rustc_target: Add alignment to indirectly-passed by-value types, correcting the

alignment of `byval` on x86 in the process.

Commit 88e4d2c from five years ago removed
support for alignment on indirectly-passed arguments because of problems with
the `i686-pc-windows-msvc` target. Unfortunately, the `memcpy` optimizations I
recently added to LLVM 16 depend on this to forward `memcpy`s. This commit
attempts to fix the problems with `byval` parameters on that target and now
correctly adds the `align` attribute.

The problem is summarized in [this comment] by @eddyb. Briefly, 32-bit x86 has
special alignment rules for `byval` parameters: for the most part, their
alignment is forced to 4. This is not well-documented anywhere but in the Clang
source. I looked at the logic in Clang `TargetInfo.cpp` and tried to replicate
it here. The relevant methods in that file are
`X86_32ABIInfo::getIndirectResult()` and
`X86_32ABIInfo::getTypeStackAlignInBytes()`. The `align` parameter attribute
for `byval` parameters in LLVM must match the platform ABI, or miscompilations
will occur. Note that this doesn't use the approach suggested by eddyb, because
I felt it was overkill to store the alignment in `on_stack` when special
handling is really only needed for 32-bit x86.

As a side effect, this should fix rust-lang#80127, because it will make the `align`
parameter attribute for `byval` parameters match the platform ABI on LLVM
x86-64.

[this comment]: rust-lang#80822 (comment)

Author: pcwalton

compiler/rustc_target/src/abi/call/m68k.rs

@@ -10,7 +10,7 @@ fn classify_ret<Ty>(ret: &mut ArgAbi<'_, Ty>) {
 
 fn classify_arg<Ty>(arg: &mut ArgAbi<'_, Ty>) {
     if arg.layout.is_aggregate() {
-        arg.make_indirect_byval();
+        arg.make_indirect_byval(None);
     } else {
         arg.extend_integer_width_to(32);
     }

compiler/rustc_target/src/abi/call/mod.rs

@@ -494,9 +494,7 @@ impl<'a, Ty> ArgAbi<'a, Ty> {
             .set(ArgAttribute::NonNull)
             .set(ArgAttribute::NoUndef);
         attrs.pointee_size = layout.size;
-        // FIXME(eddyb) We should be doing this, but at least on
-        // i686-pc-windows-msvc, it results in wrong stack offsets.
-        // attrs.pointee_align = Some(layout.align.abi);
+        attrs.pointee_align = Some(layout.align.abi);
 
         let extra_attrs = layout.is_unsized().then_some(ArgAttributes::new());
 
@@ -513,11 +511,19 @@ impl<'a, Ty> ArgAbi<'a, Ty> {
         self.mode = Self::indirect_pass_mode(&self.layout);
     }
 
-    pub fn make_indirect_byval(&mut self) {
+    pub fn make_indirect_byval(&mut self, byval_align: Option<Align>) {
         self.make_indirect();
         match self.mode {
-            PassMode::Indirect { attrs: _, extra_attrs: _, ref mut on_stack } => {
+            PassMode::Indirect { ref mut attrs, extra_attrs: _, ref mut on_stack } => {
                 *on_stack = true;
+
+                // Some platforms, like 32-bit x86, change the alignment of the type when passing
+                // `byval`. Account for that.
+                if let Some(byval_align) = byval_align {
+                    // On all targets with byval align this is currently true, so let's assert it.
+                    debug_assert!(byval_align >= Align::from_bytes(4).unwrap());
+                    attrs.pointee_align = Some(byval_align);
+                }
             }
             _ => unreachable!(),
         }
@@ -644,7 +650,8 @@ impl<'a, Ty> FnAbi<'a, Ty> {
     {
         if abi == spec::abi::Abi::X86Interrupt {
             if let Some(arg) = self.args.first_mut() {
-                arg.make_indirect_byval();
+                // FIXME(pcwalton): This probably should use the x86 `byval` ABI...
+                arg.make_indirect_byval(None);
             }
             return Ok(());
         }

