llvm dev - Jan 2020 - How to find out the default CPU / Features String for a given triple?

When I pass an empty string for cpu and features to createTargetMachine,
and then use LLVMGetTargetMachineCPU() and
LLVMGetTargetMachineFeatureString() to get the strings back, they are
still empty. Is there a way to have llvm compute the effective
cpu/features string, and provide it so that I can inspect it?

I'm trying to figure out how the cpu/features string that I am
explicitly passing, which I am expecting to be equivalent to passing
empty string, is different.

As an example, I have a test passing in the CI using the triple
"aarch64v8.1a-unknown-linux-unknown", and "" for both CPU
name and
target features string. However when I pass the following target
specific features string, I get qemu crashing in the CI:

-a35,-a53,-a55,-a57,-a72,-a73,-a75,-a76,-aes,-aggressive-fma,-alternate-sextload-cvt-f32-pattern,+altnzcv,+am,-arith-bcc-fusion,-arith-cbz-fusion,-balance-fp-ops,+bti,-call-saved-x10,-call-saved-x11,-call-saved-x12,-call-saved-x13,-call-saved-x14,-call-saved-x15,-call-saved-x18,-call-saved-x8,-call-saved-x9,+ccdp,+ccidx,+ccpp,+complxnum,+crc,-crypto,-custom-cheap-as-move,-cyclone,-disable-latency-sched-heuristic,+dit,+dotprod,-exynos-cheap-as-move,-exynosm1,-exynosm2,-exynosm3,-exynosm4,-falkor,+fmi,-force-32bit-jump-tables,+fp-armv8,-fp16fml,+fptoint,-fullfp16,-fuse-address,+fuse-aes,-fuse-arith-logic,-fuse-crypto-eor,-fuse-csel,-fuse-literals,+jsconv,-kryo,+lor,+lse,-lsl-fast,+mpam,-mte,+neon,-no-neg-immediates,+nv,+pa,+pan,+pan-rwv,+perfmon,-predictable-select-expensive,+predres,-rand,+ras,+rasv8_4,+rcpc,+rcpc-immo,+rdm,-reserve-x1,-reserve-x10,-reserve-x11,-reserve-x12,-reserve-x13,-reserve-x14,-reserve-x15,-reserve-x18,-reserve-x2,-reserve-x20,-reserve-x21,-reserve-x22,-reserve-x23,-reserve-x24,-reserve-x25,-reserve-x26,-reserve-x27,-reserve-x28,-reserve-x3,-reserve-x4,-reserve-x5,-reserve-x6,-reserve-x7,-reserve-x9,-saphira,+sb,+sel2,-sha2,-sha3,-slow-misaligned-128store,-slow-paired-128,-slow-strqro-store,-sm4,-spe,+specrestrict,+ssbs,-strict-align,-sve,-sve2,-sve2-aes,-sve2-bitperm,-sve2-sha3,-sve2-sm4,-thunderx,-thunderx2t99,-thunderxt81,-thunderxt83,-thunderxt88,+tlb-rmi,-tpidr-el1,-tpidr-el2,-tpidr-el3,+tracev8.4,-tsv110,+uaops,-use-aa,+use-postra-scheduler,-use-reciprocal-square-root,+v8.1a,+v8.2a,+v8.3a,+v8.4a,+v8.5a,+vh,-zcm,-zcz,-zcz-fp,-zcz-fp-workaround,-zcz-gp

Now, I understand that qemu crashing can be fixed by using a newer qemu
version. And, indeed, on my laptop with a newer qemu, the test passes.
However, this test was previously passing on the CI, which means that
with this explicit string of target specific features, *something is
different*. I am trying to determine what that different thing is.

