[llvm] 67b789b - [NVPTX] Fix global destructor handling not working with a single dtor (#162537)

[via llvm-commits]

Author: Joseph Huber
Date: 2025-10-08T18:13:48-05:00
New Revision: 67b789baf47d4e496f9f97912ea03048b63b851b

URL: https://github.com/llvm/llvm-project/commit/67b789baf47d4e496f9f97912ea03048b63b851b
DIFF: https://github.com/llvm/llvm-project/commit/67b789baf47d4e496f9f97912ea03048b63b851b.diff

LOG: [NVPTX] Fix global destructor handling not working with a single dtor (#162537)

Summary:
Very small typo here caused it to not enter the main loop if the size
was not greater than 1. The inner loop is "less than" so the inverse
should be
"greater than or equal" but we were only doing "greater than". Did not
notice this because the libc handling does its own implementation and
the OpenMP tests only tested multiple destructors, and most people only
ever use global constructors when they show up on the GPU.

Added: 
    

Modified: 
    llvm/lib/Target/NVPTX/NVPTXCtorDtorLowering.cpp
    llvm/test/CodeGen/NVPTX/lower-ctor-dtor.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/NVPTX/NVPTXCtorDtorLowering.cpp b/llvm/lib/Target/NVPTX/NVPTXCtorDtorLowering.cpp
index bb8cec05f4d84..4e069398d540f 100644
--- a/llvm/lib/Target/NVPTX/NVPTXCtorDtorLowering.cpp
+++ b/llvm/lib/Target/NVPTX/NVPTXCtorDtorLowering.cpp
@@ -88,7 +88,7 @@ static Function *createInitOrFiniKernelFunction(Module &M, bool IsCtor) {
 //     reinterpret_cast<InitCallback *>(*start)();
 // }
 //
-// void call_init_array_callbacks() {
+// void call_fini_array_callbacks() {
 //   size_t fini_array_size = __fini_array_end - __fini_array_start;
 //   for (size_t i = fini_array_size; i > 0; --i)
 //     reinterpret_cast<FiniCallback *>(__fini_array_start[i - 1])();
@@ -153,7 +153,7 @@ static void createInitOrFiniCalls(Function &F, bool IsCtor) {
         "start");
   }
   IRB.CreateCondBr(
-      IRB.CreateCmp(IsCtor ? ICmpInst::ICMP_NE : ICmpInst::ICMP_UGT, BeginVal,
+      IRB.CreateCmp(IsCtor ? ICmpInst::ICMP_NE : ICmpInst::ICMP_UGE, BeginVal,
                     EndVal),
       LoopBB, ExitBB);
   IRB.SetInsertPoint(LoopBB);

diff  --git a/llvm/test/CodeGen/NVPTX/lower-ctor-dtor.ll b/llvm/test/CodeGen/NVPTX/lower-ctor-dtor.ll
index 02118fbf741bf..b503da4a0bfde 100644
--- a/llvm/test/CodeGen/NVPTX/lower-ctor-dtor.ll
+++ b/llvm/test/CodeGen/NVPTX/lower-ctor-dtor.ll
@@ -72,7 +72,7 @@ define internal void @bar() {
 ; CHECK-NEXT:    [[OFFSET:%.*]] = ashr exact i64 [[TMP2]], 3
 ; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr ptr, ptr addrspace(1) [[BEGIN]], i64 [[OFFSET]]
 ; CHECK-NEXT:    [[START:%.*]] = getelementptr inbounds ptr, ptr addrspace(1) [[TMP3]], i64 -1
-; CHECK-NEXT:    [[TMP4:%.*]] = icmp ugt ptr addrspace(1) [[START]], [[BEGIN]]
+; CHECK-NEXT:    [[TMP4:%.*]] = icmp uge ptr addrspace(1) [[START]], [[BEGIN]]
 ; CHECK-NEXT:    br i1 [[TMP4]], label [[WHILE_ENTRY:%.*]], label [[WHILE_END:%.*]]
 ; CHECK:       while.entry:
 ; CHECK-NEXT:    [[PTR:%.*]] = phi ptr addrspace(1) [ [[START]], [[ENTRY:%.*]] ], [ [[NEXT:%.*]], [[WHILE_ENTRY]] ]