compiler/rustc_target/src/abi/call/wasm.rs

@@ -36,7 +36,7 @@ where
 {
     arg.extend_integer_width_to(32);
     if arg.layout.is_aggregate() && !unwrap_trivial_aggregate(cx, arg) {
-        arg.make_indirect_byval();
+        arg.make_indirect_byval(None);
     }
 }
 

compiler/rustc_target/src/abi/call/x86.rs

@@ -1,5 +1,5 @@
 use crate::abi::call::{ArgAttribute, FnAbi, PassMode, Reg, RegKind};
-use crate::abi::{HasDataLayout, TyAbiInterface};
+use crate::abi::{Align, HasDataLayout, TyAbiInterface};
 use crate::spec::HasTargetSpec;
 
 #[derive(PartialEq)]
@@ -53,11 +53,38 @@ where
         if arg.is_ignore() {
             continue;
         }
-        if arg.layout.is_aggregate() {
-            arg.make_indirect_byval();
-        } else {
+        if !arg.layout.is_aggregate() {
             arg.extend_integer_width_to(32);
+            continue;
         }
+
+        // We need to compute the alignment of the `byval` argument. The rules can be found in
+        // `X86_32ABIInfo::getTypeStackAlignInBytes` in Clang's `TargetInfo.cpp`. Summarized here,
+        // they are:
+        //
+        // 1. If the natural alignment of the type is less than or equal to 4, the alignment is 4.
+        //
+        // 2. Otherwise, on Linux, the alignment of any vector type is the natural alignment.
+        // (This doesn't matter here because we ensure we have an aggregate with the check above.)
+        //
+        // 3. Otherwise, on Apple platforms, the alignment of anything that contains a vector type
+        // is 16.
+        //
+        // 4. If none of these conditions are true, the alignment is 4.
+        let t = cx.target_spec();
+        let align_4 = Align::from_bytes(4).unwrap();
+        let align_16 = Align::from_bytes(16).unwrap();
+        let byval_align = if arg.layout.align.abi < align_4 {
+            align_4
+        } else if t.is_like_osx && arg.layout.align.abi >= align_16 {
+            // FIXME(pcwalton): This is dubious--we should actually be looking inside the type to
+            // determine if it contains SIMD vector values--but I think it's fine?
+            align_16
+        } else {
+            align_4
+        };
+
+        arg.make_indirect_byval(Some(byval_align));
     }
 
     if flavor == Flavor::FastcallOrVectorcall {

compiler/rustc_target/src/abi/call/x86_64.rs

@@ -213,7 +213,7 @@ where
         match cls_or_mem {
             Err(Memory) => {
                 if is_arg {
-                    arg.make_indirect_byval();
+                    arg.make_indirect_byval(None);
                 } else {
                     // `sret` parameter thus one less integer register available
                     arg.make_indirect();

tests/codegen/align-byval.rs

@@ -0,0 +1,56 @@
+// ignore-x86
+// ignore-aarch64
+// ignore-aarch64_be
+// ignore-arm
+// ignore-armeb
+// ignore-avr
+// ignore-bpfel
+// ignore-bpfeb
+// ignore-hexagon
+// ignore-mips
+// ignore-mips64
+// ignore-msp430
+// ignore-powerpc64
+// ignore-powerpc64le
+// ignore-powerpc
+// ignore-r600
+// ignore-amdgcn
+// ignore-sparc
+// ignore-sparcv9
+// ignore-sparcel
+// ignore-s390x
+// ignore-tce
+// ignore-thumb
+// ignore-thumbeb
+// ignore-xcore
+// ignore-nvptx
+// ignore-nvptx64
+// ignore-le32
+// ignore-le64
+// ignore-amdil
+// ignore-amdil64
+// ignore-hsail
+// ignore-hsail64
+// ignore-spir
+// ignore-spir64
+// ignore-kalimba
+// ignore-shave
+//
+// Tests that `byval` alignment is properly specified (#80127).
+// The only targets that use `byval` are m68k, wasm, x86-64, and x86. Note that
+// x86 has special rules (see #103830), and it's therefore ignored here.
+
+#[repr(C)]
+#[repr(align(16))]
+struct Foo {
+    a: [i32; 16],
+}
+
+extern "C" {
+    // CHECK: declare void @f({{.*}}byval(%Foo) align 16{{.*}})
+    fn f(foo: Foo);
+}
+
+pub fn main() {
+    unsafe { f(Foo { a: [1; 16] }) }
+}

