The following adjusts the cost model of invariant motion to consider
[VEC_]COND_EXPRs and comparisons producing a data value as expensive.
For 503.bwaves_r this avoids an unnecessarily high vectorization
factor because of an integer comparison besides data operations on
double.
Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.
PR tree-optimization/110451
* tree-ssa-loop-im.cc (stmt_cost): [VEC_]COND_EXPR and
tcc_comparison are expensive.
* gfortran.dg/vect/pr110451.f: New testcase.
---
gcc/testsuite/gfortran.dg/vect/pr110451.f | 51 +++++++++++++++++++++++
gcc/tree-ssa-loop-im.cc | 11 ++++-
2 files changed, 61 insertions(+), 1 deletion(-)
create mode 100644 gcc/testsuite/gfortran.dg/vect/pr110451.f
diff --git a/gcc/testsuite/gfortran.dg/vect/pr110451.f
b/gcc/testsuite/gfortran.dg/vect/pr110451.f
new file mode 100644
index 00000000000..ba77b0dd174
--- /dev/null
+++ b/gcc/testsuite/gfortran.dg/vect/pr110451.f
@@ -0,0 +1,51 @@
+! { dg-do compile }
+! { dg-require-effective-target vect_condition }
+! { dg-require-effective-target vect_double }
+! { dg-additional-options "-ffast-math -floop-interchange
-fdump-tree-linterchange-details -fdump-tree-vect-details" }
+! { dg-additional-options "-mprefer-vector-width=128" { target x86_64-*-*
i?86-*-* } }
+
+ subroutine mat_times_vec(y,x,a,axp,ayp,azp,axm,aym,azm,
+ $ nb,nx,ny,nz)
+ implicit none
+ integer nb,nx,ny,nz,i,j,k,m,l,kit,im1,ip1,jm1,jp1,km1,kp1
+
+ real*8 y(nb,nx,ny,nz),x(nb,nx,ny,nz)
+
+ real*8 a(nb,nb,nx,ny,nz),
+ 1 axp(nb,nb,nx,ny,nz),ayp(nb,nb,nx,ny,nz),azp(nb,nb,nx,ny,nz),
+ 2 axm(nb,nb,nx,ny,nz),aym(nb,nb,nx,ny,nz),azm(nb,nb,nx,ny,nz)
+
+
+ do k=1,nz
+ km1=mod(k+nz-2,nz)+1
+ kp1=mod(k,nz)+1
+ do j=1,ny
+ jm1=mod(j+ny-2,ny)+1
+ jp1=mod(j,ny)+1
+ do i=1,nx
+ im1=mod(i+nx-2,nx)+1
+ ip1=mod(i,nx)+1
+ do l=1,nb
+ y(l,i,j,k)=0.0d0
+ do m=1,nb
+ y(l,i,j,k)=y(l,i,j,k)+
+ 1 a(l,m,i,j,k)*x(m,i,j,k)+
+ 2 axp(l,m,i,j,k)*x(m,ip1,j,k)+
+ 3 ayp(l,m,i,j,k)*x(m,i,jp1,k)+
+ 4 azp(l,m,i,j,k)*x(m,i,j,kp1)+
+ 5 axm(l,m,i,j,k)*x(m,im1,j,k)+
+ 6 aym(l,m,i,j,k)*x(m,i,jm1,k)+
+ 7 azm(l,m,i,j,k)*x(m,i,j,km1)
+ enddo
+ enddo
+ enddo
+ enddo
+ enddo
+ return
+ end
+
+! loop interchange adds a conditional on m != 1 in the innermost loop
+! verify that is hoisted and thus not affecting the vectorization factor
+
+! { dg-final { scan-tree-dump-times "is interchanged" 1 "linterchange" } }
+! { dg-final { scan-tree-dump "vectorization factor = 2" "vect" { target
x86_64-*-* i?86-*-* } } }
diff --git a/gcc/tree-ssa-loop-im.cc b/gcc/tree-ssa-loop-im.cc
index 86ce6acb023..f5b01e986ae 100644
--- a/gcc/tree-ssa-loop-im.cc
+++ b/gcc/tree-ssa-loop-im.cc
@@ -617,7 +617,8 @@ stmt_cost (gimple *stmt)
if (gimple_code (stmt) != GIMPLE_ASSIGN)
return 1;
- switch (gimple_assign_rhs_code (stmt))
+ enum tree_code code = gimple_assign_rhs_code (stmt);
+ switch (code)
{
case MULT_EXPR:
case WIDEN_MULT_EXPR:
@@ -645,6 +646,11 @@ stmt_cost (gimple *stmt)
/* Shifts and rotates are usually expensive. */
return LIM_EXPENSIVE;
+ case COND_EXPR:
+ case VEC_COND_EXPR:
+ /* Conditionals are expensive. */
+ return LIM_EXPENSIVE;
+
case CONSTRUCTOR:
/* Make vector construction cost proportional to the number
of elements. */
@@ -658,6 +664,9 @@ stmt_cost (gimple *stmt)
return 0;
default:
+ /* Comparisons are usually expensive. */
+ if (TREE_CODE_CLASS (code) == tcc_comparison)
+ return LIM_EXPENSIVE;
return 1;
}
}
--
2.35.3
