Add llvm.sideeffect to potential infinite loops and recursions

LLVM assumes that a thread will eventually cause side effect. This is
not true in Rust if a loop or recursion does nothing in its body,
causing undefined behavior even in common cases like `loop {}`.
Inserting llvm.sideeffect fixes the undefined behavior.

As a micro-optimization, only insert llvm.sideeffect when jumping back
in blocks or calling a function.

A patch for LLVM is expected to allow empty non-terminate code by
default and fix this issue from LLVM side.

https://github.com/rust-lang/rust/issues/28728

src/librustc_codegen_llvm/context.rs

@@ -537,6 +537,7 @@ impl CodegenCx<'b, 'tcx> {
         ifn!("llvm.trap", fn() -> void);
         ifn!("llvm.debugtrap", fn() -> void);
         ifn!("llvm.frameaddress", fn(t_i32) -> i8p);
+        ifn!("llvm.sideeffect", fn() -> void);
 
         ifn!("llvm.powi.f32", fn(t_f32, t_i32) -> t_f32);
         ifn!("llvm.powi.v2f32", fn(t_v2f32, t_i32) -> t_v2f32);

src/librustc_codegen_llvm/intrinsic.rs

@@ -124,6 +124,7 @@ impl IntrinsicCallMethods<'tcx> for Builder<'a, 'll, 'tcx> {
                 self.call(expect, &[args[0].immediate(), self.const_bool(false)], None)
             }
             "try" => {
+                self.sideeffect();
                 try_intrinsic(self,
                               args[0].immediate(),
                               args[1].immediate(),
@@ -724,6 +725,11 @@ impl IntrinsicCallMethods<'tcx> for Builder<'a, 'll, 'tcx> {
         self.call(expect, &[cond, self.const_bool(expected)], None)
     }
 
+    fn sideeffect(&mut self) {
+        let fnname = self.get_intrinsic(&("llvm.sideeffect"));
+        self.call(fnname, &[], None);
+    }
+
     fn va_start(&mut self, va_list: &'ll Value) -> &'ll Value {
         let intrinsic = self.cx().get_intrinsic("llvm.va_start");
         self.call(intrinsic, &[va_list], None)

src/librustc_codegen_ssa/mir/block.rs

@@ -148,6 +148,24 @@ impl<'a, 'tcx> TerminatorCodegenHelper<'a, 'tcx> {
             }
         }
     }
+
+    // Generate sideeffect intrinsic if jumping to any of the targets can form
+    // a loop.
+    fn maybe_sideeffect<'b, 'tcx2: 'b, Bx: BuilderMethods<'b, 'tcx2>>(
+        &self,
+        mir: &'b mir::Body<'tcx>,
+        bx: &mut Bx,
+        targets: &[mir::BasicBlock],
+    ) {
+        if targets.iter().any(|target| {
+            *target <= *self.bb
+                && target
+                    .start_location()
+                    .is_predecessor_of(self.bb.start_location(), mir)
+        }) {
+            bx.sideeffect();
+        }
+    }
 }
 
 /// Codegen implementations for some terminator variants.
@@ -196,6 +214,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             let lltrue = helper.llblock(self, targets[0]);
             let llfalse = helper.llblock(self, targets[1]);
             if switch_ty == bx.tcx().types.bool {
+                helper.maybe_sideeffect(self.mir, &mut bx, targets.as_slice());
                 // Don't generate trivial icmps when switching on bool
                 if let [0] = values[..] {
                     bx.cond_br(discr.immediate(), llfalse, lltrue);
@@ -209,9 +228,11 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                 );
                 let llval = bx.const_uint_big(switch_llty, values[0]);
                 let cmp = bx.icmp(IntPredicate::IntEQ, discr.immediate(), llval);
+                helper.maybe_sideeffect(self.mir, &mut bx, targets.as_slice());
                 bx.cond_br(cmp, lltrue, llfalse);
             }
         } else {
+            helper.maybe_sideeffect(self.mir, &mut bx, targets.as_slice());
             let (otherwise, targets) = targets.split_last().unwrap();
             bx.switch(
                 discr.immediate(),
@@ -310,6 +331,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
 
         if let ty::InstanceDef::DropGlue(_, None) = drop_fn.def {
             // we don't actually need to drop anything.
+            helper.maybe_sideeffect(self.mir, &mut bx, &[target]);
             helper.funclet_br(self, &mut bx, target);
             return
         }
@@ -340,6 +362,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                  FnType::of_instance(&bx, drop_fn))
             }
         };
+        bx.sideeffect();
         helper.do_call(self, &mut bx, fn_ty, drop_fn, args,
                        Some((ReturnDest::Nothing, target)),
                        unwind);
@@ -375,6 +398,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
 
