Skip to content

[Mono]: Reduce Mono AOT cross compiler x64 memory footprint. #97096

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
steveisok merged 3 commits into from
Feb 9, 2024
Merged

[Mono]: Reduce Mono AOT cross compiler x64 memory footprint. #97096

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f090a34
Reduce Mono AOT cross compiler x64 memory footprint.
lateralusX Jan 17, 2024
cfaf8d9
Fix review feedback.
lateralusX Jan 22, 2024
5ab0d94
Minor cleanups.
vargaz Feb 9, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
30 changes: 20 additions & 10 deletions src/libraries/System.Private.CoreLib/src/System.Private.CoreLib.Shared.projitems
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2576,17 +2576,26 @@
</ItemGroup>
<ItemGroup Condition="'$(SupportsX86Intrinsics)' == 'true'">
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Aes.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx2.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx512BW.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx512CD.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx512DQ.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx512F.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx512Vbmi.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\AvxVnni.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx.cs" Condition="'$(FeatureMono)' != 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx.PlatformNotSupported.cs" Condition="'$(FeatureMono)' == 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx2.cs" Condition="'$(FeatureMono)' != 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx2.PlatformNotSupported.cs" Condition="'$(FeatureMono)' == 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx512BW.cs" Condition="'$(FeatureMono)' != 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx512BW.PlatformNotSupported.cs" Condition="'$(FeatureMono)' == 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx512CD.cs" Condition="'$(FeatureMono)' != 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx512CD.PlatformNotSupported.cs" Condition="'$(FeatureMono)' == 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx512DQ.cs" Condition="'$(FeatureMono)' != 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx512DQ.PlatformNotSupported.cs" Condition="'$(FeatureMono)' == 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx512F.cs" Condition="'$(FeatureMono)' != 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx512F.PlatformNotSupported.cs" Condition="'$(FeatureMono)' == 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx512Vbmi.cs" Condition="'$(FeatureMono)' != 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Avx512Vbmi.PlatformNotSupported.cs" Condition="'$(FeatureMono)' == 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\AvxVnni.cs" Condition="'$(FeatureMono)' != 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\AvxVnni.PlatformNotSupported.cs" Condition="'$(FeatureMono)' == 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Bmi1.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Bmi2.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Fma.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Fma.cs" Condition="'$(FeatureMono)' != 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Fma.PlatformNotSupported.cs" Condition="'$(FeatureMono)' == 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Lzcnt.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Pclmulqdq.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Popcnt.cs" />
Expand All @@ -2597,7 +2606,8 @@
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Sse42.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Ssse3.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\X86Base.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\X86Serialize.cs" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\X86Serialize.cs" Condition="'$(FeatureMono)' != 'true'" />
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\X86Serialize.PlatformNotSupported.cs" Condition="'$(FeatureMono)' == 'true'" />
</ItemGroup>
<ItemGroup Condition="'$(SupportsX86Intrinsics)' != 'true'">
<Compile Include="$(MSBuildThisFileDirectory)System\Runtime\Intrinsics\X86\Aes.PlatformNotSupported.cs" />
Expand Down
2 changes: 2 additions & 0 deletions src/mono/System.Private.CoreLib/System.Private.CoreLib.csproj
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -151,6 +151,8 @@

<ILLinkSubstitutionsXmls Include="$(ILLinkDirectory)ILLink.Substitutions.$(Platform).xml" Condition="Exists('$(ILLinkDirectory)ILLink.Substitutions.$(Platform).xml')" />
<ILLinkSubstitutionsXmls Include="$(ILLinkDirectory)ILLink.Substitutions.iOS.xml" Condition="'$(TargetsiOS)' == 'true' or '$(TargetstvOS)' == 'true'" />
<ILLinkSubstitutionsXmls Include="$(ILLinkDirectory)ILLink.Substitutions.Intrinsics.x86.xml" />
<ILLinkSubstitutionsXmls Include="$(ILLinkDirectory)ILLink.Substitutions.Intrinsics.Vectors.xml" />

