Re: [PATCH] tree-optimization/107852 - missed optimization with PHIs

[Richard Biener via Gcc-patches]

On Mon, 5 Dec 2022, Jan-Benedict Glaw wrote:

> On Tue, 2022-11-29 14:30:22 +0100, Richard Biener via Gcc-patches 
>  wrote:
> > Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.
> > 
> > PR tree-optimization/107852
> > * tree-ssa-sccvn.cc (visit_phi): Use equivalences recorded
> > as predicated values to elide more redundant PHIs.
> > 
> > * gcc.dg/tree-ssa/ssa-fre-101.c: New testcase.
> 
> This seems to trigger an issue when building the Linux powerpc kernel
> for the skiroot_defconfig:
> 
> [mk all 2022-12-05 19:50:10]   powerpc64-linux-gcc 
> -Wp,-MMD,drivers/dma-buf/.dma-fence-array.o.d -nostdinc 
> -I./arch/powerpc/include -I./arch/powerpc/include/generated  -I./include 
> -I./arch/powerpc/include/uapi -I./arch/powerpc/include/generated/uapi 
> -I./include/uapi -I./include/generated/uapi -include 
> ./include/linux/compiler-version.h -include ./include/linux/kconfig.h 
> -include ./include/linux/compiler_types.h -D__KERNEL__ -I ./arch/powerpc 
> -DHAVE_AS_ATHIGH=1 -fmacro-prefix-map=./= -Wall -Wundef 
> -Werror=strict-prototypes -Wno-trigraphs -fno-strict-aliasing -fno-common 
> -fshort-wchar -fno-PIE -Werror=implicit-function-declaration 
> -Werror=implicit-int -Werror=return-type -Wno-format-security -std=gnu11 
> -mlittle-endian -m64 -msoft-float -pipe -mtraceback=no -mabi=elfv2 
> -mcmodel=medium -mno-pointers-to-nested-functions -mcpu=power8 -mtune=power10 
> -mno-prefixed -mno-pcrel -mno-altivec -mno-vsx -mno-mma 
> -fno-asynchronous-unwind-tables -mno-string -Wa,-maltivec -Wa,-mpower4 
> -Wa,-many -mno
 -strict-align -mlittle-endian -mstack-protector-guard=tls 
-mstack-protector-guard-reg=r13 -fno-delete-null-pointer-checks 
-Wno-frame-address -Wno-format-truncation -Wno-format-overflow 
-Wno-address-of-packed-member -Os -fno-allow-store-data-races 
-Wframe-larger-than=2048 -fstack-protector-strong -Wno-main 
-Wno-unused-but-set-variable -Wno-unused-const-variable -Wno-dangling-pointer 
-fomit-frame-pointer -ftrivial-auto-var-init=zero -fno-stack-clash-protection 
-Wdeclaration-after-statement -Wvla -Wno-pointer-sign -Wcast-function-type 
-Wno-stringop-truncation -Wno-stringop-overflow -Wno-restrict 
-Wno-maybe-uninitialized -Wno-alloc-size-larger-than -Wimplicit-fallthrough=5 
-fno-strict-overflow -fno-stack-check -fconserve-stack -Werror=date-time 
-Werror=incompatible-pointer-types -Werror=designated-init 
-Wno-packed-not-aligned -mstack-protector-guard-offset=2800
-DKBUILD_MODFILE='"drivers/dma-buf/dma-fence-array"' 
-DKBUILD_BASENAME='"dma_fence_array"' -DKBUILD_MODNAME='"dma_fence_arra
 y"' -D__KBUILD_MODNAME=kmod_dma_fence_array -c -o 
drivers/dma-buf/dma-fence-array.o drivers/dma-buf/dma-fence-array.c  
> [mk all 2022-12-05 19:50:10] drivers/dma-buf/dma-fence-array.c: In function 
> 'dma_fence_array_create':
> [mk all 2022-12-05 19:50:10] drivers/dma-buf/dma-fence-array.c:154:25: error: 
> control flow in the middle of basic block 12
> [mk all 2022-12-05 19:50:10]   154 | struct dma_fence_array 
> *dma_fence_array_create(int num_fences,
> [mk all 2022-12-05 19:50:10]   | 
> ^~
> [mk all 2022-12-05 19:50:10] during GIMPLE pass: ivopts
> [mk all 2022-12-05 19:50:10] drivers/dma-buf/dma-fence-array.c:154:25: 
> internal compiler error: verify_flow_info failed
> [mk all 2022-12-05 19:50:10] 0x19ea876 internal_error(char const*, ...)
> [mk all 2022-12-05 19:50:10]???:0
> [mk all 2022-12-05 19:50:10] 0x94b00e verify_flow_info()
> [mk all 2022-12-05 19:50:10]???:0
> [mk all 2022-12-05 19:50:10] Please submit a full bug report, with 
> preprocessed source (by using -freport-bug).
> [mk all 2022-12-05 19:50:10] Please include the complete backtrace with any 
> bug report.
> [mk all 2022-12-05 19:50:10] See  for instructions.
> 
> Maybe you've got an idea, otherwise I'll try to reproduce it manually.
> (That's all automated building.)

