update autodiff flags

compiler/rustc_codegen_llvm/messages.ftl

@@ -1,3 +1,4 @@
+codegen_llvm_autodiff_without_enable = using the autodiff feature requires -Z autodiff=Enable
 codegen_llvm_autodiff_without_lto = using the autodiff feature requires using fat-lto
 
 codegen_llvm_copy_bitcode = failed to copy bitcode to object file: {$err}

compiler/rustc_codegen_llvm/src/back/lto.rs

@@ -586,6 +586,42 @@ fn thin_lto(
     }
 }
 
+fn enable_autodiff_settings(ad: &[config::AutoDiff], module: &mut ModuleCodegen<ModuleLlvm>) {
+    for &val in ad {
+        match val {
+            config::AutoDiff::PrintModBefore => {
+                unsafe { llvm::LLVMDumpModule(module.module_llvm.llmod()) };
+            }
+            config::AutoDiff::PrintPerf => {
+                llvm::set_print_perf(true);
+            }
+            config::AutoDiff::PrintAA => {
+                llvm::set_print_activity(true);
+            }
+            config::AutoDiff::PrintTA => {
+                llvm::set_print_type(true);
+            }
+            config::AutoDiff::Inline => {
+                llvm::set_inline(true);
+            }
+            config::AutoDiff::LooseTypes => {
+                llvm::set_loose_types(false);
+            }
+            config::AutoDiff::PrintSteps => {
+                llvm::set_print(true);
+            }
+            // We handle this below
+            config::AutoDiff::PrintModAfter => {}
+            // This is required and already checked
+            config::AutoDiff::Enable => {}
+        }
+    }
+    // This helps with handling enums for now.
+    llvm::set_strict_aliasing(false);
+    // FIXME(ZuseZ4): Test this, since it was added a long time ago.
+    llvm::set_rust_rules(true);
+}
+
 pub(crate) fn run_pass_manager(
     cgcx: &CodegenContext<LlvmCodegenBackend>,
     dcx: DiagCtxtHandle<'_>,
@@ -604,34 +640,37 @@ pub(crate) fn run_pass_manager(
     let opt_stage = if thin { llvm::OptStage::ThinLTO } else { llvm::OptStage::FatLTO };
     let opt_level = config.opt_level.unwrap_or(config::OptLevel::No);
 
-    // If this rustc version was build with enzyme/autodiff enabled, and if users applied the
-    // `#[autodiff]` macro at least once, then we will later call llvm_optimize a second time.
-    debug!("running llvm pm opt pipeline");
+    // The PostAD behavior is the same that we would have if no autodiff was used.
+    // It will run the default optimization pipeline. If AD is enabled we select
+    // the DuringAD stage, which will disable vectorization and loop unrolling, and
+    // schedule two autodiff optimization + differentiation passes.
+    // We then run the llvm_optimize function a second time, to optimize the code which we generated
+    // in the enzyme differentiation pass.
+    let enable_ad = config.autodiff.contains(&config::AutoDiff::Enable);
+    let stage =
+        if enable_ad { write::AutodiffStage::DuringAD } else { write::AutodiffStage::PostAD };
+
+    if enable_ad {
+        enable_autodiff_settings(&config.autodiff, module);
+    }
+
     unsafe {
-        write::llvm_optimize(
-            cgcx,
-            dcx,
-            module,
-            config,
-            opt_level,
-            opt_stage,
-            write::AutodiffStage::DuringAD,
-        )?;
+        write::llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage, stage)?;
     }
-    // FIXME(ZuseZ4): Make this more granular
-    if cfg!(llvm_enzyme) && !thin {
+
+    if cfg!(llvm_enzyme) && enable_ad {
+        let opt_stage = llvm::OptStage::FatLTO;
+        let stage = write::AutodiffStage::PostAD;
         unsafe {
-            write::llvm_optimize(
-                cgcx,
-                dcx,
-                module,
-                config,
-                opt_level,
-                llvm::OptStage::FatLTO,
-                write::AutodiffStage::PostAD,
-            )?;
+            write::llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage, stage)?;
+        }
+
+        // This is the final IR, so people should be able to inspect the optimized autodiff output.
+        if config.autodiff.contains(&config::AutoDiff::PrintModAfter) {
+            unsafe { llvm::LLVMDumpModule(module.module_llvm.llmod()) };
         }
     }
+
     debug!("lto done");
     Ok(())
 }

compiler/rustc_codegen_llvm/src/builder/autodiff.rs

@@ -10,7 +10,7 @@ use crate::back::write::llvm_err;
 use crate::builder::SBuilder;
 use crate::context::SimpleCx;
 use crate::declare::declare_simple_fn;
-use crate::errors::LlvmError;
+use crate::errors::{AutoDiffWithoutEnable, LlvmError};
 use crate::llvm::AttributePlace::Function;
 use crate::llvm::{Metadata, True};
 use crate::value::Value;
@@ -46,9 +46,6 @@ fn generate_enzyme_call<'ll>(
     let output = attrs.ret_activity;
 
