| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
This extension defines vector instructions to extract a pair of FP16 data from
a floating-point register. Multiply the top FP16 data with the FP16 elements
and add the result with the bottom FP16 data.
gcc/ChangeLog:
* common/config/riscv/riscv-common.cc:
Turn on VECTOR_ELEN_FP_16 for XAndesvpackfph.
* config/riscv/andes-vector-builtins-bases.cc (nds_vfpmad): New class.
* config/riscv/andes-vector-builtins-bases.h: New def.
* config/riscv/andes-vector-builtins-functions.def (nds_vfpmadt): Ditto.
(nds_vfpmadb): Ditto.
(nds_vfpmadt_frm): Ditto.
(nds_vfpmadb_frm): Ditto.
* config/riscv/andes-vector.md (@pred_nds_vfpmad<nds_tb><mode>):
New pattern.
* config/riscv/riscv-vector-builtins-types.def
(DEF_RVV_F16_OPS): New def.
* config/riscv/riscv-vector-builtins.cc (f16_ops): Ditto
* config/riscv/riscv-vector-builtins.def (float32_type_node): Ditto.
* config/riscv/riscv-vector-builtins.h (XANDESVPACKFPH_EXT): Ditto.
(required_ext_to_isa_name): Add case XANDESVPACKFPH_EXT.
(required_extensions_specified): Ditto.
* config/riscv/vector-iterators.md (VHF): New iterator.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/xandesvector/non-policy/non-overloaded/nds_vfpmadb.c: New test.
* gcc.target/riscv/rvv/xandesvector/non-policy/non-overloaded/nds_vfpmadt.c: New test.
* gcc.target/riscv/rvv/xandesvector/non-policy/overloaded/nds_vfpmadb.c: New test.
* gcc.target/riscv/rvv/xandesvector/non-policy/overloaded/nds_vfpmadt.c: New test.
* gcc.target/riscv/rvv/xandesvector/policy/non-overloaded/nds_vfpmadb.c: New test.
* gcc.target/riscv/rvv/xandesvector/policy/non-overloaded/nds_vfpmadt.c: New test.
* gcc.target/riscv/rvv/xandesvector/policy/overloaded/nds_vfpmadb.c: New test.
* gcc.target/riscv/rvv/xandesvector/policy/overloaded/nds_vfpmadt.c: New test.
|
|
gcc/
PR target/121794
* config/avr/avr.md (cmpqi3): Use cpi R,0 if possible.
|
|
This patch would like to combine the vec_duplicate + vnmsub.vv to the
vnmsub.vx. From example as below code. The related pattern will depend
on the cost of vec_duplicate from GR2VR. Then the late-combine will
take action if the cost of GR2VR is zero, and reject the combination
if the GR2VR cost is greater than zero.
Assume we have example code like below, GR2VR cost is 0.
Before this patch:
11 │ beq a3,zero,.L8
12 │ vsetvli a5,zero,e32,m1,ta,ma
13 │ vmv.v.x v2,a2
...
16 │ .L3:
17 │ vsetvli a5,a3,e32,m1,ta,ma
...
22 │ vnmsub.vv v1,v2,v3
...
25 │ bne a3,zero,.L3
After this patch:
11 │ beq a3,zero,.L8
...
14 │ .L3:
15 │ vsetvli a5,a3,e32,m1,ta,ma
...
20 │ vnmsub.vx v1,a2,v3
...
23 │ bne a3,zero,.L3
gcc/ChangeLog:
* config/riscv/autovec-opt.md (*vnmsac_vx_<mode>): Rename from.
(*mul_minus_vx_<mode>): Rename to and add nmsub support.
* config/riscv/vector.md (@pred_vnmsac_vx_<mode>): Rename from.
(@pred_mul_minus_vx_<mode>): Rename to and add nmsub support.
(*pred_nmsac_<mode>_scalar_undef): Rename from.
(*pred_mul_minus_vx<mode>_undef): Rename to and add nmsub support.
Signed-off-by: Pan Li <pan2.li@intel.com>
|
|
This extension defines vector load instructions to move sign-extended or
zero-extended INT4 data into 8-bit vector register elements.
gcc/ChangeLog:
* config/riscv/andes-vector-builtins-bases.cc
(nds_nibbleload): New class.
* config/riscv/andes-vector-builtins-bases.h (nds_vln8): New def.
(nds_vlnu8): Ditto.
* config/riscv/andes-vector-builtins-functions.def (nds_vln8): Ditto.
(nds_vlnu8): Ditto.
* config/riscv/andes-vector.md (@pred_intload_mov<su><mode>): New pattern.
* config/riscv/riscv-vector-builtins-types.def (DEF_RVV_Q_OPS): New def.
(DEF_RVV_QU_OPS): Ditto.
* config/riscv/riscv-vector-builtins.cc
(q_v_void_const_ptr_ops): New operand information.
(qu_v_void_const_ptr_ops): Ditto.
* config/riscv/riscv-vector-builtins.def (void_const_ptr): New def.
* config/riscv/riscv-vector-builtins.h (enum required_ext): Ditto.
(required_ext_to_isa_name): Add case XANDESVSINTLOAD_EXT.
(required_extensions_specified): Ditto.
* config/riscv/vector-iterators.md (NDS_QVI): New iterator.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/xandesvector/non-policy/non-overloaded/nds_vln8.c: New test.
* gcc.target/riscv/rvv/xandesvector/non-policy/overloaded/nds_vln8.c: New test.
* gcc.target/riscv/rvv/xandesvector/policy/non-overloaded/nds_vln8.c: New test.
* gcc.target/riscv/rvv/xandesvector/policy/overloaded/nds_vln8.c: New test.
|
|
This patch add support for XAndesvbfhcvt ISA extension.
This extension defines instructions to perform vector floating-point
conversion between the BFLOAT16 floating-point data and the IEEE-754 32-bit
single-precision floating-point (SP) data in a vector register.
gcc/ChangeLog:
* common/config/riscv/riscv-common.cc:
Turn on VECTOR_ELEN_BF_16 for XAndesvbfhcvt.
* config.gcc: Add extra_objs andes-vector-builtins-bases.o
and extra_headers andes_vector.h.
* config/riscv/riscv-vector-builtins-shapes.cc
(BASE_NAME_MAX_LEN): Increase size to 20.
* config/riscv/riscv-vector-builtins.cc
(f32_to_bf16_nf_w_ops): New operand information.
(f32_to_bf16_nf_w_ops): New operand information.
(DEF_RVV_FUNCTION): New def.
* config/riscv/riscv-vector-builtins.def (bf16): Ditto.
* config/riscv/riscv-vector-builtins.h (enum required_ext): Ditto.
(required_ext_to_isa_name): Add case XANDESVBFHCVT_EXT.
(required_extensions_specified): Ditto.
* config/riscv/t-riscv: Add andes-vector-builtins-functions.def,
andes-vector-builtins-bases.h and andes-vector-builtins-bases.o.
* config/riscv/vector-iterators.md (NDS_VWEXTBF): New iterator.
(NDS_V_DOUBLE_TRUNC_BF): New attr.
* config/riscv/andes-vector-builtins-bases.cc: New file.
* config/riscv/andes-vector-builtins-bases.h: New file.
* config/riscv/andes-vector-builtins-functions.def: New file.
* config/riscv/andes_vector.h: New file.
* config/riscv/andes-vector.md: New file.
* config/riscv/vector.md: Include andes_vector.md.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/rvv.exp: Add regression for xandesvector.
* gcc.target/riscv/rvv/xandesvector/non-policy/non-overloaded/nds_vfncvtbf16s.c: New test.
* gcc.target/riscv/rvv/xandesvector/non-policy/non-overloaded/nds_vfwcvtsbf16.c: New test.
* gcc.target/riscv/rvv/xandesvector/non-policy/overloaded/nds_vfncvtbf16s.c: New test.
* gcc.target/riscv/rvv/xandesvector/non-policy/overloaded/nds_vfwcvtsbf16.c: New test.
* gcc.target/riscv/rvv/xandesvector/policy/non-overloaded/nds_vfncvtbf16s.c: New test.
* gcc.target/riscv/rvv/xandesvector/policy/non-overloaded/nds_vfwcvtsbf16.c: New test.
* gcc.target/riscv/rvv/xandesvector/policy/overloaded/nds_vfncvtbf16s.c: New test.
* gcc.target/riscv/rvv/xandesvector/policy/overloaded/nds_vfwcvtsbf16.c: New test.
|
|
Add pipeline description for the Tenstorrent Ascalon 8 wide CPU.
gcc/ChangeLog
* config/riscv/riscv-cores.def (RISCV_TUNE): Update.