tests/codegen/function-arguments-noopt.rs

@@ -42,7 +42,7 @@ pub fn borrow_call(x: &i32, f: fn(&i32) -> &i32) -> &i32 {
   f(x)
 }
 
-// CHECK: void @struct_({{%S\*|ptr}} sret(%S){{( %_0)?}}, {{%S\*|ptr}} %x)
+// CHECK: void @struct_({{%S\*|ptr}} sret(%S) align 4{{( %_0)?}}, {{%S\*|ptr}} align 4 %x)
 #[no_mangle]
 pub fn struct_(x: S) -> S {
   x
@@ -51,7 +51,7 @@ pub fn struct_(x: S) -> S {
 // CHECK-LABEL: @struct_call
 #[no_mangle]
 pub fn struct_call(x: S, f: fn(S) -> S) -> S {
-  // CHECK: call void %f({{%S\*|ptr}} sret(%S){{( %_0)?}}, {{%S\*|ptr}} %{{.+}})
+  // CHECK: call void %f({{%S\*|ptr}} sret(%S) align 4{{( %_0)?}}, {{%S\*|ptr}} align 4 %{{.+}})
   f(x)
 }
 

tests/codegen/function-arguments.rs

@@ -142,7 +142,7 @@ pub fn mutable_notunpin_borrow(_: &mut NotUnpin) {
 pub fn notunpin_borrow(_: &NotUnpin) {
 }
 
-// CHECK: @indirect_struct({{%S\*|ptr}} noalias nocapture noundef readonly dereferenceable(32) %_1)
+// CHECK: @indirect_struct({{%S\*|ptr}} noalias nocapture noundef readonly align 4 dereferenceable(32) %_1)
 #[no_mangle]
 pub fn indirect_struct(_: S) {
 }
@@ -188,7 +188,7 @@ pub fn notunpin_box(x: Box<NotUnpin>) -> Box<NotUnpin> {
   x
 }
 
-// CHECK: @struct_return({{%S\*|ptr}} noalias nocapture noundef sret(%S) dereferenceable(32){{( %_0)?}})
+// CHECK: @struct_return({{%S\*|ptr}} noalias nocapture noundef sret(%S) align 4 dereferenceable(32){{( %_0)?}})
 #[no_mangle]
 pub fn struct_return() -> S {
   S {

tests/run-make/extern-fn-explicit-align/Makefile

@@ -0,0 +1,5 @@
+include ../tools.mk
+
+all: $(call NATIVE_STATICLIB,test)
+	$(RUSTC) test.rs
+	$(call RUN,test) || exit 1

tests/run-make/extern-fn-explicit-align/test.c

@@ -0,0 +1,35 @@
+#include <assert.h>
+#include <stdbool.h>
+#include <stdint.h>
+#include <string.h>
+
+struct TwoU64s
+{
+    uint64_t a;
+    uint64_t b;
+} __attribute__((aligned(16)));
+
+struct BoolAndU32
+{
+    bool a;
+    uint32_t b;
+};
+
+int32_t many_args(
+    void *a,
+    void *b,
+    const char *c,
+    uint64_t d,
+    bool e,
+    struct BoolAndU32 f,
+    void *g,
+    struct TwoU64s h,
+    void *i,
+    void *j,
+    void *k,
+    void *l,
+    const char *m)
+{
+    assert(strcmp(m, "Hello world") == 0);
+    return 0;
+}