I'll note the above ICE is quite a few passes later during IVOPTs so
the change triggered a latent issue.  Wild guessing makes me think
it's some asm goto being mis-handled.

Can you please open a bugreport and provide preprocessed source so
one can reproduce this with a cc1 cross?

Thanks,
Richard.

Re: [PATCH] tree-optimization/107852 - missed optimization with PHIs

[Jan-Benedict Glaw]

On Tue, 2022-11-29 14:30:22 +0100, Richard Biener via Gcc-patches 
 wrote:
> Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.
> 
>   PR tree-optimization/107852
>   * tree-ssa-sccvn.cc (visit_phi): Use equivalences recorded
>   as predicated values to elide more redundant PHIs.
> 
>   * gcc.dg/tree-ssa/ssa-fre-101.c: New testcase.

This seems to trigger an issue when building the Linux powerpc kernel
for the skiroot_defconfig:

[mk all 2022-12-05 19:50:10]   powerpc64-linux-gcc 
-Wp,-MMD,drivers/dma-buf/.dma-fence-array.o.d -nostdinc 
-I./arch/powerpc/include -I./arch/powerpc/include/generated  -I./include 
-I./arch/powerpc/include/uapi -I./arch/powerpc/include/generated/uapi 
-I./include/uapi -I./include/generated/uapi -include 
./include/linux/compiler-version.h -include ./include/linux/kconfig.h -include 
./include/linux/compiler_types.h -D__KERNEL__ -I ./arch/powerpc 
-DHAVE_AS_ATHIGH=1 -fmacro-prefix-map=./= -Wall -Wundef 
-Werror=strict-prototypes -Wno-trigraphs -fno-strict-aliasing -fno-common 
-fshort-wchar -fno-PIE -Werror=implicit-function-declaration 
-Werror=implicit-int -Werror=return-type -Wno-format-security -std=gnu11 
-mlittle-endian -m64 -msoft-float -pipe -mtraceback=no -mabi=elfv2 
-mcmodel=medium -mno-pointers-to-nested-functions -mcpu=power8 -mtune=power10 
-mno-prefixed -mno-pcrel -mno-altivec -mno-vsx -mno-mma 
-fno-asynchronous-unwind-tables -mno-string -Wa,-maltivec -Wa,-mpower4 
-Wa,-many -mno-strict-align -mlittle-endian -mstack-protector-guard=tls 
-mstack-protector-guard-reg=r13 -fno-delete-null-pointer-checks 
-Wno-frame-address -Wno-format-truncation -Wno-format-overflow 
-Wno-address-of-packed-member -Os -fno-allow-store-data-races 
-Wframe-larger-than=2048 -fstack-protector-strong -Wno-main 
-Wno-unused-but-set-variable -Wno-unused-const-variable -Wno-dangling-pointer 
-fomit-frame-pointer -ftrivial-auto-var-init=zero -fno-stack-clash-protection 
-Wdeclaration-after-statement -Wvla -Wno-pointer-sign -Wcast-function-type 
-Wno-stringop-truncation -Wno-stringop-overflow -Wno-restrict 
-Wno-maybe-uninitialized -Wno-alloc-size-larger-than -Wimplicit-fallthrough=5 
-fno-strict-overflow -fno-stack-check -fconserve-stack -Werror=date-time 
-Werror=incompatible-pointer-types -Werror=designated-init 
-Wno-packed-not-aligned -mstack-protector-guard-offset=2800
-DKBUILD_MODFILE='"drivers/dma-buf/dma-fence-array"' 
-DKBUILD_BASENAME='"dma_fence_array"' -DKBUILD_MODNAME='"dma_fence_array"' 
-D__KBUILD_MODNAME=kmod_dma_fence_array -c -o drivers/dma-buf/dma-fence-array.o 
drivers/dma-buf/dma-fence-array.c  
[mk all 2022-12-05 19:50:10] drivers/dma-buf/dma-fence-array.c: In function 
'dma_fence_array_create':
[mk all 2022-12-05 19:50:10] drivers/dma-buf/dma-fence-array.c:154:25: error: 
control flow in the middle of basic block 12
[mk all 2022-12-05 19:50:10]   154 | struct dma_fence_array 
*dma_fence_array_create(int num_fences,
[mk all 2022-12-05 19:50:10]   | 
^~
[mk all 2022-12-05 19:50:10] during GIMPLE pass: ivopts
[mk all 2022-12-05 19:50:10] drivers/dma-buf/dma-fence-array.c:154:25: internal 
compiler error: verify_flow_info failed
[mk all 2022-12-05 19:50:10] 0x19ea876 internal_error(char const*, ...)
[mk all 2022-12-05 19:50:10]???:0
[mk all 2022-12-05 19:50:10] 0x94b00e verify_flow_info()
[mk all 2022-12-05 19:50:10]???:0
[mk all 2022-12-05 19:50:10] Please submit a full bug report, with preprocessed 
source (by using -freport-bug).
[mk all 2022-12-05 19:50:10] Please include the complete backtrace with any bug 
report.
[mk all 2022-12-05 19:50:10] See  for instructions.