* config/riscv/riscv-opts.h (enum riscv_microarchitecture_type):
Add tt_ascalon_d8.
* config/riscv/riscv.md: Update tune attribute and include
tt-ascalon-d8.md.
* config/riscv/tt-ascalon-d8.md: New file.
|
|
For Zvfhmin a vector mode exists but the corresponding vec_extract does
not. This patch checks that a vec_extract is available and otherwise
falls back to standard handling.
PR target/121510
gcc/ChangeLog:
* config/riscv/riscv.cc (riscv_legitimize_move): Check if we can
vec_extract.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/pr121510.c: New test.
|
|
There are some cases where involing zero_reg is not needed and
where there are other sequences with the same efficiency.
An example is to use SBCI R,0 instead of SBC R,__zero_reg__
when R >= R16. This may turn out to be better for small ISRs.
PR target/121794
gcc/
* config/avr/avr.cc (avr_out_compare): Only use zero_reg
when there is no other sequence of the same length.
(avr_out_plus_ext): Same.
(avr_out_plus_1): Same.
|
|
Unlike Advanced SIMD, SVE has instruction to perform smin, smax, umin, umax
on 64-bit elements. Thus, we can use them with the fixed-width V2DImode
expander. Most of the machinery is already there on the define_insn side,
supporting V2DImode operands of the SVE pattern. We just need to wire up
the RTL emission to the v2di standard names for the TARGET_SVE case.
So for the smin case we now generate:
min_di:
ldr q30, [x0]
ptrue p7.b, all
ldr q31, [x1]
smin z30.d, p7/m, z30.d, z31.d
str q30, [x2]
ret
min_imm_di:
ldr q31, [x0]
smin z31.d, z31.d, #5
str q31, [x2]
ret
instead of the previous:
min_di:
ldr q30, [x0]
ldr q31, [x1]
cmgt v29.2d, v30.2d, v31.2d
bsl v29.16b, v31.16b, v30.16b
str q29, [x2]
ret
min_imm_di:
ldr q31, [x0]
mov z30.d, #5
cmgt v29.2d, v30.2d, v31.2d
bsl v29.16b, v31.16b, v30.16b
str q29, [x2]
ret
The register operand case is the same length, though the new ptrue can now be
shared and moved away. But the immediate operand case is obviously better
as the SVE immediate form doesn't require a predicate operand.
Bootstrapped and tested on aarch64-none-linux-gnu.
Signed-off-by: Kyrylo Tkachov <ktkachov@nvidia.com>
gcc/
* config/aarch64/iterators.md (sve_di_suf): New mode attribute.
* config/aarch64/aarch64-sve.md (<optab><mode>3 SVE_INT_BINARY_MULTI):
Rename to...
(<optab><mode>3<sve_di_suf>): ... This. Use SVE_I_SIMD_DI mode
iterator.
* config/aarch64/aarch64-simd.md (<su><maxmin>v2di3): Use the above
for TARGET_SVE.
gcc/testsuite/
* gcc.target/aarch64/sve/usminmax_di.c: New test.
|
|
This patch would like to combine the vec_duplicate + vmadd.vv to the
vmadd.vx. From example as below code. The related pattern will depend
on the cost of vec_duplicate from GR2VR. Then the late-combine will
take action if the cost of GR2VR is zero, and reject the combination
if the GR2VR cost is greater than zero.
Assume we have example code like below, GR2VR cost is 0.
Before this patch:
11 │ beq a3,zero,.L8
12 │ vsetvli a5,zero,e32,m1,ta,ma
13 │ vmv.v.x v2,a2
...
16 │ .L3:
17 │ vsetvli a5,a3,e32,m1,ta,ma
...
22 │ vmadd.vv v1,v2,v3
...
25 │ bne a3,zero,.L3
After this patch:
11 │ beq a3,zero,.L8
...
14 │ .L3:
15 │ vsetvli a5,a3,e32,m1,ta,ma
...
20 │ vmadd.vx v1,a2,v3
...
23 │ bne a3,zero,.L3
gcc/ChangeLog:
* config/riscv/autovec-opt.md (*vmacc_vx_<mode>): Rename to
handle both the macc and madd.
(*mul_plus_vx_<mode>): Add madd pattern.
* config/riscv/vector.md (@pred_mul_plus_vx_<mode>): Rename to
handle both the macc and madd.
(*pred_macc_<mode>_scalar_undef): Remove.
(*pred_nmsac_<mode>_scalar_undef): Remove.
(*pred_mul_plus_vx<mode>_undef): Add new pattern to handle
both the vmacc and vmadd.
(@pred_mul_plus_vx<mode>): Ditto.
Signed-off-by: Pan Li <pan2.li@intel.com>
|
|
Insufficient validation of the operands in vec_set_<mode>_internal
means that the optimizers can transform the exanded code into
something that is invalid. We then emit code based on the incorrect
RTL assuming that it is still valid. A valid pattern can only have a
single bit set in the immediate operand, representing the lane to be
written.
gcc/ChangeLog:
PR target/121775
* config/arm/neon.md (vec_set<mode>_internal, all variants):
validate the immediate operand that indicates the lane to
modify.
gcc/testsuite/ChangeLog:
PR target/121775
* gcc.target/arm/simd/vset_lane_u8.c: New test.
|
|
This fixes a glaring mistake in yesterday's change to the expansion of
vec_perm. We should of course move tmp_target into the real target
and not the other way around. I wonder why my testing hasn't
caught this...
PR target/121742
PR target/121780
PR target/121781
gcc/ChangeLog:
* config/riscv/riscv-v.cc (expand_vec_perm): Swap target and
tmp_target.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/pr121780.c: New test.
* gcc.target/riscv/rvv/autovec/pr121781.c: New test.
|
|
Previously, vector built-in functions were not properly registered during
the LTO pipeline, causing link failures when vector intrinsics were used
in LTO builds with mixed architecture options. This patch ensures all
vector built-in functions are always registered during LTO compilation.
The key changes include:
- Moving pragma intrinsic flag manipulation from riscv-c.cc to
riscv-vector-builtins.cc for better encapsulation
- Registering all vector built-in functions regardless of current ISA
extensions, deferring the actual extension checking to expansion time
- Adding proper support for built-in type registration during LTO
This approach is safe because we already perform extension requirement
checking at expansion time. The trade-off is a slight increase in
bootstrap time for LTO builds due to registering more built-in functions.
PR target/110812
gcc/ChangeLog:
* config/riscv/riscv-c.cc (pragma_intrinsic_flags): Remove struct.
(riscv_pragma_intrinsic_flags_pollute): Remove function.
(riscv_pragma_intrinsic_flags_restore): Remove function.
(riscv_pragma_intrinsic): Simplify to only call handle_pragma_vector.
* config/riscv/riscv-vector-builtins.cc (pragma_intrinsic_flags):
Move struct definition here from riscv-c.cc.
(riscv_pragma_intrinsic_flags_pollute): Move and adapt from
riscv-c.cc, add zvfbfmin, zvfhmin and vector_elen_bf_16 support.
(riscv_pragma_intrinsic_flags_restore): Move from riscv-c.cc.
(rvv_switcher::rvv_switcher): Add pollute_flags parameter to
control flag manipulation.
(rvv_switcher::~rvv_switcher): Restore flags conditionally.
(register_builtin_types): Use rvv_switcher without polluting flags.
(get_required_extensions): Remove function.
(check_required_extensions): Simplify to only check type validity.
(function_instance::function_returns_void_p): Move implementation
from header.
(function_builder::add_function): Register placeholder for LTO.
(init_builtins): Simplify and handle LTO case.
(reinit_builtins): Remove function.
(handle_pragma_vector): Remove extension checking.
* config/riscv/riscv-vector-builtins.h
(function_instance::function_returns_void_p): Add declaration.
(function_call_info::function_returns_void_p): Remove inline
implementation.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/lto/pr110812_0.c: New test.
* gcc.target/riscv/lto/pr110812_1.c: New test.
* gcc.target/riscv/lto/riscv-lto.exp: New test driver.
* gcc.target/riscv/lto/riscv_vector.h: New header wrapper.
|
|
This extension defines instructions to perform scalar floating-point
conversion between the BFLOAT16 floating-point data and the IEEE-754
32-bit single-precision floating-point (SP) data in a scalar
floating point register.
gcc/ChangeLog:
* config/riscv/andes.def: Add nds_fcvt_s_bf16 and nds_fcvt_bf16_s.
* config/riscv/riscv.md (truncsfbf2): Add TARGET_XANDESBFHCVT support.
(extendbfsf2): Ditto.
* config/riscv/riscv-builtins.cc: New AVAIL andesbfhcvt.
