add batching backend

ZuseZ4 · ZuseZ4 · commit f6ac98cb7abc · 2025-04-03T02:46:18.000-04:00
diff --git a/compiler/rustc_codegen_llvm/src/builder/autodiff.rs b/compiler/rustc_codegen_llvm/src/builder/autodiff.rs
@@ -3,8 +3,8 @@ use std::ptr;
 use rustc_ast::expand::autodiff_attrs::{AutoDiffAttrs, AutoDiffItem, DiffActivity, DiffMode};
 use rustc_codegen_ssa::ModuleCodegen;
 use rustc_codegen_ssa::back::write::ModuleConfig;
-use rustc_codegen_ssa::traits::BaseTypeCodegenMethods as _;
 use rustc_errors::FatalError;
+use rustc_middle::bug;
 use tracing::{debug, trace};
 
 use crate::back::write::llvm_err;
@@ -28,11 +28,32 @@ fn get_params(fnc: &Value) -> Vec<&Value> {
     }
 }
 
+fn has_sret(fnc: &Value) -> bool {
+    let num_args = unsafe { llvm::LLVMCountParams(fnc) as usize };
+    if num_args == 0 {
+        false
+    } else {
+        unsafe { llvm::LLVMRustHasAttributeAtIndex(fnc, 0, llvm::AttributeKind::StructRet) }
+    }
+}
+
+// When we call the `__enzyme_autodiff` or `__enzyme_fwddiff` function, we need to pass all the
+// original inputs, as well as metadata and the additional shadow arguments.
+// This function matches the arguments from the outer function to the inner enzyme call.
+//
+// This function also considers that Rust level arguments not always match the llvm-ir level
+// arguments. A slice, `&[f32]`, for example, is represented as a pointer and a length on
+// llvm-ir level. The number of activities matches the number of Rust level arguments, so we
+// need to match those.
+// FIXME(ZuseZ4): This logic is a bit more complicated than it should be, can we simplify it
+// using iterators and peek()?
 fn match_args_from_caller_to_enzyme<'ll>(
     cx: &SimpleCx<'ll>,
+    width: u32,
     args: &mut Vec<&'ll llvm::Value>,
     inputs: &[DiffActivity],
     outer_args: &[&'ll llvm::Value],
+    has_sret: bool,
 ) {
     debug!("matching autodiff arguments");
     // We now handle the issue that Rust level arguments not always match the llvm-ir level
@@ -44,6 +65,14 @@ fn match_args_from_caller_to_enzyme<'ll>(
     let mut outer_pos: usize = 0;
     let mut activity_pos = 0;
 
