Fix LLVM 23 compatibility: replace removed NoNaNsFPMath/NoInfsFPMath TargetOptions fields (#9029)

* Initial plan

* Fix LLVM 23 compatibility: guard NoNaNsFPMath usage with #if LLVM_VERSION < 230

Co-authored-by: alexreinking <169273+alexreinking@users.noreply.github.com>

* Add no-nans-fp-math/no-infs-fp-math function attrs as LLVM 23 replacement for removed TargetOptions fields

Co-authored-by: alexreinking <169273+alexreinking@users.noreply.github.com>

---------

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: alexreinking <169273+alexreinking@users.noreply.github.com>

Author: Copilot

src/CodeGen_Internal.cpp

@@ -607,15 +607,16 @@ void get_target_options(const llvm::Module &module, llvm::TargetOptions &options
     bool use_soft_float_abi = get_modflag_bool(module, "halide_use_soft_float_abi");
     std::string mabi = get_modflag_string(module, "halide_mabi");
 
-    // FIXME: can this be migrated into `set_function_attributes_from_halide_target_options()`?
     bool per_instruction_fast_math_flags = get_modflag_bool(module, "halide_per_instruction_fast_math_flags");
 
     options = llvm::TargetOptions();
     options.AllowFPOpFusion = per_instruction_fast_math_flags ? llvm::FPOpFusion::Strict : llvm::FPOpFusion::Fast;
 #if LLVM_VERSION < 230
     options.NoInfsFPMath = !per_instruction_fast_math_flags;
 #endif
+#if LLVM_VERSION < 230
     options.NoNaNsFPMath = !per_instruction_fast_math_flags;
+#endif
     options.HonorSignDependentRoundingFPMathOption = !per_instruction_fast_math_flags;
     options.NoZerosInBSS = false;
     options.GuaranteedTailCallOpt = false;
@@ -720,6 +721,16 @@ void set_function_attributes_from_halide_target_options(llvm::Function &fn) {
         fn.addFnAttr(llvm::Attribute::getWithVScaleRangeArgs(
             module.getContext(), vscale_range, vscale_range));
     }
+
+    // When not using per-instruction fast-math flags (i.e., the whole module
+    // is in fast-math mode), propagate the fast-math assumptions as function
+    // attributes. In LLVM 23+, NoNaNsFPMath and NoInfsFPMath were removed from
+    // TargetOptions in favor of these per-function attributes.
+    bool per_instruction_fast_math_flags = get_modflag_bool(module, "halide_per_instruction_fast_math_flags");
+    if (!per_instruction_fast_math_flags) {
+        fn.addFnAttr("no-nans-fp-math", "true");
+        fn.addFnAttr("no-infs-fp-math", "true");
+    }
 }
 
 void embed_bitcode(llvm::Module *M, const string &halide_command) {

src/CodeGen_PTX_Dev.cpp

@@ -230,6 +230,11 @@ void CodeGen_PTX_Dev::init_module() {
 
     module = get_initial_module_for_ptx_device(target, context);
 
+    // Propagate the strict-float flag as a module flag so that
+    // set_function_attributes_from_halide_target_options can read it.
+    module->addModuleFlag(llvm::Module::Warning, "halide_per_instruction_fast_math_flags",
+                          CodeGen_GPU_Dev::any_strict_float ? 1 : 0);
+
     struct Intrinsic {
         const char *name;
         Type ret_type;
@@ -631,7 +636,9 @@ vector<char> CodeGen_PTX_Dev::compile_to_src() {
 #if LLVM_VERSION < 230
     options.NoInfsFPMath = !CodeGen_GPU_Dev::any_strict_float;
 #endif
+#if LLVM_VERSION < 230
     options.NoNaNsFPMath = !CodeGen_GPU_Dev::any_strict_float;
+#endif
     options.HonorSignDependentRoundingFPMathOption = !CodeGen_GPU_Dev::any_strict_float;
     options.NoZerosInBSS = false;
     options.GuaranteedTailCallOpt = false;