Add new define RISCV_ATYPE_BF and RISCV_ATYPE_SF.
* config/riscv/riscv-ftypes.def: New DEF_RISCV_FTYPE.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/xandes/xandesbfhcvt-1.c: New test.
* gcc.target/riscv/xandes/xandesbfhcvt-2.c: New test.
|
|
This patch adds support for the XAndesperf ISA extension.
The 32-bit AndeStar V5 extension includes branch instructions,
load effective address instructions, and string processing
instructions for performance improvement.
New INSN patterns are added into the new file andes.md
as a seprated vender extension.
gcc/ChangeLog:
* config/riscv/constraints.md (Ou07): New constraint.
(ads_Bext): New constraint.
* config/riscv/iterators.md (ANYLE32): New iterator.
(sizen): New iterator.
(sh_limit): New iterator.
(sh_bit): New iterator.
(cs): New iterator.
* config/riscv/predicates.md (ads_branch_bbcs_operand): New predicate.
(ads_branch_bimm_operand): New predicate.
(ads_imm_extract_operand): New predicate.
(ads_extract_size_imm_si): New predicate.
(ads_extract_size_imm_di): New predicate.
(const_int5_operand): New predicate.
* config/riscv/riscv-builtins.cc:
Add new AVAIL andesperf32 and andesperf64.
Add new define RISCV_ATYPE_DI.
* config/riscv/riscv-ftypes.def: New DEF_RISCV_FTYPE.
* config/riscv/riscv.cc
(riscv_extend_cost): Cost for pattern 'bfo'.
(riscv_rtx_costs): Cost for XAndesperf extension.
* config/riscv/riscv.md: Add support for XAndesperf to patterns
zero_extendsidi2_internal, zero_extendhi2, extendsidi2_internal,
extend<SHORT:mode><SUPERQI:mode>2, <any_extract:optab><GPR:mode>3
and branch_on_bit.
* config/riscv/vector-iterators.md
(sz): Add sign_extract and zero_extract.
* config/riscv/andes.def: New file for vender Andes.
* config/riscv/andes.md: New file for vender Andes.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/riscv.exp: Add runtest for subdir xandes.
* gcc.target/riscv/xandes/xandesperf-1.c: New test.
* gcc.target/riscv/xandes/xandesperf-10.c: New test.
* gcc.target/riscv/xandes/xandesperf-2.c: New test.
* gcc.target/riscv/xandes/xandesperf-3.c: New test.
* gcc.target/riscv/xandes/xandesperf-4.c: New test.
* gcc.target/riscv/xandes/xandesperf-5.c: New test.
* gcc.target/riscv/xandes/xandesperf-6.c: New test.
* gcc.target/riscv/xandes/xandesperf-7.c: New test.
* gcc.target/riscv/xandes/xandesperf-8.c: New test.
* gcc.target/riscv/xandes/xandesperf-9.c: New test.
|
|
This patch add basic support for the following XAndes ISA extensions:
XANDESPERF
XANDESBFHCVT
XANDESVBFHCVT
XANDESVSINTLOAD
XANDESVPACKFPH
XANDESVDOT
gcc/ChangeLog:
* config/riscv/riscv-ext.def: Include riscv-ext-andes.def.
* config/riscv/riscv-ext.opt (riscv_xandes_subext): New variable.
(XANDESPERF) : New mask.
(XANDESBFHCVT): Ditto.
(XANDESVBFHCVT): Ditto.
(XANDESVSINTLOAD): Ditto.
(XANDESVPACKFPH): Ditto.
(XANDESVDOT): Ditto.
* config/riscv/t-riscv: Add riscv-ext-andes.def.
* doc/riscv-ext.texi: Regenerated.
* config/riscv/riscv-ext-andes.def: New file.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/xandes/xandes-predef-1.c: New test.
* gcc.target/riscv/xandes/xandes-predef-2.c: New test.
* gcc.target/riscv/xandes/xandes-predef-3.c: New test.
* gcc.target/riscv/xandes/xandes-predef-4.c: New test.
* gcc.target/riscv/xandes/xandes-predef-5.c: New test.
* gcc.target/riscv/xandes/xandes-predef-6.c: New test.
Co-author: Lino Hsing-Yu Peng (linopeng@andestech.com)
Co-author: Kai Kai-Yi Weng (kaiweng@andestech.com).
|
|
This pattern enables the combine pass (or late-combine, depending on the case)
to merge a vec_duplicate into an smax RTL instruction.
Before this patch, we have two instructions, e.g.:
vfmv.v.f v2,fa0
vfmax.vv v1,v1,v2
After, we get only one:
vfmax.vf v1,v1,fa0
In some cases, it also shaves off one vsetvli.
gcc/ChangeLog:
* config/riscv/autovec-opt.md (*vfmax_vf_<mode>): Rename into...
(*vf<optab>_vf_<mode>): New pattern to combine vec_duplicate +
vf{min,max}.vv into vf{max,min}.vf.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/vls/floating-point-max-2.c: Adjust scan
dump.
* gcc.target/riscv/rvv/autovec/vls/floating-point-max-4.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-1-f16.c: Add vfmax. Also add
missing scan-dump for vfmul.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-1-f32.c: Add vfmax.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-1-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-2-f16.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-2-f32.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-2-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-3-f16.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-3-f32.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-3-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-4-f16.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-4-f32.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-4-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_binop.h: Add max functions.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_binop_data.h: Add data for
vfmax.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_vfmax-run-1-f16.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_vfmax-run-1-f32.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_vfmax-run-1-f64.c: New test.
|
|
This is Austin's work to remove the redundant sign extension seen in pr121213.
--
The .w form of amoswap will sign extend its result from 32 to 64 bits, thus any
explicit sign extension insn doing the same is redundant.
This uses Jivan's approach of allocating a DI temporary for an extended result
and using a promoted subreg extraction to get that result into the final
destination.
Tested with no regressions on riscv32-elf and riscv64-elf and bootstrapped on
the BPI and pioneer systems.
PR target/121213
gcc/
* config/riscv/sync.md (amo_atomic_exchange_extended<mode>):
Separate insn with sign extension for 64 bit targets.
gcc/testsuite
* gcc.target/riscv/amo/pr121213.c: Remove xfail.
|
|
With g:d20b2ad845876eec0ee80a3933ad49f9f6c4ee30 the narrowing shift instructions
are now represented with standard RTL and more merging optimisations occur.
This exposed a wrong predicate for the shift amount operand.
The shift amount is the number of bits of the narrow destination, not the input
sources.
Correct this by using the vn_mode attribute when specifying the predicate, which
exists for this purpose.
I've spotted a few more narrowing shift patterns that need the restriction, so
they are updated as well.
Bootstrapped and tested on aarch64-none-linux-gnu.
Signed-off-by: Kyrylo Tkachov <ktkachov@nvidia.com>
gcc/
PR target/121749
* config/aarch64/aarch64-simd.md (aarch64_<shrn_op>shrn_n<mode>):
Use aarch64_simd_shift_imm_offset_<vn_mode> instead of
aarch64_simd_shift_imm_offset_<ve_mode> predicate.
(aarch64_<shrn_op>shrn_n<mode> VQN define_expand): Likewise.
(*aarch64_<shrn_op>rshrn_n<mode>_insn): Likewise.
(aarch64_<shrn_op>rshrn_n<mode>): Likewise.
(aarch64_<shrn_op>rshrn_n<mode> VQN define_expand): Likewise.
(aarch64_sqshrun_n<mode>_insn): Likewise.
(aarch64_sqshrun_n<mode>): Likewise.
(aarch64_sqshrun_n<mode> VQN define_expand): Likewise.
(aarch64_sqrshrun_n<mode>_insn): Likewise.
(aarch64_sqrshrun_n<mode>): Likewise.
(aarch64_sqrshrun_n<mode>): Likewise.
* config/aarch64/iterators.md (vn_mode): Handle DI, SI, HI modes.
gcc/testsuite/
PR target/121749
* gcc.target/aarch64/simd/pr121749.c: New test.
|
|
This patch changes is_vlmax_len_p to handle VLS modes properly.
Before we would check if len == GET_MODE_NUNITS (mode). This works vor
VLA modes but not necessarily for VLS modes. We regularly have e.g.
small VLS modes where LEN equals their number of units but which do not
span a full vector. Therefore now check if len * GET_MODE_UNIT_SIZE
(mode) equals BYTES_PER_RISCV_VECTOR * TARGET_MAX_LMUL.
Changing this uncovered an oversight in avlprop where we used
GET_MODE_NUNITS as AVL when GET_MODE_NUNITS / NF would be correct.
