Re: [PATCH, ARM] Improve robustness of -mslow-flash-data

[Christophe Lyon <christophe dot lyon at linaro dot org>]

On Mon, 19 Nov 2018 at 18:56, Thomas Preudhomme
<thomas.preudhomme@foss.arm.com> wrote:
>
> Hi,
>
> Current code to handle -mslow-flash-data in machine description files
> suffers from a number of issues which this patch fixes:
>
> 1) The insn_and_split in vfp.md to load a generic floating-point
> constant via GPR first and move it to VFP register are guarded by
> !reload_completed which is forbidden explicitely in the GCC internals
> documentation section 17.2 point 3;
>
> 2) A number of testcase in the testsuite ICEs under -mslow-flash-data
> when targeting the hardfloat ABI [1];
>
> 3) Instructions performing load from literal pool are not disabled.
>
> These problems are addressed by 2 separate actions:
>
> 1) Making the splitters take a clobber and changing the expanders
> accordingly to generate a mov with clobber in cases where a literal
> pool would be used. The splitter can thus be enabled after reload since
> it does not call gen_reg_rtx anymore;
>
> 2) Adding new predicates and constraints to disable literal pool loads
> in existing instructions when -mslow-flash-data is in effect.
>
> The patch also rework the splitter for DFmode slightly to generate an
> intermediate DI load instead of 2 intermediate SI loads, thus relying on
> the existing DI splitters instead of redoing their job. At last, the
> patch adds some missing arm_fp_ok effective target to some of the
> slow-flash-data testcases.
>
> [1]
> c-c++-common/Wunused-var-3.c
> gcc.c-torture/compile/pr72771.c
> gcc.c-torture/compile/vector-5.c
> gcc.c-torture/compile/vector-6.c
> gcc.c-torture/execute/20030914-1.c
> gcc.c-torture/execute/20050316-1.c
> gcc.c-torture/execute/pr59643.c
> gcc.dg/builtin-tgmath-1.c
> gcc.dg/debug/pr55730.c
> gcc.dg/graphite/interchange-7.c
> gcc.dg/pr56890-2.c
> gcc.dg/pr68474.c
> gcc.dg/pr80286.c
> gcc.dg/torture/pr35227.c
> gcc.dg/torture/pr65077.c
> gcc.dg/torture/pr86363.c
> g++.dg/torture/pr81112.C
> g++.dg/torture/pr82985.C
> g++.dg/warn/Wunused-var-7.C
> and a lot more in libstdc++ in special_functions/*_comp_ellint_* and
> special_functions/*_ellint_* directories.
>
> ChangeLog entries are as follows:
>
> *** gcc/ChangeLog ***
>
> 2018-11-14  Thomas Preud'homme  <thomas.preudhomme@arm.com>
>
>         * config/arm/arm.md (arm_movdi): Split if -mslow-flash-data and
>         source is a constant that would be loaded by literal pool.
>         (movsf expander): Generate a no_literal_pool_sf_immediate insn if
>         -mslow-flash-data is present, targeting hardfloat ABI and source is a
>         float constant that cannot be loaded via vmov.
>         (movdf expander): Likewise but generate a no_literal_pool_df_immediate
>         insn.
>         (arm_movsf_soft_insn): Split if -mslow-flash-data and source is a
>         float constant that would be loaded by literal pool.
>         (softfloat constant movsf splitter): Splitter for the above case.
>         (movdf_soft_insn): Split if -mslow-flash-data and source is a float
>         constant that would be loaded by literal pool.
>         (softfloat constant movdf splitter): Splitter for the above case.
>         * config/arm/constraints.md (Pz): Document existing constraint.
>         (Ha): Define constraint.
>         (Tu): Likewise.
>         * config/arm/predicates.md (hard_sf_operand): New predicate.
>         (hard_df_operand): Likewise.
>         * config/arm/thumb2.md (thumb2_movsi_insn): Split if
>         -mslow-flash-data and constant would be loaded by literal pool.
>         * constant/arm/vfp.md (thumb2_movsi_vfp): Likewise and disable constant
>         load in VFP register.
>         (movdi_vfp): Likewise.
>         (thumb2_movsf_vfp): Use hard_sf_operand as predicate for source to
>         prevent match for a constant load if -mslow-flash-data and constant
>         cannot be loaded via vmov.  Adapt constraint accordingly by
>         using Ha instead of E for generic floating-point constant load.
>         (thumb2_movdf_vfp): Likewise using hard_df_operand predicate instead.
>         (no_literal_pool_df_immediate): Add a clobber to use as the
>         intermediate general purpose register and also enable it after reload
>         but disable it constant is a valid FP constant.  Add constraints and
>         generate a DI intermediate load rather than 2 SI loads.
>         (no_literal_pool_sf_immediate): Add a clobber to use as the
>         intermediate general purpose register and also enable it after
>         reload.
>
> *** gcc/testsuite/ChangeLog ***
>
> 2018-11-14  Thomas Preud'homme  <thomas.preudhomme@arm.com>
>
>         * gcc.target/arm/thumb2-slow-flash-data-2.c: Require arm_fp_ok
>         effective target.
>         * gcc.target/arm/thumb2-slow-flash-data-3.c: Likewise.
>         * gcc.target/arm/thumb2-slow-flash-data-4.c: Likewise.
>         * gcc.target/arm/thumb2-slow-flash-data-5.c: Likewise.
>
> Testing: Built arm-none-eabi cross compilers for Armv7E-M defaulting to
> softfloat and hardfloat ABI which showed no regression and some
> FAIL->PASS for hardfloat ABI. Bootstraped on Arm and Thumb-2 without any
> regression. Compiled SPEC2k6 without -mslow-flash-data and checked that
> code generation didn't change.
>

FWIW, it's OK for my validations: I do not see the improvements you mention,
because I do not have this very configuration. On arm-eabi --with-cpu=cortex-m3,
I only see the above tests PASS -> UNSUPPORTED, which is OK
(the reason being "-mfloat-abi=hard: selected procesor lacks an FPU"_

Thanks,

Christophe


> Is this ok for stage3?
>
> Best regards,
>
> Thomas