
From-SVN: r260397
|
|
gcc/
* config/nds32/nds32-md-auxiliary.c
(nds32_valid_smw_lwm_base_p): Refine.
(nds32_output_smw_single_word): Refine.
(nds32_output_smw_double_word): New.
* config/nds32/nds32-protos.h (nds32_output_smw_double_word): New.
From-SVN: r260396
|
|
gcc/
* config/nds32/nds32-md-auxiliary.c (nds32_output_stack_push): Refine.
(nds32_output_stack_pop): Refine.
(nds32_expand_unaligned_load): Refine.
(nds32_expand_unaligned_store): Refine.
From-SVN: r260394
|
|
gcc/
* config/nds32/constants.md: Add TP_REGNUM constant.
(unspec_element): Add UNSPEC_GOTINIT, UNSPEC_GOT, UNSPEC_GOTOFF,
UNSPEC_PLT, UNSPEC_TLSGD, UNSPEC_TLSLD, UNSPEC_TLSIE, UNSPEC_TLSLE and
UNSPEC_ADD32.
* config/nds32/nds32-doubleword.md: Consider flag_pic.
* config/nds32/nds32-dspext.md (mov<mode>): Expand TLS and PIC cases.
* config/nds32/nds32-predicates.c (nds32_const_unspec_p): New.
* config/nds32/nds32-md-auxiliary.c: Implementation that support TLS
and PIC code generation.
* config/nds32/nds32-protos.h: Declarations that support TLS and PIC
code generation.
* config/nds32/nds32-relax-opt.c: Consider TLS and PIC for relax
optimization.
* config/nds32/nds32.md: Support TLS and PIC.
* config/nds32/nds32.c: Support TLS and PIC.
* config/nds32/nds32.h (nds32_relax_insn_type): New enum type.
* config/nds32/predicates.md (nds32_nonunspec_symbolic_operand): New
predicate.
Co-Authored-By: Chung-Ju Wu <jasonwucj@gmail.com>
From-SVN: r260393
|
|
gcc/
* config/nds32/nds32-predicates.c (const_vector_to_hwint): Use machine
mode with E_ prefix.
From-SVN: r260391
|
|
* config/nds32/constants.md (unspec_element): Add UNSPEC_ICT.
* config/nds32/nds32-md-auxiliary.c
(symbolic_reference_mentioned_p): New.
(nds32_legitimize_ict_address): New.
(nds32_expand_ict_move): New.
(nds32_indirect_call_referenced_p): New.
(nds32_symbol_binds_local_p): Delete.
(nds32_long_call_p): Modify.
* config/nds32/nds32-opts.h (nds32_ict_model_type): New enum type.
* config/nds32/nds32-protos.h
(symbolic_reference_mentioned_p): Declare.
(nds32_legitimize_ict_address): Declare.
(nds32_expand_ict_move): Declare.
(nds32_indirect_call_referenced_p): Declare.
* config/nds32/nds32-relax-opt.c (nds32_ict_const_p): New.
(nds32_relax_group): Use nds32_ict_const_p as condition.
* config/nds32/nds32.c (nds32_attribute_table): Add "indirect_call".
(nds32_asm_file_start): Output ict_model directive in asm code.
(nds32_legitimate_address_p): Consider indirect call.
(nds32_print_operand): Consider indirect call.
(nds32_print_operand_address): Consider indirect call.
(nds32_insert_attributes): Handle "indirect_call" attribute.
(TARGET_LEGITIMATE_ADDRESS_P): Define.
(TARGET_LEGITIMATE_CONSTANT_P): Define.
(TARGET_CANNOT_FORCE_CONST_MEM): Define.
(TARGET_DELEGITIMIZE_ADDRESS): Define.
(TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA): Define.
* config/nds32/nds32.h (SYMBOLIC_CONST_P): Define.
(TARGET_ICT_MODEL_SMALL): Define.
(TARGET_ICT_MODEL_LARGE): Define.
* config/nds32/nds32.md (movsi): Consider ict model.
(call, call_value): Consider ict model.
(sibcall, sibcall_value): Consider ict model.
* config/nds32/nds32.opt (mict-model): New option.
* config/nds32/predicates.md (nds32_symbolic_operand): Consider ict
model.