The testsuite is unchanged. I didn't bother to add a dedicated test
because we would have seen the fallout any way once the gather patch
lands.
gcc/ChangeLog:
* config/riscv/riscv-v.cc (is_vlmax_len_p): Properly handle VLS
modes.
(imm_avl_p): Fix VLS length check.
(expand_strided_load): Use is_vlmax_len_p.
(expand_strided_store): Ditto.
* config/riscv/riscv-avlprop.cc (pass_avlprop::execute):
Use GET_MODE_NUNITS / NF as avl.
|
|
In a two-source gather we unconditionally overwrite target with the
first gather's result already. If op1 == target this clobbers the
source operand for the second gather. This patch uses a temporary in
that case.
PR target/121742
gcc/ChangeLog:
* config/riscv/riscv-v.cc (expand_vec_perm): Use temporary if
op1 and target overlap.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/pr121742.c: New test.
|
|
The print_ext_doc_entry function and associated version_t struct in
gen-riscv-ext-opt.cc were not being used anywhere in the codebase.
Remove them to clean up the code.
gcc/
* config/riscv/gen-riscv-ext-opt.cc (version_t): Remove unused
struct.
(print_ext_doc_entry): Remove unused function.
|
|
The problem was caused by an erroneous note about creating a stack frame,
which caused the cur_cfa reg to fail to assert with a value other than
the frame pointer.
This fix will generate notes that correctly update cur_cfa.
v2 changes.
Add testcase.
All tests that failed with
"internal compiler error: in dwarf2out_frame_debug_adjust_cfa, at dwarf2cfi.cc"
now pass.
PR target/89828
gcc
* config/rx/rx.cc (add_pop_cfi_notes): Release the frame pointer if it is
used.
(rx_expand_prologue): Redesigned stack pointer and frame pointer update
process.
gcc/testsuite/
* gcc.dg/pr89828.c: New.
|
|
While we have already the accessor info_for_reduction, its result
is a plain stmt_vec_info. The following turns that into a class
for the purpose of changing accesses to reduction info to a new
set of accessors prefixed with VECT_REDUC_INFO and removes
the corresponding STMT_VINFO prefixed accessors where possible.
There is few reduction related things that are used by scalar
cycle detection and thus have to stay as-is for now and as
copies in future.
This also separates reduction info into one object per reduction
and associate it with SLP nodes, splitting it out from
stmt_vec_info, retaining (and duplicating) parts used by scalar
cycle analysis. The data is then associated with SLP nodes
forming reduction cycles and accessible via info_for_reduction.
The data is created at SLP discovery time as we look at it even
pre-vectorizable_reduction analysis, but most of the data is
only populated by the latter. There is no reduction info with
nested cycles that are not part of an outer reduction.
In the process this adds cycle info to each SLP tree, notably
the reduc-idx and a way to identify the reduction info.
* tree-vectorizer.h (vect_reduc_info): New.
(create_info_for_reduction): Likewise.
(VECT_REDUC_INFO_TYPE): Likewise.
(VECT_REDUC_INFO_CODE): Likewise.
(VECT_REDUC_INFO_FN): Likewise.
(VECT_REDUC_INFO_SCALAR_RESULTS): Likewise.
(VECT_REDUC_INFO_INITIAL_VALUES): Likewise.
(VECT_REDUC_INFO_REUSED_ACCUMULATOR): Likewise.
(VECT_REDUC_INFO_INDUC_COND_INITIAL_VAL): Likewise.
(VECT_REDUC_INFO_EPILOGUE_ADJUSTMENT): Likewise.
(VECT_REDUC_INFO_FORCE_SINGLE_CYCLE): Likewise.
(VECT_REDUC_INFO_RESULT_POS): Likewise.
(VECT_REDUC_INFO_VECTYPE): Likewise.
(STMT_VINFO_VEC_INDUC_COND_INITIAL_VAL): Remove.
(STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT): Likewise.
(STMT_VINFO_FORCE_SINGLE_CYCLE): Likewise.
(STMT_VINFO_REDUC_FN): Likewise.
(STMT_VINFO_REDUC_VECTYPE): Likewise.
(vect_reusable_accumulator::reduc_info): Adjust.
(vect_reduc_type): Adjust.
(_slp_tree::cycle_info): New member.
(SLP_TREE_REDUC_IDX): Likewise.
(vect_reduc_info_s): Move/copy data from ...
(_stmt_vec_info): ... here.
(_loop_vec_info::redcu_infos): New member.
(info_for_reduction): Adjust to take SLP node.
(vect_reduc_type): Adjust.
(vect_is_reduction): Add overload for SLP node.
* tree-vectorizer.cc (vec_info::new_stmt_vec_info):
Do not initialize removed members.
(vec_info::free_stmt_vec_info): Do not release them.
* tree-vect-stmts.cc (vectorizable_condition): Adjust.
* tree-vect-slp.cc (_slp_tree::_slp_tree): Initialize
cycle info.
(vect_build_slp_tree_2): Compute SLP reduc_idx and store
it. Create, populate and propagate reduction info.
(vect_print_slp_tree): Print cycle info.
(vect_analyze_slp_reduc_chain): Set cycle info on the
manual added conversion node.
(vect_optimize_slp_pass::start_choosing_layouts): Adjust.
* tree-vect-loop.cc (_loop_vec_info::~_loop_vec_info):
Release reduction infos.
(info_for_reduction): Get the reduction info from
the vector in the loop_vinfo.
(vect_create_epilog_for_reduction): Adjust.
(vectorizable_reduction): Likewise.
(vect_transform_reduction): Likewise.
(vect_transform_cycle_phi): Likewise, deal with nested
cycles not part of a double reduction have no reduction info.
* config/aarch64/aarch64.cc (aarch64_force_single_cycle):
Use VECT_REDUC_INFO_FORCE_SINGLE_CYCLE, get SLP node and use
that.
(aarch64_vector_costs::count_ops): Adjust.
|
|
1) Fix predicate of operands[3] in cond_<insn><mode> since only
const_vec_dup_operand is excepted for masked operations, and pass real
count to ix86_vgf2p8affine_shift_matrix.
2) Pass operands[2] instead of operands[1] to
gen_vgf2p8affineqb_<mode>_mask which excepted the operand to shifted,
but operands[1] is mask operand in cond_<insn><mode>.
gcc/ChangeLog:
PR target/121699
* config/i386/predicates.md (const_vec_dup_operand): New
predicate.
* config/i386/sse.md (cond_<insn><mode>): Fix predicate of
operands[3], and fix wrong operands passed to
ix86_vgf2p8affine_shift_matrix and
gen_vgf2p8affineqb_<mode>_mask.
gcc/testsuite/ChangeLog:
* gcc.target/i386/pr121699.c: New test.
|
|
CLAMPS instruction
The CLAMPS instruction in Xtensa ISA, provided when the TARGET_CLAMPS
configuration is enabled (and also requires TARGET_MINMAX), returns a
value clamped the number in the specified register to between -(1<<N) and
(1<<N)-1 inclusive, where N is an immediate value from 7 to 22.
Therefore, when the above configurations are met, by comparing the clamped
result with the original value for equality, branching whether the value
is within the range mentioned above or not is implemented with fewer
instructions, especially when the upper and lower bounds of the range are
too large to fit into a single immediate assignment.
/* example (TARGET_MINMAX and TARGET_CLAMPS) */
extern void foo(void);
void test0(int a) {
if (a >= -(1 << 9) && a < (1 << 9))
foo();
}
void test1(int a) {
if (a < -(1 << 20) || a >= (1 << 20))
foo();
}
;; before
test0:
entry sp, 32
addmi a2, a2, 0x200
movi a8, 0x3ff
bltu a8, a2, .L1
call8 foo
.L1:
retw.n
test1:
entry sp, 32
movi.n a9, 1
movi.n a8, -1
slli a9, a9, 20
srli a8, a8, 11
add.n a2, a2, a9
bgeu a8, a2, .L4
call8 foo
.L4:
retw.n
;; after
test0:
entry sp, 32
clamps a8, a2, 9
bne a2, a8, .L1
call8 foo
.L1:
retw.n
test1:
entry sp, 32
clamps a8, a2, 20
beq a2, a8, .L4
call8 foo
.L4:
retw.n
(note: Currently, in the RTL instruction combination pass, the possible
const_int values are fundamentally constrained by
TARGET_LEGITIMATE_CONSTANT_P() if no bare large constant assignments are
possible (i.e., neither -mconst16 nor -mauto-litpools), so limiting N to
a range of 7 to only 10 instead of to 22. A series of forthcoming
patches will introduce an entirely new "xt_largeconst" pass that will
solve several issues including this.)
gcc/ChangeLog:
* config/xtensa/predicates.md (alt_ubranch_operator):
New predicate.