+    if has_sret {
+        // Then the first outer arg is the sret pointer. Enzyme doesn't know about sret, so the
+        // inner function will still return something. We increase our outer_pos by one,
+        // and once we're done with all other args we will take the return of the inner call and
+        // update the sret pointer with it
+        outer_pos = 1;
+    }
+
     let enzyme_const = cx.create_metadata("enzyme_const".to_string()).unwrap();
     let enzyme_out = cx.create_metadata("enzyme_out".to_string()).unwrap();
     let enzyme_dup = cx.create_metadata("enzyme_dup".to_string()).unwrap();
@@ -95,20 +124,25 @@ fn match_args_from_caller_to_enzyme<'ll>(
                 assert!(unsafe {
                     llvm::LLVMRustGetTypeKind(next_outer_ty) == llvm::TypeKind::Integer
                 });
-                let next_outer_arg2 = outer_args[outer_pos + 2];
-                let next_outer_ty2 = cx.val_ty(next_outer_arg2);
-                assert!(unsafe {
-                    llvm::LLVMRustGetTypeKind(next_outer_ty2) == llvm::TypeKind::Pointer
-                });
-                let next_outer_arg3 = outer_args[outer_pos + 3];
-                let next_outer_ty3 = cx.val_ty(next_outer_arg3);
-                assert!(unsafe {
-                    llvm::LLVMRustGetTypeKind(next_outer_ty3) == llvm::TypeKind::Integer
-                });
-                args.push(next_outer_arg2);
+
+                for _ in 0..width {
+                    let next_outer_arg2 = outer_args[outer_pos + 2];
+                    let next_outer_ty2 = cx.val_ty(next_outer_arg2);
+                    assert!(
+                        unsafe { llvm::LLVMRustGetTypeKind(next_outer_ty2) }
+                            == llvm::TypeKind::Pointer
+                    );
+                    let next_outer_arg3 = outer_args[outer_pos + 3];
+                    let next_outer_ty3 = cx.val_ty(next_outer_arg3);
+                    assert!(
+                        unsafe { llvm::LLVMRustGetTypeKind(next_outer_ty3) }
+                            == llvm::TypeKind::Integer
+                    );
+                    args.push(next_outer_arg2);
+                }
                 args.push(cx.get_metadata_value(enzyme_const));
                 args.push(next_outer_arg);
-                outer_pos += 4;
+                outer_pos += 2 + 2 * width as usize;
                 activity_pos += 2;
             } else {
                 // A duplicated pointer will have the following two outer_fn arguments:
@@ -125,6 +159,13 @@ fn match_args_from_caller_to_enzyme<'ll>(
                 args.push(next_outer_arg);
                 outer_pos += 2;
                 activity_pos += 1;
+
+                // Now, if width > 1, we need to account for that
+                for _ in 1..width {
+                    let next_outer_arg = outer_args[outer_pos];
+                    args.push(next_outer_arg);
+                    outer_pos += 1;
+                }
             }
         } else {
             // We do not differentiate with resprect to this argument.
@@ -135,6 +176,74 @@ fn match_args_from_caller_to_enzyme<'ll>(
     }
 }
 
+// On LLVM-IR, we can luckily declare __enzyme_ functions without specifying the input
+// arguments. We do however need to declare them with their correct return type.
+// We already figured the correct return type out in our frontend, when generating the outer_fn,
+// so we can now just go ahead and use that. This is not always trivial, e.g. because sret.
+// Beyond sret, this article describes our challenges nicely:
+// <https://yorickpeterse.com/articles/the-mess-that-is-handling-structure-arguments-and-returns-in-llvm/>
+// I.e. (i32, f32) will get merged into i64, but we don't handle that yet.
+fn compute_enzyme_fn_ty<'ll>(
+    cx: &SimpleCx<'ll>,
+    attrs: &AutoDiffAttrs,
+    fn_to_diff: &'ll Value,
+    outer_fn: &'ll Value,
+) -> &'ll llvm::Type {
+    let fn_ty = cx.get_type_of_global(outer_fn);
+    let mut ret_ty = cx.get_return_type(fn_ty);
+
+    let has_sret = has_sret(outer_fn);
+
+    if has_sret {
+        // Now we don't just forward the return type, so we have to figure it out based on the
+        // primal return type, in combination with the autodiff settings.
+        let fn_ty = cx.get_type_of_global(fn_to_diff);
+        let inner_ret_ty = cx.get_return_type(fn_ty);
+
+        let void_ty = unsafe { llvm::LLVMVoidTypeInContext(cx.llcx) };
+        if inner_ret_ty == void_ty {
+            // This indicates that even the inner function has an sret.
+            // Right now I only look for an sret in the outer function.
+            // This *probably* needs some extra handling, but I never ran
+            // into such a case. So I'll wait for user reports to have a test case.
+            bug!("sret in inner function");
+        }
+
+        if attrs.width == 1 {
+            todo!("Handle sret for scalar ad");
+        } else {
+            // First we check if we also have to deal with the primal return.
+            if attrs.mode.is_fwd() {
+                match attrs.ret_activity {
+                    DiffActivity::Dual => {
+                        let arr_ty =
+                            unsafe { llvm::LLVMArrayType2(inner_ret_ty, attrs.width as u64 + 1) };
+                        ret_ty = arr_ty;
+                    }
+                    DiffActivity::DualOnly => {
+                        let arr_ty =
+                            unsafe { llvm::LLVMArrayType2(inner_ret_ty, attrs.width as u64) };
+                        ret_ty = arr_ty;
+                    }
+                    DiffActivity::Const => {
+                        todo!("Not sure, do we need to do something here?");
+                    }
+                    _ => {
+                        bug!("unreachable");
+                    }
+                }
+            } else if attrs.mode.is_rev() {
+                todo!("Handle sret for reverse mode");
+            } else {
+                bug!("unreachable");
+            }
+        }
+    }
+
+    // LLVM can figure out the input types on it's own, so we take a shortcut here.
+    unsafe { llvm::LLVMFunctionType(ret_ty, ptr::null(), 0, True) }
+}
+
 /// When differentiating `fn_to_diff`, take a `outer_fn` and generate another
 /// function with expected naming and calling conventions[^1] which will be
 /// discovered by the enzyme LLVM pass and its body populated with the differentiated
@@ -197,17 +306,9 @@ fn generate_enzyme_call<'ll>(
     // }
     // ```
     unsafe {
-        // On LLVM-IR, we can luckily declare __enzyme_ functions without specifying the input
-        // arguments. We do however need to declare them with their correct return type.
-        // We already figured the correct return type out in our frontend, when generating the outer_fn,
-        // so we can now just go ahead and use that. FIXME(ZuseZ4): This doesn't handle sret yet.
-        let fn_ty = llvm::LLVMGlobalGetValueType(outer_fn);
-        let ret_ty = llvm::LLVMGetReturnType(fn_ty);
+        let enzyme_ty = compute_enzyme_fn_ty(cx, &attrs, fn_to_diff, outer_fn);
 