Co-Authored-By: Chung-Ju Wu <jasonwucj@gmail.com>
From-SVN: r260390
|
|
Kito Cheng <kito.cheng@gmail.com>
Monk Chiang <sh.chiang04@gmail.com>
gcc/
* common/config/riscv/riscv-common.c (riscv_parse_arch_string):
Add support to parse rv32e*. Clear MASK_RVE for rv32i and rv64i.
* config.gcc (riscv*-*-*): Add support for rv32e* and ilp32e.
* config/riscv/riscv-c.c (riscv_cpu_cpp_builtins): Define
__riscv_32e when TARGET_RVE. Handle ABI_ILP32E as soft-float ABI.
* config/riscv/riscv-opts.h (riscv_abi_type): Add ABI_ILP32E.
* config/riscv/riscv.c (riscv_compute_frame_info): When TARGET_RVE,
compute save_libcall_adjustment properly.
(riscv_option_override): Call error if TARGET_RVE and not ABI_ILP32E.
(riscv_conditional_register_usage): Handle TARGET_RVE and ABI_ILP32E.
* config/riscv/riscv.h (UNITS_PER_FP_ARG): Handle ABI_ILP32E.
(STACK_BOUNDARY, ABI_STACK_BOUNDARY): Handle TARGET_RVE.
(GP_REG_LAST, MAX_ARGS_IN_REGISTERS): Likewise.
(ABI_SPEC): Handle mabi=ilp32e.
* config/riscv/riscv.opt (abi_type): Add ABI_ILP32E.
(RVE): Add RVE mask.
* doc/invoke.texi (RISC-V options) <-mabi>: Add ilp32e info.
<-march>: Add rv32e as an example.
gcc/testsuite/
* gcc.dg/stack-usage-1.c: Add support for rv32e.
libgcc/
* config/riscv/save-restore.S: Add support for rv32e.
Co-Authored-By: Jim Wilson <jimw@sifive.com>
Co-Authored-By: Monk Chiang <sh.chiang04@gmail.com>
From-SVN: r260384
|
|
(sparc_expand_builtin) breaks SPARC bootstrap)
PR bootstrap/85838
* config/sparc/sparc.c (sparc_expand_builtin): Always initialize op[0].
From-SVN: r260374
|
|
We deprecated architecture versions earlier than Armv4T in GCC 6 [1].
This patch removes support for architectures lower than Armv4.
That is the -march values armv2, armv2a, armv3, armv3m are removed
with this patch. I did not remove armv4 because it's a bit more
involved code-wise and there has been some pushback on the implications
for -mcpu=strongarm support.
Removing armv3m and earlier though is pretty straightforward.
This allows us to get rid of the armv3m and mode32 feature bits
in arm-cpus.in as they can be assumed to be universally available.
Consequently the mcpu values arm2, arm250, arm3, arm6, arm60, arm600, arm610, arm620, arm7, arm7d, arm7di, arm70, arm700, arm700i, arm710, arm720, arm710c, arm7100, arm7500, arm7500fe, arm7m, arm7dm, arm7dm are now also removed.
Bootstrapped and tested on arm-none-linux-gnueabihf and on arm-none-eabi
with an aprofile multilib configuration (which builds quite a lot of library
configurations).
[1] https://gcc.gnu.org/gcc-6/changes.html#arm
* config/arm/arm-cpus.in (armv3m, mode32): Delete features.
(ARMv4): Update.
(ARMv2, ARMv3, ARMv3m): Delete fgroups.
(ARMv6m): Update.
(armv2, armv2a, armv3, armv3m): Delete architectures.
(arm2, arm250, arm3, arm6, arm60, arm600, arm610, arm620,
arm7, arm7d, arm7di, arm70, arm700, arm700i, arm710, arm720,
arm710c, arm7100, arm7500, arm7500fe, arm7m, arm7dm, arm7dmi):
Delete cpus.
* config/arm/arm.md (maddsidi4): Remove check for arm_arch3m.
(*mulsidi3adddi): Likewise.
(mulsidi3): Likewise.
(*mulsidi3_nov6): Likewise.
(umulsidi3): Likewise.
(umulsidi3_nov6): Likewise.
(umaddsidi4): Likewise.
(*umulsidi3adddi): Likewise.
(smulsi3_highpart): Likewise.
(*smulsi3_highpart_nov6): Likewise.
