Implement coloring::iterate_solution().

It can happen that the currently live registers are blocking a smaller
register class completely, so the only way of solving the allocation
problem is to turn some of the live-through registers into solver
variables.

When the quick_solve attempt fails, try to free up registers in the
critical register class by turning live-through values into solver
variables.

cranelift/filetests/regalloc/iterate.cton

@@ -0,0 +1,107 @@
+test compile
+set is_64bit
+isa intel haswell
+
+function #00000009(i64 [%rdi], f32 [%xmm0], f64 [%xmm1], i32 [%rsi], i32 [%rdx], i64 vmctx [%r14]) -> i64 [%rax] spiderwasm {
+ebb0(v0: i64, v1: f32, v2: f64, v3: i32, v4: i32, v5: i64):
+    v32 = iconst.i32 0
+    v6 = bitcast.f32 v32
+    v7 = iconst.i64 0
+    v33 = iconst.i64 0
+    v8 = bitcast.f64 v33
+    v34 = iconst.i32 0xbe99_999a
+    v9 = bitcast.f32 v34
+    v10 = iconst.i32 40
+    v11 = iconst.i32 -7
+    v35 = iconst.i32 0x40b0_0000
+    v12 = bitcast.f32 v35
+    v13 = iconst.i64 6
+    v36 = iconst.i64 0x4020_0000_0000_0000
+    v14 = bitcast.f64 v36
+    v44 = iconst.i64 0
+    v37 = icmp slt v0, v44
+    brnz v37, ebb2
+    v38 = fcvt_from_sint.f64 v0
+    jump ebb3(v38)
+
+ebb2:
+    v45 = iconst.i32 1
+    v39 = ushr.i64 v0, v45
+    v40 = band_imm.i64 v0, 1
+    v41 = bor v39, v40
+    v42 = fcvt_from_sint.f64 v41
+    v43 = fadd v42, v42
+    jump ebb3(v43)
+
+ebb3(v15: f64):
+    v16 = fpromote.f64 v9
+    v46 = uextend.i64 v10
+    v17 = fcvt_from_sint.f64 v46
+    v18 = fcvt_from_sint.f64 v11
+    v19 = fpromote.f64 v12
+    v54 = iconst.i64 0
+    v47 = icmp.i64 slt v13, v54
+    brnz v47, ebb4
+    v48 = fcvt_from_sint.f64 v13
+    jump ebb5(v48)
+
+ebb4:
+    v55 = iconst.i32 1
+    v49 = ushr.i64 v13, v55
+    v50 = band_imm.i64 v13, 1
+    v51 = bor v49, v50
+    v52 = fcvt_from_sint.f64 v51
+    v53 = fadd v52, v52
+    jump ebb5(v53)
+
+ebb5(v20: f64):
+    v63 = iconst.i64 0
+    v56 = icmp.i64 slt v7, v63
+    brnz v56, ebb6
+    v57 = fcvt_from_sint.f64 v7
+    jump ebb7(v57)
+
+ebb6:
+    v64 = iconst.i32 1
+    v58 = ushr.i64 v7, v64
+    v59 = band_imm.i64 v7, 1
+    v60 = bor v58, v59
+    v61 = fcvt_from_sint.f64 v60
+    v62 = fadd v61, v61
+    jump ebb7(v62)
+
+ebb7(v21: f64):
+    v22 = fadd v21, v14
+    v23 = fadd.f64 v20, v22
+    v24 = fadd.f64 v19, v23
+    v25 = fadd.f64 v18, v24
+    v26 = fadd.f64 v17, v25
+    v27 = fadd.f64 v2, v26
+    v28 = fadd.f64 v16, v27
+    v29 = fadd.f64 v15, v28
+    v30 = x86_cvtt2si.i64 v29
+    v69 = iconst.i64 0x8000_0000_0000_0000
+    v65 = icmp ne v30, v69
+    brnz v65, ebb8
+    v66 = fcmp uno v29, v29
+    brz v66, ebb9
+    trap bad_toint
+
+ebb9:
+    v70 = iconst.i64 0xc3e0_0000_0000_0000
+    v67 = bitcast.f64 v70
+    v68 = fcmp gt v67, v29
+    brz v68, ebb10
+    trap int_ovf
+
+ebb10:
+    jump ebb8
+
+ebb8:
+    jump ebb1(v30)
+
+ebb1(v31: i64):
+    return v31
+}
+
+

lib/cretonne/meta/isa/intel/encodings.py

@@ -119,6 +119,7 @@ enc_i32_i64(x86.udivmodx, r.div, 0xf7, rrr=6)
 
 enc_i32_i64(base.copy, r.umr, 0x89)
 enc_i32_i64(base.regmove, r.rmov, 0x89)
+enc_flt(base.regmove.b1, r.rmov, 0x89)
 