     // We have to pick the name depending on whether we want forward or reverse mode autodiff.
-    // FIXME(ZuseZ4): The new pass based approach should not need the {Forward/Reverse}First method anymore, since
-    // it will handle higher-order derivatives correctly automatically (in theory). Currently
-    // higher-order derivatives fail, so we should debug that before adjusting this code.
     let mut ad_name: String = match attrs.mode {
         DiffMode::Forward => "__enzyme_fwddiff",
         DiffMode::Reverse => "__enzyme_autodiff",
@@ -291,6 +288,14 @@ pub(crate) fn differentiate<'ll>(
     let diag_handler = cgcx.create_dcx();
     let cx = SimpleCx { llmod: module.module_llvm.llmod(), llcx: module.module_llvm.llcx };
 
+    // First of all, did the user try to use autodiff without using the -Zautodiff=Enable flag?
+    if !diff_items.is_empty()
+        && !cgcx.opts.unstable_opts.autodiff.contains(&rustc_session::config::AutoDiff::Enable)
+    {
+        let dcx = cgcx.create_dcx();
+        return Err(dcx.handle().emit_almost_fatal(AutoDiffWithoutEnable));
+    }
+
     // Before dumping the module, we want all the TypeTrees to become part of the module.
     for item in diff_items.iter() {
         let name = item.source.clone();

compiler/rustc_codegen_llvm/src/errors.rs

@@ -92,9 +92,12 @@ impl<G: EmissionGuarantee> Diagnostic<'_, G> for ParseTargetMachineConfig<'_> {
 
 #[derive(Diagnostic)]
 #[diag(codegen_llvm_autodiff_without_lto)]
-#[note]
 pub(crate) struct AutoDiffWithoutLTO;
 
+#[derive(Diagnostic)]
+#[diag(codegen_llvm_autodiff_without_enable)]
+pub(crate) struct AutoDiffWithoutEnable;
+
 #[derive(Diagnostic)]
 #[diag(codegen_llvm_lto_disallowed)]
 pub(crate) struct LtoDisallowed;

compiler/rustc_codegen_llvm/src/llvm/enzyme_ffi.rs

@@ -35,3 +35,97 @@ pub enum LLVMRustVerifierFailureAction {
     LLVMPrintMessageAction = 1,
     LLVMReturnStatusAction = 2,
 }
+
+#[cfg(llvm_enzyme)]
+pub use self::Enzyme_AD::*;
+
+#[cfg(llvm_enzyme)]
+pub mod Enzyme_AD {
+    use libc::c_void;
+    extern "C" {
+        pub fn EnzymeSetCLBool(arg1: *mut ::std::os::raw::c_void, arg2: u8);
+    }
+    extern "C" {
+        static mut EnzymePrintPerf: c_void;
+        static mut EnzymePrintActivity: c_void;
+        static mut EnzymePrintType: c_void;
+        static mut EnzymePrint: c_void;
+        static mut EnzymeStrictAliasing: c_void;
+        static mut looseTypeAnalysis: c_void;
+        static mut EnzymeInline: c_void;
+        static mut RustTypeRules: c_void;
+    }
+    pub fn set_print_perf(print: bool) {
+        unsafe {
+            EnzymeSetCLBool(std::ptr::addr_of_mut!(EnzymePrintPerf), print as u8);
+        }
+    }
+    pub fn set_print_activity(print: bool) {
+        unsafe {
+            EnzymeSetCLBool(std::ptr::addr_of_mut!(EnzymePrintActivity), print as u8);
+        }
+    }
+    pub fn set_print_type(print: bool) {
+        unsafe {
+            EnzymeSetCLBool(std::ptr::addr_of_mut!(EnzymePrintType), print as u8);
+        }
+    }
+    pub fn set_print(print: bool) {
+        unsafe {
+            EnzymeSetCLBool(std::ptr::addr_of_mut!(EnzymePrint), print as u8);
+        }
+    }
+    pub fn set_strict_aliasing(strict: bool) {
+        unsafe {
+            EnzymeSetCLBool(std::ptr::addr_of_mut!(EnzymeStrictAliasing), strict as u8);
+        }
+    }
+    pub fn set_loose_types(loose: bool) {
+        unsafe {
+            EnzymeSetCLBool(std::ptr::addr_of_mut!(looseTypeAnalysis), loose as u8);
+        }
+    }
+    pub fn set_inline(val: bool) {
+        unsafe {
+            EnzymeSetCLBool(std::ptr::addr_of_mut!(EnzymeInline), val as u8);
+        }
+    }
+    pub fn set_rust_rules(val: bool) {
+        unsafe {
+            EnzymeSetCLBool(std::ptr::addr_of_mut!(RustTypeRules), val as u8);
+        }
+    }
+}
+
+#[cfg(not(llvm_enzyme))]
+pub use self::Fallback_AD::*;
+
+#[cfg(not(llvm_enzyme))]
+pub mod Fallback_AD {
+    #![allow(unused_variables)]
+
+    pub fn set_inline(val: bool) {
+        unimplemented!()
+    }
+    pub fn set_print_perf(print: bool) {
+        unimplemented!()
+    }
+    pub fn set_print_activity(print: bool) {
+        unimplemented!()
+    }
+    pub fn set_print_type(print: bool) {
+        unimplemented!()
+    }
+    pub fn set_print(print: bool) {
+        unimplemented!()
+    }
+    pub fn set_strict_aliasing(strict: bool) {
+        unimplemented!()
+    }
+    pub fn set_loose_types(loose: bool) {
+        unimplemented!()
+    }
+    pub fn set_rust_rules(val: bool) {
+        unimplemented!()
+    }
+}

