diff mbox series

Add option to force indirect calls for x86

Message ID 20171108143222.4212-1-andi@firstfloor.org
State New
Headers show
Series Add option to force indirect calls for x86 | expand

Commit Message

Andi Kleen Nov. 8, 2017, 2:32 p.m. UTC 
From: Andi Kleen <ak@linux.intel.com>

This patch adds a -mforce-indirect-call option to force all calls
or tail calls on x86_64 between functions to indirect. This is similar to the
large code model, but doesn't affect jumps inside functions, so has much
less run time overhead.

This is useful with Intel Processor Trace (PT). PT has precise timing
for indirect calls/jumps, but not for direct ones. So if we can force
them to indirect it allows to time every function relatively accurately
(minus the overhead of the indirect branch)

Without this short functions often don't see a timing update and cannot
be measured.

The timing requires at least Skylake or Goldmont based CPUs.

I made it an option. Originally I tried to make it a new code model,
but since it can be combined with other code models (medium, pic, kernel
etc.) this turned out to be too many combinations.

For example with gcc. This first column is a ns time stamp for the
functions.

$ perf record -e intel_pt/noretcomp=1,cyc=1,cyc_thresh=1/u ./cc1 -O3 hello.c
$ perf script --itrace=cr -F callindent,time,sym,addr --ns  | sed -n 180000,182000p | less
...
1184596.432756920:                             build_int_cst                        =>           79c9de c_common_nodes_and_builtins
1184596.432756921:                             tree_cons                            =>           ee2080 tree_cons
1184596.432756938:                                 ggc_internal_alloc               =>           80f3e0 ggc_internal_alloc
1184596.432756951:                                     memset@plt                   =>           598af0 memset@plt
1184596.432756967:                                     __memset_avx2_unaligned_erms =>           80f605 ggc_internal_alloc
1184596.432756969:                                 ggc_internal_alloc               =>           ee20a2 tree_cons
1184596.432756973:                             tree_cons                            =>           79c9f4 c_common_nodes_and_builtins
1184596.432756974:                             build_int_cst                        =>           ef9a40 build_int_cst
1184596.432756996:                                 wide_int_to_tree                 =>           ef93a0 wide_int_to_tree
1184596.432757000:                                     wi::force_to_size            =>           f48f70 wi::force_to_size
1184596.432757005:                                     canonize                     =>           ef94de wide_int_to_tree
1184596.432757021:                                     get_int_cst_ext_nunits       =>           ee1960 get_int_cst_ext_nunits
1184596.432757026:                                     get_int_cst_ext_nunits       =>           ef94fe wide_int_to_tree
1184596.432757042:                                     tree_int_cst_elt_check       =>           83e310 tree_int_cst_elt_check
1184596.432757044:                                     tree_int_cst_elt_check       =>           ef9761 wide_int_to_tree
1184596.432757046:                                 wide_int_to_tree                 =>           ef9a9b build_int_cst

Passes bootstrap and test suite with x86_64, also a gcc itself built
with the option bootstraps.

gcc/:
2017-11-08  Andi Kleen  <ak@linux.intel.com>

	* config/i386/i386.opt: Add -mforce-indirect-call.
	* config/i386/predicates.md: Check for flag_force_indirect_call.
	* doc/invoke.texi: Document -mforce-indirect-call

gcc/testsuite/:
2017-11-08  Andi Kleen  <ak@linux.intel.com>

	* gcc.target/i386/force-indirect-call-1.c: New test.
	* gcc.target/i386/force-indirect-call-2.c: New test.
	* gcc.target/i386/force-indirect-call-3.c: New test.
---
 gcc/config/i386/i386.opt                           |  4 ++++
 gcc/config/i386/predicates.md                      |  3 ++-
 gcc/doc/invoke.texi                                |  8 +++++++-
 .../gcc.target/i386/force-indirect-call-1.c        | 23 ++++++++++++++++++++++
 .../gcc.target/i386/force-indirect-call-2.c        |  5 +++++
 .../gcc.target/i386/force-indirect-call-3.c        |  5 +++++
 6 files changed, 46 insertions(+), 2 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/i386/force-indirect-call-1.c
 create mode 100644 gcc/testsuite/gcc.target/i386/force-indirect-call-2.c
 create mode 100644 gcc/testsuite/gcc.target/i386/force-indirect-call-3.c

Comments

Rainer Orth Nov. 10, 2017, 2:13 p.m. UTC | #1 
Hi Andi,

> This patch adds a -mforce-indirect-call option to force all calls
> or tail calls on x86_64 between functions to indirect. This is similar to the
> large code model, but doesn't affect jumps inside functions, so has much
> less run time overhead.
>
> This is useful with Intel Processor Trace (PT). PT has precise timing
> for indirect calls/jumps, but not for direct ones. So if we can force
> them to indirect it allows to time every function relatively accurately
> (minus the overhead of the indirect branch)
>
> Without this short functions often don't see a timing update and cannot
> be measured.
[...]
> gcc/testsuite/:
> 2017-11-08  Andi Kleen  <ak@linux.intel.com>
>
> 	* gcc.target/i386/force-indirect-call-1.c: New test.
> 	* gcc.target/i386/force-indirect-call-2.c: New test.
> 	* gcc.target/i386/force-indirect-call-3.c: New test.

two of those new tests FAIL for 32-bit x86 (seen on both
i386-pc-solaris2.11 and x86_64-pc-linux-gnu):

+FAIL: gcc.target/i386/force-indirect-call-2.c scan-assembler-times call[ \\\\t]
+\\\\*% 2 (found 3 times)
+FAIL: gcc.target/i386/force-indirect-call-2.c scan-assembler-times jmp[ \\\\t]+
\\\\*% 1 (found 0 times)

Also, the test uses -fPIC, thus requires target fpic.

