On 20/10/2020 12:22, Richard Earnshaw wrote:
> On 19/10/2020 17:32, Christophe Lyon via Gcc-patches wrote:
>> On Mon, 19 Oct 2020 at 16:39, Richard Earnshaw
>> <[email protected]> wrote:
>>>
>>> On 12/10/2020 08:59, Christophe Lyon via Gcc-patches wrote:
>>>> On Thu, 8 Oct 2020 at 11:58, Richard Earnshaw
>>>> <[email protected]> wrote:
>>>>>
>>>>> On 08/10/2020 10:07, Christophe Lyon via Gcc-patches wrote:
>>>>>> On Tue, 6 Oct 2020 at 18:02, Richard Earnshaw
>>>>>> <[email protected]> wrote:
>>>>>>>
>>>>>>> On 29/09/2020 20:50, Christophe Lyon via Gcc-patches wrote:
>>>>>>>> When mi_delta is > 255 and -mpure-code is used, we cannot load delta
>>>>>>>> from code memory (like we do without -mpure-code).
>>>>>>>>
>>>>>>>> This patch builds the value of mi_delta into r3 with a series of
>>>>>>>> movs/adds/lsls.
>>>>>>>>
>>>>>>>> We also do some cleanup by not emitting the function address and delta
>>>>>>>> via .word directives at the end of the thunk since we don't use them
>>>>>>>> with -mpure-code.
>>>>>>>>
>>>>>>>> No need for new testcases, this bug was already identified by
>>>>>>>> eg. pr46287-3.C
>>>>>>>>
>>>>>>>> 2020-09-29 Christophe Lyon <[email protected]>
>>>>>>>>
>>>>>>>> gcc/
>>>>>>>> * config/arm/arm.c (arm_thumb1_mi_thunk): Build mi_delta in r3
>>>>>>>> and
>>>>>>>> do not emit function address and delta when -mpure-code is used.
>>>>>>>
>>>>>> Hi Richard,
>>>>>>
>>>>>> Thanks for your comments.
>>>>>>
>>>>>>> There are some optimizations you can make to this code.
>>>>>>>
>>>>>>> Firstly, for values between 256 and 510 (inclusive), it would be better
>>>>>>> to just expand a mov of 255 followed by an add.
>>>>>> I now see the splitted for the "Pe" constraint which I hadn't noticed
>>>>>> before, so I can write something similar indeed.
>>>>>>
>>>>>> However, I'm note quite sure to understand the benefit in the split
>>>>>> when -mpure-code is NOT used.
>>>>>> Consider:
>>>>>> int f3_1 (void) { return 510; }
>>>>>> int f3_2 (void) { return 511; }
>>>>>> Compile with -O2 -mcpu=cortex-m0:
>>>>>> f3_1:
>>>>>> movs r0, #255
>>>>>> lsls r0, r0, #1
>>>>>> bx lr
>>>>>> f3_2:
>>>>>> ldr r0, .L4
>>>>>> bx lr
>>>>>>
>>>>>> The splitter makes the code bigger, does it "compensate" for this by
>>>>>> not having to load the constant?
>>>>>> Actually the constant uses 4 more bytes, which should be taken into
>>>>>> account when comparing code size,
>>>>>
>>>>> Yes, the size of the literal pool entry needs to be taken into account.
>>>>> It might happen that the entry could be shared with another use of that
>>>>> literal, but in general that's rare.
>>>>>
>>>>>> so f3_1 uses 6 bytes, and f3_2 uses 8, so as you say below three
>>>>>> thumb1 instructions would be equivalent in size compared to loading
>>>>>> from the literal pool. Should the 256-510 range be extended?
>>>>>
>>>>> It's a bit borderline at three instructions when literal pools are not
>>>>> expensive to use, but in thumb1 literal pools tend to be quite small due
>>>>> to the limited pc offsets we can use. I think on balance we probably
>>>>> want to use the instruction sequence unless optimizing for size.
>>>>>
>>>>>>
>>>>>>
>>>>>>> This is also true for
>>>>>>> the literal pools alternative as well, so should be handled before all
>>>>>>> this.
>>>>>> I am not sure what you mean: with -mpure-code, the above sample is
>>>>>> compiled as:
>>>>>> f3_1:
>>>>>> movs r0, #255
>>>>>> lsls r0, r0, #1
>>>>>> bx lr
>>>>>> f3_2:
>>>>>> movs r0, #1
>>>>>> lsls r0, r0, #8
>>>>>> adds r0, r0, #255
>>>>>> bx lr
>>>>>>
>>>>>> so the "return 510" case is already handled as without -mpure-code.