 # Immediate instructions with sign-extended 8-bit and 32-bit immediate.
 for inst,               rrr in [

lib/cretonne/src/isa/registers.rs

@@ -167,7 +167,7 @@ impl RegClassData {
     }
 
     /// Returns true if `other` is a subclass of this register class.
-    /// A register class is considerd to be a subclass of itself.
+    /// A register class is considered to be a subclass of itself.
     pub fn has_subclass<RCI: Into<RegClassIndex>>(&self, other: RCI) -> bool {
         self.subclasses & (1 << other.into().0) != 0
     }
@@ -276,6 +276,11 @@ impl RegInfo {
     pub fn rc(&self, idx: RegClassIndex) -> RegClass {
         &self.classes[idx.index()]
     }
+
+    /// Get the top-level register class containing `rc`.
+    pub fn toprc(&self, rc: RegClass) -> RegClass {
+        &self.classes[rc.toprc as usize]
+    }
 }
 
 /// Temporary object that holds enough information to print a register unit.

lib/cretonne/src/regalloc/coloring.rs

@@ -367,9 +367,10 @@ impl<'a> Context<'a> {
 
         // Finally, we've fully programmed the constraint solver.
         // We expect a quick solution in most cases.
-        let mut output_regs = self.solver.quick_solve().unwrap_or_else(
+        let mut output_regs = self.solver.quick_solve().unwrap_or_else(|rc| {
-            |_| self.iterate_solution(),
+            dbg!("quick_solve needs more registers in {}", rc);
-        );
+            self.iterate_solution(throughs, locations)
+        });
 
 
         // The solution and/or fixed input constraints may require us to shuffle the set of live
@@ -731,8 +732,52 @@ impl<'a> Context<'a> {
     ///
     /// We may need to move more registers around before a solution is possible. Use an iterative
     /// algorithm that adds one more variable until a solution can be found.
-    fn iterate_solution(&self) -> AllocatableSet {
+    fn iterate_solution(
-        unimplemented!();
+        &mut self,
+        throughs: &[LiveValue],
+        locations: &mut ValueLocations,
+    ) -> AllocatableSet {
+        loop {
+            dbg!("real_solve for {} variables", self.solver.vars().len());
+            let rc = match self.solver.real_solve() {
+                Ok(regs) => return regs,
+                Err(rc) => rc,
+            };
+
+            // Do we have any live-through `rc` registers that are not already variables?
+            assert!(
+                self.try_add_var(rc, throughs, locations),
+                "Ran out of registers in {}",
+                rc
+            );
+        }
+    }
+
+    /// Try to add an `rc` variable to the solver from the `throughs` set.
+    fn try_add_var(
+        &mut self,
+        rc: RegClass,
+        throughs: &[LiveValue],
+        locations: &mut ValueLocations,
+    ) -> bool {
+        dbg!("Trying to add a {} reg from {} values", rc, throughs.len());
+
+        for lv in throughs {
+            if let Affinity::Reg(rci) = lv.affinity {
+                let rc2 = self.reginfo.rc(rci);
+                let reg2 = self.divert.reg(lv.value, locations);
+                if rc.contains(reg2) && self.solver.can_add_var(lv.value, rc2, reg2) {
+                    // The new variable gets to roam the whole top-level register class because
+                    // it is not actually constrained by the instruction. We just want it out
+                    // of the way.
+                    let toprc = self.reginfo.toprc(rc2);
+                    self.solver.add_var(lv.value, toprc, reg2, &self.reginfo);
+                    return true;
+                }
+            }
+        }
+
+        false
     }
 
     /// Emit `regmove` instructions as needed to move the live registers into place before the

lib/cretonne/src/regalloc/solver.rs

@@ -590,6 +590,18 @@ impl Solver {
         self.find_solution()
     }
 
+    /// Try harder to find a solution.
+    ///
+    /// Call this method after `quick_solve()` fails.
+    ///
+    /// This may return an error with a register class that has run out of registers. If registers
+    /// can be freed up in the starving class, this method can be called again after adding
+    /// variables for the freed registers.
+    pub fn real_solve(&mut self) -> Result<AllocatableSet, RegClass> {
+        // TODO: Sort variables to assign smallest register classes first.
+        self.find_solution()
+    }
+
     /// Search for a solution with the current list of variables.
     ///
     /// If a solution was found, returns `Ok(regs)` with the set of available registers on the
@@ -623,6 +635,11 @@ impl Solver {
     pub fn vars(&self) -> &[Variable] {
         &self.vars
     }
+
+    /// Check if `value` can be added as a variable to help find a solution.
+    pub fn can_add_var(&mut self, _value: Value, constraint: RegClass, from: RegUnit) -> bool {
+        !self.regs_in.is_avail(constraint, from)
+    }
 }
 
 /// Interface for working with parallel copies once a solution has been found.