Reorganise optimisation level settings, and make the insn shrink pass optional (#1044)

This patch:

* removes the "default" opt level, on the basis that it has no definition and
  is referred to nowhere in the compiler.

* renames the "fastest" level to "none".  The resulting set of transformations
  is unchanged.

* renames the "best" level to "speed_and_size".  The resulting set of
  transformations is unchanged.

* adds a new level, "speed".  This is the same as "speed_and_size" except that
  it omits transformations aimed only at reducing code size.  Currently it
  omits only the insn shrinking pass.

cranelift/codegen/meta/src/shared/settings.rs

@@ -8,11 +8,12 @@ pub fn define() -> SettingGroup {
         r#"
         Optimization level:
 
-        - default: Very profitable optimizations enabled, none slow.
+        - none: Minimise compile time by disabling most optimizations.
-        - best: Enable all optimizations
+        - speed: Generate the fastest possible code
-        - fastest: Optimize for compile time by disabling most optimizations.
+        - speed_and_size: like "speed", but also perform transformations
+          aimed at reducing code size.
         "#,
-        vec!["default", "best", "fastest"],
+        vec!["none", "speed", "speed_and_size"],
     );
 
     settings.add_bool(

cranelift/codegen/src/context.rs

@@ -132,18 +132,18 @@ impl Context {
         self.verify_if(isa)?;
         debug!("Compiling:\n{}", self.func.display(isa));
 
+        let opt_level = isa.flags().opt_level();
+
         self.compute_cfg();
-        if isa.flags().opt_level() != OptLevel::Fastest {
+        if opt_level != OptLevel::None {
             self.preopt(isa)?;
         }
         if isa.flags().enable_nan_canonicalization() {
             self.canonicalize_nans(isa)?;
         }
         self.legalize(isa)?;
-        if isa.flags().opt_level() != OptLevel::Fastest {
+        if opt_level != OptLevel::None {
             self.postopt(isa)?;
-        }
-        if isa.flags().opt_level() == OptLevel::Best {
             self.compute_domtree();
             self.compute_loop_analysis();
             self.licm(isa)?;
@@ -151,13 +151,15 @@ impl Context {
         }
         self.compute_domtree();
         self.eliminate_unreachable_code(isa)?;
-        if isa.flags().opt_level() != OptLevel::Fastest {
+        if opt_level != OptLevel::None {
             self.dce(isa)?;
         }
         self.regalloc(isa)?;
         self.prologue_epilogue(isa)?;
-        if isa.flags().opt_level() == OptLevel::Best {
+        if opt_level == OptLevel::Speed || opt_level == OptLevel::SpeedAndSize {
             self.redundant_reload_remover(isa)?;
+        }
+        if opt_level == OptLevel::SpeedAndSize {
             self.shrink_instructions(isa)?;
         }
         let result = self.relax_branches(isa);

cranelift/codegen/src/settings.rs

@@ -14,10 +14,10 @@
 //! use cranelift_codegen::settings::{self, Configurable};
 //!
 //! let mut b = settings::builder();
-//! b.set("opt_level", "fastest");
+//! b.set("opt_level", "speed_and_size");
 //!
 //! let f = settings::Flags::new(b);
-//! assert_eq!(f.opt_level(), settings::OptLevel::Fastest);
+//! assert_eq!(f.opt_level(), settings::OptLevel::SpeedAndSize);
 //! ```
 