(umulsi3_highpart): Likewise.
(*umulsi3_highpart_nov6): Likewise.
* config/arm/arm.h (arm_arch3m): Delete.
* config/arm/arm.c (arm_arch3m): Delete.
(arm_option_override_internal): Update armv3-related comment.
(arm_configure_build_target): Delete use of isa_bit_mode32.
(arm_option_reconfigure_globals): Delete set of arm_ach3m.
(arm_rtx_costs_internal): Delete check of arm_arch3m.
* config/arm/arm-fixed.md (mulsq3): Delete check for arm_arch3m.
(mulsa3): Likewise.
(mulusa3): Likewise.
* config/arm/arm-protos.h (arm_arch3m): Delete.
* config/arm/arm-tables.opt: Regenerate.
* config/arm/arm-tune.md: Likewise.
* config/arm/t-arm-elf (all_early_nofp): Delete mentions of
deleted architectures.
* gcc.target/arm/pr62554.c: Delete.
* gcc.target/arm/pr69610-1.c: Likewise.
* gcc.target/arm/pr69610-2.c: Likewise.
From-SVN: r260363
|
|
The -march=armv5 and armv5e options have been deprecated in GCC 7 [1].
This patch removes support for them.
It's mostly mechanical stuff. The functionality that was previously
gated on arm_arch5 is now gated on arm_arch5t and the functionality
that was gated on arm_arch5e is now gated on arm_arch5te.
A path in TARGET_OS_CPP_BUILTINS for VxWorks is now unreachable and
therefore is deleted.
References to armv5 and armv5e are deleted/updated throughout the
source tree and testsuite.
Bootstrapped and tested on arm-none-linux-gnueabihf.
Also built a cc1 for arm-wrs-vxworks as a sanity check.
* config/arm/arm-cpus.in (armv5, armv5e): Delete features.
(armv5t, armv5te): New features.
(ARMv5, ARMv5e): Delete fgroups.
(ARMv5t, ARMv5te): Adjust for above changes.
(ARMv6m): Likewise.
(armv5, armv5e): Delete arches.
* config/arm/arm.md (*call_reg_armv5): Use arm_arch5t instead of
arm_arch5.
(*call_reg_arm): Likewise.
(*call_value_reg_armv5): Likewise.
(*call_value_reg_arm): Likewise.
(*call_symbol): Likewise.
(*call_value_symbol): Likewise.
(*sibcall_insn): Likewise.
(*sibcall_value_insn): Likewise.
(clzsi2): Likewise.
(prefetch): Likewise.
(define_split and define_peephole2 dependent on arm_arch5):
Likewise.
* config/arm/arm.h (TARGET_LDRD): Use arm_arch5te instead of
arm_arch5e.
(TARGET_ARM_QBIT): Likewise.
(TARGET_DSP_MULTIPLY): Likewise.
(enum base_architecture): Delete BASE_ARCH_5, BASE_ARCH_5E.
(arm_arch5, arm_arch5e): Delete.
(arm_arch5t, arm_arch5te): Declare.
* config/arm/arm.c (arm_arch5, arm_arch5e): Delete.
(arm_arch5t): Declare.
(arm_option_reconfigure_globals): Update for the above.
(arm_options_perform_arch_sanity_checks): Update comment, replace
use of arm_arch5 with arm_arch5t.
(use_return_insn): Likewise.
(arm_emit_call_insn): Likewise.
(output_return_instruction): Likewise.
(arm_final_prescan_insn): Likewise.
(arm_coproc_builtin_available): Likewise.
* config/arm/arm-c.c (arm_cpu_builtins): Replace arm_arch5 and
arm_arch5e with arm_arch5t and arm_arch5te.
* config/arm/arm-protos.h (arm_arch5, arm_arch5e): Delete.
(arm_arch5t, arm_arch5te): Declare.
* config/arm/arm-tables.opt: Regenerate.
* config/arm/t-arm-elf: Remove references to armv5, armv5e.
* config/arm/t-multilib: Likewise.
* config/arm/thumb1.md (*call_reg_thumb1_v5): Check arm_arch5t
instead of arm_arch5.
(*call_reg_thumb1): Likewise.
(*call_value_reg_thumb1_v5): Likewise.
(*call_value_reg_thumb1): Likewise.
