[PATCH] D107850: [asan] Implemented intrinsic for the custom calling convention similar used by HWASan for X86.

[Vitaly Buka via Phabricator via llvm-commits]

vitalybuka added inline comments.


================
Comment at: llvm/include/llvm/IR/Intrinsics.td:1640
 
+def int_asan_check_memaccess :
+  Intrinsic<[],[llvm_ptr_ty, llvm_i8_ty, llvm_i8_ty],
----------------
eugenis wrote:
> vitalybuka wrote:
> > pcc wrote:
> > > eugenis wrote:
> > > > vitalybuka wrote:
> > > > > kstoimenov wrote:
> > > > > > vitalybuka wrote:
> > > > > > > vitalybuka wrote:
> > > > > > > > PTAL at lvm.read_register.i32
> > > > > > > > 
> > > > > > > > How about:
> > > > > > > > 
> > > > > > > > llvm.asan.check.memaccess ->
> > > > > > > >   lvm.asan.check_read
> > > > > > > >   lvm.asan.check_write
> > > > > > > >   lvm.asan.kernel.check_read
> > > > > > > >   lvm.asan.kernel.check_write
> > > > > > > > 
> > > > > > > > Even better
> > > > > > > >   lvm.asan.check_read.{i8, i16, i32, ...}
> > > > > > > >   lvm.asan.check_write.{i8, i16, i32, ...}
> > > > > > > >   lvm.asan.kernel.check_read.{i8, i16, i32, ...}
> > > > > > > >   lvm.asan.kernel.check_write.{i8, i16, i32, ...}
> > > > > > > > 
> > > > > > > Looks like underscore is not used in intrinsic names, so essentially the same with dots.
> > > > > > Sounds good to me. I do the full expansion so there will be 20 intrinsics altogether. I will update the code and ping you when done. 
> > > > > @pcc @eugenis 
> > > > > WDYT, I think later we can do the same for HWASAN?
> > > > I don't see what these multiple intrinsics give us that a single memaccess one does not provide?
> > > > 
> > > > As long as access type and similar arguments are immediates.
> > > > 
> > > Agree with @eugenis - these sorts of intrinsic variants are typically used for distinguishing different pointer element types and we're in the process of getting rid of those anyway.
> > @pcc @eugenis Then do you prefer to encode is_write+size+kernel into non-human unreadable AccessInfo, like hwasan, or separate 0/1 arguments.
> > I probably prefer AccessInfo, as they both unreadable, but the hwasan version is shorter.
> don't have a strong opinion, but sometimes I wish that hwasan outlined function names were more readable. The magic number in the names takes effort to decode.
> 
AccessInfo


================
Comment at: llvm/include/llvm/IR/Intrinsics.td:1640
 
+def int_asan_check_memaccess :
+  Intrinsic<[],[llvm_ptr_ty, llvm_i8_ty, llvm_i8_ty],
----------------
pcc wrote:
> kstoimenov wrote:
> > vitalybuka wrote:
> > > eugenis wrote:
> > > > vitalybuka wrote:
> > > > > pcc wrote:
> > > > > > eugenis wrote:
> > > > > > > vitalybuka wrote:
> > > > > > > > kstoimenov wrote:
> > > > > > > > > vitalybuka wrote:
> > > > > > > > > > vitalybuka wrote:
> > > > > > > > > > > PTAL at lvm.read_register.i32
> > > > > > > > > > > 
> > > > > > > > > > > How about:
> > > > > > > > > > > 
> > > > > > > > > > > llvm.asan.check.memaccess ->
> > > > > > > > > > >   lvm.asan.check_read
> > > > > > > > > > >   lvm.asan.check_write
> > > > > > > > > > >   lvm.asan.kernel.check_read
> > > > > > > > > > >   lvm.asan.kernel.check_write
> > > > > > > > > > > 
> > > > > > > > > > > Even better
> > > > > > > > > > >   lvm.asan.check_read.{i8, i16, i32, ...}
> > > > > > > > > > >   lvm.asan.check_write.{i8, i16, i32, ...}
> > > > > > > > > > >   lvm.asan.kernel.check_read.{i8, i16, i32, ...}
> > > > > > > > > > >   lvm.asan.kernel.check_write.{i8, i16, i32, ...}
> > > > > > > > > > > 
> > > > > > > > > > Looks like underscore is not used in intrinsic names, so essentially the same with dots.
> > > > > > > > > Sounds good to me. I do the full expansion so there will be 20 intrinsics altogether. I will update the code and ping you when done. 
> > > > > > > > @pcc @eugenis 
> > > > > > > > WDYT, I think later we can do the same for HWASAN?
> > > > > > > I don't see what these multiple intrinsics give us that a single memaccess one does not provide?
> > > > > > > 
> > > > > > > As long as access type and similar arguments are immediates.
> > > > > > > 
> > > > > > Agree with @eugenis - these sorts of intrinsic variants are typically used for distinguishing different pointer element types and we're in the process of getting rid of those anyway.
> > > > > @pcc @eugenis Then do you prefer to encode is_write+size+kernel into non-human unreadable AccessInfo, like hwasan, or separate 0/1 arguments.
> > > > > I probably prefer AccessInfo, as they both unreadable, but the hwasan version is shorter.
> > > > don't have a strong opinion, but sometimes I wish that hwasan outlined function names were more readable. The magic number in the names takes effort to decode.
> > > > 
> > > AccessInfo
> > I think I am gonna go with int_asan_check(?_kernel)_(load|store) and pass the size as parameter. What do you think? 
> You mean in a register? I think that could mean more register pressure -> higher code size.
> 
> The magic numbers are unfortunate but they aren't that hard to decode (maybe we should be printing them as hex to make it a bit easier). I suppose we could pretty print the access info into the symbol name but only a few people will be looking at these so I'm not sure it's worth the effort.
> I think I am gonna go with int_asan_check(?_kernel)_(load|store) and pass the size as parameter. What do you think? 
Looks like both @pcc and @eugenis do not like multiple intrinsic.

