Auto merge of #121577 - erikdesjardins:struct, r=<try>

Stop using LLVM struct types for alloca, byval, sret, and many GEPs

This is an extension of #98615, extending the removal from field offsets to most places that it's feasible right now. (It might make sense to split this PR up, but I want to test perf with everything.)

For `alloca`, `byval`, and `sret`, the type has no semantic meaning, only the size matters\*†. Using `[N x i8]` is a more direct way to specify that we want `N` bytes, and avoids relying on LLVM's layout algorithm. Particularly for `alloca`, it is likely that a future LLVM will change to a representation where you only specify the size.

For GEPs, upstream LLVM is in the beginning stages of [migrating to `ptradd`](https://discourse.llvm.org/t/rfc-replacing-getelementptr-with-ptradd/68699). LLVM 19 will [canonicalize](llvm/llvm-project#68882) all constant-offset GEPs to i8, which is the same thing we do here.

\*: Since we always explicitly specify the alignment. For `byval`, this wasn't the case until #112157.

†: For `byval`, the hidden copy may be impacted by padding in the LLVM struct type, i.e. padding bytes may not be copied. (I'm not sure if this is done today, but I think it would be legal.) But we manually pad our LLVM struct types specifically to avoid there ever being LLVM-visible padding, so that shouldn't be an issue here.

r? `@ghost`

compiler/rustc_codegen_gcc/src/builder.rs

@@ -734,7 +734,18 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
         self.gcc_checked_binop(oop, typ, lhs, rhs)
     }
 
-    fn alloca(&mut self, ty: Type<'gcc>, align: Align) -> RValue<'gcc> {
+    fn alloca(&mut self, size: Size, align: Align) -> RValue<'gcc> {
+        let ty = self.cx.type_array(self.cx.type_i8(), size.bytes()).get_aligned(align.bytes());
+        // TODO(antoyo): It might be better to return a LValue, but fixing the rustc API is non-trivial.
+        self.stack_var_count.set(self.stack_var_count.get() + 1);
+        self.current_func().new_local(None, ty, &format!("stack_var_{}", self.stack_var_count.get())).get_address(None)
+    }
+
+    fn dynamic_alloca(&mut self, _len: RValue<'gcc>, _align: Align) -> RValue<'gcc> {
+        unimplemented!();
+    }
+
+    fn typed_alloca(&mut self, ty: Type<'gcc>, align: Align) -> RValue<'gcc> {
         // FIXME(antoyo): this check that we don't call get_aligned() a second time on a type.
         // Ideally, we shouldn't need to do this check.
         let aligned_type =
@@ -749,10 +760,6 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
         self.current_func().new_local(None, aligned_type, &format!("stack_var_{}", self.stack_var_count.get())).get_address(None)
     }
 
-    fn byte_array_alloca(&mut self, _len: RValue<'gcc>, _align: Align) -> RValue<'gcc> {
-        unimplemented!();
-    }
-
     fn load(&mut self, pointee_ty: Type<'gcc>, ptr: RValue<'gcc>, align: Align) -> RValue<'gcc> {
         let block = self.llbb();
         let function = block.get_function();
@@ -834,10 +841,17 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
             }
             else if let abi::Abi::ScalarPair(ref a, ref b) = place.layout.abi {
                 let b_offset = a.size(self).align_to(b.align(self).abi);
-                let pair_type = place.layout.gcc_type(self);
 
                 let mut load = |i, scalar: &abi::Scalar, align| {
-                    let llptr = self.struct_gep(pair_type, place.llval, i as u64);
+                    let llptr = if i == 0 {
+                        place.llval
+                    } else {
+                        self.inbounds_gep(
+                            self.type_i8(),
+                            place.llval,
+                            &[self.const_usize(b_offset.bytes())],
+                        )
+                    };
                     let llty = place.layout.scalar_pair_element_gcc_type(self, i);
                     let load = self.load(llty, llptr, align);
                     scalar_load_metadata(self, load, scalar);
@@ -971,33 +985,6 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
         result.get_address(None)
     }
 