         // Don't codegen the panic block if success if known.
         if const_cond == Some(expected) {
+            helper.maybe_sideeffect(self.mir, &mut bx, &[target]);
             helper.funclet_br(self, &mut bx, target);
             return;
         }
@@ -385,6 +409,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         // Create the failure block and the conditional branch to it.
         let lltarget = helper.llblock(self, target);
         let panic_block = self.new_block("panic");
+        helper.maybe_sideeffect(self.mir, &mut bx, &[target]);
         if expected {
             bx.cond_br(cond, lltarget, panic_block.llbb());
         } else {
@@ -437,6 +462,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         let fn_ty = FnType::of_instance(&bx, instance);
         let llfn = bx.get_fn(instance);
 
+        bx.sideeffect();
         // Codegen the actual panic invoke/call.
         helper.do_call(self, &mut bx, fn_ty, llfn, &args, None, cleanup);
     }
@@ -488,6 +514,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             if let Some(destination_ref) = destination.as_ref() {
                 let &(ref dest, target) = destination_ref;
                 self.codegen_transmute(&mut bx, &args[0], dest);
+                helper.maybe_sideeffect(self.mir, &mut bx, &[target]);
                 helper.funclet_br(self, &mut bx, target);
             } else {
                 // If we are trying to transmute to an uninhabited type,
@@ -518,6 +545,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             Some(ty::InstanceDef::DropGlue(_, None)) => {
                 // Empty drop glue; a no-op.
                 let &(_, target) = destination.as_ref().unwrap();
+                helper.maybe_sideeffect(self.mir, &mut bx, &[target]);
                 helper.funclet_br(self, &mut bx, target);
                 return;
             }
@@ -554,6 +582,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                 let fn_ty = FnType::of_instance(&bx, instance);
                 let llfn = bx.get_fn(instance);
 
+                bx.sideeffect();
                 // Codegen the actual panic invoke/call.
                 helper.do_call(
                     self,
@@ -566,7 +595,9 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                 );
             } else {
                 // a NOP
-                helper.funclet_br(self, &mut bx, destination.as_ref().unwrap().1)
+                let target = destination.as_ref().unwrap().1;
+                helper.maybe_sideeffect(self.mir, &mut bx, &[target]);
+                helper.funclet_br(self, &mut bx, target);
             }
             return;
         }
@@ -675,6 +706,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             }
 
             if let Some((_, target)) = *destination {
+                helper.maybe_sideeffect(self.mir, &mut bx, &[target]);
                 helper.funclet_br(self, &mut bx, target);
             } else {
                 bx.unreachable();
@@ -786,6 +818,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             _ => span_bug!(span, "no llfn for call"),
         };
 
+        bx.sideeffect();
         helper.do_call(self, &mut bx, fn_ty, fn_ptr, &llargs,
                        destination.as_ref().map(|&(_, target)| (ret_dest, target)),
                        cleanup);
@@ -835,6 +868,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             }
 
             mir::TerminatorKind::Goto { target } => {
+                helper.maybe_sideeffect(self.mir, &mut bx, &[target]);
                 helper.funclet_br(self, &mut bx, target);
             }
 

src/librustc_codegen_ssa/mir/rvalue.rs

@@ -486,6 +486,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                 };
                 let instance = ty::Instance::mono(bx.tcx(), def_id);
                 let r = bx.cx().get_fn(instance);
+                bx.sideeffect();
                 let call = bx.call(r, &[llsize, llalign], None);
                 let val = bx.pointercast(call, llty_ptr);
 

