On Fri, 12 Aug 2016, Richard Biener wrote:
>
> The following fixes a missed optimization in PRE which was over-cautionous
> with rejecting simplifications to SSA names during phi-translation.
>
> Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
Requires some testsuite adjustments as below as PRE now catches quite
some predictive commoning cases.
Tested on x86_64_unknown-linux-gnu, applied.
Richard.
2016-08-12 Richard Biener <rguent...@suse.de>
PR tree-optimization/57326
* tree-ssa-pre.c (fully_constant_expression): Handle simplification
returning an SSA name.
(phi_translate_1): When fully_constant_expression returns a NAME
make sure we have a leader for it.
* gcc.dg/tree-ssa/ssa-pre-32.c: New testcase.
* gcc.dg/tree-ssa/loadpre14.c: Adjust.
* gcc.dg/tree-ssa/pr35287.c: Likewise.
* gcc.target/i386/pr45685.c: Likewise.
* gcc.dg/tree-ssa/predcom-1.c: Disable PRE.
* gcc.dg/tree-ssa/predcom-2.c: Likewise.
* gcc.dg/tree-ssa/predcom-3.c: Likewise.
* gcc.dg/tree-ssa/ssa-sink-10.c: Likewise.
* gfortran.dg/pr34163.f90: Likewise.
Index: gcc/tree-ssa-pre.c
===================================================================
*** gcc/tree-ssa-pre.c (revision 239361)
--- gcc/tree-ssa-pre.c (working copy)
*************** get_or_alloc_expr_for (tree t)
*** 1201,1207 ****
}
/* Return the folded version of T if T, when folded, is a gimple
! min_invariant. Otherwise, return T. */
static pre_expr
fully_constant_expression (pre_expr e)
--- 1201,1207 ----
}
/* Return the folded version of T if T, when folded, is a gimple
! min_invariant or an SSA name. Otherwise, return T. */
static pre_expr
fully_constant_expression (pre_expr e)
*************** fully_constant_expression (pre_expr e)
*** 1218,1227 ****
return e;
if (is_gimple_min_invariant (res))
return get_or_alloc_expr_for_constant (res);
! /* We might have simplified the expression to a
! SSA_NAME for example from x_1 * 1. But we cannot
! insert a PHI for x_1 unconditionally as x_1 might
! not be available readily. */
return e;
}
case REFERENCE:
--- 1218,1225 ----
return e;
if (is_gimple_min_invariant (res))
return get_or_alloc_expr_for_constant (res);
! if (TREE_CODE (res) == SSA_NAME)
! return get_or_alloc_expr_for_name (res);
return e;
}
case REFERENCE:
*************** phi_translate_1 (pre_expr expr, bitmap_s
*** 1464,1470 ****
constant = fully_constant_expression (expr);
PRE_EXPR_NARY (expr) = nary;
if (constant != expr)
! return constant;
tree result = vn_nary_op_lookup_pieces (newnary->length,
newnary->opcode,
--- 1462,1481 ----
constant = fully_constant_expression (expr);
PRE_EXPR_NARY (expr) = nary;
if (constant != expr)
! {
! /* For non-CONSTANTs we have to make sure we can eventually
! insert the expression. Which means we need to have a
! leader for it. */
! if (constant->kind != CONSTANT)
! {
! unsigned value_id = get_expr_value_id (constant);
! constant = find_leader_in_sets (value_id, set1, set2);
! if (constant)
! return constant;
! }
! else
! return constant;
! }
tree result = vn_nary_op_lookup_pieces (newnary->length,
newnary->opcode,
Index: gcc/testsuite/gcc.dg/tree-ssa/ssa-pre-32.c
===================================================================
*** gcc/testsuite/gcc.dg/tree-ssa/ssa-pre-32.c (revision 0)
--- gcc/testsuite/gcc.dg/tree-ssa/ssa-pre-32.c (working copy)
***************
*** 0 ****
--- 1,11 ----
+ /* { dg-do compile } */
+ /* { dg-options "-O2 -fdump-tree-pre" } */
+
+ unsigned f(unsigned x, unsigned y, _Bool b)
+ {
+ #define m (b?-1:0)
+ return (x&m)|(y&~m);
+ #undef m
+ }
+
+ /* { dg-final { scan-tree-dump "# prephitmp_\[0-9\]+ = PHI
<\[xy\]_\[0-9\]+\\(D\\)\[^,\]*, \[xy\]_\[0-9\]+\\(D\\)" "pre" } } */
Index: gcc/testsuite/gcc.dg/tree-ssa/loadpre14.c
===================================================================
--- gcc/testsuite/gcc.dg/tree-ssa/loadpre14.c (revision 239406)
+++ gcc/testsuite/gcc.dg/tree-ssa/loadpre14.c (working copy)
@@ -6,7 +6,8 @@ int foo(type *a, int argc)
type c = {0, 1};
int d, e;
- /* Should be able to eliminate the second load of *a along the main path. */
+ /* Should be able to eliminate the second load of *a and the add of zero
+ along the main path. */
d = (*a)[0];
if (argc)
{
@@ -15,4 +16,4 @@ int foo(type *a, int argc)
e = (*a)[0];
return d + e;
}
-/* { dg-final { scan-tree-dump-times "Eliminated: 1" 1 "pre"} } */