>>>>>
>>>>> I was thinking specifically of the thunk sequence where you seem to be
>>>>> emitting instructions directly rather than generating RTL. The examples
>>>>> you show here are not thunks.
>>>>>
>>>> OK thanks for the clarification.
>>>>
>>>> Here is an updated version, split into 3 patches to hopefully make
>>>> review easier.
>>>> They apply on top of my other mpure-code patches for PR96967 and PR96770:
>>>> https://gcc.gnu.org/pipermail/gcc-patches/2020-September/554956.html
>>>> https://gcc.gnu.org/pipermail/gcc-patches/2020-September/554957.html
>>>>
>>>> I kept it this way to make incremental changes easier to understand.
>>>>
>>>> Patch 1: With the hope to avoid confusion and make maintenance easier,
>>>> I have updated thumb1_gen_const_int() so that it can generate either RTL or
>>>> asm. This way, all the code used to build thumb-1 constants is in the
>>>> same place,
>>>> in case we need to improve/fix it later. We now generate shorter
>>>> sequences in
>>>> several cases matching your comments.
>>>>
>>>> Patch 2: Removes the equivalent loop from thumb1_movsi_insn pattern and
>>>> calls thumb1_gen_const_int.
>>>>
>>>> Patch 3: Update of the original patch in this thread, now calls
>>>> thumb1_gen_const_int.
>>>
>>> Yuk! Those changes to thumb1_gen_const_int are horrible.
>>>
>>> I think we should be able to leverage the fact that the compiler can use
>>> C++ now to do much better than that, for example by making that function
>>> a template. For example (and this is just a sketch):
>>>
>>
>> Indeed! I didn't think about it since there is no other use of
>> templates in arm.c yet.
>> I'll send an update soon.
>>
>> Other than that, does the approach look OK to you?
>>
>
> Yes, I think this is heading in the right direction. Bringing the two
> immediate generating operations into a single function can only be a
> good thing.
>
> Looking again at your example constant sequences, I see:
>
> 0x1000010:
> movs r3, #16
> lsls r3, #16
> adds r3, #1
> lsls r3, #4
> 0x1000011:
> movs r3, #1
> lsls r3, #24
> adds r3, #17
>
> The first of these looks odd, given the second sequence. Why doesn't
> the first expand to
>
> 0x1000010:
> movs r3, #16
> lsls r3, #16
> adds r3, #16
>
Err, I mean to:
0x1000010:
movs r3, #1
lsls r3, #24
adds r3, #16
> ?
>
> R.
>
>> Thanks,
>>
>> Christophe
>>
>>> class t1_rtl
>>> {
>>> public:
>>> void ashift(int a) { gen_rtx_ASHIFT(a); }
>>> void rshift(int b) { gen_rtx_SHIFTRT(b); }
>>> };
>>>
>>> class t1_print
>>> {
>>> public:
>>> t1_print (FILE *f) : t_file(f) {}
>>> void ashift (int a) { fprintf (t_file, "a shift %d\n", a); }
>>> void rshift (int b) { fprintf (t_file, "r shift %d\n", b); }
>>> private:
>>> FILE *t_file;
>>> };
>>>
>>> template <class T>
>>> void thumb1_gen_const_int(T t, int f)
>>> {
>>> // Expansion of thumb1_gen_const_int ...
>>> t.ashift(f);
>>> }
>>>
>>> // Usage...
>>> void f1()
>>> {
>>> // Use the RTL expander
>>> t1_rtl g;
>>> thumb1_gen_const_int (g, 3);
>>> }
>>>
>>> void f2()
>>> {
>>> // Use the printf expander writing to stdout
>>> t1_print g(stdout);
>>> thumb1_gen_const_int (g, 3);
>>> }
>>>
>>> With this you can write thumb1_gen_const_int without having to worry
>>> about which expander is being used in each instance and the template
>>> expansion will use the right version.
>>>
>>> R.
>>>
>>>>
>>>>>
>>>>>>
>>>>>>> I also suspect (but haven't check) that the base adjustment will
>>>>>>> most commonly be a multiple of the machine word size (ie 4). If that is
>>>>>>> the case then you could generate n/4 and then shift it left by 2 for an
>>>>>>> even greater range of literals.
>>>>>> I can see there is provision for this in the !TARGET_THUMB1_ONLY case,
>>>>>> I'll update my patch.
>>>>>>
>>>>>>> More generally, any sequence of up to
>>>>>>> three thumb1 instructions will be no larger, and probably as fast as the
>>>>>>> existing literal pool fall back.
