diff --git a/compiler/rustc_span/src/symbol.rs b/compiler/rustc_span/src/symbol.rs
index 382754be2ca83..f021243b11aee 100644
--- a/compiler/rustc_span/src/symbol.rs
+++ b/compiler/rustc_span/src/symbol.rs
@@ -1489,6 +1489,7 @@ symbols! {
         sreg,
         sreg_low16,
         sse,
+        sse2,
         sse4a_target_feature,
         stable,
         staged_api,
diff --git a/compiler/rustc_target/src/spec/i686_apple_darwin.rs b/compiler/rustc_target/src/spec/i686_apple_darwin.rs
index b5103d15db695..3bf2177fb5ad4 100644
--- a/compiler/rustc_target/src/spec/i686_apple_darwin.rs
+++ b/compiler/rustc_target/src/spec/i686_apple_darwin.rs
@@ -9,6 +9,7 @@ pub fn target() -> Target {
     base.add_pre_link_args(LinkerFlavor::Darwin(Cc::Yes, Lld::No), &["-m32"]);
     base.stack_probes = StackProbeType::X86;
     base.frame_pointer = FramePointer::Always;
+    base.x86_use_xmm0 = true;
 
     Target {
         // Clang automatically chooses a more specific target based on
diff --git a/compiler/rustc_target/src/spec/i686_linux_android.rs b/compiler/rustc_target/src/spec/i686_linux_android.rs
index c7c30c23901d3..65ea366fb6114 100644
--- a/compiler/rustc_target/src/spec/i686_linux_android.rs
+++ b/compiler/rustc_target/src/spec/i686_linux_android.rs
@@ -12,6 +12,7 @@ pub fn target() -> Target {
     base.cpu = "pentiumpro".into();
     base.features = "+mmx,+sse,+sse2,+sse3,+ssse3".into();
     base.stack_probes = StackProbeType::X86;
+    base.x86_use_xmm0 = true;
 
     Target {
         llvm_target: "i686-linux-android".into(),
diff --git a/compiler/rustc_target/src/spec/i686_pc_windows_gnu.rs b/compiler/rustc_target/src/spec/i686_pc_windows_gnu.rs
index 7a11138754fa8..bfbc1f58fab62 100644
--- a/compiler/rustc_target/src/spec/i686_pc_windows_gnu.rs
+++ b/compiler/rustc_target/src/spec/i686_pc_windows_gnu.rs
@@ -14,6 +14,7 @@ pub fn target() -> Target {
         &["-m", "i386pe", "--large-address-aware"],
     );
     base.add_pre_link_args(LinkerFlavor::Gnu(Cc::Yes, Lld::No), &["-Wl,--large-address-aware"]);
+    base.x86_use_xmm0 = true;
 
     Target {
         llvm_target: "i686-pc-windows-gnu".into(),
diff --git a/compiler/rustc_target/src/spec/i686_pc_windows_gnullvm.rs b/compiler/rustc_target/src/spec/i686_pc_windows_gnullvm.rs
index 3154b512a5202..f5cd9f76929f3 100644
--- a/compiler/rustc_target/src/spec/i686_pc_windows_gnullvm.rs
+++ b/compiler/rustc_target/src/spec/i686_pc_windows_gnullvm.rs
@@ -13,6 +13,7 @@ pub fn target() -> Target {
         LinkerFlavor::Gnu(Cc::No, Lld::No),
         &["-m", "i386pe", "--large-address-aware"],
     );
+    base.x86_use_xmm0 = true;
 
     Target {
         llvm_target: "i686-pc-windows-gnu".into(),
diff --git a/compiler/rustc_target/src/spec/i686_pc_windows_msvc.rs b/compiler/rustc_target/src/spec/i686_pc_windows_msvc.rs
index db4c00dc697d7..c2aaac285dc07 100644
--- a/compiler/rustc_target/src/spec/i686_pc_windows_msvc.rs
+++ b/compiler/rustc_target/src/spec/i686_pc_windows_msvc.rs
@@ -19,6 +19,7 @@ pub fn target() -> Target {
     );
     // Workaround for #95429
     base.has_thread_local = false;
+    base.x86_use_xmm0 = true;
 