Thanks in advance for the help,
Andrew

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: OpenPGP digital signature
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20200123/899e818a/attachment.sig>

Hi Andrew,

What tools are you passing this triple to, and where did you get the list of
features you showed? (looks like an MC level list to me)

Depending on the tool you'll get different levels of expansion of the
features, "clang -###" is the most minimal then some things are
expanded later in llvm-mc. (where is somewhat inconsistent)

David Spickett.
________________________________
From: llvm-dev <llvm-dev-bounces at lists.llvm.org> on behalf of Andrew
Kelley via llvm-dev <llvm-dev at lists.llvm.org>
Sent: 23 January 2020 06:49
To: LLVM Dev <llvm-dev at lists.llvm.org>
Subject: [llvm-dev] How to find out the default CPU / Features String for a
given triple?

When I pass an empty string for cpu and features to createTargetMachine,
and then use LLVMGetTargetMachineCPU() and
LLVMGetTargetMachineFeatureString() to get the strings back, they are
still empty. Is there a way to have llvm compute the effective
cpu/features string, and provide it so that I can inspect it?

I'm trying to figure out how the cpu/features string that I am
explicitly passing, which I am expecting to be equivalent to passing
empty string, is different.

As an example, I have a test passing in the CI using the triple
"aarch64v8.1a-unknown-linux-unknown", and "" for both CPU
name and
target features string. However when I pass the following target
specific features string, I get qemu crashing in the CI:

-a35,-a53,-a55,-a57,-a72,-a73,-a75,-a76,-aes,-aggressive-fma,-alternate-sextload-cvt-f32-pattern,+altnzcv,+am,-arith-bcc-fusion,-arith-cbz-fusion,-balance-fp-ops,+bti,-call-saved-x10,-call-saved-x11,-call-saved-x12,-call-saved-x13,-call-saved-x14,-call-saved-x15,-call-saved-x18,-call-saved-x8,-call-saved-x9,+ccdp,+ccidx,+ccpp,+complxnum,+crc,-crypto,-custom-cheap-as-move,-cyclone,-disable-latency-sched-heuristic,+dit,+dotprod,-exynos-cheap-as-move,-exynosm1,-exynosm2,-exynosm3,-exynosm4,-falkor,+fmi,-force-32bit-jump-tables,+fp-armv8,-fp16fml,+fptoint,-fullfp16,-fuse-address,+fuse-aes,-fuse-arith-logic,-fuse-crypto-eor,-fuse-csel,-fuse-literals,+jsconv,-kryo,+lor,+lse,-lsl-fast,+mpam,-mte,+neon,-no-neg-immediates,+nv,+pa,+pan,+pan-rwv,+perfmon,-predictable-select-expensive,+predres,-rand,+ras,+rasv8_4,+rcpc,+rcpc-immo,+rdm,-reserve-x1,-reserve-x10,-reserve-x11,-reserve-x12,-reserve-x13,-reserve-x14,-reserve-x15,-reserve-x18,-reserve-x2,-reserve-x20,-reserve-x21,-reserve-x22,-reserve-x23,-reserve-x24,-reserve-x25,-reserve-x26,-reserve-x27,-reserve-x28,-reserve-x3,-reserve-x4,-reserve-x5,-reserve-x6,-reserve-x7,-reserve-x9,-saphira,+sb,+sel2,-sha2,-sha3,-slow-misaligned-128store,-slow-paired-128,-slow-strqro-store,-sm4,-spe,+specrestrict,+ssbs,-strict-align,-sve,-sve2,-sve2-aes,-sve2-bitperm,-sve2-sha3,-sve2-sm4,-thunderx,-thunderx2t99,-thunderxt81,-thunderxt83,-thunderxt88,+tlb-rmi,-tpidr-el1,-tpidr-el2,-tpidr-el3,+tracev8.4,-tsv110,+uaops,-use-aa,+use-postra-scheduler,-use-reciprocal-square-root,+v8.1a,+v8.2a,+v8.3a,+v8.4a,+v8.5a,+vh,-zcm,-zcz,-zcz-fp,-zcz-fp-workaround,-zcz-gp

Now, I understand that qemu crashing can be fixed by using a newer qemu
version. And, indeed, on my laptop with a newer qemu, the test passes.
However, this test was previously passing on the CI, which means that
with this explicit string of target specific features, *something is
different*. I am trying to determine what that different thing is.

