[lno] [RFC] if-conversion and auto vectorizer
Dorit Naishlos
DORIT@il.ibm.com
Sun Mar 14 10:59:00 GMT 2004
> I am not at all sure what you want to achieve with this transformation;
> it seems to me that you haven't done anything that could possibly make
> vectorizer's work easier.
I think Devang's example showed an intermediate stage in the development of
the transformation. The final goal of the transformation is to transform a
multi-block loop like the following:
LOOP:
if (A[i] > 0)
{
A[i] = A[i] + C[i];
B[i] = A[i] + 1;
}
else
{
A[i] = D[i];
}
into straight-line single block loop:
LOOP:
[1] C = CMP A[i] > 0
[2] A_then = A[i] + c[i]
[3] A[i] = "SELECT" C ? A_then : A[i]
[4] B_then = A[i] + 1
[5] B[i] = "SELECT" C ? B_then : B[i]
[6] A_else = D[i]
[7] A[i] = "SELECT" ^C ? A_else : A[i]
[3]&[7] would be merged (by an additional pass, or directly) to obtain the
following optimized code:
LOOP:
[1] C = CMP A[i] > 0
[2] A_then = A[i] + c[i]
[3] B_then = A[i] + 1
[4] B[i] = "SELECT" C ? B_then : B[i]
[5] A_else = D[i]
[6] A[i] = "SELECT" C ? A_then : A_else
(This is also where special idioms like saturation, would be recognized.)
dorit
Zdenek Dvorak
<rakdver@atrey.karlin.m To: Devang Patel <dpatel@apple.com>
ff.cuni.cz> cc: "gcc@gcc.gnu.org list" <gcc@gcc.gnu.org>, Jason Merrill
<jason@redhat.com>, Dorit Naishlos/Haifa/IBM@IBMIL, Richard Henderson
13/03/2004 11:24 <rth@redhat.com>, Diego Novillo <dnovillo@redhat.com>
Subject: Re: [lno] [RFC] if-conversion and auto vectorizer
Hello,
> OK. So far I have coded transformation to transform
>
> bar ()
> {
> int A[N+1], B[N+1], C[N+1], D[N+1];
> int i;
>
> for (i = 1; i<N; i++)
> {
> if (A[i] > 0)
> {
> A[i] = A[i] + C[i];
> B[i] = A[i] + 1;
> }
> else
> {
> A[i] = D[i];
> B[i] = D[i] + 1;
> }
> }
> ibar(A);
> ibar(B);
> }
>
> into :
>
> bar ()
> {
> int i;
> int D[17];
> int C[17];
> int B[17];
> int A[17];
> int T.5;
> int T.4;
> int T.3;
> int T.2;
> int T.1;
> int T.0;
>
> # BLOCK 0
> # PRED: ENTRY [100.0%] (fallthru,exec)
> # SUCC: 1 [100.0%] (fallthru,exec)
>
> # BLOCK 1
> # PRED: 6 [100.0%] (fallthru) 0 [100.0%] (fallthru,exec)
> # i_1 = PHI <1(0), i_15(6)>;
> <L0>:;
> T.0_3 = if (1)
> {
> A[i_1];
> };
> if (T.0_3 > 0) goto <L1>; else goto <L2>;
> # SUCC: 3 [21.0%] (false,exec) 2 [79.0%] (true,exec)
>
> # BLOCK 2
> # PRED: 1 [79.0%] (true,exec)
> <L1>:;
> T.1_9 = if (1 && T.0_3 > 0)
why this "1 &&" part?
> {
> C[i_1];
> };
> T.2_10 = if (1 && T.0_3 > 0)
> {
> T.0_3 + T.1_9;
> };
> A[i_1] = if (1 && T.0_3 > 0)
> {
> T.2_10
> };
> T.3_12 = if (1 && T.0_3 > 0)
> {
> T.2_10 + 1;
> };
> B[i_1] = if (1 && T.0_3 > 0)
> {
> T.3_12
> };
> goto <bb 4> (<L3>);
> # SUCC: 4 [100.0%] (fallthru,exec)
>
> # BLOCK 3
> # PRED: 1 [21.0%] (false,exec)
> <L2>:;
> T.4_5 = if (1 && !(T.0_3 > 0))
> {
> D[i_1];
> };
> A[i_1] = if (1 && !(T.0_3 > 0))
> {
> T.4_5
> };
> T.5_7 = if (1 && !(T.0_3 > 0))
> {
> T.4_5 + 1;
> };
> B[i_1] = if (1 && !(T.0_3 > 0))
> {
> T.5_7
> };
> # SUCC: 4 [100.0%] (fallthru,exec)
>
> # BLOCK 4
> # PRED: 3 [100.0%] (fallthru,exec) 2 [100.0%] (fallthru,exec)
> <L3>:;
> i_15 = i_1 + 1;
> if (i_15 <= 15) goto <L8>; else goto <L5>;
> # SUCC: 5 [11.0%] (false,exec) 6 [89.0%] (true,exec)
>
> # BLOCK 6
> # PRED: 4 [89.0%] (true,exec)
> <L8>:;
> goto <bb 1> (<L0>);
> # SUCC: 1 [100.0%] (fallthru)
>
> # BLOCK 5
> # PRED: 4 [11.0%] (false,exec)
> <L5>:;
> ibar (&A);
> ibar (&B);
> return;
> # SUCC: EXIT [100.0%]
>
> }
>
> I spent lot of time unnecessary to create super block inside loop by
> merging block 2 and 3 into block 1.
>
> If vectorizer can not vectorize this loop then it will be reverted to
> original form using loop versioning. Am I on the right track ?
I am not at all sure what you want to achieve with this transformation;
it seems to me that you haven't done anything that could possibly make
vectorizer's work easier.
Zdenek
More information about the Gcc
mailing list