-        // LLVM can figure out the input types on it's own, so we take a shortcut here.
-        let enzyme_ty = llvm::LLVMFunctionType(ret_ty, ptr::null(), 0, True);
-
-        //FIXME(ZuseZ4): the CC/Addr/Vis values are best effort guesses, we should look at tests and
+        // FIXME(ZuseZ4): the CC/Addr/Vis values are best effort guesses, we should look at tests and
         // think a bit more about what should go here.
         let cc = llvm::LLVMGetFunctionCallConv(outer_fn);
         let ad_fn = declare_simple_fn(
@@ -240,14 +341,27 @@ fn generate_enzyme_call<'ll>(
         if matches!(attrs.ret_activity, DiffActivity::Dual | DiffActivity::Active) {
             args.push(cx.get_metadata_value(enzyme_primal_ret));
         }
+        if attrs.width > 1 {
+            let enzyme_width = cx.create_metadata("enzyme_width".to_string()).unwrap();
+            args.push(cx.get_metadata_value(enzyme_width));
+            args.push(cx.get_const_i64(attrs.width as u64));
+        }
 
+        let has_sret = has_sret(outer_fn);
         let outer_args: Vec<&llvm::Value> = get_params(outer_fn);
-        match_args_from_caller_to_enzyme(&cx, &mut args, &attrs.input_activity, &outer_args);
+        match_args_from_caller_to_enzyme(
+            &cx,
+            attrs.width,
+            &mut args,
+            &attrs.input_activity,
+            &outer_args,
+            has_sret,
+        );
 
         let call = builder.call(enzyme_ty, ad_fn, &args, None);
 
         // This part is a bit iffy. LLVM requires that a call to an inlineable function has some
-        // metadata attachted to it, but we just created this code oota. Given that the
+        // metadata attached to it, but we just created this code oota. Given that the
         // differentiated function already has partly confusing metadata, and given that this
         // affects nothing but the auttodiff IR, we take a shortcut and just steal metadata from the
         // dummy code which we inserted at a higher level.
@@ -268,7 +382,22 @@ fn generate_enzyme_call<'ll>(
         // Now that we copied the metadata, get rid of dummy code.
         llvm::LLVMRustEraseInstUntilInclusive(entry, last_inst);
 