* config/xtensa/xtensa.md (*eqne_in_range):
New insn_and_split pattern.
|
|
So this is the next chunk of Shreya's work to adjust our add expanders. In this
patch we're adding support for adding a 2*s12 immediate in SI for rv64.
To recap, the basic idea is reduce our reliance on the define_insn_and_split
that was added a year or so ago by synthesizing the more efficient sequence at
expansion time. By handling this early rather than late the synthesized
sequence participates in the various optimizer passes in the natural way. In
contrast using the define_insn_and_split bypasses the cost modeling in combine
and hides the synthesis until after reload as completed (which in turn leads to
the problems seen in pr120811).
This doesn't solve pr120811, but it is the last prerequisite patch before
directly tackling pr120811.
This has been bootstrapped & regression tested on the pioneer & bpi and been
through the usual testing on riscv32-elf and riscv64-elf. Waiting on
pre-commit CI before moving forward.
gcc/
* config/riscv/riscv-protos.h (synthesize_add_extended): Prototype.
* config/riscv/riscv.cc (synthesize_add_extended): New function.
* config/riscv/riscv.md (addsi3): For RV64, try synthesize_add_extended.
gcc/testsuite/
* gcc.target/riscv/add-synthesis-2.c: New test.
|
|
recent libstdc++ changes [PR121698]
libstdc++ changed its ABI in <compare> for C++20 recently (under the
C++20 is still experimental rule). In addition to the -1, 0, 1 values
for less, equal, greater it now uses -128 for unordered instead of
former 2 and changes some of the operators, instead of checks like
(_M_value & ~1) == _M_value in some cases it now uses _M_reverse()
which is negation in unsigned char type + conversion back to the original
type. _M_reverse() thus turns the -1, 0, 1, -128 values into
1, 0, -1, -128. Note libc++ uses value -127 instead of 2/-128.
Now, the middle-end has some optimizations which rely on the particular
implementation and don't optimize if not. One is optimize_spaceship
which on some targets (currently x86, aarch64 and s390) attempts to use
better comparison instructions (ideally just one floating point comparison
to get all 4 possible outcomes plus some flag tests or magic instead of
2 or 3 floating point comparisons). This one can actually handle
arbitrary int non-[-1,1] values for unordered but still has a default
of 2. The patch changes that default to -128 so that even if something
is expanded as branches if it is later during RTL optimizations determined
to convert that into partial_ordering we get better code.
The other optimization (phiopt one) is about optimizing (x <=> y) < 0
etc. into just x < y. This one actually relies on the exact unordered
value (2) and has code to deal with that (_M_value & ~1) == _M_value
kind of tests and whatever match.pd lowers it. So, this patch partially
rewrites it to look for -128 instead of 2, drop those
(_M_value & ~1) == _M_value pattern recognitions and instead introduces
pattern recognition of _M_reverse(), i.e. cast to unsigned char, negation
in that type and cast back to the original signed type.
With all these changes we get back the desired optimizations for all
the cases we could optimize previously (note, for HONOR_NANS case
we don't try to optimize say (x <=> y) == 0 because the original
will raise exception if either x or y is a NaN, while turning it into
x == y will not, but (x <=> y) <= 0 is fine (x <= y), because it
does raise those exceptions.
2025-08-30 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/121698
* tree-ssa-phiopt.cc (spaceship_replacement): Adjust
to handle spaceship unordered value -128 rather than 2 and
stmts from the new std::partial_order::_M_reverse() instead
of (_M_value & ~1) == _M_value etc.
* doc/md.texi (spaceship@var{m}4): Use -128 instead of 2.
* tree-ssa-math-opts.cc (optimize_spaceship): Adjust comments
that libstdc++ unordered value is -128 rather than 2 and use
that as the default unordered value.
* config/i386/i386-expand.cc (ix86_expand_fp_spaceship): Use
GEN_INT (-128) instead of const2_rtx and adjust comment accordingly.
* config/aarch64/aarch64.cc (aarch64_expand_fp_spaceship): Likewise.
* config/s390/s390.cc (s390_expand_fp_spaceship): Likewise.
* gcc.dg/pr94589-2.c: Adjust for expected unordered value -128
rather than 2 and negations in unsigned char instead of and with
~1 and comparison against original value.
* gcc.dg/pr94589-4.c: Likewise.
* gcc.dg/pr94589-5.c: Likewise.
* gcc.dg/pr94589-6.c: Likewise.
|
|
Since the first operand of PLUS in the source of TLS64_COMBINE pattern:
(set (reg/f:DI 128)
(plus:DI (unspec:DI [
(symbol_ref:DI ("_TLS_MODULE_BASE_") [flags 0x10])
(reg:DI 126)
(reg/f:DI 7 sp)
] UNSPEC_TLSDESC)
(const:DI (unspec:DI [
(symbol_ref:DI ("bfd_error") [flags 0x1a] <var_decl 0x7fffe99d6e40 bfd_error>)
] UNSPEC_DTPOFF))))
is unused, use the second operand of PLUS:
(const:DI (unspec:DI [
(symbol_ref:DI ("bfd_error") [flags 0x1a] <var_decl 0x7fffe99d6e40 bfd_error>)
] UNSPEC_DTPOFF))
to check if 2 TLS_COMBINE patterns have the same source.
gcc/
PR target/121725
* config/i386/i386-features.cc
(pass_x86_cse::candidate_gnu2_tls_p): Use the UNSPEC_DTPOFF
operand to check source operand in TLS64_COMBINE pattern.
gcc/testsuite/
PR target/121725
* gcc.target/i386/pr121725-1a.c: New test.
* gcc.target/i386/pr121725-1b.c: Likewise.
Signed-off-by: H.J. Lu <hjl.tools@gmail.com>
|
|
Also, the omission of the instruction that sets the shift amount register
(SAR) to 8 is now more efficient: it is omitted if there was a previous
bswapsi2 in the same BB, but not omitted if no bswapsi2 is found or another
insn that modifies SAR is found first (see below).
Note that the five instructions for writing to SAR are as follows, along
with the insns that use them (except for bswapsi2_internal itself):
- SSA8B
*shift_per_byte, *shlrd_per_byte
- SSA8L
*shift_per_byte, *shlrd_per_byte
- SSR
ashrsi3 (alt 1), lshrsi3 (alt 1), *shlrd_reg, rotrsi3 (alt 1)
- SSL
ashlsi3_internal (alt 1), *shlrd_reg, rotlsi3 (alt 1)
- SSAI
*shlrd_const, rotlsi3 (alt 0), rotrsi3 (alt 0)
gcc/ChangeLog:
* config/xtensa/xtensa-protos.h (xtensa_bswapsi2_output):
New function prototype.
* config/xtensa/xtensa.cc
(xtensa_bswapsi2_output_1, xtensa_bswapsi2_output):
New functions.
* config/xtensa/xtensa.md (bswapsi2_internal):
Rewrite in compact syntax and use xtensa_bswapsi2_output() as asm
output.
gcc/testsuite/ChangeLog:
* gcc.target/xtensa/bswap-SSAI8.c: New.
|
|
So the RISC-V port has attributes which indicate the index within the
recog_data where certain operands will be found.
For this BZ the default value for the merge_op_idx attribute on the given insn
is "2". But the insn only has operands 0 & 1. So we do an out of bounds array
access and boom the ICE/valgrind failure.
As we discussed in the patchwork meeting, this is all a bit clunky and has been
fairly error prone. This doesn't add any massive checking, but does introduce
some asserts to help catch problems a bit earlier and clearer.
In particular in cases where we're already asserting that the returned index is
valid (!= INVALID_ATTRIBUTE) we also assert that the index is less than the
total number of operands.
In the get_vlmax_ta_preferred_avl routine it appears like we need to handle
these two cases more gracefully as we apparently legitimately query for the
merge_op_idx on a fairly arbitrary insn. We just have to make sure to not
*use* the result if it's INVALID_ATTRIBUTE. So for that code we assert that
merge_op_idx is either INVALID_ATTRIBUTE or smaller than the number of
operands.
This patch also adds overrides for 3 patterns to return INVALID_ATTRIBUTE for
merge_op_idx, similar to how they already do for mode_idx and avl_type_idx.
This has been bootstrapped and regression tested on the bpi & pioneer systems
and regression tested for riscv32-elf and riscv64-elf. Waiting on CI before
pushing.
PR target/121548
gcc/
* config/riscv/riscv-avlprop.cc (get_insn_vtype_mode): Assert
MODE_IDX is smaller than the number of operands.
(simplify_replace_vlmax_avl): Similarly.