* config/arm/vxworks.h (TARGET_OS_CPP_BUILTINS): Remove now
unreachable path.
* doc/invoke.texi (ARM Options): Remove references to armv5, armv5e.
* gcc.target/arm/pr40887.c: Update comment.
* lib/target-supports.exp: Don't generate effective target checks
and related helpers for armv5. Update comment.
* gcc.target/arm/armv5_thumb_isa.c: Delete.
* gcc.target/arm/di-longlong64-sync-withhelpers.c: Update effective
target check and options.
* config/arm/libunwind.S: Update comment relating to armv5.
From-SVN: r260362
|
|
We've a deficiency in our vec_set family of patterns.
We don't support directly loading a vector lane using LD1 for V2DImode and all the vector floating-point modes.
We do do it correctly for the other integer vector modes (V4SI, V8HI etc) though.
The alternatives on the relative floating-point patterns only allow a register-to-register INS instruction.
That means if we want to load a value into a vector lane we must first load it into a scalar register and then
perform an INS, which is wasteful.
There is also an explicit V2DI vec_set expander dangling around for no reason that I can see. It seems to do the
exact same things as the other vec_set expanders. This patch removes that.
It now unifies all vec_set expansions into a single "vec_set<mode>" define_expand using the catch-all VALL_F16 iterator.
With this patch we avoid loading values into scalar registers and then doing an explicit INS on them to move them into
the desired vector lanes. For example for:
typedef float v4sf __attribute__ ((vector_size (16)));
typedef long long v2di __attribute__ ((vector_size (16)));
v2di
foo_v2di (long long *a, long long *b)
{
v2di res = { *a, *b };
return res;
}
v4sf
foo_v4sf (float *a, float *b, float *c, float *d)
{
v4sf res = { *a, *b, *c, *d };
return res;
}
we currently generate:
foo_v2di:
ldr d0, [x0]
ldr x0, [x1]
ins v0.d[1], x0
ret
foo_v4sf:
ldr s0, [x0]
ldr s3, [x1]
ldr s2, [x2]
ldr s1, [x3]
ins v0.s[1], v3.s[0]
ins v0.s[2], v2.s[0]
ins v0.s[3], v1.s[0]
ret
but with this patch we generate the much cleaner:
foo_v2di:
ldr d0, [x0]
ld1 {v0.d}[1], [x1]
ret
foo_v4sf:
ldr s0, [x0]
ld1 {v0.s}[1], [x1]
ld1 {v0.s}[2], [x2]
ld1 {v0.s}[3], [x3]
ret
* config/aarch64/aarch64-simd.md (vec_set<mode>): Use VALL_F16 mode
iterator. Delete separate integer-mode vec_set<mode> expander.
(aarch64_simd_vec_setv2di): Delete.
(vec_setv2di): Delete.
(aarch64_simd_vec_set<mode>): Delete all other patterns with that name.
Use VALL_F16 mode iterator. Add LD1 alternative and use vwcore for
the "w, r" alternative.
* gcc.target/aarch64/vect-init-ld1.c: New test.
From-SVN: r260351
|
|
There are four optabs for various forms of fused multiply-add:
fma, fms, fnma and fnms. Of these, only fma had a direct gimple
representation. For the other three we relied on special pattern-
matching during expand, although tree-ssa-math-opts.c did have
some code to try to second-guess what expand would do.
This patch removes the old FMA_EXPR representation of fma and
introduces four new internal functions, one for each optab.
IFN_FMA is tied to BUILT_IN_FMA* while the other three are
independent directly-mapped internal functions. It's then
possible to do the pattern-matching in match.pd and
tree-ssa-math-opts.c (via folding) can select the exact
FMA-based operation.
The BRIG & HSA parts are a best guess, but seem relatively simple.
2018-05-18 Richard Sandiford <richard.sandiford@linaro.org>
gcc/
* doc/sourcebuild.texi (scalar_all_fma): Document.
* tree.def (FMA_EXPR): Delete.
* internal-fn.def (FMA, FMS, FNMA, FNMS): New internal functions.
* internal-fn.c (ternary_direct): New macro.
(expand_ternary_optab_fn): Likewise.
(direct_ternary_optab_supported_p): Likewise.
* Makefile.in (build/genmatch.o): Depend on case-fn-macros.h.
