[dfsan] Fix Endianess issue

[anoopkg6]

Fix Endianess issue with getting shadow 4 bytes corresponding to the first origin pointer.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

@llvm/pr-subscribers-compiler-rt-sanitizer

Author: None (anoopkg6)

Changes

Fix Endianess issue with getting shadow 4 bytes corresponding to the first origin pointer.

---

Full diff: https://github.com/llvm/llvm-project/pull/162881.diff

1 Files Affected:

• (modified) llvm/lib/Transforms/Instrumentation/DataFlowSanitizer.cpp (+8-2)

diff --git a/llvm/lib/Transforms/Instrumentation/DataFlowSanitizer.cpp b/llvm/lib/Transforms/Instrumentation/DataFlowSanitizer.cpp
index 5ba2167859490..b4f88779b00c0 100644
--- a/llvm/lib/Transforms/Instrumentation/DataFlowSanitizer.cpp
+++ b/llvm/lib/Transforms/Instrumentation/DataFlowSanitizer.cpp
@@ -2187,8 +2187,14 @@ std::pair<Value *, Value *> DFSanFunction::loadShadowFast(
       // and then the entire shadow for the second origin pointer (which will be
       // chosen by combineOrigins() iff the least-significant half of the wide
       // shadow was empty but the other half was not).
-      Value *WideShadowLo = IRB.CreateShl(
-          WideShadow, ConstantInt::get(WideShadowTy, WideShadowBitWidth / 2));
+      Value *WideShadowLo =
+          F->getParent()->getDataLayout().isLittleEndian()
+              ? IRB.CreateShl(
+                    WideShadow,
+                    ConstantInt::get(WideShadowTy, WideShadowBitWidth / 2))
+              : IRB.CreateAnd(
+                    WideShadow,
+                    ConstantInt::get(WideShadowTy, 0xFFFFFFFF00000000ULL));
       Shadows.push_back(WideShadow);
       Origins.push_back(DFS.loadNextOrigin(Pos, OriginAlign, &OriginAddr));
 

[thurstond]

Please add "[dfsan]" to the title

[thurstond]

Can this be rewritten in terms of WideShadowBitWidth?

(I think it's something like (1 - (1 << (WideShadowBitWidth / 2))) << (WideShadowBitWidth / 2))

[anoopkg6]

Right way of writing this expression. Thanks.

[vitalybuka]

I don't believe we support other that little endian. So there is no point to complicate code for that.

[thurstond]

I don't believe we support other that little endian. So there is no point to complicate code for that.

They are adding support for SystemZ (#162195).

[thurstond]

Also, RISC-V has an exciting big-endian extension (/cc @fmayer)

[vitalybuka]

I don't believe we support other that little endian. So there is no point to complicate code for that.

They are adding support for SystemZ (#162195).

Is #162195 enough?
I'd expect sanitizers does a lot of assumptions on little-endianness.

[anoopkg6]

Dfsan tests passed, which include extensive custom library functions like memcpy, memmove ... etc. Other sanitizers - asan,lsan, tsan, ubsan - have successfully been supported on SystemZ. Tysan, rtsan have already been tested and are in the process of pr. If some issue crops up, it will be fixed.

…

________________________________ From: Vitaly Buka ***@***.***> Sent: Friday, October 10, 2025 3:05 PM To: llvm/llvm-project ***@***.***> Cc: Anoop Kumar ***@***.***>; Author ***@***.***> Subject: [EXTERNAL] Re: [llvm/llvm-project] Fix Endianess issue (PR #162881) vitalybuka left a comment (llvm/llvm-project#162881) I don't believe we support other that little endian. So there is no point to complicate code for that. They are adding support for SystemZ (#162195). Is #162195 enough? I'd expect sanitizers [https://avatars.githubusercontent.com/u/1531415?s=20&v=4%5Dvitalybuka left a comment (llvm/llvm-project#162881)<#162881 (comment) > I don't believe we support other that little endian. So there is no point to complicate code for that. They are adding support for SystemZ (#162195<#162195 >). Is #162195<#162195 > enough? I'd expect sanitizers does a lot of assumptions on little-endianness. — Reply to this email directly, view it on GitHub<#162881 (comment) >, or unsubscribe<https://github.com/notifications/unsubscribe-auth/BM5K4GXOGNHB3NB43QBTZFL3XAGQBAVCNFSM6AAAAACI3L4A5CVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTGOJSGE3DENZQGM >. You are receiving this because you authored the thread.Message ID: ***@***.***>

[vitalybuka]

Can you please add a test?

[anoopkg6]

Added origin_endianness.c. This is a reduced test from existing test origin_ldst.c in dfsan testsuite. There are several tests in origin_ldst.c for combinations of partial load/store. Reduced test in origin_endianness.c is the only test case that
fails on SystemZ. Added new test origin_endianness.c does not test anything new than existing dfsan tests. It only highlights the issue of endianness.

[thurstond]

Reduced test in origin_endianness.c is the only test case that
fails on SystemZ

Does the test case need an XFAIL for SystemZ? Otherwise it will break SystemZ buildbots/anyone running tests on SystemZ.

[anoopkg6]

I think XFAIL is not needed for this test. This is a subset of already existing test origin_ldst.c in dfsan tests. Moreover, DFSan has not yet been enabled for SystemZ.

[thurstond]

I think XFAIL is not needed for this test. This is a subset of already existing test origin_ldst.c in dfsan tests. Moreover, DFSan has not yet been enabled for SystemZ.

Oh right, thanks!

[anoopkg6]

@uweigand