-        if cx.val_ty(call) == cx.type_void() {
+        if cx.val_ty(call) == cx.type_void() || has_sret {
+            if has_sret {
+                // This is what we already have in our outer_fn (shortened):
+                // define void @_foo(ptr <..> sret([32 x i8]) initializes((0, 32)) %0, <...>) {
+                //   %7 = call [4 x double] (...) @__enzyme_fwddiff_foo(ptr @square, metadata !"enzyme_width", i64 4, <...>)
+                //   <Here we are, we want to add the following two lines>
+                //   store [4 x double] %7, ptr %0, align 8
+                //   ret void
+                // }
+
+                // now store the result of the enzyme call into the sret pointer.
+                let sret_ptr = outer_args[0];
+                let call_ty = cx.val_ty(call);
+                assert!(llvm::LLVMRustIsArrayTy(call_ty));
+                llvm::LLVMBuildStore(&builder.llbuilder, call, sret_ptr);
+            }
             builder.ret_void();
         } else {
             builder.ret(call);
@@ -300,8 +429,7 @@ pub(crate) fn differentiate<'ll>(
     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));
+        return Err(diag_handler.handle().emit_almost_fatal(AutoDiffWithoutEnable));
     }
 
     // Before dumping the module, we want all the TypeTrees to become part of the module.
diff --git a/compiler/rustc_codegen_llvm/src/consts.rs b/compiler/rustc_codegen_llvm/src/consts.rs
@@ -425,7 +425,7 @@ impl<'ll> CodegenCx<'ll, '_> {
             let val_llty = self.val_ty(v);
 
             let g = self.get_static_inner(def_id, val_llty);
-            let llty = llvm::LLVMGlobalGetValueType(g);
+            let llty = self.get_type_of_global(g);
 
             let g = if val_llty == llty {
                 g
diff --git a/compiler/rustc_codegen_llvm/src/context.rs b/compiler/rustc_codegen_llvm/src/context.rs
@@ -38,7 +38,7 @@ use crate::debuginfo::metadata::apply_vcall_visibility_metadata;
 use crate::llvm::Metadata;
 use crate::type_::Type;
 use crate::value::Value;
-use crate::{attributes, coverageinfo, debuginfo, llvm, llvm_util};
+use crate::{attributes, common, coverageinfo, debuginfo, llvm, llvm_util};
 
 /// `TyCtxt` (and related cache datastructures) can't be move between threads.
 /// However, there are various cx related functions which we want to be available to the builder and
@@ -643,7 +643,18 @@ impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> {
         llvm::set_section(g, c"llvm.metadata");
     }
 }
-
+impl<'ll> SimpleCx<'ll> {
+    pub(crate) fn get_return_type(&self, ty: &'ll Type) -> &'ll Type {
+        assert!(unsafe { llvm::LLVMRustIsFunctionTy(ty) });
+        unsafe { llvm::LLVMGetReturnType(ty) }
+    }
+    pub(crate) fn get_type_of_global(&self, val: &'ll Value) -> &'ll Type {
+        unsafe { llvm::LLVMGlobalGetValueType(val) }
+    }
+    pub(crate) fn val_ty(&self, v: &'ll Value) -> &'ll Type {
+        common::val_ty(v)
+    }
+}
 impl<'ll> SimpleCx<'ll> {
     pub(crate) fn new(
         llmod: &'ll llvm::Module,
@@ -660,6 +671,11 @@ impl<'ll, CX: Borrow<SCx<'ll>>> GenericCx<'ll, CX> {
         llvm::LLVMMetadataAsValue(self.llcx(), metadata)
     }
 