     Target {
         llvm_target: "i686-pc-windows-msvc".into(),
diff --git a/compiler/rustc_target/src/spec/i686_unknown_freebsd.rs b/compiler/rustc_target/src/spec/i686_unknown_freebsd.rs
index 35ca78034f170..edf0c62d44c21 100644
--- a/compiler/rustc_target/src/spec/i686_unknown_freebsd.rs
+++ b/compiler/rustc_target/src/spec/i686_unknown_freebsd.rs
@@ -6,6 +6,7 @@ pub fn target() -> Target {
     base.max_atomic_width = Some(64);
     base.add_pre_link_args(LinkerFlavor::Gnu(Cc::Yes, Lld::No), &["-m32", "-Wl,-znotext"]);
     base.stack_probes = StackProbeType::X86;
+    base.x86_use_xmm0 = true;
 
     Target {
         llvm_target: "i686-unknown-freebsd".into(),
diff --git a/compiler/rustc_target/src/spec/i686_unknown_haiku.rs b/compiler/rustc_target/src/spec/i686_unknown_haiku.rs
index e6b72336c5cf5..41c7ddef06aa5 100644
--- a/compiler/rustc_target/src/spec/i686_unknown_haiku.rs
+++ b/compiler/rustc_target/src/spec/i686_unknown_haiku.rs
@@ -6,6 +6,7 @@ pub fn target() -> Target {
     base.max_atomic_width = Some(64);
     base.add_pre_link_args(LinkerFlavor::Gnu(Cc::Yes, Lld::No), &["-m32"]);
     base.stack_probes = StackProbeType::X86;
+    base.x86_use_xmm0 = true;
 
     Target {
         llvm_target: "i686-unknown-haiku".into(),
diff --git a/compiler/rustc_target/src/spec/i686_unknown_linux_gnu.rs b/compiler/rustc_target/src/spec/i686_unknown_linux_gnu.rs
index 73e536a7e4d93..b579e70f63717 100644
--- a/compiler/rustc_target/src/spec/i686_unknown_linux_gnu.rs
+++ b/compiler/rustc_target/src/spec/i686_unknown_linux_gnu.rs
@@ -7,6 +7,7 @@ pub fn target() -> Target {
     base.supported_sanitizers = SanitizerSet::ADDRESS;
     base.add_pre_link_args(LinkerFlavor::Gnu(Cc::Yes, Lld::No), &["-m32"]);
     base.stack_probes = StackProbeType::X86;
+    base.x86_use_xmm0 = true;
 
     Target {
         llvm_target: "i686-unknown-linux-gnu".into(),
diff --git a/compiler/rustc_target/src/spec/i686_unknown_linux_musl.rs b/compiler/rustc_target/src/spec/i686_unknown_linux_musl.rs
index 3825082ba25e4..8c2b9e9f7014d 100644
--- a/compiler/rustc_target/src/spec/i686_unknown_linux_musl.rs
+++ b/compiler/rustc_target/src/spec/i686_unknown_linux_musl.rs
@@ -20,6 +20,7 @@ pub fn target() -> Target {
     // This may or may not be related to this bug:
     // https://llvm.org/bugs/show_bug.cgi?id=30879
     base.frame_pointer = FramePointer::Always;
+    base.x86_use_xmm0 = true;
 
     Target {
         llvm_target: "i686-unknown-linux-musl".into(),
diff --git a/compiler/rustc_target/src/spec/i686_unknown_netbsd.rs b/compiler/rustc_target/src/spec/i686_unknown_netbsd.rs
index b191996c7de0d..205c6a2067a96 100644
--- a/compiler/rustc_target/src/spec/i686_unknown_netbsd.rs
+++ b/compiler/rustc_target/src/spec/i686_unknown_netbsd.rs
@@ -6,6 +6,7 @@ pub fn target() -> Target {
     base.max_atomic_width = Some(64);
     base.add_pre_link_args(LinkerFlavor::Gnu(Cc::Yes, Lld::No), &["-m32"]);
     base.stack_probes = StackProbeType::X86;
+    base.x86_use_xmm0 = true;
 
     Target {
         llvm_target: "i686-unknown-netbsdelf".into(),
diff --git a/compiler/rustc_target/src/spec/i686_unknown_openbsd.rs b/compiler/rustc_target/src/spec/i686_unknown_openbsd.rs
index 8babe55971280..34699b942990f 100644
--- a/compiler/rustc_target/src/spec/i686_unknown_openbsd.rs
+++ b/compiler/rustc_target/src/spec/i686_unknown_openbsd.rs
@@ -6,6 +6,7 @@ pub fn target() -> Target {
     base.max_atomic_width = Some(64);
     base.add_pre_link_args(LinkerFlavor::Gnu(Cc::Yes, Lld::No), &["-m32", "-fuse-ld=lld"]);
     base.stack_probes = StackProbeType::X86;
+    base.x86_use_xmm0 = true;
 