+/* { dg-final { scan-tree-dump-times "Eliminated: 2" 1 "pre"} } */
Index: gcc/testsuite/gcc.dg/tree-ssa/pr35287.c
===================================================================
--- gcc/testsuite/gcc.dg/tree-ssa/pr35287.c (revision 239406)
+++ gcc/testsuite/gcc.dg/tree-ssa/pr35287.c (working copy)
@@ -10,5 +10,6 @@ int foo(int p)
return (*gp + t);
}
-/* We will eliminate one load of gp and one indirect load of *gp. */
-/* { dg-final { scan-tree-dump-times "Eliminated: 2" 1 "pre"} } */
+/* We will eliminate one load of gp and one indirect load of *gp and
+ the add of zero. */
+/* { dg-final { scan-tree-dump-times "Eliminated: 3" 1 "pre"} } */
Index: gcc/testsuite/gcc.dg/tree-ssa/predcom-1.c
===================================================================
--- gcc/testsuite/gcc.dg/tree-ssa/predcom-1.c (revision 239406)
+++ gcc/testsuite/gcc.dg/tree-ssa/predcom-1.c (working copy)
@@ -1,6 +1,6 @@
/* { dg-do compile } */
/* { dg-do run } */
-/* { dg-options "-O2 -funroll-loops --param max-unroll-times=8
-fpredictive-commoning -fdump-tree-pcom-details" } */
+/* { dg-options "-O2 -funroll-loops --param max-unroll-times=8
-fpredictive-commoning -fdump-tree-pcom-details -fno-tree-pre" } */
void abort (void);
@@ -45,6 +45,6 @@ int main(void)
/* Verify that both loops were transformed and unrolled. */
/* { dg-final { scan-tree-dump-times "Unrolling 2 times." 2 "pcom"} } */
-/* Also check that we undid the transformation previously made by PRE. */
-/* { dg-final { scan-tree-dump-times "looparound ref" 1 "pcom"} } */
-
+/* Also check that we undid the transformation previously made by PRE.
+ ??? PRE now does the predictive commoning in count_averages. */
+/* dg-final { scan-tree-dump-times "looparound ref" 1 "pcom" } */
Index: gcc/testsuite/gcc.dg/tree-ssa/predcom-2.c
===================================================================
--- gcc/testsuite/gcc.dg/tree-ssa/predcom-2.c (revision 239406)
+++ gcc/testsuite/gcc.dg/tree-ssa/predcom-2.c (working copy)
@@ -1,5 +1,5 @@
/* { dg-do run } */
-/* { dg-options "-O2 -funroll-loops --param max-unroll-times=8
-fpredictive-commoning -fdump-tree-pcom-details" } */
+/* { dg-options "-O2 -funroll-loops --param max-unroll-times=8
-fpredictive-commoning -fdump-tree-pcom-details -fno-tree-pre" } */
void abort (void);
Index: gcc/testsuite/gcc.dg/tree-ssa/predcom-3.c
===================================================================
--- gcc/testsuite/gcc.dg/tree-ssa/predcom-3.c (revision 239406)
+++ gcc/testsuite/gcc.dg/tree-ssa/predcom-3.c (working copy)
@@ -1,5 +1,5 @@
/* { dg-do compile } */
-/* { dg-options "-O2 -funroll-loops --param max-unroll-times=8
-fpredictive-commoning -fdump-tree-pcom-details" } */
+/* { dg-options "-O2 -funroll-loops --param max-unroll-times=8
-fpredictive-commoning -fdump-tree-pcom-details -fno-tree-pre" } */
int a[1000], b[1000];
Index: gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-10.c
===================================================================
--- gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-10.c (revision 239406)
+++ gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-10.c (working copy)
@@ -1,5 +1,5 @@
/* { dg-do compile } */
-/* { dg-options "-O2 -fdump-tree-sink-details" } */
+/* { dg-options "-O2 -fdump-tree-sink-details -fno-tree-pre" } */
int x[1024], y[1024], z[1024], w[1024];
void foo (void)
Index: gcc/testsuite/gcc.target/i386/pr45685.c
===================================================================
--- gcc/testsuite/gcc.target/i386/pr45685.c (revision 239406)
+++ gcc/testsuite/gcc.target/i386/pr45685.c (working copy)
@@ -36,4 +36,4 @@ int summation_helper_2(int64_t* products
return s;
}
-/* { dg-final { scan-assembler-times "cmov" 4 } } */
+/* { dg-final { scan-assembler-times "cmov" 6 } } */
Index: gcc/testsuite/gfortran.dg/pr34163.f90
===================================================================
--- gcc/testsuite/gfortran.dg/pr34163.f90 (revision 239406)
+++ gcc/testsuite/gfortran.dg/pr34163.f90 (working copy)
@@ -1,5 +1,5 @@
! { dg-do compile }
-! { dg-options "-O2 -fpredictive-commoning -fdump-tree-pcom-details" }
+! { dg-options "-O2 -fno-tree-pre -fpredictive-commoning
-fdump-tree-pcom-details" }
subroutine trisolve2(x,i1,i2,nxyz)
integer :: nxyz
real,dimension(nxyz):: au1