(pass_avlprop::get_vlmax_ta_preferred_avl): Similarly.
* config/riscv/vector.md: Override merge_op_idx computation
for simple moves, just like is done for avl_type_idx and mode_idx.
|
|
This pattern enables the combine pass (or late-combine, depending on the case)
to merge a vec_duplicate into an unspec_vfmin RTL instruction.
Before this patch, we have two instructions, e.g.:
vfmv.v.f v2,fa0
vfmin.vv v1,v1,v2
After, we get only one:
vfmin.vf v1,v1,fa0
gcc/ChangeLog:
* config/riscv/autovec-opt.md
(*vfmin_vf_ieee_<mode>): Add new patterns to combine vec_duplicate +
vfmin.vv (unspec) into vfmin.vf.
(*vfmul_vf_<mode>, *vfrdiv_vf_<mode>, *vfmin_vf_<mode>): Fix attribute
types.
* config/riscv/vector.md (@pred_<ieee_fmaxmin_op><mode>_scalar): Allow
VLS modes.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/vx_vf/vf-4-f16.c: Add vfmin.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-4-f32.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-4-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-5-f16.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-5-f32.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-5-f64.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-6-f16.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-6-f32.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-6-f64.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-7-f16.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-7-f32.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-7-f64.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-8-f16.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-8-f32.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-8-f64.c: New test.
|
|
Since
commit 401199377c50045ede560daf3f6e8b51749c2a87
Author: H.J. Lu <hjl.tools@gmail.com>
Date: Tue Jun 17 10:17:17 2025 +0800
x86: Improve vector_loop/unrolled_loop for memset/memcpy
uses move_by_pieces and store_by_pieces to expand memcpy/memset epilogue
with vector_loop even when targetm.use_by_pieces_infrastructure_p returns
false, which triggers
gcc_assert (targetm.use_by_pieces_infrastructure_p
(len, align,
memsetp ? SET_BY_PIECES : STORE_BY_PIECES,
optimize_insn_for_speed_p ()));
in store_by_pieces. Fix it by:
1. Add by_pieces_in_use to machine_function to indicate that by_pieces op
is currently in use.
2. Set and clear by_pieces_in_use when expanding memcpy/memset epilogue
with move_by_pieces and store_by_pieces.
3. Define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P to return true if
by_pieces_in_use is true.
gcc/
PR target/121096
* config/i386/i386-expand.cc (expand_cpymem_epilogue): Set and
clear by_pieces_in_use when using by_pieces op.
(expand_setmem_epilogue): Likewise.
* config/i386/i386.cc (ix86_use_by_pieces_infrastructure_p): New.
(TARGET_USE_BY_PIECES_INFRASTRUCTURE_P): Likewise.
* config/i386/i386.h (machine_function): Add by_pieces_in_use.
gcc/testsuite/
PR target/121096
* gcc.target/i386/memcpy-strategy-14.c: New test.
* gcc.target/i386/memcpy-strategy-15.c: Likewise.
* gcc.target/i386/memset-strategy-10.c: Likewise.
* gcc.target/i386/memset-strategy-11.c: Likewise.
* gcc.target/i386/memset-strategy-12.c: Likewise.
* gcc.target/i386/memset-strategy-13.c: Likewise.
* gcc.target/i386/memset-strategy-14.c: Likewise.
* gcc.target/i386/memset-strategy-15.c: Likewise.
Signed-off-by: H.J. Lu <hjl.tools@gmail.com>
|
|
Since the constant passed to setmem_epilogue_gen_val may not be in
word_mode, update setmem_epilogue_gen_val to handle any integer modes.
gcc/
PR target/121108
* config/i386/i386-expand.cc (setmem_epilogue_gen_val): Don't
assert op_mode == word_mode and handle any integer modes.
gcc/testsuite/
PR target/121108
* gcc.target/i386/memset-strategy-16.c: New test.
Signed-off-by: H.J. Lu <hjl.tools@gmail.com>
|
|
Source operands of 2 TLS_CALL patterns in
(insn 10 9 11 3 (set (reg:DI 100)
(unspec:DI [
(symbol_ref:DI ("caml_state") [flags 0x10] <var_decl 0x7fe10e1d9e40 caml_state>)
] UNSPEC_TLSDESC)) "x.c":7:16 1674 {*tls_dynamic_gnu2_lea_64_di}
(nil))
(insn 11 10 12 3 (parallel [
(set (reg:DI 99)
(unspec:DI [
(symbol_ref:DI ("caml_state") [flags 0x10] <var_decl 0x7fe10e1d9e40 caml_state>)
(reg:DI 100)
(reg/f:DI 7 sp)
] UNSPEC_TLSDESC))
(clobber (reg:CC 17 flags))
]) "x.c":7:16 1676 {*tls_dynamic_gnu2_call_64_di}
(expr_list:REG_DEAD (reg:DI 100)
(expr_list:REG_UNUSED (reg:CC 17 flags)
(nil))))
and
(insn 19 17 20 4 (set (reg:DI 104)
(unspec:DI [
(symbol_ref:DI ("caml_state") [flags 0x10] <var_decl 0x7fe10e1d9e40 caml_state>)
] UNSPEC_TLSDESC)) "x.c":6:10 discrim 1 1674 {*tls_dynamic_gnu2_lea_64_di}
(nil))
(insn 20 19 21 4 (parallel [
(set (reg:DI 103)
(unspec:DI [
(symbol_ref:DI ("caml_state") [flags 0x10] <var_decl 0x7fe10e1d9e40 caml_state>)
(reg:DI 104)
(reg/f:DI 7 sp)
] UNSPEC_TLSDESC))
(clobber (reg:CC 17 flags))
]) "x.c":6:10 discrim 1 1676 {*tls_dynamic_gnu2_call_64_di}
(expr_list:REG_DEAD (reg:DI 104)
(expr_list:REG_UNUSED (reg:CC 17 flags)
(nil))))
are the same even though rtx_equal_p returns false since (reg:DI 100)
and (reg:DI 104) are set from the same symbol. Use the UNSPEC_TLSDESC
symbol
(unspec:DI [(symbol_ref:DI ("caml_state") [flags 0x10])] UNSPEC_TLSDESC))
to check if 2 TLS_CALL patterns have the same source.
For TLS64_COMBINE, use both UNSPEC_TLSDESC and UNSPEC_DTPOFF unspecs to
check if 2 TLS64_COMBINE patterns have the same source.
gcc/
PR target/121694
* config/i386/i386-features.cc (redundant_pattern): Add
tlsdesc_val.
(pass_x86_cse): Likewise.
(pass_x86_cse::tls_set_insn_from_symbol): New member function.
(pass_x86_cse::candidate_gnu2_tls_p): Set tlsdesc_val. For
TLS64_COMBINE, match both UNSPEC_TLSDESC and UNSPEC_DTPOFF
symbols. For TLS64_CALL, match the UNSPEC_TLSDESC sumbol.
(pass_x86_cse::x86_cse): Initialize the tlsdesc_val field in
load. Pass the tlsdesc_val field to ix86_place_single_tls_call
for X86_CSE_TLSDESC.
gcc/testsuite/
PR target/121694
* gcc.target/i386/pr121668-1b.c: New test.
* gcc.target/i386/pr121694-1a.c: Likewise.
* gcc.target/i386/pr121694-1b.c: Likewise.
Signed-off-by: H.J. Lu <hjl.tools@gmail.com>
|
|
This patch would like to combine the vec_duplicate + vnmsac.vv to the
vnmsac.vx. From example as below code. The related pattern will depend
on the cost of vec_duplicate from GR2VR. Then the late-combine will
take action if the cost of GR2VR is zero, and reject the combination
if the GR2VR cost is greater than zero.
Assume we have example code like below, GR2VR cost is 0.
#define DEF_VX_TERNARY_CASE_0(T, OP_1, OP_2, NAME)
\
void
\
test_vx_ternary_##NAME##_##T##_case_0 (T * restrict vd, T * restrict
vs2, \NAME T rs1, unsigned n)
\
{
\
for (unsigned i = 0; i < n; i++)
\
vd[i] = vd[i] OP_2 vs2[i] OP_1 rs1;
\
}
DEF_VX_TERNARY_CASE_0(int32_t, *, +, macc)
Before this patch:
11 │ beq a3,zero,.L8
12 │ vsetvli a5,zero,e32,m1,ta,ma
13 │ vmv.v.x v2,a2
...
16 │ .L3:
17 │ vsetvli a5,a3,e32,m1,ta,ma
...
22 │ vnmsac.vv v1,v2,v3
...
25 │ bne a3,zero,.L3
After this patch:
11 │ beq a3,zero,.L8
...