* builtins.c (fold_builtin_fma): Delete.
(fold_builtin_3): Don't call it.
* cfgexpand.c (expand_debug_expr): Remove FMA_EXPR handling.
* expr.c (expand_expr_real_2): Likewise.
* fold-const.c (operand_equal_p): Likewise.
(fold_ternary_loc): Likewise.
* gimple-pretty-print.c (dump_ternary_rhs): Likewise.
* gimple.c (DEFTREECODE): Likewise.
* gimplify.c (gimplify_expr): Likewise.
* optabs-tree.c (optab_for_tree_code): Likewise.
* tree-cfg.c (verify_gimple_assign_ternary): Likewise.
* tree-eh.c (operation_could_trap_p): Likewise.
(stmt_could_throw_1_p): Likewise.
* tree-inline.c (estimate_operator_cost): Likewise.
* tree-pretty-print.c (dump_generic_node): Likewise.
(op_code_prio): Likewise.
* tree-ssa-loop-im.c (stmt_cost): Likewise.
* tree-ssa-operands.c (get_expr_operands): Likewise.
* tree.c (commutative_ternary_tree_code, add_expr): Likewise.
* fold-const-call.h (fold_fma): Delete.
* fold-const-call.c (fold_const_call_ssss): Handle CFN_FMS,
CFN_FNMA and CFN_FNMS.
(fold_fma): Delete.
* genmatch.c (combined_fn): New enum.
(commutative_ternary_tree_code): Remove FMA_EXPR handling.
(commutative_op): New function.
(commutate): Use it. Handle more than 2 operands.
(dt_operand::gen_gimple_expr): Use commutative_op.
(parser::parse_expr): Allow :c to be used with non-binary
operators if the commutative operand is known.
* gimple-ssa-backprop.c (backprop::process_builtin_call_use): Handle
CFN_FMS, CFN_FNMA and CFN_FNMS.
(backprop::process_assign_use): Remove FMA_EXPR handling.
* hsa-gen.c (gen_hsa_insns_for_operation_assignment): Likewise.
(gen_hsa_fma): New function.
(gen_hsa_insn_for_internal_fn_call): Use it for IFN_FMA, IFN_FMS,
IFN_FNMA and IFN_FNMS.
* match.pd: Add folds for IFN_FMS, IFN_FNMA and IFN_FNMS.
* gimple-fold.h (follow_all_ssa_edges): Declare.
* gimple-fold.c (follow_all_ssa_edges): New function.
* tree-ssa-math-opts.c (convert_mult_to_fma_1): Use the
gimple_build interface and use follow_all_ssa_edges to fold the result.
(convert_mult_to_fma): Use direct_internal_fn_suppoerted_p
instead of checking for optabs directly.
* config/i386/i386.c (ix86_add_stmt_cost): Recognize FMAs as calls
rather than FMA_EXPRs.
* config/rs6000/rs6000.c (rs6000_gimple_fold_builtin): Create a
call to IFN_FMA instead of an FMA_EXPR.
gcc/brig/
* brigfrontend/brig-function.cc
(brig_function::get_builtin_for_hsa_opcode): Use BUILT_IN_FMA
for BRIG_OPCODE_FMA.
(brig_function::get_tree_code_for_hsa_opcode): Treat BUILT_IN_FMA
as a call.
gcc/c/
* gimple-parser.c (c_parser_gimple_postfix_expression): Remove
__FMA_EXPR handlng.
gcc/cp/
* constexpr.c (cxx_eval_constant_expression): Remove FMA_EXPR handling.
(potential_constant_expression_1): Likewise.
gcc/testsuite/
* lib/target-supports.exp (check_effective_target_scalar_all_fma):
New proc.
* gcc.dg/fma-1.c: New test.
* gcc.dg/fma-2.c: Likewise.
* gcc.dg/fma-3.c: Likewise.
* gcc.dg/fma-4.c: Likewise.
* gcc.dg/fma-5.c: Likewise.
* gcc.dg/fma-6.c: Likewise.
* gcc.dg/fma-7.c: Likewise.
* gcc.dg/gimplefe-26.c: Use .FMA instead of __FMA and require
scalar_all_fma.
* gfortran.dg/reassoc_7.f: Pass -ffp-contract=off.
* gfortran.dg/reassoc_8.f: Likewise.