>>>>>>>
>>>>>>> Secondly, if the value is, for example, 65536 (0x10000), your code will
>>>>>>> emit a mov followed by two shift-by-8 instructions; the two shifts could
>>>>>>> be merged into a single shift-by-16.
>>>>>>
>>>>>> Right, I'll try to make use of thumb_shiftable_const.
>>>>>>
>>>>>>> Finally, I'd really like to see some executable tests for this, if at
>>>>>>> all possible.
>>>>>> I mentioned pr46287-3.C, but that's not the only existing testcase
>>>>>> that showed the problem. There are also:
>>>>>> g++.dg/opt/thunk1.C
>>>>>> g++.dg/ipa/pr46984.C
>>>>>> g++.dg/torture/pr46287.C
>>>>>> g++.dg/torture/pr45699.C
>>>>>>
>>>>>> Do you want that I copy one of these in the arm subdir and add
>>>>>> -mpure-code in dg-options?
>>>>>
>>>>> On reflection, probably not - that just makes things more complicated
>>>>> with all the dg-options mess (I'm worried about interactions with other
>>>>> sets of options on the command line and the fall-out from that). If
>>>>> someone cares about pure-code they should be doing full testsuite runs
>>>>> with it enabled and that should be sufficient.
>>>>
>>>> Yes, that's what I am doing manually, it's a bit tricky, and I use a
>>>> modified simulator.
>>>>
>>>> Thanks,
>>>>
>>>> Christophe
>>>>
>>>>>
>>>>> R.
>>>>>
>>>>>>
>>>>>> Thanks,
>>>>>>
>>>>>> Christophe
>>>>>>
>>>>>>> R.
>>>>>>>
>>>>>>>>
>>>>>>>> k# (use "git pull" to merge the remote branch into yours)
>>>>>>>> ---
>>>>>>>> gcc/config/arm/arm.c | 91
>>>>>>>> +++++++++++++++++++++++++++++++++++++---------------
>>>>>>>> 1 file changed, 66 insertions(+), 25 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
>>>>>>>> index ceeb91f..62abeb5 100644
>>>>>>>> --- a/gcc/config/arm/arm.c
>>>>>>>> +++ b/gcc/config/arm/arm.c
>>>>>>>> @@ -28342,9 +28342,43 @@ arm_thumb1_mi_thunk (FILE *file, tree,
>>>>>>>> HOST_WIDE_INT delta,
>>>>>>>> {
>>>>>>>> if (mi_delta > 255)
>>>>>>>> {
>>>>>>>> - fputs ("\tldr\tr3, ", file);
>>>>>>>> - assemble_name (file, label);
>>>>>>>> - fputs ("+4\n", file);
>>>>>>>> + /* With -mpure-code, we cannot load delta from the constant
>>>>>>>> + pool: we build it explicitly. */
>>>>>>>> + if (target_pure_code)
>>>>>>>> + {
>>>>>>>> + bool mov_done_p = false;
>>>>>>>> + int i;
>>>>>>>> +
>>>>>>>> + /* Emit upper 3 bytes if needed. */
>>>>>>>> + for (i = 0; i < 3; i++)
>>>>>>>> + {
>>>>>>>> + int byte = (mi_delta >> (8 * (3 - i))) & 0xff;
>>>>>>>> +
>>>>>>>> + if (byte)
>>>>>>>> + {
>>>>>>>> + if (mov_done_p)
>>>>>>>> + asm_fprintf (file, "\tadds\tr3, #%d\n", byte);
>>>>>>>> + else
>>>>>>>> + asm_fprintf (file, "\tmovs\tr3, #%d\n", byte);
>>>>>>>> + mov_done_p = true;
>>>>>>>> + }
>>>>>>>> +
>>>>>>>> + if (mov_done_p)
>>>>>>>> + asm_fprintf (file, "\tlsls\tr3, #8\n");
>>>>>>>> + }
>>>>>>>> +
>>>>>>>> + /* Emit lower byte if needed. */
>>>>>>>> + if (!mov_done_p)
>>>>>>>> + asm_fprintf (file, "\tmovs\tr3, #%d\n", mi_delta & 0xff);
