Hi,<br><br>Seems pass scalar-evolution+indvars fail to get the loop trip count of the following case:<br><br>int foo(int x, int y, int lam[256], int alp[256]) {<br> int i;<br> int z = y;<br><br> for (i = 255; i >= 0; i--) {<br>
z += x;<br> lam[i] = alp[i];<br> }<br> return z;<br>}<br><br><br>The final optimized ll code is :<br><br>define i32 @foo(i32 %x, i32 %y, i32* %lam, i32* %alp) nounwind {<br>entry:<br> br label %bb<br><br>bb: ; preds = %bb, %entry<br>
%indvar = phi i32 [ 0, %entry ], [ %indvar.next, %bb ] ; <i32> [#uses=4]<br> %i.0.reg2mem.0 = sub i32 255, %indvar ; <i32> [#uses=2]<br> %tmp7 = getelementptr i32* %alp, i32 %i.0.reg2mem.0 ; <i32*> [#uses=1]<br>
%tmp8 = load i32* %tmp7, align 4 ; <i32> [#uses=1]<br> %tmp11 = getelementptr i32* %lam, i32 %i.0.reg2mem.0 ; <i32*> [#uses=1]<br> store i32 %tmp8, i32* %tmp11, align 4<br>
%tmp13 = sub i32 254, %indvar ; <i32> [#uses=1]<br> %tmp16 = icmp slt i32 %tmp13, 0 ; <i1> [#uses=1]<br> %indvar.next = add i32 %indvar, 1 ; <i32> [#uses=1]<br>
br i1 %tmp16, label %bb17, label %bb<br><br>bb17: ; preds = %bb<br> %indvar.lcssa = phi i32 [ %indvar, %bb ] ; <i32> [#uses=1]<br> %tmp28 = mul i32 %indvar.lcssa, %x ; <i32> [#uses=1]<br>
%z.0.reg2mem.0 = add i32 %y, %x ; <i32> [#uses=1]<br> %tmp4 = add i32 %z.0.reg2mem.0, %tmp28 ; <i32> [#uses=1]<br> ret i32 %tmp4<br>}<br><br>Surely the loop trip count is 256, but the Loop::getTripCount() will return NULL.<br>
This is due to the exit condition: <br><br>%tmp16 = icmp slt i32 %tmp13, 0 ; <i1> [#uses=1]<br><br>pass indvars should transform it into a canonical exit condtion EQ or NE, but as scalarevolution returned NonComputedItCount, it failed to do that.<br>
<br>In pass ScalarEvolution three algorithms failed to compute the trip count:<br><br>1. In SCEVAddRecExpr::getNumIterationsInRange(), as the loop stride is negative, the exit value expression will be zero: (Here A is the stride value)<br>
<br> // The exit value should be (End+A)/A.<br> APInt ExitVal = (End + A).udiv(A);<br> ConstantInt *ExitValue = ConstantInt::get(ExitVal); <br><br>2. In Line 1953, the switch-case computes the trip count due to different Cond code.<br>
But in this case, as it exits on true, the slt is converted to sge, not switch-case for this condtion.<br><br><br>3. As the loop count is over 100, the Brute-Force algorithm also doent work<br><br>Can we improve this?<br>
<br><br>Thanks,<br>Sheng.<br><br> <br>