     Target {
         llvm_target: "i686-unknown-openbsd".into(),
diff --git a/compiler/rustc_target/src/spec/i686_uwp_windows_gnu.rs b/compiler/rustc_target/src/spec/i686_uwp_windows_gnu.rs
index a3e32569827fb..884f229f5ed7b 100644
--- a/compiler/rustc_target/src/spec/i686_uwp_windows_gnu.rs
+++ b/compiler/rustc_target/src/spec/i686_uwp_windows_gnu.rs
@@ -13,6 +13,7 @@ pub fn target() -> Target {
         &["-m", "i386pe", "--large-address-aware"],
     );
     base.add_pre_link_args(LinkerFlavor::Gnu(Cc::Yes, Lld::No), &["-Wl,--large-address-aware"]);
+    base.x86_use_xmm0 = true;
 
     Target {
         llvm_target: "i686-pc-windows-gnu".into(),
diff --git a/compiler/rustc_target/src/spec/i686_uwp_windows_msvc.rs b/compiler/rustc_target/src/spec/i686_uwp_windows_msvc.rs
index 4c657fe908ac4..0de04d23bf859 100644
--- a/compiler/rustc_target/src/spec/i686_uwp_windows_msvc.rs
+++ b/compiler/rustc_target/src/spec/i686_uwp_windows_msvc.rs
@@ -4,6 +4,7 @@ pub fn target() -> Target {
     let mut base = super::windows_uwp_msvc_base::opts();
     base.cpu = "pentium4".into();
     base.max_atomic_width = Some(64);
+    base.x86_use_xmm0 = true;
 
     Target {
         llvm_target: "i686-pc-windows-msvc".into(),
diff --git a/compiler/rustc_target/src/spec/i686_wrs_vxworks.rs b/compiler/rustc_target/src/spec/i686_wrs_vxworks.rs
index b5cfdfcebea90..0a20e5f4cef8f 100644
--- a/compiler/rustc_target/src/spec/i686_wrs_vxworks.rs
+++ b/compiler/rustc_target/src/spec/i686_wrs_vxworks.rs
@@ -6,6 +6,7 @@ pub fn target() -> Target {
     base.max_atomic_width = Some(64);
     base.add_pre_link_args(LinkerFlavor::Gnu(Cc::Yes, Lld::No), &["-m32"]);
     base.stack_probes = StackProbeType::X86;
+    base.x86_use_xmm0 = true;
 
     Target {
         llvm_target: "i686-unknown-linux-gnu".into(),
diff --git a/compiler/rustc_target/src/spec/mod.rs b/compiler/rustc_target/src/spec/mod.rs
index 8aa72797a0d25..6a0848b6c299b 100644
--- a/compiler/rustc_target/src/spec/mod.rs
+++ b/compiler/rustc_target/src/spec/mod.rs
@@ -1935,6 +1935,9 @@ pub struct TargetOptions {
     /// wasm32 where the whole program either has simd or not.
     pub simd_types_indirect: bool,
 
+    /// On x86, use XMM0 instead of FP0 as float return register for the Rust ABI
+    pub x86_use_xmm0: bool,
+
     /// Pass a list of symbol which should be exported in the dylib to the linker.
     pub limit_rdylib_exports: bool,
 
