beam_bsm: Improve handling of bs_start_match2 instructions

The handling of bs_start_match2 was both too conservative and too
careless. It was too conservative in that would not do the
optimization if the were copies of the match state in other
registers. It was careless in that it did not consider the
failure branch.

Reorganize the code and fix both these issues. Add a test case
to test that the failure branch is considered.

1 files changed, 36 insertions, 12 deletions

diff --git a/lib/compiler/src/beam_bsm.erl b/lib/compiler/src/beam_bsm.erl
index c9c5b59bf2..02794a8e18 100644
--- a/lib/compiler/src/beam_bsm.erl
+++ b/lib/compiler/src/beam_bsm.erl
@@ -241,21 +241,45 @@ btb_reaches_match_2([{bif,_,{f,F},Ss,Dst}=I|Is], Regs0, D0) ->
     Regs = btb_kill([Dst], Regs0),
     D = btb_follow_branch(F, Regs, D0),
     btb_reaches_match_1(Is, Regs, D);
-btb_reaches_match_2([{test,bs_start_match2,_,_,[Ctx,_],Ctx}|Is], Regs, D) ->
-    case btb_context_regs(Regs) of
-	[Ctx] ->
-	    D;
-	CtxRegs ->
-	    case member(Ctx, CtxRegs) of
-		false -> btb_reaches_match_2(Is, Regs, D);
-		true -> btb_error(unsuitable_bs_start_match)
+btb_reaches_match_2([{test,bs_start_match2,{f,F},Live,[Ctx,_],Ctx}=I|Is],
+		    Regs0, D0) ->
+    CtxRegs = btb_context_regs(Regs0),
+    case member(Ctx, CtxRegs) of
+	false ->
+	    %% This bs_start_match2 instruction does not use "our"
+	    %% match state. Therefore we can continue the search
+	    %% for another bs_start_match2 instruction.
+	    D = btb_follow_branch(F, Regs0, D0),
+	    Regs = btb_kill_not_live(Live, Regs0),
+	    btb_reaches_match_2(Is, Regs, D);
+	true ->
+	    %% OK. This instruction will use "our" match state,
+	    %% but we must make sure that all other copies of the
+	    %% match state are killed in the code that follows
+	    %% the instruction. (We know that the fail branch cannot
+	    %% be taken in this case.)
+	    OtherCtxRegs = CtxRegs -- [Ctx],
+	    case btb_are_all_unused(OtherCtxRegs, Is, D0) of
+		false -> btb_error({OtherCtxRegs,not_all_unused_after,I});
+		true -> D0
 	    end
     end;
-btb_reaches_match_2([{test,bs_start_match2,_,_,[Bin,_],Ctx}|Is], Regs, D) ->
-    CtxRegs = btb_context_regs(Regs),
+btb_reaches_match_2([{test,bs_start_match2,{f,F},Live,[Bin,_],Ctx}|Is],
+		    Regs0, D0) ->
+    CtxRegs = btb_context_regs(Regs0),
     case member(Bin, CtxRegs) orelse member(Ctx, CtxRegs) of
-	false -> btb_reaches_match_2(Is, Regs, D);
-	true -> btb_error(unsuitable_bs_start_match)
+	false ->
+	    %% This bs_start_match2 does not reference any copy of the
+	    %% match state. Therefore it can safely be passed on the
+	    %% way to another (perhaps more suitable) bs_start_match2
+	    %% instruction.
+	    D = btb_follow_branch(F, Regs0, D0),
+	    Regs = btb_kill_not_live(Live, Regs0),
+	    btb_reaches_match_2(Is, Regs, D);
+	true ->
+	    %% This variant of the bs_start_match2 instruction does
+	    %% not accept a match state as source.
+	    btb_error(unsuitable_bs_start_match)
     end;
 btb_reaches_match_2([{test,_,{f,F},Ss}=I|Is], Regs, D0) ->
     btb_ensure_not_used(Ss, I, Regs),