 use crate::constant_hash::{probe, simple_hash};
@@ -378,7 +378,7 @@ mod tests {
         assert_eq!(
             f.to_string(),
             "[shared]\n\
-             opt_level = \"default\"\n\
+             opt_level = \"none\"\n\
              libcall_call_conv = \"isa_default\"\n\
              baldrdash_prologue_words = 0\n\
              probestack_size_log2 = 12\n\
@@ -398,7 +398,7 @@ mod tests {
              probestack_func_adjusts_sp = false\n\
              jump_tables_enabled = true\n"
         );
-        assert_eq!(f.opt_level(), super::OptLevel::Default);
+        assert_eq!(f.opt_level(), super::OptLevel::None);
         assert_eq!(f.enable_simd(), false);
         assert_eq!(f.baldrdash_prologue_words(), 0);
     }
@@ -428,13 +428,15 @@ mod tests {
         );
         assert_eq!(
             b.set("opt_level", "true"),
-            Err(BadValue("any among default, best, fastest".to_string()))
+            Err(BadValue(
+                "any among none, speed, speed_and_size".to_string()
+            ))
         );
-        assert_eq!(b.set("opt_level", "best"), Ok(()));
+        assert_eq!(b.set("opt_level", "speed"), Ok(()));
         assert_eq!(b.set("enable_simd", "0"), Ok(()));
 
         let f = Flags::new(b);
         assert_eq!(f.enable_simd(), false);
-        assert_eq!(f.opt_level(), super::OptLevel::Best);
+        assert_eq!(f.opt_level(), super::OptLevel::Speed);
     }
 }

cranelift/filetests/filetests/isa/x86/allones_funcaddrs32.clif

@@ -1,6 +1,6 @@
 ; binary emission of 32-bit code.
 test binemit
-set opt_level=best
+set opt_level=speed_and_size
 set allones_funcaddrs
 target i686 haswell
 

cranelift/filetests/filetests/isa/x86/allones_funcaddrs64.clif

@@ -1,6 +1,6 @@
 ; binary emission of 64-bit code.
 test binemit
-set opt_level=best
+set opt_level=speed_and_size
 set allones_funcaddrs
 target x86_64 haswell
 

cranelift/filetests/filetests/isa/x86/baseline_clz_ctz_popcount_encoding.clif

@@ -1,5 +1,5 @@
 test binemit
-set opt_level=best
+set opt_level=speed_and_size
 target x86_64 baseline
 
 ; The binary encodings can be verified with the command:

cranelift/filetests/filetests/isa/x86/binary32.clif

@@ -1,6 +1,6 @@
 ; binary emission of x86-32 code.
 test binemit
-set opt_level=best
+set opt_level=speed_and_size
 target i686 haswell
 
 ; The binary encodings can be verified with the command:

cranelift/filetests/filetests/isa/x86/binary64-float.clif

@@ -1,6 +1,6 @@
 ; Binary emission of 64-bit floating point code.
 test binemit
-set opt_level=best
+set opt_level=speed_and_size
 target x86_64 haswell
 
 ; The binary encodings can be verified with the command:

cranelift/filetests/filetests/isa/x86/binary64-pic.clif

@@ -1,6 +1,6 @@
 ; binary emission of 64-bit code.
 test binemit
-set opt_level=best
+set opt_level=speed_and_size
 set is_pic
 target x86_64 haswell
 

cranelift/filetests/filetests/isa/x86/binary64.clif

@@ -1,6 +1,6 @@
 ; binary emission of x86-64 code.
 test binemit
-set opt_level=best
+set opt_level=speed_and_size
 target x86_64 haswell
 
 ; The binary encodings can be verified with the command:

cranelift/filetests/filetests/isa/x86/isub_imm-i8.clif

@@ -1,4 +1,5 @@
 test compile
+set opt_level=speed_and_size
 target x86_64
 
 function u0:0(i8) -> i8 fast {

cranelift/filetests/filetests/isa/x86/legalize-br-table.clif

@@ -1,4 +1,5 @@
 test compile
+set opt_level=speed_and_size
 target x86_64
 feature !"basic-blocks"
 ; regex: V=v\d+

cranelift/filetests/filetests/isa/x86/legalize-call.clif

@@ -1,6 +1,6 @@
 ; Test legalization of a non-colocated call in 64-bit non-PIC mode.
 test legalizer
-set opt_level=best
+set opt_level=speed_and_size
 target x86_64 haswell
 
 function %call() {

cranelift/filetests/filetests/isa/x86/optimized-zero-constants-32bit.clif

@@ -1,6 +1,6 @@
 ; Check that floating-point and integer constants equal to zero are optimized correctly.
 test binemit
-set opt_level=best
+set opt_level=speed_and_size
 target i686
 