+    pub(crate) fn get_const_i64(&self, n: u64) -> &'ll Value {
+        let ty = unsafe { llvm::LLVMInt64TypeInContext(self.llcx()) };
+        unsafe { llvm::LLVMConstInt(ty, n, llvm::False) }
+    }
+
     pub(crate) fn get_function(&self, name: &str) -> Option<&'ll Value> {
         let name = SmallCStr::new(name);
         unsafe { llvm::LLVMGetNamedFunction((**self).borrow().llmod, name.as_ptr()) }
diff --git a/compiler/rustc_codegen_llvm/src/llvm/enzyme_ffi.rs b/compiler/rustc_codegen_llvm/src/llvm/enzyme_ffi.rs
@@ -4,7 +4,7 @@
 use libc::{c_char, c_uint};
 
 use super::MetadataKindId;
-use super::ffi::{BasicBlock, Metadata, Module, Type, Value};
+use super::ffi::{AttributeKind, BasicBlock, Metadata, Module, Type, Value};
 use crate::llvm::Bool;
 
 #[link(name = "llvm-wrapper", kind = "static")]
@@ -17,6 +17,11 @@ unsafe extern "C" {
     pub(crate) fn LLVMRustEraseInstFromParent(V: &Value);
     pub(crate) fn LLVMRustGetTerminator<'a>(B: &BasicBlock) -> &'a Value;
     pub(crate) fn LLVMRustVerifyFunction(V: &Value, action: LLVMRustVerifierFailureAction) -> Bool;
+    pub(crate) fn LLVMRustHasAttributeAtIndex(V: &Value, i: c_uint, Kind: AttributeKind) -> bool;
+
+    pub(crate) fn LLVMRustIsFunctionTy(Ty: &Type) -> bool;
+    pub(crate) fn LLVMRustIsArrayTy(Ty: &Type) -> bool;
+    pub(crate) fn LLVMRustGetArrayNumElements(Ty: &Type) -> u64;
 }
 
 unsafe extern "C" {
diff --git a/compiler/rustc_llvm/llvm-wrapper/RustWrapper.cpp b/compiler/rustc_llvm/llvm-wrapper/RustWrapper.cpp
@@ -384,6 +384,12 @@ static inline void AddAttributes(T *t, unsigned Index, LLVMAttributeRef *Attrs,
   t->setAttributes(PALNew);
 }
 
+extern "C" bool LLVMRustHasAttributeAtIndex(LLVMValueRef Fn, unsigned Index,
+                                            LLVMRustAttributeKind RustAttr) {
+  Function *F = unwrap<Function>(Fn);
+  return F->hasParamAttribute(Index, fromRust(RustAttr));
+}
+
 extern "C" void LLVMRustAddFunctionAttributes(LLVMValueRef Fn, unsigned Index,
                                               LLVMAttributeRef *Attrs,
                                               size_t AttrsLen) {
@@ -635,6 +641,17 @@ static InlineAsm::AsmDialect fromRust(LLVMRustAsmDialect Dialect) {
     report_fatal_error("bad AsmDialect.");
   }
 }
+extern "C" bool LLVMRustIsFunctionTy(LLVMTypeRef Ty) {
+  return unwrap(Ty)->isFunctionTy();
+}
+
+extern "C" bool LLVMRustIsArrayTy(LLVMTypeRef Ty) {
+  return unwrap(Ty)->isArrayTy();
+}
+
+extern "C" uint64_t LLVMRustGetArrayNumElements(LLVMTypeRef Ty) {
+  return unwrap(Ty)->getArrayNumElements();
+}
 
 extern "C" LLVMValueRef
 LLVMRustInlineAsm(LLVMTypeRef Ty, char *AsmString, size_t AsmStringLen,

Original file line number	Diff line number	Diff line change
`@@ -425,7 +425,7 @@ impl<'ll> CodegenCx<'ll, '_> {`
`425`	`425`	`let val_llty = self.val_ty(v);`
`426`	`426`
`427`	`427`	`let g = self.get_static_inner(def_id, val_llty);`
`428`		`- let llty = llvm::LLVMGlobalGetValueType(g);`
	`428`	`+ let llty = self.get_type_of_global(g);`
`429`	`429`
`430`	`430`	`let g = if val_llty == llty {`
`431`	`431`	`g`