14 │ .L3:
15 │ vsetvli a5,a3,e32,m1,ta,ma
...
20 │ vnmsac.vx v1,a2,v3
...
23 │ bne a3,zero,.L3
gcc/ChangeLog:
* config/riscv/autovec-opt.md (*vnmsac_vx_<mode>): Add new
pattern to combine to vx.
* config/riscv/vector.md (@pred_vnmsac_vx_<mode>): Add new
pattern to generate rtl.
(*pred_nmsac_<mode>_scalar_undef): Ditto.
Signed-off-by: Pan Li <pan2.li@intel.com>
|
|
This pattern enables the combine pass (or late-combine, depending on the case)
to merge a vec_duplicate into an smin RTL instruction.
Before this patch, we have two instructions, e.g.:
vfmv.v.f v2,fa0
vfmin.vv v1,v1,v2
After, we get only one:
vfmin.vf v1,v1,fa0
gcc/ChangeLog:
* config/riscv/autovec-opt.md (*vfmin_vf_<mode>): Add new pattern to
combine vec_duplicate + vfmin.vv into vfmin.vf.
* config/riscv/vector.md (@pred_<optab><mode>_scalar): Allow VLS modes.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/vls/floating-point-min-2.c: Adjust scan
dump.
* gcc.target/riscv/rvv/autovec/vls/floating-point-min-4.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-1-f16.c: Add vfmin.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-1-f32.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-1-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-2-f16.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-2-f32.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-2-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-3-f16.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-3-f32.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-3-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_binop.h: Add support for
function variants.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_binop_data.h: Add data for
vfmin.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_vfmin-run-1-f16.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_vfmin-run-1-f32.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_vfmin-run-1-f64.c: New test.
|
|
Add an expander for isinf using integer arithmetic. This is
typically faster and avoids generating spurious exceptions on
signaling NaNs. This fixes part of PR66462.
int isinf1 (float x) { return __builtin_isinf (x); }
Before:
fabs s0, s0
mov w0, 2139095039
fmov s31, w0
fcmp s0, s31
cset w0, le
eor w0, w0, 1
ret
After:
fmov w1, s0
mov w0, -16777216
cmp w0, w1, lsl 1
cset w0, eq
ret
gcc:
PR middle-end/66462
* config/aarch64/aarch64.md (isinf<mode>2): Add new expander.
* config/aarch64/iterators.md (mantissa_bits): Add new mode_attr.
gcc/testsuite:
PR middle-end/66462
* gcc.target/aarch64/pr66462.c: Add new test.
|
|
For a basic block with only a debug marker:
(note 3 0 2 2 [bb 2] NOTE_INSN_BASIC_BLOCK)
(note 2 3 5 2 NOTE_INSN_FUNCTION_BEG)
(debug_insn 5 2 16 2 (debug_marker) "x.c":6:3 -1 (nil))
emit the TLS call after debug marker.
gcc/
PR target/121668
* config/i386/i386-features.cc (ix86_emit_tls_call): Emit the
TLS call after debug marker.
gcc/testsuite/
PR target/121668
* gcc.target/i386/pr121668-1a.c: New test.
Signed-off-by: H.J. Lu <hjl.tools@gmail.com>
|
|
More testsuite hygiene. Some of the thead tests are expecting to find
xtheadvdot in the extension set, but it's not defined as a valid extension
anywhere. I'm just removing xtheadvdot. Someone more familiar with these
cores can add it back properly if they're so inclined.
Second, there's a space after the zifencei in a couple of the thead arch
strings. Naturally that causes failures as well. That's a trivial fix, just
remove the bogus whitespace.
That gets us clean on riscv.exp on the pioneer system.
The pioneer is happy, as is riscv32-elf and riscv64-elf. Pushing to the trunk.
gcc/
* config/riscv/riscv-cores.def (xt-c908v): Drop xtheadvdot.
(xt-c910v2): Remove extraenous whitespace.
(xt-c920v2): Drop xtheadvdot and remove extraeonous whitespace.
gcc/testsuite/
* gcc.target/riscv/mcpu-xt-c908v.c: Drop xtheadvdot.
* gcc.target/riscv/mcpu-xt-c920v2.c: Drop xtheadvdot.
|
|
FMA/DOT_PROD_EXPR/SAD_EXPR
The patch is trying to unroll the vectorized loop when there're
FMA/DOT_PRDO_EXPR/SAD_EXPR reductions, it will break cross-iteration dependence
and enable more parallelism(since vectorize will also enable partial
sum).
When there's gather/scatter or scalarization in the loop, don't do the
unroll since the performance bottleneck is not at the reduction.
The unroll factor is set according to FMA/DOT_PROX_EXPR/SAD_EXPR
CEIL ((latency * throught), num_of_reduction)
.i.e
For fma, latency is 4, throught is 2, if there's 1 FMA for reduction
then unroll factor is 2 * 4 / 1 = 8.
There's also a vect_unroll_limit, the final suggested_unroll_factor is
set as MIN (vect_unroll_limix, 8).
The vect_unroll_limit is mainly for register pressure, avoid to many
spills.
Ideally, all instructions in the vectorized loop should be used to
determine unroll_factor with their (latency * throughput) / number,
but that would too much for this patch, and may just GIGO, so the
patch only considers 3 kinds of instructions: FMA, DOT_PROD_EXPR,
SAD_EXPR.
Note when DOT_PROD_EXPR is not native support,
m_num_reduction += 3 * count which almost prevents unroll.
There's performance boost for simple benchmark with DOT_PRDO_EXPR/FMA
chain, slight improvement in SPEC2017 performance.
gcc/ChangeLog:
* config/i386/i386.cc (ix86_vector_costs::ix86_vector_costs):
Addd new memeber m_num_reduc, m_prefer_unroll.
(ix86_vector_costs::add_stmt_cost): Set m_prefer_unroll and
m_num_reduc
(ix86_vector_costs::finish_cost): Determine
m_suggested_unroll_vector with consideration of
reduc_lat_mult_thr, m_num_reduction and
ix86_vect_unroll_limit.
* config/i386/i386.h (enum ix86_reduc_unroll_factor): New
enum.
(processor_costs): Add reduc_lat_mult_thr and
vect_unroll_limit.
* config/i386/x86-tune-costs.h: Initialize
reduc_lat_mult_thr and vect_unroll_limit.
* config/i386/i386.opt: Add -param=ix86-vect-unroll-limit.
gcc/testsuite/ChangeLog:
* gcc.target/i386/vect_unroll-1.c: New test.
* gcc.target/i386/vect_unroll-2.c: New test.
* gcc.target/i386/vect_unroll-3.c: New test.
* gcc.target/i386/vect_unroll-4.c: New test.
* gcc.target/i386/vect_unroll-5.c: New test.
|
|
This pattern enables the combine pass (or late-combine, depending on the case)
to merge a vec_duplicate into a div RTL instruction. The vec_duplicate is the
dividend operand.
Before this patch, we have two instructions, e.g.:
vfmv.v.f v2,fa0
vfdiv.vv v1,v2,v1
After, we get only one:
vfrdiv.vf v1,v1,fa0
gcc/ChangeLog:
* config/riscv/autovec-opt.md (*vfrdiv_vf_<mode>): Add new pattern to
combine vec_duplicate + vfdiv.vv into vfrdiv.vf.
* config/riscv/vector.md (@pred_<optab><mode>_reverse_scalar): Allow VLS
modes.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/vx_vf/vf-1-f16.c: Add vfrdiv.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-1-f32.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-1-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-2-f16.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-2-f32.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-2-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-3-f16.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-3-f32.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-3-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-4-f16.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-4-f32.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-4-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_binop.h: Add support for reverse
variants.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_binop_data.h: Add data for
reverse variants.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_vfrdiv-run-1-f16.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_vfrdiv-run-1-f32.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_vfrdiv-run-1-f64.c: New test.
|
|
invariant [PR121290]
Consider the example:
void
f (int *restrict x, int *restrict y, int *restrict z, int n)
{
for (int i = 0; i < 4; ++i)
{
int res = 0;
for (int j = 0; j < 100; ++j)
res += y[j] * z[i];
x[i] = res;
}
}
we currently vectorize as
f:
movi v30.4s, 0
ldr q31, [x2]
add x2, x1, 400
.L2:
ld1r {v29.4s}, [x1], 4
mla v30.4s, v29.4s, v31.4s
cmp x2, x1
bne .L2
str q30, [x0]
ret
which is not useful because by doing outer-loop vectorization we're performing
less work per iteration than we would had we done inner-loop vectorization and
simply unrolled the inner loop.
This patch teaches the cost model that if all your leafs are invariant, then
adjust the loop cost by * VF, since every vector iteration has at least one lane
really just doing 1 scalar.
