[test-suite] r178968 - [tests] Modify TSVC static data layout to make benchmarks more stable.

[Daniel Dunbar]

Author: ddunbar
Date: Sat Apr  6 22:21:06 2013
New Revision: 178968

URL: http://llvm.org/viewvc/llvm-project?rev=178968&view=rev
Log:
[tests] Modify TSVC static data layout to make benchmarks more stable.

As was previously written, the performance of these benchmarks can depend
significantly on the exact addresses assigned to the global data arrays at
runtime. Rewrite the tests to make the relative layout of each of the data
arrays fixed by storing them all in a single global structure. This makes the
behavior more predictable across platforms.

For example, previously on Mac OS X, two of the arrays would end up with
relative offsets at an exact multiple of 4K, which causes problems with some
Intel CPUs. This caused the performance of some benchmarks
(StatementReordering-flt in particular) to be dominated by a CPU specific hazard
instead of the actual benchmark code.

Modified:
    test-suite/trunk/MultiSource/Benchmarks/TSVC/tsc.inc

Modified: test-suite/trunk/MultiSource/Benchmarks/TSVC/tsc.inc
URL: http://llvm.org/viewvc/llvm-project/test-suite/trunk/MultiSource/Benchmarks/TSVC/tsc.inc?rev=178968&r1=178967&r2=178968&view=diff
==============================================================================
--- test-suite/trunk/MultiSource/Benchmarks/TSVC/tsc.inc (original)
+++ test-suite/trunk/MultiSource/Benchmarks/TSVC/tsc.inc Sat Apr  6 22:21:06 2013
@@ -66,10 +66,46 @@ TYPE x[LEN] __attribute__((aligned(ALIGN
 TYPE temp;
 int temp_int;
 
-
-__attribute__((aligned(ALIGNMENT))) TYPE a[LEN],b[LEN],c[LEN],d[LEN],e[LEN],
-                                   aa[LEN2][LEN2],bb[LEN2][LEN2],cc[LEN2][LEN2],tt[LEN2][LEN2];
-
+// We place all of the data into one single global structure so that we can
+// control its exact layout. Otherwise, the performance of the code can become
+// very dependent on the exact addresses assigned to arrays at compile time, and
+// probing that behavior is not the purpose of this set of benchmarks.
+//
+// We insert prime-multiple padding in between each array to help ensure that
+// the relative offsets of the arrays are unlikely to trigger unmodelled
+// architecture specific problems w.r.t. cache behavior or other CPU
+// features. For example, this ensures that no two of the arrays will be
+// 4K-aliased with each other, which can be important for some Intel processors.
+struct GlobalData {
+  __attribute__((aligned(ALIGNMENT))) TYPE a[LEN];
+  int pad1[3];
+  __attribute__((aligned(ALIGNMENT))) TYPE b[LEN];
+  int pad2[5];
+  __attribute__((aligned(ALIGNMENT))) TYPE c[LEN];
+  int pad3[7];
+  __attribute__((aligned(ALIGNMENT))) TYPE d[LEN];
+  int pad4[11];
+  __attribute__((aligned(ALIGNMENT))) TYPE e[LEN];
+
+  int pad5[13];
+  __attribute__((aligned(ALIGNMENT))) TYPE aa[LEN2][LEN2];
+  int pad6[17];
+  __attribute__((aligned(ALIGNMENT))) TYPE bb[LEN2][LEN2];
+  int pad7[19];
+  __attribute__((aligned(ALIGNMENT))) TYPE cc[LEN2][LEN2];
+  int pad8[23];
+  __attribute__((aligned(ALIGNMENT))) TYPE tt[LEN2][LEN2];
+} global_data;
+
+__attribute__((aligned(ALIGNMENT))) TYPE * const a = global_data.a;
+__attribute__((aligned(ALIGNMENT))) TYPE * const b = global_data.b;
+__attribute__((aligned(ALIGNMENT))) TYPE * const c = global_data.c;
+__attribute__((aligned(ALIGNMENT))) TYPE * const d = global_data.d;
+__attribute__((aligned(ALIGNMENT))) TYPE * const e = global_data.e;
+__attribute__((aligned(ALIGNMENT))) TYPE (* const aa)[LEN2] = global_data.aa;
+__attribute__((aligned(ALIGNMENT))) TYPE (* const bb)[LEN2] = global_data.bb;
+__attribute__((aligned(ALIGNMENT))) TYPE (* const cc)[LEN2] = global_data.cc;
+__attribute__((aligned(ALIGNMENT))) TYPE (* const tt)[LEN2] = global_data.tt;
 
 int indx[LEN] __attribute__((aligned(ALIGNMENT)));
 
@@ -5757,4 +5793,3 @@ int main(int argc, char *argv[]){
 #endif
 	return 0;
 }
-