-    fn struct_gep(&mut self, value_type: Type<'gcc>, ptr: RValue<'gcc>, idx: u64) -> RValue<'gcc> {
-        // FIXME(antoyo): it would be better if the API only called this on struct, not on arrays.
-        assert_eq!(idx as usize as u64, idx);
-        let value = ptr.dereference(None).to_rvalue();
-
-        if value_type.dyncast_array().is_some() {
-            let index = self.context.new_rvalue_from_long(self.u64_type, i64::try_from(idx).expect("i64::try_from"));
-            let element = self.context.new_array_access(None, value, index);
-            element.get_address(None)
-        }
-        else if let Some(vector_type) = value_type.dyncast_vector() {
-            let array_type = vector_type.get_element_type().make_pointer();
-            let array = self.bitcast(ptr, array_type);
-            let index = self.context.new_rvalue_from_long(self.u64_type, i64::try_from(idx).expect("i64::try_from"));
-            let element = self.context.new_array_access(None, array, index);
-            element.get_address(None)
-        }
-        else if let Some(struct_type) = value_type.is_struct() {
-            // NOTE: due to opaque pointers now being used, we need to bitcast here.
-            let ptr = self.bitcast_if_needed(ptr, value_type.make_pointer());
-            ptr.dereference_field(None, struct_type.get_field(idx as i32)).get_address(None)
-        }
-        else {
-            panic!("Unexpected type {:?}", value_type);
-        }
-    }
-
     /* Casts */
     fn trunc(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
         // TODO(antoyo): check that it indeed truncate the value.

compiler/rustc_codegen_gcc/src/intrinsic/mod.rs

@@ -490,7 +490,7 @@ impl<'gcc, 'tcx> ArgAbiExt<'gcc, 'tcx> for ArgAbi<'tcx, Ty<'tcx>> {
                 // We instead thus allocate some scratch space...
                 let scratch_size = cast.size(bx);
                 let scratch_align = cast.align(bx);
-                let llscratch = bx.alloca(cast.gcc_type(bx), scratch_align);
+                let llscratch = bx.alloca(scratch_size, scratch_align);
                 bx.lifetime_start(llscratch, scratch_size);
 
                 // ... where we first store the value...

compiler/rustc_codegen_gcc/src/intrinsic/simd.rs

@@ -16,7 +16,7 @@ use rustc_middle::span_bug;
 use rustc_middle::ty::layout::HasTyCtxt;
 use rustc_middle::ty::{self, Ty};
 use rustc_span::{sym, Span, Symbol};
-use rustc_target::abi::Align;
+use rustc_target::abi::{Align, Size};
 
 use crate::builder::Builder;
 #[cfg(feature = "master")]
@@ -363,7 +363,7 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>(
                 let ze = bx.zext(result, bx.type_ix(expected_bytes * 8));
 
                 // Convert the integer to a byte array
-                let ptr = bx.alloca(bx.type_ix(expected_bytes * 8), Align::ONE);
+                let ptr = bx.alloca(Size::from_bytes(expected_bytes), Align::ONE);
                 bx.store(ze, ptr, Align::ONE);
                 let array_ty = bx.type_array(bx.type_i8(), expected_bytes);
                 let ptr = bx.pointercast(ptr, bx.cx.type_ptr_to(array_ty));

compiler/rustc_codegen_gcc/src/type_of.rs

@@ -151,7 +151,6 @@ pub trait LayoutGccExt<'tcx> {
     fn immediate_gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc>;
     fn scalar_gcc_type_at<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, scalar: &abi::Scalar, offset: Size) -> Type<'gcc>;
     fn scalar_pair_element_gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, index: usize) -> Type<'gcc>;
-    fn gcc_field_index(&self, index: usize) -> u64;
     fn pointee_info_at<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, offset: Size) -> Option<PointeeInfo>;
 }
 
@@ -304,24 +303,6 @@ impl<'tcx> LayoutGccExt<'tcx> for TyAndLayout<'tcx> {
         self.scalar_gcc_type_at(cx, scalar, offset)
     }
 
-    fn gcc_field_index(&self, index: usize) -> u64 {
-        match self.abi {
-            Abi::Scalar(_) | Abi::ScalarPair(..) => {
-                bug!("TyAndLayout::gcc_field_index({:?}): not applicable", self)
-            }
-            _ => {}
-        }
-        match self.fields {
-            FieldsShape::Primitive | FieldsShape::Union(_) => {
-                bug!("TyAndLayout::gcc_field_index({:?}): not applicable", self)
-            }
-
-            FieldsShape::Array { .. } => index as u64,
-
-            FieldsShape::Arbitrary { .. } => 1 + (self.fields.memory_index(index) as u64) * 2,
-        }
-    }
-
     fn pointee_info_at<'a>(&self, cx: &CodegenCx<'a, 'tcx>, offset: Size) -> Option<PointeeInfo> {
         if let Some(&pointee) = cx.pointee_infos.borrow().get(&(self.ty, offset)) {
             return pointee;
@@ -351,10 +332,6 @@ impl<'gcc, 'tcx> LayoutTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
         layout.is_gcc_scalar_pair()
     }
 