There are a couple of ways we could have solved this, one is to increase the
unroll factor to process more iterations of the inner loop. This removes the
need for the broadcast, however we don't support unrolling the inner loop within
the outer loop. We only support unrolling by increasing the VF, which would
affect the outer loop as well as the inner loop.
We also don't directly support costing inner-loop vs outer-loop vectorization,
and as such we're left trying to predict/steer the cost model ahead of time to
what we think should be profitable. This patch attempts to do so using a
heuristic which penalizes the outer-loop vectorization.
We now cost the loop as
note: Cost model analysis:
Vector inside of loop cost: 2000
Vector prologue cost: 4
Vector epilogue cost: 0
Scalar iteration cost: 300
Scalar outside cost: 0
Vector outside cost: 4
prologue iterations: 0
epilogue iterations: 0
missed: cost model: the vector iteration cost = 2000 divided by the scalar iteration cost = 300 is greater or equal to the vectorization factor = 4.
missed: not vectorized: vectorization not profitable.
missed: not vectorized: vector version will never be profitable.
missed: Loop costings may not be worthwhile.
And subsequently generate:
.L5:
add w4, w4, w7
ld1w z24.s, p6/z, [x0, #1, mul vl]
ld1w z23.s, p6/z, [x0, #2, mul vl]
ld1w z22.s, p6/z, [x0, #3, mul vl]
ld1w z29.s, p6/z, [x0]
mla z26.s, p6/m, z24.s, z30.s
add x0, x0, x8
mla z27.s, p6/m, z23.s, z30.s
mla z28.s, p6/m, z22.s, z30.s
mla z25.s, p6/m, z29.s, z30.s
cmp w4, w6
bls .L5
and avoids the load and replicate if it knows it has enough vector pipes to do
so.
gcc/ChangeLog:
PR target/121290
* config/aarch64/aarch64.cc
(class aarch64_vector_costs ): Add m_loop_fully_scalar_dup.
(aarch64_vector_costs::add_stmt_cost): Detect invariant inner loops.
(adjust_body_cost): Adjust final costing if m_loop_fully_scalar_dup.
gcc/testsuite/ChangeLog:
PR target/121290
* gcc.target/aarch64/pr121290.c: New test.
|
|
multiply
This pattern enables the combine pass (or late-combine, depending on the case)
to merge a vec_duplicate into a mult RTL instruction.
Before this patch, we have two instructions, e.g.:
vfmv.v.f v2,fa0
vfmul.vv v1,v1,v2
After, we get only one:
vfmul.vf v2,v2,fa0
gcc/ChangeLog:
* config/riscv/autovec-opt.md (*vfmul_vf_<mode>): Add new pattern to
combine vec_duplicate + vfmul.vv into vfmul.vf.
* config/riscv/vector.md (@pred_<optab><mode>_scalar): Allow VLS modes.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/vx_vf/vf-1-f16.c: Add vfmul.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-1-f32.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-1-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-2-f16.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-2-f32.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-2-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-3-f16.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-3-f32.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-3-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-4-f16.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-4-f32.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf-4-f64.c: Likewise.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_binop.h: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_binop_data.h: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_binop_run.h: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_vfmul-run-1-f16.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_vfmul-run-1-f32.c: New test.
* gcc.target/riscv/rvv/autovec/vx_vf/vf_vfmul-run-1-f64.c: New test.
* gcc.target/riscv/rvv/autovec/vls/floating-point-mul-2.c: Adjust scan
dump.
* gcc.target/riscv/rvv/autovec/vls/floating-point-mul-3.c: Likewise.
|
|
Recent changes from Kito have an unused parameter. On the assumption that he's
going to likely want it as part of the API, I've simply removed the parameter's
name until such time as Kito needs it.
This should restore bootstrapping to the RISC-V port. Committing now rather
than waiting for the CI system given bootstrap builds currently fail.
* config/riscv/riscv.cc (riscv_arg_partial_bytes): Remove name
from unused parameter.
|
|
The following changes the vect_reduc_type API to work on the SLP node.
The API is only used from the aarch64 backend, so all changes are there.
In particular I noticed aarch64_force_single_cycle is invoked even
for scalar costing (where the flag tested isn't computed yet), I
figured in scalar costing all reductions are a single cycle.
* tree-vectorizer.h (vect_reduc_type): Get SLP node as argument.
* config/aarch64/aarch64.cc (aarch64_sve_in_loop_reduction_latency):
Take SLO node as argument and adjust.
(aarch64_in_loop_reduction_latency): Likewise.
(aarch64_detect_vector_stmt_subtype): Adjust.
(aarch64_vector_costs::count_ops): Likewise. Treat reductions
during scalar costing as single-cycle.
|
|
The vgf2p8affineqb_<mode><mask_name> pattern uses "register_operand"
predicate for the first input operand, so using "general_operand"
for the rotate operand passed to it leads to ICEs, and so does
the "nonimmediate_operand" in the <insn>v16qi3 define_expand.
The following patch fixes it by using "register_operand" in the former
case (that pattern is TARGET_GFNI only) and using force_reg in
the latter case (the pattern is TARGET_XOP || TARGET_GFNI and for XOP
we can handle MEM operand).
The rest of the changes are small formatting tweaks or use of const0_rtx
instead of GEN_INT (0).
2025-08-26 Jakub Jelinek <jakub@redhat.com>
PR target/121658
* config/i386/sse.md (<insn><mode>3 any_shift): Use const0_rtx
instead of GEN_INT (0).
(cond_<insn><mode> any_shift): Likewise. Formatting fix.
(<insn><mode>3 any_rotate): Use register_operand predicate instead of
general_operand for match_operand 1. Use const0_rtx instead of
GEN_INT (0).
(<insn>v16qi3 any_rotate): Use force_reg on operands[1]. Formatting
fix.
* config/i386/i386.cc (ix86_shift_rotate_cost): Comment formatting
fixes.
* gcc.target/i386/pr121658.c: New test.
|
|
This patch would like to combine the vec_duplicate + vmacc.vv to the
vmacc.vx. From example as below code. The related pattern will depend
on the cost of vec_duplicate from GR2VR. Then the late-combine will
take action if the cost of GR2VR is zero, and reject the combination
if the GR2VR cost is greater than zero.
Assume we have example code like below, GR2VR cost is 0.
#define DEF_VX_TERNARY_CASE_0(T, OP_1, OP_2, NAME) \
void \
test_vx_ternary_##NAME##_##T##_case_0 (T * restrict vd, T * restrict vs2, \
T rs1, unsigned n) \
{ \
for (unsigned i = 0; i < n; i++) \
vd[i] = vd[i] OP_2 vs2[i] OP_1 rs1; \
}
DEF_VX_TERNARY_CASE_0(int32_t, *, +, macc)
Before this patch:
11 │ beq a3,zero,.L8
12 │ vsetvli a5,zero,e32,m1,ta,ma
13 │ vmv.v.x v2,a2
...
16 │ .L3:
17 │ vsetvli a5,a3,e32,m1,ta,ma
...
22 │ vmacc.vv v1,v2,v3
...
25 │ bne a3,zero,.L3
After this patch:
11 │ beq a3,zero,.L8
...
14 │ .L3:
15 │ vsetvli a5,a3,e32,m1,ta,ma
...
20 │ vmacc.vx v1,a2,v3
...
23 │ bne a3,zero,.L3
gcc/ChangeLog:
* config/riscv/vector.md (@pred_mul_plus_vx_<mode>): Add new pattern to
generate vmacc rtl.
(*pred_macc_<mode>_scalar_undef): Ditto.
* config/riscv/autovec-opt.md (*vmacc_vx_<mode>): Add new
pattern to match the vmacc vx combine.
Signed-off-by: Pan Li <pan2.li@intel.com>
|
|
multiple alternatives
The use of compact syntax makes the relationship between asm output,
operand constraints, and insn attributes easier to understand and modify,
especially for "mov<mode>_internal".
gcc/ChangeLog:
* config/xtensa/xtensa.md (addsi3, <u>mulhisi3, andsi3,
zero_extend<mode>si2, extendhisi2_internal, movsi_internal,
movhi_internal, movqi_internal, movsf_internal, ashlsi3_internal,
ashrsi3, lshrsi3, rotlsi3, rotrsi3):
Rewrite in compact syntax.
|
|
gcc/ChangeLog:
* config/xtensa/xtensa.md
(The auxiliary define_split for *masktrue_const_bitcmpl):
Use a more concise function call, i.e.,
(1 << GET_MODE_BITSIZE (mode)) - 1 is equivalent to
GET_MODE_MASK (mode).
|