 function %foo() -> f32 fast {

cranelift/filetests/filetests/isa/x86/optimized-zero-constants.clif

@@ -1,6 +1,6 @@
 ; Check that floating-point constants equal to zero are optimized correctly.
 test binemit
-set opt_level=best
+set opt_level=speed_and_size
 target x86_64
 
 function %zero_const_32bit_no_rex() -> f32 fast {

cranelift/filetests/filetests/isa/x86/pinned-reg.clif

@@ -2,7 +2,7 @@ test compile
 
 set enable_pinned_reg=true
 set use_pinned_reg_as_heap_base=true
-set opt_level=best
+set opt_level=speed_and_size
 
 target x86_64
 

cranelift/filetests/filetests/isa/x86/prologue-epilogue.clif

@@ -1,5 +1,5 @@
 test compile
-set opt_level=best
+set opt_level=speed_and_size
 set is_pic
 target x86_64 haswell
 

cranelift/filetests/filetests/isa/x86/relax_branch.clif

@@ -1,5 +1,5 @@
 test binemit
-set opt_level=best
+set opt_level=speed_and_size
 set avoid_div_traps
 set baldrdash_prologue_words=3
 set allones_funcaddrs

cranelift/filetests/filetests/isa/x86/scalar_to_vector-binemit.clif

@@ -1,5 +1,5 @@
 test binemit
-set opt_level=best
+set opt_level=speed_and_size
 set enable_simd
 target x86_64
 

cranelift/filetests/filetests/isa/x86/scalar_to_vector-compile.clif

@@ -1,5 +1,5 @@
 test compile
-set opt_level=best
+set opt_level=speed_and_size
 set probestack_enabled=false
 set enable_simd
 target x86_64

cranelift/filetests/filetests/isa/x86/shrink-multiple-uses.clif

@@ -1,5 +1,5 @@
 test shrink
-set opt_level=best
+set opt_level=speed_and_size
 target x86_64
 
 function %test_multiple_uses(i32 [%rdi]) -> i32 {

cranelift/filetests/filetests/isa/x86/shrink.clif

@@ -1,5 +1,5 @@
 test binemit
-set opt_level=best
+set opt_level=speed_and_size
 target x86_64
 
 ; Test that instruction shrinking eliminates REX prefixes when possible.

cranelift/filetests/filetests/isa/x86/stack-addr64.clif

@@ -1,6 +1,6 @@
 ; binary emission of stack address instructions on x86-64.
 test binemit
-set opt_level=fastest
+set opt_level=none
 target x86_64 haswell
 
 ; The binary encodings can be verified with the command:

cranelift/filetests/filetests/isa/x86/stack-load-store64.clif

@@ -1,6 +1,6 @@
 ; legalization of stack load and store instructions on x86-64.
 test legalizer
-set opt_level=fastest
+set opt_level=none
 target x86_64 haswell
 
 function %stack_load_and_store() {

cranelift/filetests/filetests/isa/x86/vconst-binemit.clif

@@ -1,5 +1,5 @@
 test binemit
-set opt_level=best
+set opt_level=speed_and_size
 set enable_simd
 target x86_64
 

cranelift/filetests/filetests/isa/x86/windows_fastcall_x64.clif

@@ -1,5 +1,5 @@
 test compile
-set opt_level=best
+set opt_level=speed_and_size
 set is_pic
 target x86_64 haswell
 

cranelift/filetests/src/test_binemit.rs

@@ -162,7 +162,7 @@ impl SubTest for TestBinEmit {
                                 recipe_constraints.satisfied(inst, &divert, &func)
                             });
 
-                        if opt_level == OptLevel::Best {
+                        if opt_level == OptLevel::SpeedAndSize {
                             // Get the smallest legal encoding
                             legal_encodings
                                 .min_by_key(|&e| encinfo.byte_size(e, inst, &divert, &func))