diff --git a/src/cargo/core/compiler/standard_lib.rs b/src/cargo/core/compiler/standard_lib.rs
index 6b76a5681be..7248495cec9 100644
--- a/src/cargo/core/compiler/standard_lib.rs
+++ b/src/cargo/core/compiler/standard_lib.rs
@@ -8,6 +8,7 @@ use crate::core::resolver::HasDevUnits;
 use crate::core::{Dependency, PackageId, PackageSet, Resolve, SourceId, Workspace};
 use crate::ops::{self, Packages};
 use crate::util::errors::CargoResult;
+use std::collections::hash_map::Entry;
 use std::collections::{HashMap, HashSet};
 use std::env;
 use std::path::PathBuf;
@@ -133,7 +134,8 @@ pub fn generate_std_roots(
     crates: &[String],
     std_resolve: &Resolve,
     std_features: &ResolvedFeatures,
-    kinds: &[CompileKind],
+    requested_kinds: &[CompileKind],
+    explicit_host_kind: CompileKind,
     package_set: &PackageSet<'_>,
     interner: &UnitInterner,
     profiles: &Profiles,
@@ -147,41 +149,58 @@ pub fn generate_std_roots(
     let std_pkgs = package_set.get_many(std_ids)?;
     // Generate a map of Units for each kind requested.
     let mut ret = HashMap::new();
-    for pkg in std_pkgs {
-        let lib = pkg
-            .targets()
-            .iter()
-            .find(|t| t.is_lib())
-            .expect("std has a lib");
-        let unit_for = UnitFor::new_normal();
-        // I don't think we need to bother with Check here, the difference
-        // in time is minimal, and the difference in caching is
-        // significant.
-        let mode = CompileMode::Build;
-        let features = std_features.activated_features(pkg.package_id(), FeaturesFor::NormalOrDev);
 
-        for kind in kinds {
-            let list = ret.entry(*kind).or_insert_with(Vec::new);
+    let std_pkg_infos: Vec<_> = std_pkgs
+        .iter()
+        .map(|pkg| {
+            let lib = pkg
+                .targets()
+                .iter()
+                .find(|t| t.is_lib())
+                .expect("std has a lib");
+            let unit_for = UnitFor::new_normal();
+            let features =
+                std_features.activated_features(pkg.package_id(), FeaturesFor::NormalOrDev);
+            (pkg, lib, unit_for, features)
+        })
+        .collect();
+
+    for kind in package_set
+        .packages()
+        .flat_map(|pkg| pkg.explicit_kinds(requested_kinds, explicit_host_kind))
+    {
+        let e = match ret.entry(kind) {
+            Entry::Vacant(e) => e,
+            Entry::Occupied(_) => continue,
+        };
+        let units = std_pkg_infos.iter().map(|(pkg, lib, unit_for, features)| {
+            // I don't think we need to bother with Check here, the difference
+            // in time is minimal, and the difference in caching is
+            // significant.
+            let mode = CompileMode::Build;
             let profile = profiles.get_profile(
                 pkg.package_id(),
                 /*is_member*/ false,
                 /*is_local*/ false,
-                unit_for,
+                *unit_for,
                 mode,
-                *kind,
+                kind,
             );
-            list.push(interner.intern(
+            interner.intern(
                 pkg,
                 lib,
                 profile,
-                *kind,
+                kind,
                 mode,
                 features.clone(),
                 /*is_std*/ true,
                 /*dep_hash*/ 0,
-            ));
-        }
+            )
+        });
+
+        e.insert(units.collect());
     }
+
     Ok(ret)
 }
 
diff --git a/src/cargo/core/package.rs b/src/cargo/core/package.rs
index 0ebe0277e78..e7bb403aca6 100644
--- a/src/cargo/core/package.rs
+++ b/src/cargo/core/package.rs
@@ -186,6 +186,34 @@ impl Package {
         self.targets().iter().any(|t| t.is_custom_build())
     }
 