compiler/rustc_codegen_ssa/src/back/write.rs

@@ -405,7 +405,8 @@ fn generate_lto_work<B: ExtraBackendMethods>(
             B::run_fat_lto(cgcx, needs_fat_lto, import_only_modules).unwrap_or_else(|e| e.raise());
         if cgcx.lto == Lto::Fat && !autodiff.is_empty() {
             let config = cgcx.config(ModuleKind::Regular);
-            module = unsafe { module.autodiff(cgcx, autodiff, config).unwrap() };
+            module =
+                unsafe { module.autodiff(cgcx, autodiff, config).unwrap_or_else(|e| e.raise()) };
         }
         // We are adding a single work item, so the cost doesn't matter.
         vec![(WorkItem::LTO(module), 0)]

compiler/rustc_interface/src/tests.rs

@@ -759,7 +759,7 @@ fn test_unstable_options_tracking_hash() {
     tracked!(allow_features, Some(vec![String::from("lang_items")]));
     tracked!(always_encode_mir, true);
     tracked!(assume_incomplete_release, true);
-    tracked!(autodiff, vec![AutoDiff::Print]);
+    tracked!(autodiff, vec![AutoDiff::Enable]);
     tracked!(binary_dep_depinfo, true);
     tracked!(box_noalias, false);
     tracked!(

compiler/rustc_session/src/config.rs

@@ -198,33 +198,26 @@ pub enum CoverageLevel {
 /// The different settings that the `-Z autodiff` flag can have.
 #[derive(Clone, Copy, PartialEq, Hash, Debug)]
 pub enum AutoDiff {
+    /// Enable the autodiff opt pipeline
+    Enable,
+
     /// Print TypeAnalysis information
     PrintTA,
     /// Print ActivityAnalysis Information
     PrintAA,
     /// Print Performance Warnings from Enzyme
     PrintPerf,
-    /// Combines the three print flags above.
-    Print,
+    /// Print intermediate IR generation steps
+    PrintSteps,
     /// Print the whole module, before running opts.
     PrintModBefore,
-    /// Print the whole module just before we pass it to Enzyme.
-    /// For Debug purpose, prefer the OPT flag below
-    PrintModAfterOpts,
     /// Print the module after Enzyme differentiated everything.
-    PrintModAfterEnzyme,
+    PrintModAfter,
 
-    /// Enzyme's loose type debug helper (can cause incorrect gradients)
+    /// Enzyme's loose type debug helper (can cause incorrect gradients!!)
+    /// Usable in cases where Enzyme errors with `can not deduce type of X`.
     LooseTypes,
-
-    /// More flags
-    NoModOptAfter,
-    /// Tell Enzyme to run LLVM Opts on each function it generated. By default off,
-    /// since we already optimize the whole module after Enzyme is done.
-    EnableFncOpt,
-    NoVecUnroll,
-    RuntimeActivity,
-    /// Runs Enzyme specific Inlining
+    /// Runs Enzyme's aggressive inlining
     Inline,
 }
 