+FAIL: gcc.target/i386/force-indirect-call-3.c (test for excess errors)

cc1: error: code model 'medium' not supported in the 32 bit mode

+UNRESOLVED: gcc.target/i386/force-indirect-call-3.c scan-assembler-times call[ 
\\\\t]+\\\\*% 2
+UNRESOLVED: gcc.target/i386/force-indirect-call-3.c scan-assembler-times jmp[ \
\\\t]+\\\\*% 1

Please fix.

	Rainer
diff mbox series

Patch

diff --git a/gcc/config/i386/i386.opt b/gcc/config/i386/i386.opt
index 7c9dd471686..df727eee468 100644
--- a/gcc/config/i386/i386.opt
+++ b/gcc/config/i386/i386.opt
@@ -977,3 +977,7 @@  mcet-switch
 Target Report Undocumented Var(flag_cet_switch) Init(0)
 Turn on CET instrumentation for switch statements, which use jump table and
 indirect jump.
+
+mforce-indirect-call
+Target Report Var(flag_force_indirect_call) Init(0)
+Make all function calls indirect.
diff --git a/gcc/config/i386/predicates.md b/gcc/config/i386/predicates.md
index c3f442eb8ac..c6e6e980959 100644
--- a/gcc/config/i386/predicates.md
+++ b/gcc/config/i386/predicates.md
@@ -600,7 +600,8 @@ 
 (define_predicate "constant_call_address_operand"
   (match_code "symbol_ref")
 {
-  if (ix86_cmodel == CM_LARGE || ix86_cmodel == CM_LARGE_PIC)
+  if (ix86_cmodel == CM_LARGE || ix86_cmodel == CM_LARGE_PIC
+      || flag_force_indirect_call)
     return false;
   if (TARGET_DLLIMPORT_DECL_ATTRIBUTES && SYMBOL_REF_DLLIMPORT_P (op))
     return false;
diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi
index 2ef88e081f9..e897d93070a 100644
--- a/gcc/doc/invoke.texi
+++ b/gcc/doc/invoke.texi
@@ -1205,7 +1205,7 @@  See RS/6000 and PowerPC Options.
 -msse4a  -m3dnow  -m3dnowa  -mpopcnt  -mabm  -mbmi  -mtbm  -mfma4  -mxop @gol
 -mlzcnt  -mbmi2  -mfxsr  -mxsave  -mxsaveopt  -mrtm  -mlwp  -mmpx  @gol
 -mmwaitx  -mclzero  -mpku  -mthreads @gol
--mcet -mibt -mshstk @gol
+-mcet -mibt -mshstk -mforce-indirect-call @gol
 -mms-bitfields  -mno-align-stringops  -minline-all-stringops @gol
 -minline-stringops-dynamically  -mstringop-strategy=@var{alg} @gol
 -mmemcpy-strategy=@var{strategy}  -mmemset-strategy=@var{strategy} @gol
@@ -26175,6 +26175,12 @@  You can control this behavior for specific functions by
 using the function attributes @code{ms_abi} and @code{sysv_abi}.
 @xref{Function Attributes}.
 
+@item -mforce-indirect-call
+@opindex mforce-indirect-call
+Force all calls to functions to be indirect. This is useful
+when using Intel Processor Trace where it generates more precise timing
+information for function calls.
+
 @item -mcall-ms2sysv-xlogues
 @opindex mcall-ms2sysv-xlogues
 @opindex mno-call-ms2sysv-xlogues
diff --git a/gcc/testsuite/gcc.target/i386/force-indirect-call-1.c b/gcc/testsuite/gcc.target/i386/force-indirect-call-1.c
new file mode 100644
index 00000000000..be1be2c879e
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/force-indirect-call-1.c
@@ -0,0 +1,23 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O2 -mforce-indirect-call" } */
+/* { dg-final { scan-assembler-times "call\[ \\t\]+\\*%" 2 } } */
+/* { dg-final { scan-assembler-times "jmp\[ \\t\]+\\*%" 1 } } */
+int x;
+int y;
+
+void __attribute__((noinline)) f1(void)
+{
+	x++;
+}
+
+static __attribute__((noinline)) void f3(void)
+{
+	y++;
+}
+
+void f2()
+{
+	f1();
+	f3();
+	f1();
+}
diff --git a/gcc/testsuite/gcc.target/i386/force-indirect-call-2.c b/gcc/testsuite/gcc.target/i386/force-indirect-call-2.c
new file mode 100644
index 00000000000..dd0df259ab8
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/force-indirect-call-2.c
@@ -0,0 +1,5 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O2 -mforce-indirect-call -fPIC" } */
+/* { dg-final { scan-assembler-times "call\[ \\t\]+\\*%" 2 } } */
+/* { dg-final { scan-assembler-times "jmp\[ \\t\]+\\*%" 1 } } */
+#include "force-indirect-call-1.c"
diff --git a/gcc/testsuite/gcc.target/i386/force-indirect-call-3.c b/gcc/testsuite/gcc.target/i386/force-indirect-call-3.c
new file mode 100644
index 00000000000..28d8c98b7b9
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/force-indirect-call-3.c
@@ -0,0 +1,5 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O2 -mforce-indirect-call -mcmodel=medium" } */
+/* { dg-final { scan-assembler-times "call\[ \\t\]+\\*%" 2 } } */
+/* { dg-final { scan-assembler-times "jmp\[ \\t\]+\\*%" 1 } } */
+#include "force-indirect-call-1.c"