Debugging for a probable AVX2-related rustc bug

This commit contains `dbg!` statements which are narrowed down to be as
close to the problem as possible.

The following commands succeed:

- `cargo test donna_self_test1 -- --test-threads=1 --nocapture`
- `RUSTFLAGS="-Ctarget-cpu=haswell" cargo test donna_self_test1 --release -- --test-threads=1 --nocapture`

The following command FAILS with a miscomputed output:

- `cargo test donna_self_test1 --release -- --test-threads=1 --nocapture`

Included in this commit are `dbg!` statements capturing the miscompiled
output and comments with instructions about small tweaks which make the
tests succeed under the conditions where it presently fails.

These include:

- Inserting a `dbg!` statement in the right place
- Commenting out a `#[target_feature(enable = "avx2")]` attribute

Author: tarcieri

poly1305/src/backend/avx2.rs

@@ -102,6 +102,10 @@ impl State {
     }
 
     /// Finalize output producing a [`Tag`]
+    //
+    // BUG!
+    // Comment out `#[target_feature(...)]` to make the tests pass
+    //
     #[target_feature(enable = "avx2")]
     pub(crate) unsafe fn finalize(&mut self) -> Tag {
         assert!(self.num_cached_blocks < 4);
@@ -131,7 +135,11 @@ impl State {
             if let Some(p) = p {
                 c = c + p;
             }
-            p = Some((c * self.r1).reduce());
+            let tmp = c * self.r1;
+            //dbg!(&tmp);
+            // NOTE: this is the call site where miscompilation is happening
+            p = Some(tmp.reduce());
+            //dbg!(&p);
             self.num_cached_blocks -= 1;
         }
 
@@ -144,15 +152,14 @@ impl State {
             p = Some((c * self.r1).reduce());
         }
 
-        dbg!(&self);
-
         // Compute tag: p + k mod 2^128
         let mut tag = GenericArray::<u8, _>::default();
         let tag_int = if let Some(p) = p {
             self.k + p
         } else {
             self.k.into()
         };
+
         tag_int.write(tag.as_mut_slice());
 
         Tag::new(tag)

poly1305/src/backend/avx2/helpers.rs

@@ -455,6 +455,11 @@ impl Unreduced130 {
 
             // Carry chain
             let adc = |v1: __m256i, v0: __m256i| -> (__m256i, __m256i) {
+                //
+                // HEISENBUG! Uncomment this line to make the tests pass!
+                //
+                //dbg!(&v1, &v0);
+
                 //   [t_3,       t_2 % 2^26, t_1 % 2^26, t_0 % 2^26]
                 // + [t_2 >> 26, t_1 >>  26, t_0 >>  26,  0        ]
                 // = [
@@ -470,6 +475,11 @@ impl Unreduced130 {
                         set02(2, 1, 0, 3),
                     ),
                 );
+
+                // BUG!
+                // This is the location closest to where miscompiled output is observed
+                dbg!(&v1, &v0);
+
                 //   [_, _, _, t_4]
                 // + [
                 //     (t_2 % 2^26 + t_1 >> 26) >> 26,