>>>>>>>> + else if (mi_delta & 0xff)
>>>>>>>> + asm_fprintf (file, "\tadds\tr3, #%d\n", mi_delta & 0xff);
>>>>>>>> + }
>>>>>>>> + else
>>>>>>>> + {
>>>>>>>> + fputs ("\tldr\tr3, ", file);
>>>>>>>> + assemble_name (file, label);
>>>>>>>> + fputs ("+4\n", file);
>>>>>>>> + }
>>>>>>>> asm_fprintf (file, "\t%ss\t%r, %r, r3\n",
>>>>>>>> mi_op, this_regno, this_regno);
>>>>>>>> }
>>>>>>>> @@ -28380,30 +28414,37 @@ arm_thumb1_mi_thunk (FILE *file, tree,
>>>>>>>> HOST_WIDE_INT delta,
>>>>>>>> fputs ("\tpop\t{r3}\n", file);
>>>>>>>>
>>>>>>>> fprintf (file, "\tbx\tr12\n");
>>>>>>>> - ASM_OUTPUT_ALIGN (file, 2);
>>>>>>>> - assemble_name (file, label);
>>>>>>>> - fputs (":\n", file);
>>>>>>>> - if (flag_pic)
>>>>>>>> +
>>>>>>>> + /* With -mpure-code, we don't need to emit literals for the
>>>>>>>> + function address and delta since we emitted code to build
>>>>>>>> + them. */
>>>>>>>> + if (!target_pure_code)
>>>>>>>> {
>>>>>>>> - /* Output ".word .LTHUNKn-[3,7]-.LTHUNKPCn". */
>>>>>>>> - rtx tem = XEXP (DECL_RTL (function), 0);
>>>>>>>> - /* For TARGET_THUMB1_ONLY the thunk is in Thumb mode, so the PC
>>>>>>>> - pipeline offset is four rather than eight. Adjust the
>>>>>>>> offset
>>>>>>>> - accordingly. */
>>>>>>>> - tem = plus_constant (GET_MODE (tem), tem,
>>>>>>>> - TARGET_THUMB1_ONLY ? -3 : -7);
>>>>>>>> - tem = gen_rtx_MINUS (GET_MODE (tem),
>>>>>>>> - tem,
>>>>>>>> - gen_rtx_SYMBOL_REF (Pmode,
>>>>>>>> - ggc_strdup
>>>>>>>> (labelpc)));
>>>>>>>> - assemble_integer (tem, 4, BITS_PER_WORD, 1);
>>>>>>>> - }
>>>>>>>> - else
>>>>>>>> - /* Output ".word .LTHUNKn". */
>>>>>>>> - assemble_integer (XEXP (DECL_RTL (function), 0), 4,
>>>>>>>> BITS_PER_WORD, 1);
>>>>>>>> + ASM_OUTPUT_ALIGN (file, 2);
>>>>>>>> + assemble_name (file, label);
>>>>>>>> + fputs (":\n", file);
>>>>>>>> + if (flag_pic)
>>>>>>>> + {
>>>>>>>> + /* Output ".word .LTHUNKn-[3,7]-.LTHUNKPCn". */
>>>>>>>> + rtx tem = XEXP (DECL_RTL (function), 0);
>>>>>>>> + /* For TARGET_THUMB1_ONLY the thunk is in Thumb mode, so
>>>>>>>> the PC
>>>>>>>> + pipeline offset is four rather than eight. Adjust the
>>>>>>>> offset
>>>>>>>> + accordingly. */
>>>>>>>> + tem = plus_constant (GET_MODE (tem), tem,
>>>>>>>> + TARGET_THUMB1_ONLY ? -3 : -7);
>>>>>>>> + tem = gen_rtx_MINUS (GET_MODE (tem),
>>>>>>>> + tem,
>>>>>>>> + gen_rtx_SYMBOL_REF (Pmode,
>>>>>>>> + ggc_strdup
>>>>>>>> (labelpc)));
>>>>>>>> + assemble_integer (tem, 4, BITS_PER_WORD, 1);
>>>>>>>> + }
>>>>>>>> + else
>>>>>>>> + /* Output ".word .LTHUNKn". */
>>>>>>>> + assemble_integer (XEXP (DECL_RTL (function), 0), 4,
>>>>>>>> BITS_PER_WORD, 1);
>>>>>>>>
>>>>>>>> - if (TARGET_THUMB1_ONLY && mi_delta > 255)
>>>>>>>> - assemble_integer (GEN_INT(mi_delta), 4, BITS_PER_WORD, 1);
>>>>>>>> + if (TARGET_THUMB1_ONLY && mi_delta > 255)
>>>>>>>> + assemble_integer (GEN_INT(mi_delta), 4, BITS_PER_WORD, 1);
>>>>>>>> + }
>>>>>>>> }
>>>>>>>> else
>>>>>>>> {
>>>>>>>>
>>>>>>>
>>>>>
>>>
>
