[InferAlignment] Fix updating alignment when larger than i32

[jhuber6]

Summary:
The changes made in #156057
allows the alignment value to be increased. We assert effectively
infinite alignment when the pointer argument is invalid / null. The
problem is that for whatever reason the masked load / store functions
use i32 for their alignment value which means this gets truncated to
zero.

Add a special check for this, long term we probably want to just remove
this argument entirely.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Joseph Huber (jhuber6)

Changes

Summary:
The changes made in #156057
allows the alignment value to be increased. We assert effectively
infinite alignment when the pointer argument is invalid / null. The
problem is that for whatever reason the masked load / store functions
use i32 for their alignment value which means this gets truncated to
zero.

Add a special check for this, long term we probably want to just remove
this argument entirely.

---

Full diff: https://github.com/llvm/llvm-project/pull/160109.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/Scalar/InferAlignment.cpp (+2-1)
• (modified) llvm/test/Transforms/InferAlignment/masked.ll (+12)

diff --git a/llvm/lib/Transforms/Scalar/InferAlignment.cpp b/llvm/lib/Transforms/Scalar/InferAlignment.cpp
index b60b15b6c3a2b..995b80396b8af 100644
--- a/llvm/lib/Transforms/Scalar/InferAlignment.cpp
+++ b/llvm/lib/Transforms/Scalar/InferAlignment.cpp
@@ -57,7 +57,8 @@ static bool tryToImproveAlign(
         cast<ConstantInt>(II->getArgOperand(AlignOpIdx))->getAlignValue();
     Align PrefAlign = DL.getPrefTypeAlign(Type);
     Align NewAlign = Fn(PtrOp, OldAlign, PrefAlign);
-    if (NewAlign <= OldAlign)
+    if (NewAlign <= OldAlign ||
+        NewAlign.value() > std::numeric_limits<uint32_t>().max())
       return false;
 
     Value *V =
diff --git a/llvm/test/Transforms/InferAlignment/masked.ll b/llvm/test/Transforms/InferAlignment/masked.ll
index 1b8d26417d75e..13acf9b50e7e8 100644
--- a/llvm/test/Transforms/InferAlignment/masked.ll
+++ b/llvm/test/Transforms/InferAlignment/masked.ll
@@ -29,6 +29,18 @@ entry:
   ret void
 }
 
+define <2 x i32> @null(<2 x i1> %mask, <2 x i32> %val) {
+; CHECK-LABEL: define <2 x i32> @null(
+; CHECK-SAME: <2 x i1> [[MASK:%.*]], <2 x i32> [[VAL:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[MASKED_LOAD:%.*]] = tail call <2 x i32> @llvm.masked.load.v2i32.p0(ptr null, i32 1, <2 x i1> [[MASK]], <2 x i32> [[VAL]])
+; CHECK-NEXT:    ret <2 x i32> [[MASKED_LOAD]]
+;
+entry:
+  %masked_load = tail call <2 x i32> @llvm.masked.load.v2f64.p0(ptr null, i32 1, <2 x i1> %mask, <2 x i32> %val)
+  ret <2 x i32> %masked_load
+}
+
 declare void @llvm.assume(i1)
 declare <2 x i32> @llvm.masked.load.v2i32.p0(ptr, i32, <2 x i1>, <2 x i32>)
 declare void @llvm.masked.store.v2i32.p0(<2 x i32>, ptr, i32, <2 x i1>)

[nikic]

LGTM

[arsenm]

Use Value::MaxAlignmentExponent / MaximumAlignment?

[jhuber6]

I figured this was more straightforward since the problem is that Align is stored as i64 but this argument is i32, so this makes it clear that it's a size issue.

[nikic]

@arsenm We already use MaxAlignmentExponent -- the problem is that these specific intrinsics don't support the full alignment range (the maximum exponent is one less).