diff --git a/compiler/rustc_target/src/callconv/x86_win64.rs b/compiler/rustc_target/src/callconv/x86_win64.rs
index 23ef2cf828409..17a160b108497 100644
--- a/compiler/rustc_target/src/callconv/x86_win64.rs
+++ b/compiler/rustc_target/src/callconv/x86_win64.rs
@@ -1,4 +1,4 @@
-use rustc_abi::{BackendRepr, Float, Integer, Primitive, RegKind, Size};
+use rustc_abi::{BackendRepr, Float, Integer, Primitive};
 
 use crate::callconv::{ArgAbi, FnAbi, Reg};
 use crate::spec::HasTargetSpec;
@@ -23,15 +23,11 @@ pub(crate) fn compute_abi_info<Ty>(_cx: &impl HasTargetSpec, fn_abi: &mut FnAbi<
                 // (probably what clang calls "illegal vectors").
             }
             BackendRepr::Scalar(scalar) => {
-                if is_ret && matches!(scalar.primitive(), Primitive::Int(Integer::I128, _)) {
-                    // `i128` is returned in xmm0 by Clang and GCC
-                    // FIXME(#134288): This may change for the `-msvc` targets in the future.
-                    let reg = Reg { kind: RegKind::Vector, size: Size::from_bits(128) };
-                    a.cast_to(reg);
-                } else if a.layout.size.bytes() > 8
+                if a.layout.size.bytes() > 8
                     && !matches!(scalar.primitive(), Primitive::Float(Float::F128))
+                    && !(is_ret && matches!(scalar.primitive(), Primitive::Int(Integer::I128, _)))
                 {
-                    // Match what LLVM does for `f128` so that `compiler-builtins` builtins match up
+                    // Match what LLVM does for `f128` and `i128` (the latter only for return values) so that `compiler-builtins` builtins match up
                     // with what LLVM expects.
                     a.make_indirect();
                 } else {
diff --git a/tests/codegen/i128-x86-callconv.rs b/tests/codegen/i128-x86-callconv.rs
index 9a9c9002fc026..c905bbbdbd081 100644
--- a/tests/codegen/i128-x86-callconv.rs
+++ b/tests/codegen/i128-x86-callconv.rs
@@ -4,13 +4,18 @@
 //@ compile-flags: -Copt-level=1
 
 //@ add-core-stubs
-//@ revisions: MSVC MINGW
+//@ revisions: MSVC MINGW softfloat
 //@ [MSVC] needs-llvm-components: x86
-//@ [MINGW] needs-llvm-components: x86
 //@ [MSVC] compile-flags: --target x86_64-pc-windows-msvc
-//@ [MINGW] compile-flags: --target x86_64-pc-windows-gnu
+// Use `WIN` as a common prefix for MSVC and MINGW but *not* the softfloat test.
 //@ [MSVC] filecheck-flags: --check-prefix=WIN
+//@ [MINGW] needs-llvm-components: x86
+//@ [MINGW] compile-flags: --target x86_64-pc-windows-gnu
 //@ [MINGW] filecheck-flags: --check-prefix=WIN
+// The `x86_64-unknown-uefi` target also uses the Windows calling convention,
+// but does not have SSE registers available.
+//@ [softfloat] needs-llvm-components: x86
+//@ [softfloat] compile-flags: --target x86_64-unknown-uefi
 