-    fn backend_field_index(&self, layout: TyAndLayout<'tcx>, index: usize) -> u64 {
-        layout.gcc_field_index(index)
-    }
-
     fn scalar_pair_element_backend_type(&self, layout: TyAndLayout<'tcx>, index: usize, _immediate: bool) -> Type<'gcc> {
         layout.scalar_pair_element_gcc_type(self, index)
     }

compiler/rustc_codegen_llvm/src/abi.rs

@@ -234,7 +234,7 @@ impl<'ll, 'tcx> ArgAbiExt<'ll, 'tcx> for ArgAbi<'tcx, Ty<'tcx>> {
                 // We instead thus allocate some scratch space...
                 let scratch_size = cast.size(bx);
                 let scratch_align = cast.align(bx);
-                let llscratch = bx.alloca(cast.llvm_type(bx), scratch_align);
+                let llscratch = bx.alloca(scratch_size, scratch_align);
                 bx.lifetime_start(llscratch, scratch_size);
 
                 // ... where we first store the value...
@@ -424,7 +424,10 @@ impl<'ll, 'tcx> FnAbiLlvmExt<'ll, 'tcx> for FnAbi<'tcx, Ty<'tcx>> {
             PassMode::Indirect { attrs, meta_attrs: _, on_stack } => {
                 assert!(!on_stack);
                 let i = apply(attrs);
-                let sret = llvm::CreateStructRetAttr(cx.llcx, self.ret.layout.llvm_type(cx));
+                let sret = llvm::CreateStructRetAttr(
+                    cx.llcx,
+                    cx.type_array(cx.type_i8(), self.ret.layout.size.bytes()),
+                );
                 attributes::apply_to_llfn(llfn, llvm::AttributePlace::Argument(i), &[sret]);
             }
             PassMode::Cast { cast, pad_i32: _ } => {
@@ -437,7 +440,10 @@ impl<'ll, 'tcx> FnAbiLlvmExt<'ll, 'tcx> for FnAbi<'tcx, Ty<'tcx>> {
                 PassMode::Ignore => {}
                 PassMode::Indirect { attrs, meta_attrs: None, on_stack: true } => {
                     let i = apply(attrs);
-                    let byval = llvm::CreateByValAttr(cx.llcx, arg.layout.llvm_type(cx));
+                    let byval = llvm::CreateByValAttr(
+                        cx.llcx,
+                        cx.type_array(cx.type_i8(), arg.layout.size.bytes()),
+                    );
                     attributes::apply_to_llfn(llfn, llvm::AttributePlace::Argument(i), &[byval]);
                 }
                 PassMode::Direct(attrs)
@@ -486,7 +492,10 @@ impl<'ll, 'tcx> FnAbiLlvmExt<'ll, 'tcx> for FnAbi<'tcx, Ty<'tcx>> {
             PassMode::Indirect { attrs, meta_attrs: _, on_stack } => {
                 assert!(!on_stack);
                 let i = apply(bx.cx, attrs);
-                let sret = llvm::CreateStructRetAttr(bx.cx.llcx, self.ret.layout.llvm_type(bx));
+                let sret = llvm::CreateStructRetAttr(
+                    bx.cx.llcx,
+                    bx.cx.type_array(bx.cx.type_i8(), self.ret.layout.size.bytes()),
+                );
                 attributes::apply_to_callsite(callsite, llvm::AttributePlace::Argument(i), &[sret]);
             }
             PassMode::Cast { cast, pad_i32: _ } => {
@@ -513,7 +522,10 @@ impl<'ll, 'tcx> FnAbiLlvmExt<'ll, 'tcx> for FnAbi<'tcx, Ty<'tcx>> {
                 PassMode::Ignore => {}
                 PassMode::Indirect { attrs, meta_attrs: None, on_stack: true } => {
                     let i = apply(bx.cx, attrs);
-                    let byval = llvm::CreateByValAttr(bx.cx.llcx, arg.layout.llvm_type(bx));
+                    let byval = llvm::CreateByValAttr(
+                        bx.cx.llcx,
+                        bx.cx.type_array(bx.cx.type_i8(), arg.layout.size.bytes()),
+                    );
                     attributes::apply_to_callsite(
                         callsite,
                         llvm::AttributePlace::Argument(i),

compiler/rustc_codegen_llvm/src/builder.rs

@@ -466,7 +466,9 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         val
     }
 