src/librustc_codegen_ssa/traits/intrinsic.rs

@@ -20,6 +20,7 @@ pub trait IntrinsicCallMethods<'tcx>: BackendTypes {
     fn abort(&mut self);
     fn assume(&mut self, val: Self::Value);
     fn expect(&mut self, cond: Self::Value, expected: bool) -> Self::Value;
+    fn sideeffect(&mut self);
     /// Trait method used to inject `va_start` on the "spoofed" `VaListImpl` in
     /// Rust defined C-variadic functions.
     fn va_start(&mut self, val: Self::Value) -> Self::Value;

src/test/codegen/alloc-optimisation.rs

@@ -7,7 +7,8 @@
 pub fn alloc_test(data: u32) {
     // CHECK-LABEL: @alloc_test
     // CHECK-NEXT: start:
-    // CHECK-NEXT: ret void
+    // CHECK-NOT: alloc
+    // CHECK: ret void
     let x = Box::new(data);
     drop(x);
 }

src/test/codegen/dealloc-no-unwind.rs

@@ -17,7 +17,7 @@ impl Drop for A {
 pub fn a(a: Box<i32>) {
     // CHECK-LABEL: define void @a
     // CHECK: call void @__rust_dealloc
-    // CHECK-NEXT: call void @foo
+    // CHECK: call void @foo
     let _a = A;
     drop(a);
 }

src/test/codegen/issue-34947-pow-i32.rs

@@ -6,8 +6,11 @@
 #[no_mangle]
 pub fn issue_34947(x: i32) -> i32 {
     // CHECK: mul
-    // CHECK-NEXT: mul
+    // CHECK-NOT: br label
-    // CHECK-NEXT: mul
+    // CHECK: mul
-    // CHECK-NEXT: ret
+    // CHECK-NOT: br label
+    // CHECK: mul
+    // CHECK-NOT: br label
+    // CHECK: ret
     x.pow(5)
 }

src/test/codegen/issue-45222.rs

@@ -1,3 +1,9 @@
+// ignore-test
+
+// FIXME:
+// LLVM can't optimize some loops with a large number of iterations because of
+// @llvm.sideeffect() (see also #59546)
+
 // compile-flags: -O
 // ignore-debug: the debug assertions get in the way
 

src/test/codegen/naked-functions.rs

@@ -1,5 +1,3 @@
-// ignore-tidy-linelength
-
 // compile-flags: -C no-prepopulate-passes
 
 #![crate_type = "lib"]
@@ -53,7 +51,7 @@ pub fn naked_with_args_and_return(a: isize) -> isize {
 #[naked]
 pub fn naked_recursive() {
     // CHECK-NEXT: {{.+}}:
-    // CHECK-NEXT: call void @naked_empty()
+    // CHECK: call void @naked_empty()
 
     // FIXME(#39685) Avoid one block per call.
     // CHECK-NEXT: br label %bb1
@@ -61,19 +59,19 @@ pub fn naked_recursive() {
 
     naked_empty();
 
-    // CHECK-NEXT: %{{[0-9]+}} = call i{{[0-9]+}} @naked_with_return()
+    // CHECK: %{{[0-9]+}} = call i{{[0-9]+}} @naked_with_return()
 
     // FIXME(#39685) Avoid one block per call.
     // CHECK-NEXT: br label %bb2
     // CHECK: bb2:
 
-    // CHECK-NEXT: %{{[0-9]+}} = call i{{[0-9]+}} @naked_with_args_and_return(i{{[0-9]+}} %{{[0-9]+}})
+    // CHECK: %{{[0-9]+}} = call i{{[0-9]+}} @naked_with_args_and_return(i{{[0-9]+}} %{{[0-9]+}})
 
     // FIXME(#39685) Avoid one block per call.
     // CHECK-NEXT: br label %bb3
     // CHECK: bb3:
 
-    // CHECK-NEXT: call void @naked_with_args(i{{[0-9]+}} %{{[0-9]+}})
+    // CHECK: call void @naked_with_args(i{{[0-9]+}} %{{[0-9]+}})
 