* gfortran.dg/reassoc_9.f: Likewise.
* gfortran.dg/reassoc_10.f: Likewise.
From-SVN: r260348
|
|
gcc/
* expr.c (do_tablejump): When converting index to Pmode, if we have a
sign extended promoted subreg, and the range does not have the sign bit
set, then do a sign extend.
* config/riscv/riscv.c (riscv_extend_comparands): In unsigned QImode
test, check for sign extended subreg and/or constant operands, and
do a sign extend in that case.
gcc/testsuite/
* gcc.target/riscv/switch-qi.c: New.
* gcc.target/riscv/switch-si.c: New.
From-SVN: r260340
|
|
2018-05-17 Steve Ellcey <sellcey@cavium.com>
* config/aarch64/thunderx2t99.md (thunderx2t99_ls_both): Delete.
(thunderx2t99_multiple): Delete psuedo-units from used cpus.
Add untyped.
(thunderx2t99_alu_shift): Remove alu_shift_reg, alus_shift_reg.
Change logics_shift_reg to logics_shift_imm.
(thunderx2t99_fp_loadpair_basic): Delete.
(thunderx2t99_fp_storepair_basic): Delete.
(thunderx2t99_asimd_int): Add neon_sub and neon_sub_q types.
(thunderx2t99_asimd_polynomial): Delete.
(thunderx2t99_asimd_fp_simple): Add neon_fp_mul_s_scalar_q
and neon_fp_mul_d_scalar_q.
(thunderx2t99_asimd_fp_conv): Add *int_to_fp* types.
(thunderx2t99_asimd_misc): Delete neon_dup and neon_dup_q.
(thunderx2t99_asimd_recip_step): Add missing *sqrt* types.
(thunderx2t99_asimd_lut): Add missing tbl types.
(thunderx2t99_asimd_ext): Delete.
(thunderx2t99_asimd_load1_1_mult): Delete.
(thunderx2t99_asimd_load1_2_mult): Delete.
(thunderx2t99_asimd_load1_ldp): New.
(thunderx2t99_asimd_load1): New.
(thunderx2t99_asimd_load2): Add missing *load2* types.
(thunderx2t99_asimd_load3): New.
(thunderx2t99_asimd_load4): New.
(thunderx2t99_asimd_store1_1_mult): Delete.
(thunderx2t99_asimd_store1_2_mult): Delete.
(thunderx2t99_asimd_store2_mult): Delete.
(thunderx2t99_asimd_store2_onelane): Delete.
(thunderx2t99_asimd_store_stp): New.
(thunderx2t99_asimd_store1): New.
(thunderx2t99_asimd_store2): New.
(thunderx2t99_asimd_store3): New.
(thunderx2t99_asimd_store4): New.
From-SVN: r260335
|
|
2018-05-17 Jerome Lambourg <lambourg@adacore.com>
gcc/
* config/arm/arm_cmse.h (cmse_nsfptr_create, cmse_is_nsfptr): Remove
#include <stdint.h>. Replace intptr_t with __INTPTR_TYPE__.
libgcc/
* config/arm/cmse.c (cmse_check_address_range): Replace
UINTPTR_MAX with __UINTPTR_MAX__ and uintptr_t with __UINTPTR_TYPE__.
From-SVN: r260330
|
|
PR target/85698
* config/rs6000/rs6000.c (rs6000_output_move_128bit): Check dest operand.
* gcc.target/powerpc/pr85698.c: New test.
Co-Authored-By: Segher Boessenkool <segher@kernel.crashing.org>
From-SVN: r260329
|
|
PR target/85323
* config/i386/i386.c (ix86_fold_builtin): Handle masked shifts
even if the mask is not all ones.
* gcc.target/i386/pr85323-7.c: New test.
* gcc.target/i386/pr85323-8.c: New test.
* gcc.target/i386/pr85323-9.c: New test.
From-SVN: r260313
|
|
PR target/85323
* config/i386/i386.c (ix86_fold_builtin): Fold shift builtins by
vector.
(ix86_gimple_fold_builtin): Likewise.
* gcc.target/i386/pr85323-4.c: New test.
* gcc.target/i386/pr85323-5.c: New test.
* gcc.target/i386/pr85323-6.c: New test.
From-SVN: r260312
|