 #![crate_type = "lib"]
 #![no_std]
@@ -28,24 +33,23 @@ extern "C" {
 pub extern "C" fn pass(_arg0: u32, arg1: i128) {
     // CHECK-LABEL: @pass(
     // i128 is passed indirectly on Windows. It should load the pointer to the stack and pass
-    // a pointer to that allocation.
-    // WIN-SAME: %_arg0, ptr{{.*}} %arg1)
-    // WIN: [[PASS:%[_0-9]+]] = alloca [16 x i8], align 16
-    // WIN: [[LOADED:%[_0-9]+]] = load i128, ptr %arg1
-    // WIN: store i128 [[LOADED]], ptr [[PASS]]
-    // WIN: call void @extern_call
+    // a pointer to that allocation. The softfloat ABI works the same.
+    // CHECK-SAME: %_arg0, ptr{{.*}} %arg1)
+    // CHECK: [[PASS:%[_0-9]+]] = alloca [16 x i8], align 16
+    // CHECK: [[LOADED:%[_0-9]+]] = load i128, ptr %arg1
+    // CHECK: store i128 [[LOADED]], ptr [[PASS]]
+    // CHECK: call void @extern_call
     unsafe { extern_call(arg1) };
 }
 
 // Check that we produce the correct return ABI
 #[no_mangle]
 pub extern "C" fn ret(_arg0: u32, arg1: i128) -> i128 {
-    // CHECK-LABEL: @ret(
-    // i128 is returned in xmm0 on Windows
-    // FIXME(#134288): This may change for the `-msvc` targets in the future.
-    // WIN-SAME: i32{{.*}} %_arg0, ptr{{.*}} %arg1)
-    // WIN: [[LOADED:%[_0-9]+]] = load <16 x i8>, ptr %arg1
-    // WIN-NEXT: ret <16 x i8> [[LOADED]]
+    // we use the LLVM native ABI for the return value
+    // CHECK-LABEL: i128 @ret(
+    // CHECK-SAME: i32{{.*}} %_arg0, ptr{{.*}} %arg1)
+    // CHECK: [[LOADED:%[_0-9]+]] = load i128, ptr %arg1
+    // CHECK-NEXT: ret i128 [[LOADED]]
     arg1
 }
 
@@ -53,10 +57,10 @@ pub extern "C" fn ret(_arg0: u32, arg1: i128) -> i128 {
 #[no_mangle]
 pub extern "C" fn forward(dst: *mut i128) {
     // CHECK-LABEL: @forward
-    // WIN-SAME: ptr{{.*}} %dst)
-    // WIN: [[RETURNED:%[_0-9]+]] = tail call <16 x i8> @extern_ret()
-    // WIN: store <16 x i8> [[RETURNED]], ptr %dst
-    // WIN: ret void
+    // CHECK-SAME: ptr{{.*}} %dst)
+    // CHECK: [[RETURNED:%[_0-9]+]] = tail call {{.*}}i128 @extern_ret()
+    // CHECK: store i128 [[RETURNED]], ptr %dst
+    // CHECK: ret void
     unsafe { *dst = extern_ret() };
 }
 
@@ -70,10 +74,10 @@ struct RetAggregate {
 pub extern "C" fn ret_aggregate(_arg0: u32, arg1: i128) -> RetAggregate {
     // CHECK-LABEL: @ret_aggregate(
     // Aggregates should also be returned indirectly
-    // WIN-SAME: ptr{{.*}}sret([32 x i8]){{.*}}[[RET:%[_0-9]+]], i32{{.*}}%_arg0, ptr{{.*}}%arg1)
-    // WIN: [[LOADED:%[_0-9]+]] = load i128, ptr %arg1
-    // WIN: [[GEP:%[_0-9]+]] = getelementptr{{.*}}, ptr [[RET]]
-    // WIN: store i128 [[LOADED]], ptr [[GEP]]
-    // WIN: ret void
+    // CHECK-SAME: ptr{{.*}}sret([32 x i8]){{.*}}[[RET:%[_0-9]+]], i32{{.*}}%_arg0, ptr{{.*}}%arg1)
+    // CHECK: [[LOADED:%[_0-9]+]] = load i128, ptr %arg1
+    // CHECK: [[GEP:%[_0-9]+]] = getelementptr{{.*}}, ptr [[RET]]
+    // CHECK: store i128 [[LOADED]], ptr [[GEP]]
+    // CHECK: ret void
     RetAggregate { a: 1, b: arg1 }
 }