Relax ROCm restrictions with LLVM 21 (#28421)

Relaxes the restriction to use ROCm with a bundled LLVM.

The new state as of this PR

* With ROCm 6.3, you can use either the bundled LLVM or LLVM 21+
* With ROCm <6.3, you can still only use the bundled LLVM
    * This is a LLVM limitation, they require ROCm 6.3+


Note: I only ran the `COMM=none` test suite, which did not pass 100%.
Most tests passed, a few edge cases (like interop printf) did not work
quite right. I consider that future work. IMO, its more important for
things to mostly work and we can improve specific cases as users run
into issues.

Resolves #28007

[Reviewed by @e-kayrakli]

Author: jabraham17

compiler/llvm/clangUtil.cpp

@@ -3029,6 +3029,17 @@ static void helpComputeClangArgs(std::string& clangCC,
   }
 #endif
 
+#if HAVE_LLVM_VER >= 210
+  if (usingGpuLocaleModel() &&
+      getGpuCodegenType() == GpuCodegenType::GPU_CG_AMD_HIP) {
+    // this is need to make hipcub wrappers work since
+    // TempStorage is marked deprecated, but I can't seem to find an alternative
+    // the rocm docs do not mention the deprecation and I can't find proper docs
+    // for the new API that the warning refers to
+    clangCCArgs.push_back("-Wno-deprecated-declarations");
+  }
+#endif
+
   // Add debug flags
   if (fDebugSymbols) {
     clangCCArgs.push_back("-g");
@@ -4966,7 +4977,7 @@ static void makeBinaryLLVMForHIP(const std::string& artifactFilename,
   // So this loop should run exactly one iteration.
   INT_ASSERT(gpuArches.size() == 1);
 
-  std::string targets = "-targets=host-x86_64-unknown-linux";
+  std::string targets = "-targets=host-x86_64-unknown-linux-unknown";
 #if HAVE_LLVM_VER >= 150
   std::string inputs = "-input=/dev/null ";
   std::string outputs = "-output=" + fatbinFilename;

util/chplenv/chpl_gpu.py

@@ -465,23 +465,19 @@ def _validate_cuda_llvm_version_impl(gpu: gpu_type):
         )
 
 def _validate_rocm_llvm_version_impl(gpu: gpu_type):
-    major_version = get_sdk_version().split('.')[0]
+    major_version, minor_version = get_sdk_version().split('.')[:2]
 
     if major_version in ('5',) and chpl_llvm.get() == 'bundled':
         error("Cannot target AMD GPUs with CHPL_LLVM=bundled")
-    elif major_version == '6' and chpl_llvm.get() != 'bundled' and os.environ.get("CHPL_LLVM_65188_PATCH", "0") != "1":
-        # once https://github.com/llvm/llvm-project/issues/65188 is resolved,
-        # this check can removed
-        # 'CHPL_LLVM_65188_PATCH' is a temporary escape hatch to enable system
-        # llvm with rocm 6.x if the user knows what they are doing
-        error("Cannot target AMD GPUs with ROCm 6.x without CHPL_LLVM=bundled")
-        # TODO: can we use LLVM 20 with ROCm 6.3?
-    elif (
-        major_version == "6"
-        and chpl_llvm.get() == "system"
-        and int(chpl_llvm.get_llvm_version()) < 18
-    ):
-        error("Cannot target AMD GPUs with ROCm 6.x without LLVM 18+")
+    elif major_version == '6':
+        if int(minor_version) < 3:
+            # it must be bundled LLVM or patched
+            if chpl_llvm.get() != 'bundled' and os.environ.get("CHPL_LLVM_65188_PATCH", "0") != "1":
+                error("Cannot target AMD GPUs with ROCm 6.0-6.2 without CHPL_LLVM=bundled")
+        else:
+            # it must be LLVM 21+ or the bundled LLVM
+            if chpl_llvm.get() != 'bundled' and int(chpl_llvm.get_llvm_version()) < 21:
+                error("Cannot target AMD GPUs with ROCm 6.3 without CHPL_LLVM=bundled or LLVM 21+")
     elif not validateLlvmBuiltForTgt(gpu.llvm_target):
         _reportMissingGpuReq(
             "LLVM not built for %s." % gpu.llvm_target, allowExempt=False