compiler/rustc_session/src/options.rs

@@ -707,7 +707,7 @@ mod desc {
     pub(crate) const parse_list: &str = "a space-separated list of strings";
     pub(crate) const parse_list_with_polarity: &str =
         "a comma-separated list of strings, with elements beginning with + or -";
-    pub(crate) const parse_autodiff: &str = "a comma separated list of settings: `Print`, `PrintTA`, `PrintAA`, `PrintPerf`, `PrintModBefore`, `PrintModAfterOpts`, `PrintModAfterEnzyme`, `LooseTypes`, `NoModOptAfter`, `EnableFncOpt`, `NoVecUnroll`, `Inline`";
+    pub(crate) const parse_autodiff: &str = "a comma separated list of settings: `Enable`, `PrintSteps`, `PrintTA`, `PrintAA`, `PrintPerf`, `PrintModBefore`, `PrintModAfter`, `LooseTypes`, `Inline`";
     pub(crate) const parse_comma_list: &str = "a comma-separated list of strings";
     pub(crate) const parse_opt_comma_list: &str = parse_comma_list;
     pub(crate) const parse_number: &str = "a number";
@@ -1348,17 +1348,14 @@ pub mod parse {
         v.sort_unstable();
         for &val in v.iter() {
             let variant = match val {
+                "Enable" => AutoDiff::Enable,
                 "PrintTA" => AutoDiff::PrintTA,
                 "PrintAA" => AutoDiff::PrintAA,
                 "PrintPerf" => AutoDiff::PrintPerf,
-                "Print" => AutoDiff::Print,
+                "PrintSteps" => AutoDiff::PrintSteps,
                 "PrintModBefore" => AutoDiff::PrintModBefore,
-                "PrintModAfterOpts" => AutoDiff::PrintModAfterOpts,
-                "PrintModAfterEnzyme" => AutoDiff::PrintModAfterEnzyme,
+                "PrintModAfter" => AutoDiff::PrintModAfter,
                 "LooseTypes" => AutoDiff::LooseTypes,
-                "NoModOptAfter" => AutoDiff::NoModOptAfter,
-                "EnableFncOpt" => AutoDiff::EnableFncOpt,
-                "NoVecUnroll" => AutoDiff::NoVecUnroll,
                 "Inline" => AutoDiff::Inline,
                 _ => {
                     // FIXME(ZuseZ4): print an error saying which value is not recognized
@@ -2081,19 +2078,17 @@ options! {
     assume_incomplete_release: bool = (false, parse_bool, [TRACKED],
         "make cfg(version) treat the current version as incomplete (default: no)"),
     autodiff: Vec<crate::config::AutoDiff> = (Vec::new(), parse_autodiff, [TRACKED],
-        "a list of optional autodiff flags to enable
+        "a list of autodiff flags to enable
+        Mandatory setting:
+        `=Enable`
         Optional extra settings:
         `=PrintTA`
         `=PrintAA`
         `=PrintPerf`
-         `=Print`
+        `=PrintSteps`
         `=PrintModBefore`
-         `=PrintModAfterOpts`
-         `=PrintModAfterEnzyme`
+        `=PrintModAfter`
         `=LooseTypes`
-         `=NoModOptAfter`
-         `=EnableFncOpt`
-         `=NoVecUnroll`
         `=Inline`
         Multiple options can be combined with commas."),
     #[rustc_lint_opt_deny_field_access("use `Session::binary_dep_depinfo` instead of this field")]

src/doc/unstable-book/src/compiler-flags/autodiff.md

@@ -8,16 +8,13 @@ This feature allows you to differentiate functions using automatic differentiati
 Set the `-Zautodiff=<options>` compiler flag to adjust the behaviour of the autodiff feature.
 Multiple options can be separated with a comma. Valid options are:
 
+`Enable` - Required flag to enable autodiff
 `PrintTA` - print Type Analysis Information
 `PrintAA` - print Activity Analysis Information
 `PrintPerf` - print Performance Warnings from Enzyme
-`Print` - prints all intermediate transformations
+`PrintSteps` - prints all intermediate transformations
 `PrintModBefore` - print the whole module, before running opts
-`PrintModAfterOpts` - print the whole module just before we pass it to Enzyme
-`PrintModAfterEnzyme` - print the module after Enzyme differentiated everything
+`PrintModAfter` - print the module after Enzyme differentiated everything
 `LooseTypes` - Enzyme's loose type debug helper (can cause incorrect gradients)
 `Inline` - runs Enzyme specific Inlining
-`NoModOptAfter` - do not optimize the module after Enzyme is done
-`EnableFncOpt` - tell Enzyme to run LLVM Opts on each function it generated
-`NoVecUnroll` - do not unroll vectorized loops
 `RuntimeActivity` - allow specifying activity at runtime

tests/codegen/autodiff.rs

@@ -1,4 +1,4 @@
-//@ compile-flags: -C opt-level=3  -Clto=fat
+//@ compile-flags: -Zautodiff=Enable -C opt-level=3  -Clto=fat
 //@ no-prefer-dynamic
 //@ needs-enzyme
 #![feature(autodiff)]