<ILLinkLinkAttributesXmls Include="$(ILLinkDirectory)ILLink.LinkAttributes.xml" />
<ILLinkLinkAttributesXmls Include="$(ILLinkDirectory)ILLink.LinkAttributes.$(Platform).xml" Condition="Exists('$(ILLinkDirectory)ILLink.LinkAttributes.$(Platform).xml')" />
Expand Down
10 changes: 10 additions & 0 deletions src/mono/System.Private.CoreLib/src/ILLink/ILLink.Substitutions.Intrinsics.Vectors.xml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,10 @@
<linker>
<assembly fullname="System.Private.CoreLib">
<type fullname="System.Runtime.Intrinsics.Vector256">
Copy link
Contributor

@marek-safar marek-safar Jan 17, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is this true for all codegens (e.g. interpreter)

Copy link
Member Author

@lateralusX lateralusX Jan 17, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looks like interpreter mark all vectors as not being hardware accelerated:

} else if (in_corlib &&
		   (!strncmp ("System.Runtime.Intrinsics", klass_name_space, 25) &&
			!strncmp ("Vector", klass_name, 6) &&
			!strcmp (tm, "get_IsHardwareAccelerated"))) {
	*op = MINT_LDC_I4_0;
}

so for that case it should be ok.

Copy link
Member

@lewing lewing Jan 17, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

cc @kg

Copy link
Member

@kg kg Jan 18, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We mark v128 as hardware accelerated in interp in some cases, I believe.

Copy link
Member Author

@lateralusX lateralusX Jan 22, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

ok, so then this should be OK since it only affects v256/v512.

<method signature="System.Boolean get_IsHardwareAccelerated()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.Vector512">
<method signature="System.Boolean get_IsHardwareAccelerated()" body="stub" value="false" />
</type>
</assembly>
</linker>
79 changes: 79 additions & 0 deletions src/mono/System.Private.CoreLib/src/ILLink/ILLink.Substitutions.Intrinsics.x86.xml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,79 @@
<linker>
<assembly fullname="System.Private.CoreLib">
<type fullname="System.Runtime.Intrinsics.X86.Avx">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx/X64">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx2">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx2/X64">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx512BW">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx512BW/VL">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx512BW/X64">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx512CD">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx512CD/VL">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx512CD/X64">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx512DQ">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx512DQ/VL">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx512DQ/X64">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx512F">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx512F/VL">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx512F/X64">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx512Vbmi">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx512Vbmi/VL">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Avx512Vbmi/X64">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.AvxVnni">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.AvxVnni/X64">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Fma">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.Fma/X64">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.X86Serialize">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
<type fullname="System.Runtime.Intrinsics.X86.X86Serialize/X64">
<method signature="System.Boolean get_IsSupported()" body="stub" value="false" />
</type>
</assembly>
</linker>
2 changes: 1 addition & 1 deletion src/mono/mono/mini/aot-compiler.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14229,7 +14229,7 @@ static void
acfg_free (MonoAotCompile *acfg)
{
#ifdef ENABLE_LLVM
if (acfg->aot_opts.llvm)
if (mono_use_llvm || acfg->aot_opts.llvm)
mono_llvm_free_aot_module ();
#endif

Expand Down
11 changes: 10 additions & 1 deletion src/mono/mono/mini/method-to-ir.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4746,6 +4746,15 @@ mini_inline_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *
return inline_method (cfg, cmethod, fsig, sp, ip, real_offset, inline_always, NULL);
}

static gboolean
aggressive_inline_method (MonoMethod *cmethod)
{
gboolean aggressive_inline = m_method_is_aggressive_inlining (cmethod);
if (aggressive_inline)
aggressive_inline = !mono_simd_unsupported_aggressive_inline_intrinsic_type (cmethod);
return aggressive_inline;
}

