interp: do not unroll loops

This change triggers a revert whenever a basic block runs instructions at runtime twice.
As a result, a loop body with runtime-only instructions will no longer be unrolled.
This should help some extreme cases where loops can be expanded into hundreds or thousands of instructions.

interp/testdata/revert.ll

@@ -3,6 +3,8 @@ target triple = "x86_64--linux"
 
 declare void @externalCall(i64)
 
+declare i64 @ptrHash(i8* nocapture)
+
 @foo.knownAtRuntime = global i64 0
 @bar.knownAtRuntime = global i64 0
 @baz.someGlobal = external global [3 x {i64, i32}]
@@ -10,6 +12,8 @@ declare void @externalCall(i64)
 @x.atomicNum = global i32 0
 @x.volatileNum = global i32 0
 @y.ready = global i32 0
+@z.bloom = global i64 0
+@z.arr = global [32 x i8] zeroinitializer
 
 define void @runtime.initAll() unnamed_addr {
 entry:
@@ -19,6 +23,7 @@ entry:
   call void @main.init(i8* undef)
   call void @x.init(i8* undef)
   call void @y.init(i8* undef)
+  call void @z.init(i8* undef)
   ret void
 }
 
@@ -72,3 +77,43 @@ loop:
 end:
   ret void
 }
+
+define internal void @z.init(i8* %context) unnamed_addr {
+  %bloom = bitcast i64* @z.bloom to i8*
+
+  ; This can be safely expanded.
+  call void @z.setArr(i8* %bloom, i64 1, i8* %bloom)
+
+  ; This call should be reverted to prevent unrolling.
+  call void @z.setArr(i8* bitcast ([32 x i8]* @z.arr to i8*), i64 32, i8* %bloom)
+
+  ret void
+}
+
+define internal void @z.setArr(i8* %arr, i64 %n, i8* %context) unnamed_addr {
+entry:
+  br label %loop
+
+loop:
+  %prev = phi i64 [ %n, %entry ], [ %idx, %loop ]
+  %idx = sub i64 %prev, 1
+  %elem = getelementptr i8, i8* %arr, i64 %idx
+  call void @z.set(i8* %elem, i8* %context)
+  %done = icmp eq i64 %idx, 0
+  br i1 %done, label %end, label %loop
+
+end:
+  ret void
+}
+
+define internal void @z.set(i8* %ptr, i8* %context) unnamed_addr {
+  ; Insert the pointer into the Bloom filter.
+  %hash = call i64 @ptrHash(i8* %ptr)
+  %index = lshr i64 %hash, 58
+  %bit = shl i64 1, %index
+  %bloom = bitcast i8* %context to i64*
+  %old = load i64, i64* %bloom
+  %new = or i64 %old, %bit
+  store i64 %new, i64* %bloom
+  ret void
+}

interp/testdata/revert.out.ll

@@ -8,9 +8,13 @@ target triple = "x86_64--linux"
 @x.atomicNum = local_unnamed_addr global i32 0
 @x.volatileNum = global i32 0
 @y.ready = local_unnamed_addr global i32 0
+@z.bloom = global i64 0
+@z.arr = global [32 x i8] zeroinitializer
 
 declare void @externalCall(i64) local_unnamed_addr
 
+declare i64 @ptrHash(i8* nocapture) local_unnamed_addr
+
 define void @runtime.initAll() unnamed_addr {
 entry:
   call fastcc void @baz.init(i8* undef)
@@ -24,6 +28,8 @@ entry:
   %y = load volatile i32, i32* @x.volatileNum, align 4
   store volatile i32 %y, i32* @x.volatileNum, align 4
   call fastcc void @y.init(i8* undef)
+  call fastcc void @z.set(i8* bitcast (i64* @z.bloom to i8*), i8* bitcast (i64* @z.bloom to i8*))
+  call fastcc void @z.setArr(i8* getelementptr inbounds ([32 x i8], [32 x i8]* @z.arr, i32 0, i32 0), i64 32, i8* bitcast (i64* @z.bloom to i8*))
   ret void
 }
 
@@ -48,3 +54,30 @@ loop:                                             ; preds = %loop, %entry
 end:                                              ; preds = %loop
   ret void
 }
+
+define internal fastcc void @z.setArr(i8* %arr, i64 %n, i8* %context) unnamed_addr {
+entry:
+  br label %loop
+            
+loop:                                             ; preds = %loop, %entry
+  %prev = phi i64 [ %n, %entry ], [ %idx, %loop ]
+  %idx = sub i64 %prev, 1
+  %elem = getelementptr i8, i8* %arr, i64 %idx
+  call fastcc void @z.set(i8* %elem, i8* %context)
+  %done = icmp eq i64 %idx, 0
+  br i1 %done, label %end, label %loop
+
+end:                                              ; preds = %loop
+  ret void
+}
+
+define internal fastcc void @z.set(i8* %ptr, i8* %context) unnamed_addr {
+  %hash = call i64 @ptrHash(i8* %ptr)
+  %index = lshr i64 %hash, 58
+  %bit = shl i64 1, %index
+  %bloom = bitcast i8* %context to i64*
+  %old = load i64, i64* %bloom, align 8
+  %new = or i64 %old, %bit
+  store i64 %new, i64* %bloom, align 8
+  ret void
+}