+    /// Returns explicit kinds either forced by `forced-target` in `Cargo.toml`,
+    /// fallback to `default-target`, or specified in cli parameters.
+    ///
+    /// If `--target` has not been specified, then the return value
+    /// is the same as if `--target $HOST` was specified. See
+    /// `rebuild_unit_graph_shared` for why callers want to do this.
+    ///  However, if the package has its own `package.default-target`
+    ///  key, then that gets used instead of `$HOST`
+    pub fn explicit_kinds(
+        &self,
+        requested_kinds: &[CompileKind],
+        explicit_host_kind: CompileKind,
+    ) -> Vec<CompileKind> {
+        if let Some(k) = self.manifest().forced_kind() {
+            vec![k]
+        } else {
+            requested_kinds
+                .iter()
+                .map(|kind| match kind {
+                    CompileKind::Host => {
+                        self.manifest().default_kind().unwrap_or(explicit_host_kind)
+                    }
+                    CompileKind::Target(t) => CompileKind::Target(*t),
+                })
+                .collect()
+        }
+    }
+
     pub fn map_source(self, to_replace: SourceId, replace_with: SourceId) -> Package {
         Package {
             inner: Rc::new(PackageInner {
diff --git a/src/cargo/ops/cargo_compile.rs b/src/cargo/ops/cargo_compile.rs
index c34276a5fc1..c762952157b 100644
--- a/src/cargo/ops/cargo_compile.rs
+++ b/src/cargo/ops/cargo_compile.rs
@@ -404,12 +404,7 @@ pub fn create_bcx<'a, 'cfg>(
                 .shell()
                 .warn("-Zbuild-std does not currently fully support --build-plan")?;
         }
-        if build_config.requested_kinds[0].is_host() {
-            // TODO: This should eventually be fixed. Unfortunately it is not
-            // easy to get the host triple in BuildConfig. Consider changing
-            // requested_target to an enum, or some other approach.
-            anyhow::bail!("-Zbuild-std requires --target");
-        }
+
         let (std_package_set, std_resolve, std_features) =
             standard_lib::resolve_std(ws, &target_data, &build_config.requested_kinds, crates)?;
         pkg_set.add_set(std_package_set);
@@ -476,14 +471,6 @@ pub fn create_bcx<'a, 'cfg>(
     // assuming `--target $HOST` was specified. See
     // `rebuild_unit_graph_shared` for more on why this is done.
     let explicit_host_kind = CompileKind::Target(CompileTarget::new(&target_data.rustc.host)?);
-    let explicit_host_kinds: Vec<_> = build_config
-        .requested_kinds
-        .iter()
-        .map(|kind| match kind {
-            CompileKind::Host => explicit_host_kind,
-            CompileKind::Target(t) => CompileKind::Target(*t),
-        })
-        .collect();
 
     // Passing `build_config.requested_kinds` instead of
     // `explicit_host_kinds` here so that `generate_targets` can do
@@ -562,7 +549,8 @@ pub fn create_bcx<'a, 'cfg>(
             &crates,
             std_resolve,
             std_features,
-            &explicit_host_kinds,
+            &build_config.requested_kinds,
+            explicit_host_kind,
             &pkg_set,
             interner,
             &profiles,
@@ -1059,25 +1047,7 @@ fn generate_targets(
             let features_for = FeaturesFor::from_for_host(target.proc_macro());
             let features = resolved_features.activated_features(pkg.package_id(), features_for);
 
-            // If `--target` has not been specified, then the unit
-            // graph is built almost like if `--target $HOST` was
-            // specified. See `rebuild_unit_graph_shared` for more on
-            // why this is done. However, if the package has its own
-            // `package.target` key, then this gets used instead of
-            // `$HOST`
-            let explicit_kinds = if let Some(k) = pkg.manifest().forced_kind() {
-                vec![k]
-            } else {
-                requested_kinds
-                    .iter()
-                    .map(|kind| match kind {
-                        CompileKind::Host => {
-                            pkg.manifest().default_kind().unwrap_or(explicit_host_kind)
-                        }
-                        CompileKind::Target(t) => CompileKind::Target(*t),
-                    })
-                    .collect()
-            };
+            let explicit_kinds = pkg.explicit_kinds(requested_kinds, explicit_host_kind);
 
             for kind in explicit_kinds.iter() {
                 let profile = profiles.get_profile(
diff --git a/tests/build-std/main.rs b/tests/build-std/main.rs
index c1355b317ce..c682f42201b 100644
--- a/tests/build-std/main.rs
+++ b/tests/build-std/main.rs
@@ -170,6 +170,50 @@ fn cross_custom() {
         .run();
 }
 
+/// like cross-custom but uses per-package-target instead
+#[cargo_test(build_std)]
+fn per_package_target() {
+    let p = project()
+        .file(
+            "Cargo.toml",
+            r#"
+                cargo-features = ["per-package-target"]
+                [package]
+                name = "foo"
+                version = "0.1.0"
+                edition = "2018"
+                default-target = "custom-target.json"
+
+                [target.custom-target.dependencies]
+                dep = { path = "dep" }
+            "#,
+        )
+        .file(
+            "src/lib.rs",
+            "#![no_std] pub fn f() -> u32 { dep::answer() }",
+        )
+        .file("dep/Cargo.toml", &basic_manifest("dep", "0.1.0"))
+        .file("dep/src/lib.rs", "#![no_std] pub fn answer() -> u32 { 42 }")
+        .file(
+            "custom-target.json",
+            r#"
+            {
+                "llvm-target": "x86_64-unknown-none-gnu",
+                "data-layout": "e-m:e-i64:64-f80:128-n8:16:32:64-S128",
+                "arch": "x86_64",
+                "target-endian": "little",
+                "target-pointer-width": "64",
+                "target-c-int-width": "32",
+                "os": "none",
+                "linker-flavor": "ld.lld"
+            }
+            "#,
+        )
+        .build();
+
+    p.cargo("build -v").build_std_arg("core").run();
+}
+
 #[cargo_test(build_std)]
 fn custom_test_framework() {
     let p = project()