/*
* inline_method:
*
Expand Down Expand Up @@ -4871,7 +4880,7 @@ inline_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig,
cfg->disable_inline = prev_disable_inline;
cfg->inline_depth --;

if ((costs >= 0 && costs < 60) || inline_always || (costs >= 0 && (cmethod->iflags & METHOD_IMPL_ATTRIBUTE_AGGRESSIVE_INLINING))) {
if ((costs >= 0 && costs < 60) || inline_always || (costs >= 0 && aggressive_inline_method (cmethod))) {
if (cfg->verbose_level > 2)
printf ("INLINE END %s -> %s\n", mono_method_full_name (cfg->method, TRUE), mono_method_full_name (cmethod, TRUE));

Expand Down
1 change: 1 addition & 0 deletions src/mono/mono/mini/mini.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2962,6 +2962,7 @@ MonoInst* mono_emit_common_intrinsics (MonoCompile *cfg, MonoMethod *cmethod,
MonoInst* mono_emit_simd_intrinsics (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args);
MonoInst* mono_emit_simd_field_load (MonoCompile *cfg, MonoClassField *field, MonoInst *addr);
void mono_simd_intrinsics_init (void);
gboolean mono_simd_unsupported_aggressive_inline_intrinsic_type (MonoMethod *cmethod);

MonoMethod*
mini_method_to_shared (MonoMethod *method); // null if not shared
Expand Down
61 changes: 59 additions & 2 deletions src/mono/mono/mini/simd-intrinsics.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1178,6 +1178,20 @@ create_class_instance (const char* name_space, const char *name, MonoType *param
return ivector_inst;
}

static gboolean
is_supported_vector_primitive_type (MonoType *type)
{
gboolean constrained_generic_param = (type->type == MONO_TYPE_VAR || type->type == MONO_TYPE_MVAR);

if (constrained_generic_param && type->data.generic_param->gshared_constraint && MONO_TYPE_IS_VECTOR_PRIMITIVE (type->data.generic_param->gshared_constraint))
return TRUE;

if (MONO_TYPE_IS_VECTOR_PRIMITIVE (type))
return TRUE;

return FALSE;
}

static guint16 sri_vector_methods [] = {
SN_Abs,
SN_Add,
Expand Down Expand Up @@ -1423,8 +1437,8 @@ emit_sri_vector (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsi
return NULL;

if (vector_size == 256 || vector_size == 512)
return NULL;
return NULL;

// FIXME: This limitation could be removed once everything here are supported by mini JIT on arm64
#ifdef TARGET_ARM64
if (!COMPILE_LLVM (cfg)) {
Expand Down Expand Up @@ -2477,6 +2491,12 @@ emit_sri_vector_t (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *f
g_free (name);
}

if (id == SN_get_IsSupported) {
MonoInst *ins;
EMIT_NEW_ICONST (cfg, ins, is_supported_vector_primitive_type (etype) ? 1 : 0);
return ins;
}

// Apart from filtering out non-primitive types this also filters out shared generic instance types like: T_BYTE which cannot be intrinsified
if (!MONO_TYPE_IS_VECTOR_PRIMITIVE (etype)) {
// Happens often in gshared code
Expand Down Expand Up @@ -3199,6 +3219,11 @@ emit_sys_numerics_vector_t (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSig
type = m_class_get_byval_arg (klass);
etype = mono_class_get_context (klass)->class_inst->type_argv [0];

if (id == SN_get_IsSupported) {
EMIT_NEW_ICONST (cfg, ins, is_supported_vector_primitive_type (etype) ? 1 : 0);
return ins;
}

if (!MONO_TYPE_IS_VECTOR_PRIMITIVE (etype))
return NULL;

Expand Down Expand Up @@ -6118,11 +6143,37 @@ mono_simd_decompose_intrinsic (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst *i
decompose_vtype_opt_store_arg (cfg, bb, ins, &(ins->dreg));
}
}

gboolean
mono_simd_unsupported_aggressive_inline_intrinsic_type (MonoMethod *cmethod)
{
/*
* If a method has been marked with aggressive inlining, check if we support
* aggressive inlining of the intrinsics type, if not, ignore aggressive inlining
* since it could end up inlining a large amount of code that most likely will end
* up as dead code.
*/
Comment on lines +6150 to +6155
Copy link
Member

@tannergooding tannergooding Jan 17, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you help me understand why this is needed on Mono a bit?

