perf(codegen): Eliminate size_of_val == 0 for DSTs with Non-zero-sized Prefix via NUW and Assume

compiler/rustc_codegen_ssa/src/size_of_val.rs

@@ -159,7 +159,10 @@ pub fn size_and_align_of_dst<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
             // Furthermore, `align >= unsized_align`, and therefore we only need to do:
             // let full_size = (unsized_offset_unadjusted + unsized_size).align_to(full_align);
 
-            let full_size = bx.add(unsized_offset_unadjusted, unsized_size);
+            // This is the unrounded size before alignment padding. It cannot exceed the
+            // rounded size, which itself cannot exceed `isize::MAX`. Thus the addition
+            // cannot overflow `isize::MAX`, let alone `usize::MAX`.
+            let unrounded_size = bx.unchecked_suadd(unsized_offset_unadjusted, unsized_size);
 
             // Issue #27023: must add any necessary padding to `size`
             // (to make it a multiple of `align`) before returning it.
@@ -173,10 +176,18 @@ pub fn size_and_align_of_dst<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
             //   `(size + (align-1)) & -align`
             let one = bx.const_usize(1);
             let addend = bx.sub(full_align, one);
-            let add = bx.add(full_size, addend);
+            let add = bx.add(unrounded_size, addend);
             let neg = bx.neg(full_align);
             let full_size = bx.and(add, neg);
 
+            // Alignment rounding can only increase the size, never decrease it:
+            // `round_up(x, a) >= x` for power-of-two `a`. With the `nuw` on the
+            // addition above, LLVM can therefore deduce
+            // `full_size >= unrounded_size >= offset`, which proves `full_size > 0`
+            // for types with a non-zero-sized prefix (#152788).
+            let size_ge = bx.icmp(IntPredicate::IntUGE, full_size, unrounded_size);
+            bx.assume(size_ge);
+
             (full_size, full_align)
         }
         _ => bug!("size_and_align_of_dst: {t} not supported"),

tests/codegen-llvm/dst-size-of-val-not-zst.rs

@@ -0,0 +1,36 @@
+//@ compile-flags: -Copt-level=3 -Z merge-functions=disabled
+//@ needs-deterministic-layouts
+
+#![crate_type = "lib"]
+
+// Regression test for #152788: `size_of_val(p) == 0` should optimize to `false`
+// for types whose statically-known prefix makes them clearly not ZSTs.
+//
+// This works because:
+// 1. The `offset + unsized_size` addition has NUW+NSW, so LLVM knows
+//    `unrounded_size >= offset`
+// 2. An `llvm.assume` tells LLVM `aligned_size >= unrounded_size`
+// 3. Together: `aligned_size >= unrounded_size >= offset > 0`
+
+pub struct Foo<T: ?Sized>(pub [u32; 3], pub T);
+
+// CHECK-LABEL: @size_of_val_dyn_not_zero
+#[no_mangle]
+pub fn size_of_val_dyn_not_zero(p: &Foo<dyn std::fmt::Debug>) -> bool {
+    // CHECK: ret i1 false
+    std::mem::size_of_val(p) == 0
+}
+
+// CHECK-LABEL: @size_of_val_slice_u8_not_zero
+#[no_mangle]
+pub fn size_of_val_slice_u8_not_zero(p: &Foo<[u8]>) -> bool {
+    // CHECK: ret i1 false
+    std::mem::size_of_val(p) == 0
+}
+
+// CHECK-LABEL: @size_of_val_slice_i32_not_zero
+#[no_mangle]
+pub fn size_of_val_slice_i32_not_zero(p: &Foo<[i32]>) -> bool {
+    // CHECK: ret i1 false
+    std::mem::size_of_val(p) == 0
+}

tests/codegen-llvm/dst-vtable-align-nonzero.rs

@@ -30,9 +30,8 @@ pub struct Struct<W: ?Sized> {
 pub fn eliminates_runtime_check_when_align_1(
     x: &Struct<WrapperWithAlign1<dyn Trait>>,
 ) -> &WrapperWithAlign1<dyn Trait> {
-    // CHECK: load [[USIZE:i[0-9]+]], {{.+}} !range [[RANGE_META:![0-9]+]]
+    // CHECK: load [[USIZE:i[0-9]+]]
     // CHECK-NOT: llvm.umax
-    // CHECK-NOT: icmp
     // CHECK-NOT: select
     // CHECK: ret
     &x.dst
@@ -43,7 +42,7 @@ pub fn eliminates_runtime_check_when_align_1(
 pub fn does_not_eliminate_runtime_check_when_align_2(
     x: &Struct<WrapperWithAlign2<dyn Trait>>,
 ) -> &WrapperWithAlign2<dyn Trait> {
-    // CHECK: [[X0:%[0-9]+]] = load [[USIZE]], {{.+}} !range [[RANGE_META]]
+    // CHECK: load [[USIZE]]
     // CHECK: {{icmp|llvm.umax}}
     // CHECK: ret
     &x.dst
@@ -52,7 +51,7 @@ pub fn does_not_eliminate_runtime_check_when_align_2(
 // CHECK-LABEL: @align_load_from_align_of_val
 #[no_mangle]
 pub fn align_load_from_align_of_val(x: &dyn Trait) -> usize {
-    // CHECK: {{%[0-9]+}} = load [[USIZE]], {{.+}} !range [[RANGE_META]]
+    // CHECK: {{%[0-9]+}} = load [[USIZE]], {{.+}} !range [[RANGE_META:![0-9]+]]
     core::mem::align_of_val(x)
 }