Auto merge of #94130 - erikdesjardins:partially, r=nikic

Use undef for (some) partially-uninit constants

There needs to be some limit to avoid perf regressions on large arrays
with undef in each element (see comment in the code).

Fixes: #84565
Original PR: #83698

Depends on LLVM 14: #93577

compiler/rustc_codegen_llvm/src/consts.rs

@@ -2,6 +2,7 @@ use crate::base;
 use crate::common::CodegenCx;
 use crate::debuginfo;
 use crate::llvm::{self, True};
+use crate::llvm_util;
 use crate::type_::Type;
 use crate::type_of::LayoutLlvmExt;
 use crate::value::Value;
@@ -37,7 +38,7 @@ pub fn const_alloc_to_llvm<'ll>(cx: &CodegenCx<'ll, '_>, alloc: &Allocation) ->
         alloc: &'a Allocation,
         range: Range<usize>,
     ) {
-        let mut chunks = alloc
+        let chunks = alloc
             .init_mask()
             .range_as_init_chunks(Size::from_bytes(range.start), Size::from_bytes(range.end));
 
@@ -53,30 +54,26 @@ pub fn const_alloc_to_llvm<'ll>(cx: &CodegenCx<'ll, '_>, alloc: &Allocation) ->
             }
         };
 
-        // Generating partially-uninit consts inhibits optimizations, so it is disabled by default.
+        // Generating partially-uninit consts is limited to small numbers of chunks,
+        // to avoid the cost of generating large complex const expressions.
+        // For example, `[(u32, u8); 1024 * 1024]` contains uninit padding in each element,
+        // and would result in `{ [5 x i8] zeroinitializer, [3 x i8] undef, ...repeat 1M times... }`.
+        let max = if llvm_util::get_version() < (14, 0, 0) {
+            // Generating partially-uninit consts inhibits optimizations in LLVM < 14.
             // See https://github.com/rust-lang/rust/issues/84565.
-        let allow_partially_uninit =
+            1
-            match cx.sess().opts.debugging_opts.partially_uninit_const_threshold {
+        } else {
-                Some(max) => range.len() <= max,
+            cx.sess().opts.debugging_opts.uninit_const_chunk_threshold
-                None => false,
         };
+        let allow_uninit_chunks = chunks.clone().take(max.saturating_add(1)).count() <= max;
 
-        if allow_partially_uninit {
+        if allow_uninit_chunks {
             llvals.extend(chunks.map(chunk_to_llval));
         } else {
-            let llval = match (chunks.next(), chunks.next()) {
+            // If this allocation contains any uninit bytes, codegen as if it was initialized
-                (Some(chunk), None) => {
+            // (using some arbitrary value for uninit bytes).
-                    // exactly one chunk, either fully init or fully uninit
-                    chunk_to_llval(chunk)
-                }
-                _ => {
-                    // partially uninit, codegen as if it was initialized
-                    // (using some arbitrary value for uninit bytes)
             let bytes = alloc.inspect_with_uninit_and_ptr_outside_interpreter(range);
-                    cx.const_bytes(bytes)
+            llvals.push(cx.const_bytes(bytes));
-                }
-            };
-            llvals.push(llval);
         }
     }
 

compiler/rustc_interface/src/tests.rs

@@ -758,7 +758,6 @@ fn test_debugging_options_tracking_hash() {
     tracked!(osx_rpath_install_name, true);
     tracked!(panic_abort_tests, true);
     tracked!(panic_in_drop, PanicStrategy::Abort);
-    tracked!(partially_uninit_const_threshold, Some(123));
     tracked!(pick_stable_methods_before_any_unstable, false);
     tracked!(plt, Some(true));
     tracked!(polonius, true);
@@ -789,6 +788,7 @@ fn test_debugging_options_tracking_hash() {
     tracked!(trap_unreachable, Some(false));
     tracked!(treat_err_as_bug, NonZeroUsize::new(1));
     tracked!(tune_cpu, Some(String::from("abc")));
+    tracked!(uninit_const_chunk_threshold, 123);
     tracked!(unleash_the_miri_inside_of_you, true);
     tracked!(use_ctors_section, Some(true));
     tracked!(verify_llvm_ir, true);

compiler/rustc_middle/src/mir/interpret/allocation.rs

@@ -957,6 +957,7 @@ impl InitMask {
 }
 
 /// Yields [`InitChunk`]s. See [`InitMask::range_as_init_chunks`].
+#[derive(Clone)]
 pub struct InitChunkIter<'a> {
     init_mask: &'a InitMask,
     /// Whether the next chunk we will return is initialized.