The software fallback for V128/V256/V512 is currently implemented as doing 2x operations on the lower/upper halves (except for a couple of methods like Shuffle which operate on the full vector). So V512 is implemented as 2x V256, V256 is implemented as 2x V128, and V128 is implemented as 2x V64. V64 is then implemented as a loop over the scalar elements with potentially large amounts of generic code.

So I would imagine that on any platform where V128 is supported, that the codegen for V256/V512 should be generally nice/small, even if aggressively inlined and with no real dead code elimination required. Even in the case where you have an unsupported type like V256<Guid>, the first V128 operation should have a PNSE thrown (since whether a type is supported or not is typically a known constant).

So I'd only expect that this is needed for V64 on x86/x64 where there is no acceleration and it has to hit the scalar fallback with the loop over the elements. For RyuJIT this loop isn't an issue due to the generic specialization we do, allowing us to get it down to only the code path that is actually used. I believe this isn't possible for Mono today and is non-trivial to add, so the skipping of inlining does make sense in that regard.

Copy link
Member Author

@lateralusX lateralusX Jan 17, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So without disabling aggressive inlining I ended up with methods that where huge, and not doing aggressive inline on these types (but still honor inline size limits) made them sane again. I will need to re-iterate around that change in order to tell exactly what happened with our inliner in the aggressive case.

Copy link
Member Author

@lateralusX lateralusX Jan 22, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Just validated this, with disabling the aggressive inlining as above a .net9 full AOT of S.P.C takes ~1.7 GB of memory, and not doing this will consume an additional 600 MB of memory, so I believe its worth preventing aggressive inlining for these types that are not hardware accelerated on any of the Mono supported platforms. I didn't add V64 since it seems to be handled a little differently on at least ARM case. I won't have bandwidth at the moment to do more deep analysis around why aggressive inlining of these template types cause that large increase in memory so I think doing this change, at least short term is worth it, we could probably file an issue around the bloat of Vector2561 and Vector5121, but maybe the better long term solution is to actually implement intrinsic support for these types.

Copy link
Member

@BrzVlad BrzVlad Jan 25, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

For interpreter I'm handling this in a general fashion by detecting early dead pieces of code and not inlining any of the calls there. #97514. It is possible that a similar approach for jit can produce further improvements without having to special case classes.

Copy link
Member Author

@lateralusX lateralusX Jan 25, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I did try to experiment with some dead code elimination, but that cause issues and doesn't work with our llvm codegen, so we explicitly turn that pass off for code that will be passed over to llvm. Instead I made sure we could do more in linker, but still these methods still explode and survives first simple llvm optimizations pass we now do in function manager pass, so feels like something in the inlined code prevents elimination until very late in the opt chain.

Copy link
Contributor

@vargaz vargaz Jan 25, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Trying to reenable branch optimizations when using llvm in this pr:
#97189

if (!strcmp (m_class_get_name_space (cmethod->klass), "System.Runtime.Intrinsics")) {
if (!strncmp(m_class_get_name (cmethod->klass), "Vector", 6)) {
const char *vector_type = m_class_get_name (cmethod->klass) + 6;
if (!strcmp(vector_type, "256`1") || !strcmp(vector_type, "512`1"))
return TRUE;
}
}
return FALSE;
}
#else
void
mono_simd_decompose_intrinsic (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst *ins)
{
}

gboolean
mono_simd_unsupported_aggressive_inline_intrinsic_type (MonoMethod* cmethod)
{
return FALSE;
}

#endif /*defined(TARGET_WIN32) && defined(TARGET_AMD64)*/

#endif /* DISABLE_JIT */
Expand Down Expand Up @@ -6157,6 +6208,12 @@ mono_simd_decompose_intrinsic (MonoCompile *cfg, MonoBasicBlock *bb, MonoInst *i
{
}

gboolean
mono_simd_unsupported_aggressive_inline_intrinsic_type (MonoMethod* cmethod)
{
return FALSE;
}

#endif /* MONO_ARCH_SIMD_INTRINSICS */

#if defined(TARGET_AMD64)
Expand Down