[VPlan] Use pragma pack(1) for VPIRFlags on AIX.

[fhahn]

Some compilers (e.g. on AIX) do not pack by default. Use LLVM_PACKED to ensure the VPIRFlags struct is packed as expected on all platforms.

This matches what we already do in other places for AIX, e.g. in llvm/include/llvm/CodeGen/SelectionDAGNodes.h (added in 844a02e), although it uses the more general LLVM_PACKED unconditionally

[llvmbot]

@llvm/pr-subscribers-vectorizers

Author: Florian Hahn (fhahn)

Changes

Except for GCC; by default, AIX compilers store bit-fields in 4-byte words. This means that VPIRFlags is larger than expected.

Use pack(1) to ensure smaller packing, matching other platforms.

This matches what we already do in other places for AIX, e.g. in llvm/include/llvm/CodeGen/SelectionDAGNodes.h (added in 844a02e)

---

Full diff: https://github.com/llvm/llvm-project/pull/184687.diff

1 Files Affected:

• (modified) llvm/lib/Transforms/Vectorize/VPlan.h (+13)

diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index 41eef2a368343..d2ea1b362ccba 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -668,6 +668,16 @@ class VPSingleDefRecipe : public VPRecipeBase, public VPRecipeValue {
 };
 
 /// Class to record and manage LLVM IR flags.
+#if defined(_AIX) && (!defined(__GNUC__) || defined(__clang__))
+// Except for GCC; by default, AIX compilers store bit-fields in 4-byte words
+// and give the `pack` pragma push semantics.
+#define BEGIN_ONE_BYTE_PACK() _Pragma("pack(1)")
+#define END_ONE_BYTE_PACK() _Pragma("pack(pop)")
+#else
+#define BEGIN_ONE_BYTE_PACK()
+#define END_ONE_BYTE_PACK()
+#endif
+BEGIN_ONE_BYTE_PACK()
 class VPIRFlags {
   enum class OperationType : unsigned char {
     Cmp,
@@ -1070,6 +1080,9 @@ class VPIRFlags {
   void printFlags(raw_ostream &O) const;
 #endif
 };
+END_ONE_BYTE_PACK()
+#undef BEGIN_ONE_BYTE_PACK
+#undef END_ONE_BYTE_PACK
 
 static_assert(sizeof(VPIRFlags) <= 3, "VPIRFlags should not grow");
 

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Florian Hahn (fhahn)

Changes

Except for GCC; by default, AIX compilers store bit-fields in 4-byte words. This means that VPIRFlags is larger than expected.

Use pack(1) to ensure smaller packing, matching other platforms.

This matches what we already do in other places for AIX, e.g. in llvm/include/llvm/CodeGen/SelectionDAGNodes.h (added in 844a02e)

---

Full diff: https://github.com/llvm/llvm-project/pull/184687.diff

1 Files Affected:

• (modified) llvm/lib/Transforms/Vectorize/VPlan.h (+13)

diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index 41eef2a368343..d2ea1b362ccba 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -668,6 +668,16 @@ class VPSingleDefRecipe : public VPRecipeBase, public VPRecipeValue {
 };
 
 /// Class to record and manage LLVM IR flags.
+#if defined(_AIX) && (!defined(__GNUC__) || defined(__clang__))
+// Except for GCC; by default, AIX compilers store bit-fields in 4-byte words
+// and give the `pack` pragma push semantics.
+#define BEGIN_ONE_BYTE_PACK() _Pragma("pack(1)")
+#define END_ONE_BYTE_PACK() _Pragma("pack(pop)")
+#else
+#define BEGIN_ONE_BYTE_PACK()
+#define END_ONE_BYTE_PACK()
+#endif
+BEGIN_ONE_BYTE_PACK()
 class VPIRFlags {
   enum class OperationType : unsigned char {
     Cmp,
@@ -1070,6 +1080,9 @@ class VPIRFlags {
   void printFlags(raw_ostream &O) const;
 #endif
 };
+END_ONE_BYTE_PACK()
+#undef BEGIN_ONE_BYTE_PACK
+#undef END_ONE_BYTE_PACK
 
 static_assert(sizeof(VPIRFlags) <= 3, "VPIRFlags should not grow");
 

[fhahn]

@kkwli @hubert-reinterpretcast I think this should fix the build failure mentioned in #181571 for AIX, but I am not able to verify locally. Would be great if you could help

[midhuncodes7]

I added this fix to my branch to test on AIX but ended up with different errors:

/home/midhun7/llvm/forked/llvm-project/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp:151:10: error: no member named 'setStartIndex' in 'llvm::VPScalarIVStepsRecipe'
  151 |   Steps->setStartIndex(StartIndex);
      |   ~~~~~  ^
/home/midhun7/llvm/forked/llvm-project/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp:615:37: error: no member named 'getInductionOpcode' in 'llvm::VPScalarIVStepsRecipe'
  615 |           if (!StartIndex && Steps->getInductionOpcode() == Instruction::FSub)
      |                              ~~~~~  ^
/home/midhun7/llvm/forked/llvm-project/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp:619:26: error: no member named 'getInductionOpcode' in 'llvm::VPScalarIVStepsRecipe'
  619 |           AddOp = Steps->getInductionOpcode();
      |                   ~~~~~  ^
/home/midhun7/llvm/forked/llvm-project/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp:635:16: error: no member named 'setStartIndex' in 'llvm::VPScalarIVStepsRecipe'
  635 |         Steps->setStartIndex(LaneOffset);
      |         ~~~~~  ^
4 errors generated.

On commit:

commit d316fb0797045d6f85b237aac18fabbed42e5d09 (HEAD -> main, origin/main, origin/HEAD)
Author: Florian Hahn <flo@fhahn.com>
Date:   Thu Mar 5 12:42:20 2026 +0000

    [VPlan] Replicate VPScalarIVStepsRecipe by VF outside replicate regions. (#170053)

[fhahn]

@midhuncodes7 is this with a clean build? So far looks like this commit is fine on the build bots.

[midhuncodes7]

The changes introduced in this PR: #170053 along with the fix provided in the current PR was built clean and results with the posted error.

[kkwli]

LG. I can successfully build the patch on AIX without encountering the previously described error.

[fhahn]

LG. I can successfully build the patch on AIX without encountering the previously described error.

@kkwli thanks for checking, I updated the code to just use LLVM_PACKED_START/END, so this is now done unconditionally

[artagnon]

If this fixes the issue, well and good. LGTM, thanks!