compiler/rustc_session/src/options.rs

@@ -1336,9 +1336,6 @@ options! {
         "panic strategy for panics in drops"),
     parse_only: bool = (false, parse_bool, [UNTRACKED],
         "parse only; do not compile, assemble, or link (default: no)"),
-    partially_uninit_const_threshold: Option<usize> = (None, parse_opt_number, [TRACKED],
-        "allow generating const initializers with mixed init/uninit bytes, \
-        and set the maximum total size of a const allocation for which this is allowed (default: never)"),
     perf_stats: bool = (false, parse_bool, [UNTRACKED],
         "print some performance-related statistics (default: no)"),
     pick_stable_methods_before_any_unstable: bool = (true, parse_bool, [TRACKED],
@@ -1483,6 +1480,9 @@ options! {
         "in diagnostics, use heuristics to shorten paths referring to items"),
     ui_testing: bool = (false, parse_bool, [UNTRACKED],
         "emit compiler diagnostics in a form suitable for UI testing (default: no)"),
+    uninit_const_chunk_threshold: usize = (16, parse_number, [TRACKED],
+        "allow generating const initializers with mixed init/uninit chunks, \
+        and set the maximum number of chunks for which this is allowed (default: 16)"),
     unleash_the_miri_inside_of_you: bool = (false, parse_bool, [TRACKED],
         "take the brakes off const evaluation. NOTE: this is unsound (default: no)"),
     unpretty: Option<String> = (None, parse_unpretty, [UNTRACKED],

src/test/codegen/consts.rs

@@ -1,5 +1,5 @@
 // compile-flags: -C no-prepopulate-passes
-//
+// min-llvm-version: 14.0
 
 #![crate_type = "lib"]
 
@@ -43,7 +43,7 @@ pub fn inline_enum_const() -> E<i8, i16> {
 #[no_mangle]
 pub fn low_align_const() -> E<i16, [i16; 3]> {
     // Check that low_align_const and high_align_const use the same constant
-    // CHECK: memcpy.p0i8.p0i8.i{{(32|64)}}(i8* align 2 %1, i8* align 2 getelementptr inbounds (<{ [8 x i8] }>, <{ [8 x i8] }>* [[LOW_HIGH]], i32 0, i32 0, i32 0), i{{(32|64)}} 8, i1 false)
+    // CHECK: memcpy.p0i8.p0i8.i{{(32|64)}}(i8* align 2 %1, i8* align 2 getelementptr inbounds (<{ [4 x i8], [4 x i8] }>, <{ [4 x i8], [4 x i8] }>* [[LOW_HIGH]], i32 0, i32 0, i32 0), i{{(32|64)}} 8, i1 false)
     *&E::A(0)
 }
 
@@ -51,6 +51,6 @@ pub fn low_align_const() -> E<i16, [i16; 3]> {
 #[no_mangle]
 pub fn high_align_const() -> E<i16, i32> {
     // Check that low_align_const and high_align_const use the same constant
-    // CHECK: memcpy.p0i8.p0i8.i{{(32|64)}}(i8* align 4 %1, i8* align 4 getelementptr inbounds (<{ [8 x i8] }>, <{ [8 x i8] }>* [[LOW_HIGH]], i32 0, i32 0, i32 0), i{{(32|64)}} 8, i1 false)
+    // CHECK: memcpy.p0i8.p0i8.i{{(32|64)}}(i8* align 4 %1, i8* align 4 getelementptr inbounds (<{ [4 x i8], [4 x i8] }>, <{ [4 x i8], [4 x i8] }>* [[LOW_HIGH]], i32 0, i32 0, i32 0), i{{(32|64)}} 8, i1 false)
     *&E::A(0)
 }