Thanks in advance for the help,
Andrew

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20200123/14e30846/attachment.html>

On Thu, 23 Jan 2020 at 06:49, Andrew Kelley via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
> When I pass an empty string for cpu and features to createTargetMachine,
> and then use LLVMGetTargetMachineCPU() and
> LLVMGetTargetMachineFeatureString() to get the strings back, they are
> still empty. Is there a way to have llvm compute the effective
> cpu/features string, and provide it so that I can inspect it?
Hi Andrew,

These are the supported features for AArch64:
https://github.com/llvm/llvm-project/blob/master/llvm/include/llvm/Support/AArch64TargetParser.def

> As an example, I have a test passing in the CI using the triple
> "aarch64v8.1a-unknown-linux-unknown", and "" for both
CPU name and
> target features string.
"aarch64v8.1a" doesn't exist.

Target and Arch are two different things. The target is for example
"aarch64-linux-gnu" while arch (not CPU) is "armv8.1-a" (not
"v8.1a").

CPU are their literal names, like "cortex-a53".

IIRC, we allow adding modifiers like "+sve" to CPU for compatibility,
but they should be used when defining the arch, not CPU.

Unfortunately, we don't have control over the triples and arches and
have to support what vendors and prior tools do.

> However when I pass the following target
> specific features string, I get qemu crashing in the CI:
That error message has a lot more than what's accepted for
target/arch/cpu/flags. There are a lot of target specific flags in
there, which I'm not even sure they would parse in a target string at
all.

Unfortunatelly, without knowing how QEMU is using that or where the
error comes from, it's hard to tell what that means.

> Now, I understand that qemu crashing can be fixed by using a newer qemu
> version. And, indeed, on my laptop with a newer qemu, the test passes.
> However, this test was previously passing on the CI, which means that
> with this explicit string of target specific features, *something is
> different*. I am trying to determine what that different thing is.
That string is invalid. It's possible either QEMU or LLVM supported it
by either error or legacy and that's no longer true?

I'm surprised a newer QEMU somehow understands that, but without
understanding your infrastructure, what the test is and how it's
built, where does LLVM fit into all that, etc. it's hard to have any
idea what's going on.

cheers,
--renato

On 1/23/20 5:04 AM, David Spickett wrote:
> Hi Andrew,
> 
> What tools are you passing this triple to, and where did you get the
> list of features you showed? (looks like an MC level list to me)
include/llvm/Support/TargetRegistry.h

  /// createTargetMachine - Create a target specific machine implementation
  /// for the specified \p Triple.
  ///
  /// \param TT This argument is used to determine the target machine
  /// feature set; it should always be provided. Generally this should be
  /// either the target triple from the module, or the target triple of the
  /// host if that does not exist.
  TargetMachine *createTargetMachine(StringRef TT, StringRef CPU,
                            StringRef Features,
                            const TargetOptions &Options,
                            Optional<Reloc::Model> RM,
                            Optional<CodeModel::Model> CM = None,
                            CodeGenOpt::Level OL = CodeGenOpt::Default,
                            bool JIT = false) const;



The list of features I showed comes from LLVM's
lib/Target/AArch64/AArch64.td file, processed into this file:
https://github.com/ziglang/zig/blob/c86589a738e5053a26b2be7d156bcfee1565f00b/lib/std/target/aarch64.zig

Next, the "generic" CPU is chosen, the feature corresponding to the
sub-arch (v8_5a) is added to the set given by the CPU, and then the set
is recursively populated with feature dependencies.

I'm trying to find out how to check my work here, and make sure the list
of features matches what LLVM chooses when empty strings are passed for
CPU and Features.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: OpenPGP digital signature
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20200123/196c20da/attachment.sig>