-    fn alloca(&mut self, ty: &'ll Type, align: Align) -> &'ll Value {
+    fn alloca(&mut self, size: Size, align: Align) -> &'ll Value {
+        let ty = self.cx().type_array(self.cx().type_i8(), size.bytes());
+
         let mut bx = Builder::with_cx(self.cx);
         bx.position_at_start(unsafe { llvm::LLVMGetFirstBasicBlock(self.llfn()) });
         unsafe {
@@ -476,10 +478,20 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         }
     }
 
-    fn byte_array_alloca(&mut self, len: &'ll Value, align: Align) -> &'ll Value {
+    fn dynamic_alloca(&mut self, size: &'ll Value, align: Align) -> &'ll Value {
         unsafe {
             let alloca =
-                llvm::LLVMBuildArrayAlloca(self.llbuilder, self.cx().type_i8(), len, UNNAMED);
+                llvm::LLVMBuildArrayAlloca(self.llbuilder, self.cx().type_i8(), size, UNNAMED);
+            llvm::LLVMSetAlignment(alloca, align.bytes() as c_uint);
+            alloca
+        }
+    }
+
+    fn typed_alloca(&mut self, ty: &'ll Type, align: Align) -> &'ll Value {
+        let mut bx = Builder::with_cx(self.cx);
+        bx.position_at_start(unsafe { llvm::LLVMGetFirstBasicBlock(self.llfn()) });
+        unsafe {
+            let alloca = llvm::LLVMBuildAlloca(bx.llbuilder, ty, UNNAMED);
             llvm::LLVMSetAlignment(alloca, align.bytes() as c_uint);
             alloca
         }
@@ -603,11 +615,7 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
                 let llptr = if i == 0 {
                     place.llval
                 } else {
-                    self.inbounds_gep(
-                        self.type_i8(),
-                        place.llval,
-                        &[self.const_usize(b_offset.bytes())],
-                    )
+                    self.inbounds_ptradd(place.llval, self.const_usize(b_offset.bytes()))
                 };
                 let llty = place.layout.scalar_pair_element_llvm_type(self, i, false);
                 let load = self.load(llty, llptr, align);
@@ -778,11 +786,6 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         }
     }
 
-    fn struct_gep(&mut self, ty: &'ll Type, ptr: &'ll Value, idx: u64) -> &'ll Value {
-        assert_eq!(idx as c_uint as u64, idx);
-        unsafe { llvm::LLVMBuildStructGEP2(self.llbuilder, ty, ptr, idx as c_uint, UNNAMED) }
-    }
-
     /* Casts */
     fn trunc(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
         unsafe { llvm::LLVMBuildTrunc(self.llbuilder, val, dest_ty, UNNAMED) }

compiler/rustc_codegen_llvm/src/intrinsic.rs

@@ -18,7 +18,7 @@ use rustc_middle::ty::layout::{FnAbiOf, HasTyCtxt, LayoutOf};
 use rustc_middle::ty::{self, GenericArgsRef, Ty};
 use rustc_middle::{bug, span_bug};
 use rustc_span::{sym, symbol::kw, Span, Symbol};
-use rustc_target::abi::{self, Align, HasDataLayout, Primitive};
+use rustc_target::abi::{self, Align, HasDataLayout, Primitive, Size};
 use rustc_target::spec::{HasTargetSpec, PanicStrategy};
 
 use std::cmp::Ordering;
@@ -565,7 +565,8 @@ fn codegen_msvc_try<'ll>(
         //
         // More information can be found in libstd's seh.rs implementation.
         let ptr_align = bx.tcx().data_layout.pointer_align.abi;
-        let slot = bx.alloca(bx.type_ptr(), ptr_align);
+        let ptr_size = bx.tcx().data_layout.pointer_size;
+        let slot = bx.alloca(ptr_size, ptr_align);
         let try_func_ty = bx.type_func(&[bx.type_ptr()], bx.type_void());
         bx.invoke(try_func_ty, None, None, try_func, &[data], normal, catchswitch, None);
 