@@ -2213,6 +2216,7 @@ impl Default for TargetOptions {
             requires_uwtable: false,
             default_uwtable: false,
             simd_types_indirect: true,
+            x86_use_xmm0: false,
             limit_rdylib_exports: true,
             override_export_symbols: None,
             merge_functions: MergeFunctions::Aliases,
@@ -2878,6 +2882,7 @@ impl Target {
         key!(requires_uwtable, bool);
         key!(default_uwtable, bool);
         key!(simd_types_indirect, bool);
+        key!(x86_use_xmm0, bool);
         key!(limit_rdylib_exports, bool);
         key!(override_export_symbols, opt_list);
         key!(merge_functions, MergeFunctions)?;
@@ -3135,6 +3140,7 @@ impl ToJson for Target {
         target_option_val!(requires_uwtable);
         target_option_val!(default_uwtable);
         target_option_val!(simd_types_indirect);
+        target_option_val!(x86_use_xmm0);
         target_option_val!(limit_rdylib_exports);
         target_option_val!(override_export_symbols);
         target_option_val!(merge_functions);
diff --git a/compiler/rustc_ty_utils/src/abi.rs b/compiler/rustc_ty_utils/src/abi.rs
index 16183403d67aa..b3beac89acc73 100644
--- a/compiler/rustc_ty_utils/src/abi.rs
+++ b/compiler/rustc_ty_utils/src/abi.rs
@@ -7,6 +7,7 @@ use rustc_middle::ty::layout::{
 use rustc_middle::ty::{self, InstanceDef, Ty, TyCtxt};
 use rustc_session::config::OptLevel;
 use rustc_span::def_id::DefId;
+use rustc_span::symbol::sym;
 use rustc_target::abi::call::{
     ArgAbi, ArgAttribute, ArgAttributes, ArgExtension, Conv, FnAbi, PassMode, Reg, RegKind,
     RiscvInterruptKind,
@@ -371,6 +372,9 @@ fn fn_abi_new_uncached<'tcx>(
     let target = &cx.tcx.sess.target;
     let target_env_gnu_like = matches!(&target.env[..], "gnu" | "musl" | "uclibc");
     let win_x64_gnu = target.os == "windows" && target.arch == "x86_64" && target.env == "gnu";
+    let x86_sse2 = target.arch == "x86"
+        && target.x86_use_xmm0
+        && cx.tcx.sess.parse_sess.config.contains(&(sym::target_feature, Some(sym::sse2)));
     let linux_s390x_gnu_like =
         target.os == "linux" && target.arch == "s390x" && target_env_gnu_like;
     let linux_sparc64_gnu_like =
@@ -415,6 +419,10 @@ fn fn_abi_new_uncached<'tcx>(
                 is_return,
                 drop_target_pointee,
             );
+            // Use XMM0 instead of FP0 to preserve NaN payloads
+            if x86_sse2 && rust_abi && is_return && matches!(scalar.primitive(), F32 | F64) {
+                attrs.set(ArgAttribute::InReg);
+            }
             attrs
         });
 
diff --git a/library/std/src/f32/tests.rs b/library/std/src/f32/tests.rs
index 9ca4e8f2f45fe..c6fdf660f2fb3 100644
--- a/library/std/src/f32/tests.rs
+++ b/library/std/src/f32/tests.rs
@@ -328,7 +328,7 @@ macro_rules! assert_f32_biteq {
 
 // Ignore test on x87 floating point, these platforms do not guarantee NaN
 // payloads are preserved and flush denormals to zero, failing the tests.
-#[cfg(not(target_arch = "x86"))]
+#[cfg(not(all(target_arch = "x86", not(target_feature = "sse2"))))]
 #[test]
 fn test_next_up() {
     let tiny = f32::from_bits(1);
@@ -361,7 +361,7 @@ fn test_next_up() {
 
 // Ignore test on x87 floating point, these platforms do not guarantee NaN
 // payloads are preserved and flush denormals to zero, failing the tests.
-#[cfg(not(target_arch = "x86"))]
+#[cfg(not(all(target_arch = "x86", not(target_feature = "sse2"))))]
 #[test]
 fn test_next_down() {
     let tiny = f32::from_bits(1);
diff --git a/library/std/src/f64/tests.rs b/library/std/src/f64/tests.rs
index f88d01593b5e4..3833bd71986be 100644
--- a/library/std/src/f64/tests.rs
+++ b/library/std/src/f64/tests.rs
@@ -318,7 +318,7 @@ macro_rules! assert_f64_biteq {
 
 // Ignore test on x87 floating point, these platforms do not guarantee NaN
 // payloads are preserved and flush denormals to zero, failing the tests.
-#[cfg(not(target_arch = "x86"))]
+#[cfg(not(all(target_arch = "x86", not(target_feature = "sse2"))))]
 #[test]
 fn test_next_up() {
     let tiny = f64::from_bits(1);
@@ -350,7 +350,7 @@ fn test_next_up() {
 
 // Ignore test on x87 floating point, these platforms do not guarantee NaN
 // payloads are preserved and flush denormals to zero, failing the tests.
-#[cfg(not(target_arch = "x86"))]
+#[cfg(not(all(target_arch = "x86", not(target_feature = "sse2"))))]
 #[test]
 fn test_next_down() {
     let tiny = f64::from_bits(1);