Also I chatted with @eugenis to clarify the last comment. He is OK with AccessInfo like in hwasan.




================
Comment at: llvm/include/llvm/IR/Intrinsics.td:1640
 
+def int_asan_check_memaccess :
+  Intrinsic<[],[llvm_ptr_ty, llvm_i8_ty, llvm_i8_ty],
----------------
vitalybuka wrote:
> pcc wrote:
> > kstoimenov wrote:
> > > vitalybuka wrote:
> > > > eugenis wrote:
> > > > > vitalybuka wrote:
> > > > > > pcc wrote:
> > > > > > > eugenis wrote:
> > > > > > > > vitalybuka wrote:
> > > > > > > > > kstoimenov wrote:
> > > > > > > > > > vitalybuka wrote:
> > > > > > > > > > > vitalybuka wrote:
> > > > > > > > > > > > PTAL at lvm.read_register.i32
> > > > > > > > > > > > 
> > > > > > > > > > > > How about:
> > > > > > > > > > > > 
> > > > > > > > > > > > llvm.asan.check.memaccess ->
> > > > > > > > > > > >   lvm.asan.check_read
> > > > > > > > > > > >   lvm.asan.check_write
> > > > > > > > > > > >   lvm.asan.kernel.check_read
> > > > > > > > > > > >   lvm.asan.kernel.check_write
> > > > > > > > > > > > 
> > > > > > > > > > > > Even better
> > > > > > > > > > > >   lvm.asan.check_read.{i8, i16, i32, ...}
> > > > > > > > > > > >   lvm.asan.check_write.{i8, i16, i32, ...}
> > > > > > > > > > > >   lvm.asan.kernel.check_read.{i8, i16, i32, ...}
> > > > > > > > > > > >   lvm.asan.kernel.check_write.{i8, i16, i32, ...}
> > > > > > > > > > > > 
> > > > > > > > > > > Looks like underscore is not used in intrinsic names, so essentially the same with dots.
> > > > > > > > > > Sounds good to me. I do the full expansion so there will be 20 intrinsics altogether. I will update the code and ping you when done. 
> > > > > > > > > @pcc @eugenis 
> > > > > > > > > WDYT, I think later we can do the same for HWASAN?
> > > > > > > > I don't see what these multiple intrinsics give us that a single memaccess one does not provide?
> > > > > > > > 
> > > > > > > > As long as access type and similar arguments are immediates.
> > > > > > > > 
> > > > > > > Agree with @eugenis - these sorts of intrinsic variants are typically used for distinguishing different pointer element types and we're in the process of getting rid of those anyway.
> > > > > > @pcc @eugenis Then do you prefer to encode is_write+size+kernel into non-human unreadable AccessInfo, like hwasan, or separate 0/1 arguments.
> > > > > > I probably prefer AccessInfo, as they both unreadable, but the hwasan version is shorter.
> > > > > don't have a strong opinion, but sometimes I wish that hwasan outlined function names were more readable. The magic number in the names takes effort to decode.
> > > > > 
> > > > AccessInfo
> > > I think I am gonna go with int_asan_check(?_kernel)_(load|store) and pass the size as parameter. What do you think? 
> > You mean in a register? I think that could mean more register pressure -> higher code size.
> > 
> > The magic numbers are unfortunate but they aren't that hard to decode (maybe we should be printing them as hex to make it a bit easier). I suppose we could pretty print the access info into the symbol name but only a few people will be looking at these so I'm not sure it's worth the effort.
> > I think I am gonna go with int_asan_check(?_kernel)_(load|store) and pass the size as parameter. What do you think? 
> Looks like both @pcc and @eugenis do not like multiple intrinsic.
> 
> Also I chatted with @eugenis to clarify the last comment. He is OK with AccessInfo like in hwasan.
> 
> 
> I think I am gonna go with int_asan_check(?_kernel)_(load|store) and pass the size as parameter. What do you think? 




Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D107850/new/

https://reviews.llvm.org/D107850