     // FIXME(#39685) Avoid one block per call.
     // CHECK-NEXT: br label %bb4

src/test/codegen/non-terminate/infinite-loop-1.rs

@@ -0,0 +1,17 @@
+// compile-flags: -C opt-level=3
+
+#![crate_type = "lib"]
+
+fn infinite_loop() -> u8 {
+    loop {}
+}
+
+// CHECK-LABEL: @test
+#[no_mangle]
+fn test() -> u8 {
+    // CHECK-NOT: unreachable
+    // CHECK: br label %{{.+}}
+    // CHECK-NOT: unreachable
+    let x = infinite_loop();
+    x
+}

src/test/codegen/non-terminate/infinite-loop-2.rs

@@ -0,0 +1,19 @@
+// compile-flags: -C opt-level=3
+
+#![crate_type = "lib"]
+
+fn infinite_loop() -> u8 {
+    let i = 2;
+    while i > 1 {}
+    1
+}
+
+// CHECK-LABEL: @test
+#[no_mangle]
+fn test() -> u8 {
+    // CHECK-NOT: unreachable
+    // CHECK: br label %{{.+}}
+    // CHECK-NOT: unreachable
+    let x = infinite_loop();
+    x
+}

src/test/codegen/non-terminate/infinite-recursion.rs

@@ -0,0 +1,14 @@
+// compile-flags: -C opt-level=3
+
+#![crate_type = "lib"]
+
+#![allow(unconditional_recursion)]
+
+// CHECK-LABEL: @infinite_recursion
+#[no_mangle]
+fn infinite_recursion() -> u8 {
+    // CHECK-NOT: ret i8 undef
+    // CHECK: br label %{{.+}}
+    // CHECK-NOT: ret i8 undef
+    infinite_recursion()
+}

src/test/codegen/repeat-trusted-len.rs

@@ -14,6 +14,10 @@ pub fn helper(_: usize) {
 // CHECK-LABEL: @repeat_take_collect
 #[no_mangle]
 pub fn repeat_take_collect() -> Vec<u8> {
-// CHECK: call void @llvm.memset.p0i8.[[USIZE]](i8* {{(nonnull )?}}align 1 %{{[0-9]+}}, i8 42, [[USIZE]] 100000, i1 false)
+// FIXME: At the time of writing LLVM transforms this loop into a single
+// `store` and then a `memset` with size = 99999. The correct check should be:
+//        call void @llvm.memset.p0i8.[[USIZE]](i8* {{(nonnull )?}}align 1 %{{[a-z0-9.]+}}, i8 42, [[USIZE]] 100000, i1 false)
+
+// CHECK: call void @llvm.memset.p0i8.[[USIZE]](i8* {{(nonnull )?}}align 1 %{{[a-z0-9.]+}}, i8 42, [[USIZE]] 99999, i1 false)
     iter::repeat(42).take(100000).collect()
 }

src/test/codegen/vec-clear.rs

@@ -1,3 +1,9 @@
+// ignore-test
+
+// FIXME:
+// LLVM can't optimize some loops with unknown number of iterations because of
+// @llvm.sideeffect() (see also #59546)
+
 // ignore-debug: the debug assertions get in the way
 // compile-flags: -O
 

src/test/codegen/vec-iter-collect-len.rs

@@ -5,8 +5,6 @@
 
 #[no_mangle]
 pub fn get_len() -> usize {
-    // CHECK-LABEL: @get_len
+    // CHECK-COUNT-1: {{^define}}
-    // CHECK-NOT: call
-    // CHECK-NOT: invoke
     [1, 2, 3].iter().collect::<Vec<_>>().len()
 }

src/test/codegen/vec-optimizes-away.rs

@@ -8,6 +8,6 @@
 pub fn sum_me() -> i32 {
     // CHECK-LABEL: @sum_me
     // CHECK-NEXT: {{^.*:$}}
-    // CHECK-NEXT: ret i32 6
+    // CHECK: ret i32 6
     vec![1, 2, 3].iter().sum::<i32>()
 }

src/test/run-make-fulldeps/inline-always-many-cgu/Makefile

@@ -2,7 +2,7 @@
 
 all:
 	$(RUSTC) foo.rs --emit llvm-ir -C codegen-units=2
-	if cat $(TMPDIR)/*.ll | $(CGREP) -e '\bcall\b'; then \
+	if cat $(TMPDIR)/*.ll | grep -v 'call void @llvm.sideeffect()' | $(CGREP) -e '\bcall\b'; then \
 		echo "found call instruction when one wasn't expected"; \
 		exit 1; \
 	fi