Maybe you've got an idea, otherwise I'll try to reproduce it manually.
(That's all automated building.)

Thanks,
  Jan-Benedict
-- 


signature.asc
Description: PGP signature

[PATCH] tree-optimization/107852 - missed optimization with PHIs

[Richard Biener via Gcc-patches]

The following deals with the situation where we have

 [local count: 1073741824]:
_5 = bytes.D.25336._M_impl.D.24643._M_start;
_6 = bytes.D.25336._M_impl.D.24643._M_finish;
pretmp_66 = bytes.D.25336._M_impl.D.24643._M_end_of_storage;
if (_5 != _6)
  goto ; [70.00%]
else
  goto ; [30.00%]

...

 [local count: 329045359]:
_89 = operator new (4);
_43 = bytes.D.25336._M_impl.D.24643._M_start;
_Num_44 = _137 - _43;
if (_Num_44 != 0)

but fail to see that _137 is equal to _5 and thus eventually _Num_44
is zero if not operator new would possibly clobber the global
bytes variable.

The following resolves this in value-numbering by using the
predicated values for _5 == _6 recorded for the dominating
condition.

Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.

PR tree-optimization/107852
* tree-ssa-sccvn.cc (visit_phi): Use equivalences recorded
as predicated values to elide more redundant PHIs.

* gcc.dg/tree-ssa/ssa-fre-101.c: New testcase.
---
 gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-101.c | 47 +++
 gcc/tree-ssa-sccvn.cc   | 51 -
 2 files changed, 97 insertions(+), 1 deletion(-)
 create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-101.c

diff --git a/gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-101.c 
b/gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-101.c
new file mode 100644
index 000..c67f211dcf6
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-101.c
@@ -0,0 +1,47 @@
+/* { dg-do compile } */
+/* { dg-options "-O -fdump-tree-fre1-details" } */
+
+int test1 (int i, int j)
+{
+  int k;
+  if (i != j)
+k = i;
+  else
+k = j;
+  return k;
+}
+
+int test2 (int i, int j)
+{
+  int k;
+  if (i != j)
+k = j;
+  else
+k = i;
+  return k;
+}
+
+int test3 (int i, int j)
+{
+  int k;
+  if (i == j)
+k = j;
+  else
+k = i;
+  return k;
+}
+
+int test4 (int i, int j)
+{
+  int k;
+  if (i == j)
+k = i;
+  else
+k = j;
+  return k;
+}
+
+/* We'd expect 4 hits but since we only keep one forwarder the
+   VN predication machinery cannot record something for the entry
+   block since it doesn't work on edges but on their source.  */
+/* { dg-final { scan-tree-dump-times "equal on edge" 2 "fre1" } } */
diff --git a/gcc/tree-ssa-sccvn.cc b/gcc/tree-ssa-sccvn.cc
index 1f9c6c53b52..6895ae84d13 100644
--- a/gcc/tree-ssa-sccvn.cc
+++ b/gcc/tree-ssa-sccvn.cc
@@ -5814,7 +5814,8 @@ visit_phi (gimple *phi, bool *inserted, bool 
backedges_varying_p)
 
   /* See if all non-TOP arguments have the same value.  TOP is
  equivalent to everything, so we can ignore it.  */
-  FOR_EACH_EDGE (e, ei, gimple_bb (phi)->preds)
+  basic_block bb = gimple_bb (phi);
+  FOR_EACH_EDGE (e, ei, bb->preds)
 if (e->flags & EDGE_EXECUTABLE)
   {
tree def = PHI_ARG_DEF_FROM_EDGE (phi, e);
@@ -5859,6 +5860,54 @@ visit_phi (gimple *phi, bool *inserted, bool 
backedges_varying_p)
 && known_eq (soff, doff))
  continue;
  }
+   /* There's also the possibility to use equivalences.  */
+   if (!FLOAT_TYPE_P (TREE_TYPE (def)))
+ {
+   vn_nary_op_t vnresult;
+   tree ops[2];
+   ops[0] = def;
+   ops[1] = sameval;
+   tree val = vn_nary_op_lookup_pieces (2, EQ_EXPR,
+boolean_type_node,
+ops, );
+   if (! val && vnresult && vnresult->predicated_values)
+ {
+   val = vn_nary_op_get_predicated_value (vnresult, e->src);
+   if (val && integer_truep (val))
+ {
+   if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+   fprintf (dump_file, "Predication says ");
+   print_generic_expr (dump_file, def, TDF_NONE);
+   fprintf (dump_file, " and ");
+   print_generic_expr (dump_file, sameval, TDF_NONE);
+   fprintf (dump_file, " are equal on edge %d -> %d\n",
+e->src->index, e->dest->index);
+ }
+   continue;
+ }
+   /* If on all previous edges the value was equal to def
+  we can change sameval to def.  */
+   if (EDGE_COUNT (bb->preds) == 2
+   && (val = vn_nary_op_get_predicated_value
+   (vnresult, EDGE_PRED (bb, 0)->src))
+   && integer_truep (val))
+ {
+   if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+   fprintf (dump_file, "Predication says ");
+   print_generic_expr (dump_file, def, TDF_NONE);
+