On Tue, 22 Jan 2019, Jakub Jelinek wrote:
> Hi!
>
> SCEV can analyze not just integral/pointer IVs, but (scalar) float ones as
> well. Calling build_int_cst on such types results in ICE, build_zero_cst
> works. Though the loop invariant PHI IVs, if we represent them as using
> +0.0 step, aren't correct if honoring signed zeros, as + 0.0 will make
> a +0.0 out of -0.0.
>
> Bootstrapped/regtested on {x86_64,i686,powerpc64{,le}}-linux, ok for trunk?
OK.
Richard.
> 2019-01-22 Jakub Jelinek <ja...@redhat.com>
>
> PR tree-optimization/88964
> * gimple-loop-interchange.cc (loop_cand::analyze_induction_var): Use
> build_zero_cst instead of build_int_cst. Return false for loop
> invariants which honor signed zeros.
>
> * gfortran.dg/pr88964.f90: New test.
>
> --- gcc/gimple-loop-interchange.cc.jj 2019-01-01 12:37:17.416970701 +0100
> +++ gcc/gimple-loop-interchange.cc 2019-01-22 11:34:42.303796570 +0100
> @@ -688,11 +688,16 @@ loop_cand::analyze_induction_var (tree v
> /* Var is loop invariant, though it's unlikely to happen. */
> if (tree_does_not_contain_chrecs (chrec))
> {
> + /* Punt on floating point invariants if honoring signed zeros,
> + representing that as + 0.0 would change the result if init
> + is -0.0. */
> + if (HONOR_SIGNED_ZEROS (chrec))
> + return false;
> struct induction *iv = XCNEW (struct induction);
> iv->var = var;
> iv->init_val = init;
> iv->init_expr = chrec;
> - iv->step = build_int_cst (TREE_TYPE (chrec), 0);
> + iv->step = build_zero_cst (TREE_TYPE (chrec));
> m_inductions.safe_push (iv);
> return true;
> }
> --- gcc/testsuite/gfortran.dg/pr88964.f90.jj 2019-01-22 11:54:46.046154065
> +0100
> +++ gcc/testsuite/gfortran.dg/pr88964.f90 2019-01-22 11:54:23.172526590
> +0100
> @@ -0,0 +1,57 @@
> +! PR tree-optimization/88964
> +! { dg-do compile }
> +! { dg-options "-O3 -fno-tree-forwprop --param
> sccvn-max-alias-queries-per-access=1" }
> +
> +MODULE pr88964
> + INTEGER, PARAMETER :: dp=8
> + REAL(KIND=dp) :: p, q, o
> +CONTAINS
> + SUBROUTINE foo(a,b,c,f,h)
> + IMPLICIT NONE
> + INTEGER :: a, b, c
> + REAL(KIND=dp) :: f(b*c), h(a*c)
> + CALL bar(h)
> + CALL baz(f)
> + CALL qux(h)
> + END SUBROUTINE foo
> + SUBROUTINE bar(h)
> + IMPLICIT NONE
> + REAL(KIND=dp) :: h(1*1)
> + INTEGER :: r, s, t, u
> + DO u = 1,3
> + DO t = 1,1
> + DO s = 1,3
> + DO r = 1,1
> + h((t-1)*1+r) = h((t-1)*1+r)-p*o
> + END DO
> + END DO
> + END DO
> + END DO
> + END SUBROUTINE bar
> + SUBROUTINE baz(f)
> + IMPLICIT NONE
> + REAL(KIND=dp) :: f(3*1)
> + INTEGER :: s, t, u
> + DO u = 1,4
> + DO t = 1,1
> + DO s = 1,3
> + f((t-1)*3+s) = f((t-1)*3+s) - q
> + END DO
> + END DO
> + END DO
> + END SUBROUTINE baz
> + SUBROUTINE qux(h)
> + IMPLICIT NONE
> + REAL(KIND=dp) :: h(1*1)
> + INTEGER :: r, s, t, u
> + DO u = 1,5
> + DO t = 1,1
> + DO s = 1,3
> + DO r = 1,1
> + h((t-1)*1+r) = h((t-1)*1+r)-p*o
> + END DO
> + END DO
> + END DO
> + END DO
> + END SUBROUTINE qux
> +END MODULE pr88964
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB
21284 (AG Nuernberg)
