Use available_registers() to set_hw properties consistently across LoopSet generating functions.

Author: chriselrod

src/condense_loopset.jl

@@ -220,6 +220,7 @@ function check_if_empty(ls::LoopSet, q::Expr)
 end
 
 val(x) = Expr(:call, Expr(:curly, :Val, x))
+
 # Try to condense in type stable manner
 function generate_call(ls::LoopSet, inline_unroll::NTuple{3,Int8}, debug::Bool = false)
     operation_descriptions = Expr(:tuple)
@@ -243,13 +244,7 @@ function generate_call(ls::LoopSet, inline_unroll::NTuple{3,Int8}, debug::Bool =
         :tuple, inline, u₁, u₂,
         Expr(:call, lv(:unwrap), VECTORWIDTHSYMBOL),
         Expr(:call, lv(:unwrap), Expr(:call, lv(:register_size))),
-        Expr(:call, lv(:unwrap),
-            Expr(:call, lv(:ifelse),
-                Expr(:call, lv(:unwrap), Expr(:call, lv(:has_opmask_registers))),
-                Expr(:call, lv(:unwrap), Expr(:call, lv(:register_count))),
-                Expr(:call, lv(:unwrap), Expr(:call, :(-), Expr(:call, lv(:register_count)), Expr(:call, lv(:One))))
-            )
-        ),
+        Expr(:call, lv(:unwrap), Expr(:call, lv(:available_registers))),
         Expr(:call, lv(:unwrap), Expr(:call, lv(:cache_linesize)))
     )
     q = Expr(

src/constructors.jl

@@ -31,7 +31,7 @@ function add_ci_call!(q::Expr, @nospecialize(f), args, syms, i, mod = nothing)
     end
     if mod !== nothing # indicates it's `vmaterialize(!)`
         reg_size = Expr(:call, lv(:register_size))
-        reg_count = Expr(:call, lv(:register_count))
+        reg_count = Expr(:call, lv(:available_registers))
         cache_lnsze = Expr(:call, lv(:cache_linesize))
         push!(call.args, Expr(:call, Expr(:curly, :Val, QuoteNode(mod))), reg_size, reg_count, cache_lnsze)
     end

src/graphs.jl

@@ -305,9 +305,10 @@ function set_hw!(ls::LoopSet, rs::Int, rc::Int, cls::Int)
     ls.cache_linesize[] = cls
     # ls.opmask_register[] = omr
 end
+available_registers() = ifelse(has_opmask_registers(), register_count(), register_count() - One())
 function set_hw!(ls::LoopSet)
     rs = Int(register_size())
-    rc = Int(register_count())
+    rc = Int(available_registers())
     cls = Int(cache_linesize())
     # omr = Bool(VectorizationBase.has_opmask_registers())
     set_hw!(ls, rs, rc, cls)