@@ -838,7 +839,7 @@ fn codegen_emcc_try<'ll>(
         let ptr_align = bx.tcx().data_layout.pointer_align.abi;
         let i8_align = bx.tcx().data_layout.i8_align.abi;
         let catch_data_type = bx.type_struct(&[bx.type_ptr(), bx.type_bool()], false);
-        let catch_data = bx.alloca(catch_data_type, ptr_align);
+        let catch_data = bx.typed_alloca(catch_data_type, ptr_align);
         let catch_data_0 =
             bx.inbounds_gep(catch_data_type, catch_data, &[bx.const_usize(0), bx.const_usize(0)]);
         bx.store(ptr, catch_data_0, ptr_align);
@@ -1289,7 +1290,7 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
                 let ze = bx.zext(i_, bx.type_ix(expected_bytes * 8));
 
                 // Convert the integer to a byte array
-                let ptr = bx.alloca(bx.type_ix(expected_bytes * 8), Align::ONE);
+                let ptr = bx.alloca(Size::from_bytes(expected_bytes), Align::ONE);
                 bx.store(ze, ptr, Align::ONE);
                 let array_ty = bx.type_array(bx.type_i8(), expected_bytes);
                 return Ok(bx.load(array_ty, ptr, Align::ONE));

compiler/rustc_codegen_llvm/src/llvm/ffi.rs

@@ -1301,13 +1301,6 @@ extern "C" {
         NumIndices: c_uint,
         Name: *const c_char,
     ) -> &'a Value;
-    pub fn LLVMBuildStructGEP2<'a>(
-        B: &Builder<'a>,
-        Ty: &'a Type,
-        Pointer: &'a Value,
-        Idx: c_uint,
-        Name: *const c_char,
-    ) -> &'a Value;
 
     // Casts
     pub fn LLVMBuildTrunc<'a>(

compiler/rustc_codegen_llvm/src/type_.rs

@@ -250,9 +250,6 @@ impl<'ll, 'tcx> LayoutTypeMethods<'tcx> for CodegenCx<'ll, 'tcx> {
     fn is_backend_scalar_pair(&self, layout: TyAndLayout<'tcx>) -> bool {
         layout.is_llvm_scalar_pair()
     }
-    fn backend_field_index(&self, layout: TyAndLayout<'tcx>, index: usize) -> u64 {
-        layout.llvm_field_index(self, index)
-    }
     fn scalar_pair_element_backend_type(
         &self,
         layout: TyAndLayout<'tcx>,

compiler/rustc_codegen_llvm/src/type_of.rs

@@ -174,7 +174,6 @@ pub trait LayoutLlvmExt<'tcx> {
         index: usize,
         immediate: bool,
     ) -> &'a Type;
-    fn llvm_field_index<'a>(&self, cx: &CodegenCx<'a, 'tcx>, index: usize) -> u64;
     fn scalar_copy_llvm_type<'a>(&self, cx: &CodegenCx<'a, 'tcx>) -> Option<&'a Type>;
 }
 
@@ -324,42 +323,6 @@ impl<'tcx> LayoutLlvmExt<'tcx> for TyAndLayout<'tcx> {
         self.scalar_llvm_type_at(cx, scalar)
     }
 
-    fn llvm_field_index<'a>(&self, cx: &CodegenCx<'a, 'tcx>, index: usize) -> u64 {
-        match self.abi {
-            Abi::Scalar(_) | Abi::ScalarPair(..) => {
-                bug!("TyAndLayout::llvm_field_index({:?}): not applicable", self)
-            }
-            _ => {}
-        }
-        match self.fields {
-            FieldsShape::Primitive | FieldsShape::Union(_) => {
-                bug!("TyAndLayout::llvm_field_index({:?}): not applicable", self)
-            }
-
-            FieldsShape::Array { .. } => index as u64,
-
-            FieldsShape::Arbitrary { .. } => {
-                let variant_index = match self.variants {
-                    Variants::Single { index } => Some(index),
-                    _ => None,
-                };
-
-                // Look up llvm field if indexes do not match memory order due to padding. If
-                // `field_remapping` is `None` no padding was used and the llvm field index
-                // matches the memory index.
-                match cx.type_lowering.borrow().get(&(self.ty, variant_index)) {
-                    Some(TypeLowering { field_remapping: Some(ref remap), .. }) => {
-                        remap[index] as u64
-                    }
-                    Some(_) => self.fields.memory_index(index) as u64,
-                    None => {
-                        bug!("TyAndLayout::llvm_field_index({:?}): type info not found", self)
-                    }
-                }
-            }
-        }
-    }
-
     fn scalar_copy_llvm_type<'a>(&self, cx: &CodegenCx<'a, 'tcx>) -> Option<&'a Type> {
         debug_assert!(self.is_sized());