src/test/codegen/uninit-consts-allow-partially-uninit.rs

@@ -1,35 +0,0 @@
-// compile-flags: -C no-prepopulate-passes -Z partially_uninit_const_threshold=1024
-
-// Like uninit-consts.rs, but tests that we correctly generate partially-uninit consts
-// when the (disabled by default) partially_uninit_const_threshold flag is used.
-
-#![crate_type = "lib"]
-
-use std::mem::MaybeUninit;
-
-pub struct PartiallyUninit {
-    x: u32,
-    y: MaybeUninit<[u8; 10]>
-}
-
-// This should be partially undef.
-// CHECK: [[PARTIALLY_UNINIT:@[0-9]+]] = private unnamed_addr constant <{ [4 x i8], [12 x i8] }> <{ [4 x i8] c"\EF\BE\AD\DE", [12 x i8] undef }>, align 4
-
-// This shouldn't contain undef, since it's larger than the 1024 byte limit.
-// CHECK: [[UNINIT_PADDING_HUGE:@[0-9]+]] = private unnamed_addr constant <{ [32768 x i8] }> <{ [32768 x i8] c"{{.+}}" }>, align 4
-
-// CHECK-LABEL: @partially_uninit
-#[no_mangle]
-pub const fn partially_uninit() -> PartiallyUninit {
-    const X: PartiallyUninit = PartiallyUninit { x: 0xdeadbeef, y: MaybeUninit::uninit() };
-    // CHECK: call void @llvm.memcpy.p0i8.p0i8.i{{(32|64)}}(i8* align 4 %1, i8* align 4 getelementptr inbounds (<{ [4 x i8], [12 x i8] }>, <{ [4 x i8], [12 x i8] }>* [[PARTIALLY_UNINIT]], i32 0, i32 0, i32 0), i{{(32|64)}} 16, i1 false)
-    X
-}
-
-// CHECK-LABEL: @uninit_padding_huge
-#[no_mangle]
-pub const fn uninit_padding_huge() -> [(u32, u8); 4096] {
-    const X: [(u32, u8); 4096] = [(123, 45); 4096];
-    // CHECK: call void @llvm.memcpy.p0i8.p0i8.i{{(32|64)}}(i8* align 4 %1, i8* align 4 getelementptr inbounds (<{ [32768 x i8] }>, <{ [32768 x i8] }>* [[UNINIT_PADDING_HUGE]], i32 0, i32 0, i32 0), i{{(32|64)}} 32768, i1 false)
-    X
-}

src/test/codegen/uninit-consts.rs

@@ -1,4 +1,5 @@
 // compile-flags: -C no-prepopulate-passes
+// min-llvm-version: 14.0
 
 // Check that we use undef (and not zero) for uninitialized bytes in constants.
 
@@ -12,7 +13,13 @@ pub struct PartiallyUninit {
 }
 
 // CHECK: [[FULLY_UNINIT:@[0-9]+]] = private unnamed_addr constant <{ [10 x i8] }> undef
-// CHECK: [[PARTIALLY_UNINIT:@[0-9]+]] = private unnamed_addr constant <{ [16 x i8] }> <{ [16 x i8] c"\EF\BE\AD\DE\00\00\00\00\00\00\00\00\00\00\00\00" }>, align 4
+
+// CHECK: [[PARTIALLY_UNINIT:@[0-9]+]] = private unnamed_addr constant <{ [4 x i8], [12 x i8] }> <{ [4 x i8] c"\EF\BE\AD\DE", [12 x i8] undef }>, align 4
+
+// This shouldn't contain undef, since it contains more chunks
+// than the default value of uninit_const_chunk_threshold.
+// CHECK: [[UNINIT_PADDING_HUGE:@[0-9]+]] = private unnamed_addr constant <{ [32768 x i8] }> <{ [32768 x i8] c"{{.+}}" }>, align 4
+
 // CHECK: [[FULLY_UNINIT_HUGE:@[0-9]+]] = private unnamed_addr constant <{ [16384 x i8] }> undef
 
 // CHECK-LABEL: @fully_uninit
@@ -27,7 +34,15 @@ pub const fn fully_uninit() -> MaybeUninit<[u8; 10]> {
 #[no_mangle]
 pub const fn partially_uninit() -> PartiallyUninit {
     const X: PartiallyUninit = PartiallyUninit { x: 0xdeadbeef, y: MaybeUninit::uninit() };
-    // CHECK: call void @llvm.memcpy.p0i8.p0i8.i{{(32|64)}}(i8* align 4 %1, i8* align 4 getelementptr inbounds (<{ [16 x i8] }>, <{ [16 x i8] }>* [[PARTIALLY_UNINIT]], i32 0, i32 0, i32 0), i{{(32|64)}} 16, i1 false)
+    // CHECK: call void @llvm.memcpy.p0i8.p0i8.i{{(32|64)}}(i8* align 4 %1, i8* align 4 getelementptr inbounds (<{ [4 x i8], [12 x i8] }>, <{ [4 x i8], [12 x i8] }>* [[PARTIALLY_UNINIT]], i32 0, i32 0, i32 0), i{{(32|64)}} 16, i1 false)
+    X
+}
+
+// CHECK-LABEL: @uninit_padding_huge
+#[no_mangle]
+pub const fn uninit_padding_huge() -> [(u32, u8); 4096] {
+    const X: [(u32, u8); 4096] = [(123, 45); 4096];
+    // CHECK: call void @llvm.memcpy.p0i8.p0i8.i{{(32|64)}}(i8* align 4 %1, i8* align 4 getelementptr inbounds (<{ [32768 x i8] }>, <{ [32768 x i8] }>* [[UNINIT_PADDING_HUGE]], i32 0, i32 0, i32 0), i{{(32|64)}} 32768, i1 false)
     X
 }