[InstCombine] Don't consider aligned_alloc removable if icmp uses result (#69474)

At the moment, all alloc-like functions are assumed to return non-null
pointers, if their return value is only used in a compare. This is based
on being allowed to substitute the allocation function with one that
doesn't fail to allocate the required memory.

aligned_alloc however must also return null if the required alignment
cannot be satisfied, so I don't think the same reasoning as above can be
applied to it.

This patch adds a bail-out for aligned_alloc calls to
isAllocSiteRemovable.

Author: fhahn

llvm/lib/Transforms/InstCombine/InstructionCombining.cpp

@@ -2430,6 +2430,26 @@ static bool isAllocSiteRemovable(Instruction *AI,
         unsigned OtherIndex = (ICI->getOperand(0) == PI) ? 1 : 0;
         if (!isNeverEqualToUnescapedAlloc(ICI->getOperand(OtherIndex), TLI, AI))
           return false;
+
+        // Do not fold compares to aligned_alloc calls, as they may have to
+        // return null in case the required alignment cannot be satisfied,
+        // unless we can prove that both alignment and size are valid.
+        auto AlignmentAndSizeKnownValid = [](CallBase *CB) {
+          // Check if alignment and size of a call to aligned_alloc is valid,
+          // that is alignment is a power-of-2 and the size is a multiple of the
+          // alignment.
+          const APInt *Alignment;
+          const APInt *Size;
+          return match(CB->getArgOperand(0), m_APInt(Alignment)) &&
+                 match(CB->getArgOperand(1), m_APInt(Size)) &&
+                 Alignment->isPowerOf2() && Size->urem(*Alignment).isZero();
+        };
+        auto *CB = dyn_cast<CallBase>(AI);
+        LibFunc TheLibFunc;
+        if (CB && TLI.getLibFunc(*CB->getCalledFunction(), TheLibFunc) &&
+            TLI.has(TheLibFunc) && TheLibFunc == LibFunc_aligned_alloc &&
+            !AlignmentAndSizeKnownValid(CB))
+          return false;
         Users.emplace_back(I);
         continue;
       }

llvm/test/Transforms/InstCombine/malloc-free.ll

@@ -26,9 +26,11 @@ define i32 @dead_aligned_alloc(i32 %size, i32 %alignment, i8 %value) {
   ret i32 0
 }
 
-define i1 @aligned_alloc_pointer_only_used_by_cmp(i32 %size, i32 %alignment, i8 %value) {
-; CHECK-LABEL: @aligned_alloc_pointer_only_used_by_cmp(
-; CHECK-NEXT:    ret i1 true
+define i1 @aligned_alloc_only_pointe(i32 %size, i32 %alignment, i8 %value) {
+; CHECK-LABEL: @aligned_alloc_only_pointe(
+; CHECK-NEXT:    [[ALIGNED_ALLOCATION:%.*]] = tail call ptr @aligned_alloc(i32 [[ALIGNMENT:%.*]], i32 [[SIZE:%.*]])
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ne ptr [[ALIGNED_ALLOCATION]], null
+; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %aligned_allocation = tail call ptr @aligned_alloc(i32 %alignment, i32 %size)
   %cmp = icmp ne ptr %aligned_allocation, null
@@ -46,7 +48,9 @@ define i1 @aligned_alloc_pointer_only_used_by_cmp_alignment_and_value_known_ok(i
 
 define i1 @aligned_alloc_pointer_only_used_by_cmp_alignment_no_power_of_2(i32 %size, i32 %alignment, i8 %value) {
 ; CHECK-LABEL: @aligned_alloc_pointer_only_used_by_cmp_alignment_no_power_of_2(
-; CHECK-NEXT:    ret i1 true
+; CHECK-NEXT:    [[ALIGNED_ALLOCATION:%.*]] = tail call dereferenceable_or_null(32) ptr @aligned_alloc(i32 3, i32 32)
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ne ptr [[ALIGNED_ALLOCATION]], null
+; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %aligned_allocation = tail call ptr @aligned_alloc(i32 3, i32 32)
   %cmp = icmp ne ptr %aligned_allocation, null
@@ -55,7 +59,9 @@ define i1 @aligned_alloc_pointer_only_used_by_cmp_alignment_no_power_of_2(i32 %s
 
 define i1 @aligned_alloc_pointer_only_used_by_cmp_size_not_multiple_of_alignment(i32 %size, i32 %alignment, i8 %value) {
 ; CHECK-LABEL: @aligned_alloc_pointer_only_used_by_cmp_size_not_multiple_of_alignment(
-; CHECK-NEXT:    ret i1 true
+; CHECK-NEXT:    [[ALIGNED_ALLOCATION:%.*]] = tail call dereferenceable_or_null(31) ptr @aligned_alloc(i32 8, i32 31)
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ne ptr [[ALIGNED_ALLOCATION]], null
+; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %aligned_allocation = tail call ptr @aligned_alloc(i32 8, i32 31)
   %cmp = icmp ne ptr %aligned_allocation, null