From cd8a42e01f1c869190f725ff880fd96364772318 Mon Sep 17 00:00:00 2001
From: Benjamin Bouvier <public@benj.me>
Date: Tue, 12 Mar 2019 12:06:11 +0100
Subject: [PATCH] Fix #686: Allow code shrink in relaxation if the shrinking
 pass hasn't been run;

Also:
- make sure to apply diversions when determining offsets for code
relaxation.
- select the smallest encoding when selecting a relaxed branch
instruction.
---
 cranelift/codegen/src/binemit/relaxation.rs   | 11 ++--
 .../regress/allow-relaxation-shrink.clif      | 53 +++++++++++++++++++
 2 files changed, 58 insertions(+), 6 deletions(-)
 create mode 100644 cranelift/filetests/filetests/regress/allow-relaxation-shrink.clif

diff --git a/cranelift/codegen/src/binemit/relaxation.rs b/cranelift/codegen/src/binemit/relaxation.rs
index 5bfa323646..c354a763cd 100644
--- a/cranelift/codegen/src/binemit/relaxation.rs
+++ b/cranelift/codegen/src/binemit/relaxation.rs
@@ -62,6 +62,7 @@ pub fn relax_branches(func: &mut Function, isa: &TargetIsa) -> CodegenResult<Cod
             divert.clear();
             cur.func.offsets[ebb] = offset;
             while let Some(inst) = cur.next_inst() {
+                divert.apply(&cur.func.dfg[inst]);
                 let enc = cur.func.encodings[inst];
                 offset += encinfo.byte_size(enc, inst, &divert, &cur.func);
             }
@@ -81,10 +82,6 @@ pub fn relax_branches(func: &mut Function, isa: &TargetIsa) -> CodegenResult<Cod
 
             // Record the offset for `ebb` and make sure we iterate until offsets are stable.
             if cur.func.offsets[ebb] != offset {
-                debug_assert!(
-                    cur.func.offsets[ebb] < offset,
-                    "Code shrinking during relaxation"
-                );
                 cur.func.offsets[ebb] = offset;
                 go_again = true;
             }
@@ -173,12 +170,12 @@ fn relax_branch(
         dest_offset
     );
 
-    // Pick the first encoding that can handle the branch range.
+    // Pick the smallest encoding that can handle the branch range.
     let dfg = &cur.func.dfg;
     let ctrl_type = dfg.ctrl_typevar(inst);
     if let Some(enc) = isa
         .legal_encodings(cur.func, &dfg[inst], ctrl_type)
-        .find(|&enc| {
+        .filter(|&enc| {
             let range = encinfo.branch_range(enc).expect("Branch with no range");
             if !range.contains(offset, dest_offset) {
                 debug!("  trying [{}]: out of range", encinfo.display(enc));
@@ -198,7 +195,9 @@ fn relax_branch(
                 true
             }
         })
+        .min_by_key(|&enc| encinfo.byte_size(enc, inst, &divert, &cur.func))
     {
+        debug_assert!(enc != cur.func.encodings[inst]);
         cur.func.encodings[inst] = enc;
         return encinfo.byte_size(enc, inst, &divert, &cur.func);
     }
diff --git a/cranelift/filetests/filetests/regress/allow-relaxation-shrink.clif b/cranelift/filetests/filetests/regress/allow-relaxation-shrink.clif
new file mode 100644
index 0000000000..21c0f847b5
--- /dev/null
+++ b/cranelift/filetests/filetests/regress/allow-relaxation-shrink.clif
@@ -0,0 +1,53 @@
+test compile
+target x86_64
+
+; This checks that code shrink is allowed while relaxing code, when code shrink
+; has not run.
+
+function u0:0(i64, i64) -> i64 system_v {
+    ss1 = explicit_slot 8
+    sig0 = (i64) -> i64 system_v
+    fn0 = u0:8 sig0
+
+ebb0(v0: i64, v1: i64):
+    v3 = stack_addr.i64 ss1
+    v5 = call fn0(v1)
+    v6 = iconst.i64 0
+    v8 = iconst.i64 0
+    jump ebb3(v6, v1, v8)
+
+ebb3(v39: i64, v40: i64, v42: i64):
+    v9 = load.i64 v3
+    v11 = icmp_imm ugt v9, 1
+    v12 = bint.i8 v11
+    v13 = uextend.i32 v12
+    v14 = icmp_imm eq v13, 0
+    brnz v14, ebb4
+    jump ebb5
+
+ebb4:
+    v18 = icmp_imm.i64 eq v40, 0
+    v19 = bint.i8 v18
+    v20 = uextend.i32 v19
+    brz v20, ebb6
+    trap user0
+
+ebb5:
+    v22 = iconst.i32 1
+    v23 = ishl.i64 v39, v22
+    v25 = iconst.i64 1
+    v26 = band.i64 v42, v25
+    v27 = bor v23, v26
+    v28 = iconst.i32 1
+    v29 = ushr.i64 v42, v28
+    v30 = load.i64 v3
+    v31 = iconst.i32 1
+    v32 = ushr v30, v31
+    store v32, v3
+    jump ebb3(v27, v40, v29)
+
+ebb6:
+    v38 = iconst.i64 0
+    return v38
+}
+