New approach: let the LLVM auto-upgrade mechanism take care of selecting the correct barrier intrinsic.

Author: mcourteaux

.gitignore

@@ -240,6 +240,9 @@ xcuserdata
 # NeoVim + clangd
 .cache
 
+# CCLS
+.ccls-cache/
+
 # Emacs
 tags
 TAGS

src/CodeGen_PTX_Dev.cpp

@@ -262,17 +262,9 @@ void CodeGen_PTX_Dev::visit(const Call *op) {
         auto fence_type_ptr = as_const_int(op->args[0]);
         internal_assert(fence_type_ptr) << "gpu_thread_barrier() parameter is not a constant integer.\n";
 
-        llvm::Function *barrier0 = module->getFunction("llvm.nvvm.barrier0");
-        if (barrier0) {
-            internal_assert(barrier0) << "Could not find PTX barrier intrinsic (llvm.nvvm.barrier0)\n";
-
-            builder->CreateCall(barrier0);
-        } else {
-            // Changed in LLVM 20.
-            barrier0 = module->getFunction("llvm.nvvm.barrier.cta.sync.aligned.all");
-            internal_assert(barrier0) << "Could not find PTX barrier intrinsic (llvm.nvvm.barrier.cta.sync.aligned.all)\n";
-            builder->CreateCall(barrier0, builder->getInt32(0));
-        }
+        llvm::Function *barrier = module->getFunction("halide_ptx_barrier0");
+        internal_assert(barrier) << "Could not find PTX barrier intrinsic: halide_ptx_barrier0";
+        builder->CreateCall(barrier);
         value = ConstantInt::get(i32_t, 0);
         return;
     }

src/runtime/ptx_dev.ll

@@ -1,5 +1,3 @@
-declare void @llvm.nvvm.barrier0()
-declare void @llvm.nvvm.barrier.cta.sync.aligned.all(i32)
 declare  i32 @llvm.nvvm.read.ptx.sreg.tid.x()
 declare  i32 @llvm.nvvm.read.ptx.sreg.ctaid.x()
 declare  i32 @llvm.nvvm.read.ptx.sreg.ntid.x()
@@ -18,6 +16,17 @@ declare  i32 @llvm.nvvm.read.ptx.sreg.ntid.w()
 declare  i32 @llvm.nvvm.read.ptx.sreg.nctaid.w()
 declare  i32 @llvm.nvvm.read.ptx.sreg.warpsize()
 
+
+; We define our own wrapper function that calls the intrinsic above.
+; Thanks to this wrapper, it's automatically upgraded in the new LLVM to the
+; new name of the intrinsic.
+; Once we drop support for LLVM 20.1.5, we can immediately call the new intrinsic ourselves.
+declare void @llvm.nvvm.barrier0()
+define weak_odr void @halide_ptx_barrier0() nounwind uwtable readnone alwaysinline {
+       call void @llvm.nvvm.barrier0()
+       ret void
+}
+
 ; Remove these two once the minimum required llvm version is 9.0
 declare float @llvm.nvvm.atomic.load.add.f32.p0f32(float*, float)
 declare double @llvm.nvvm.atomic.load.add.f64.p0f64(double *, double)