InstCombine: Increase threadlocal.address alignment if pointee is more aligned

[MatzeB]

Increase alignment of llvm.threadlocal.address if the pointed to global has higher alignment.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

@llvm/pr-subscribers-llvm-analysis

Author: Matthias Braun (MatzeB)

Changes

Look through llvm.threadlocal.address intrinsic when computing known bits in ValueTracking.

This improves issue #87437

---

Full diff: https://github.com/llvm/llvm-project/pull/88435.diff

2 Files Affected:

• (modified) llvm/lib/Analysis/ValueTracking.cpp (+4)
• (added) llvm/test/Analysis/ValueTracking/knownbits-tls.ll (+32)

diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index 3a10de72a27562..c7aec8518044f3 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -1686,6 +1686,10 @@ static void computeKnownBitsFromOperator(const Operator *I,
           Known.Zero.setBitsFrom(KnownZeroFirstBit);
         break;
       }
+      case Intrinsic::threadlocal_address: {
+        computeKnownBits(II->getArgOperand(0), Known, Depth + 1, Q);
+        break;
+      }
       case Intrinsic::vscale: {
         if (!II->getParent() || !II->getFunction())
           break;
diff --git a/llvm/test/Analysis/ValueTracking/knownbits-tls.ll b/llvm/test/Analysis/ValueTracking/knownbits-tls.ll
new file mode 100644
index 00000000000000..203777c33f8486
--- /dev/null
+++ b/llvm/test/Analysis/ValueTracking/knownbits-tls.ll
@@ -0,0 +1,32 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
+; RUN: opt -S -passes=instcombine < %s | FileCheck %s
+
+target datalayout = "p:64:64"
+
+@tlsvar_align32 = thread_local global i32 3, align 32
+
+define i1 @lowbits_zero() {
+; CHECK-LABEL: define i1 @lowbits_zero() {
+; CHECK-NEXT:    ret i1 true
+;
+  %p = call ptr @llvm.threadlocal.address(ptr @tlsvar_align32)
+  %p_int = ptrtoint ptr %p to i64
+  %lowbits = and i64 %p_int, 31
+  %lowbits_zero = icmp eq i64 %lowbits, 0
+  ret i1 %lowbits_zero
+}
+
+define i1 @lowbits_unknown() {
+; CHECK-LABEL: define i1 @lowbits_unknown() {
+; CHECK-NEXT:    [[P:%.*]] = call ptr @llvm.threadlocal.address.p0(ptr @tlsvar_align32)
+; CHECK-NEXT:    [[P_INT:%.*]] = ptrtoint ptr [[P]] to i64
+; CHECK-NEXT:    [[LOWBITS:%.*]] = and i64 [[P_INT]], 32
+; CHECK-NEXT:    [[LOWBITS_ZERO:%.*]] = icmp eq i64 [[LOWBITS]], 0
+; CHECK-NEXT:    ret i1 [[LOWBITS_ZERO]]
+;
+  %p = call ptr @llvm.threadlocal.address(ptr @tlsvar_align32)
+  %p_int = ptrtoint ptr %p to i64
+  %lowbits = and i64 %p_int, 63
+  %lowbits_zero = icmp eq i64 %lowbits, 0
+  ret i1 %lowbits_zero
+}

[nikic]

An alternative, and I think maybe better, approach to this would be to infer an align return attribute on threadlocal.address in InstCombine. This way, not only will computeKnownBits() automatically handle it, but it will also benefit other APIs that directly look at alignment, such as getPointerAlign().

[MatzeB]

Yep, I also just noticed that IRBuilder::CreateThreadLocalAddress already transfers the alignment of the target which I missed when I manually created LLVM-IR for tests. Anyway having an InstCombine rule that transfers minimum alignment is a good idea as well. Changed.

[nikic]

LGTM