X86 machine model: reduce SandyBridge and Haswell ILPWindow.

The initial values were arbitrary. I want them to be more
conservative. This represents the number of latency cycles hidden by
OOO execution. In practice, I think it should be within a small factor
of the complex floating point operation latency so the scheduler can
make some attempt to hide latency even for smallish blocks.

These are by no means the best values, just a starting point for
tuning heuristics. Some benchmarks such as TSVC run faster with this
lower value for SandyBridge. I haven't run anything on Haswell, but
it's shouldn't be 2x SB.

llvm-svn: 179450

diff --git a/llvm/lib/Target/X86/X86SchedHaswell.td b/llvm/lib/Target/X86/X86SchedHaswell.td
index 7de6791..84c9203 100644 (file)
--- a/llvm/lib/Target/X86/X86SchedHaswell.td
+++ b/llvm/lib/Target/X86/X86SchedHaswell.td
@@ -18,7 +18,7 @@ def HaswellModel : SchedMachineModel {
   let IssueWidth = 4;
   let MinLatency = 0; // 0 = Out-of-order execution.
   let LoadLatency = 4;
-  let ILPWindow = 40;
+  let ILPWindow = 30;
   let MispredictPenalty = 16;
 }
 

diff --git a/llvm/lib/Target/X86/X86SchedSandyBridge.td b/llvm/lib/Target/X86/X86SchedSandyBridge.td
index 74d5f1b..b36b3ad 100644 (file)
--- a/llvm/lib/Target/X86/X86SchedSandyBridge.td
+++ b/llvm/lib/Target/X86/X86SchedSandyBridge.td
@@ -19,7 +19,7 @@ def SandyBridgeModel : SchedMachineModel {
   let IssueWidth = 4;
   let MinLatency = 0; // 0 = Out-of-order execution.
   let LoadLatency = 4;
-  let ILPWindow = 30;
+  let ILPWindow = 20;
   let MispredictPenalty = 16;
 }