[llvm-commits] Patch for Bug7696

Villmow, Micah Micah.Villmow at amd.com
Mon Nov 29 08:27:11 PST 2010


Ping. Anyone?

From: llvm-commits-bounces at cs.uiuc.edu [mailto:llvm-commits-bounces at cs.uiuc.edu] On Behalf Of Villmow, Micah
Sent: Tuesday, November 23, 2010 1:39 PM
To: llvm-commits
Subject: [llvm-commits] Patch for Bug7696

Reposting bug report here, any chance that it can get reviewed?
Patch that adds a boolean where a backend can specify whether to disable
certain optimizations because jumps are expensive.

The selection dag builder has an 'optimization' added into the visitBr function
which makes assumptions that are not valid on all architectures. The problem is
this.
The following function
kernel void cf_test(global int* a, int b, int c, int e)
{
        int d = 0;
        if (!b && c < e) {
                d = a + b;
        }
        *a = d;
}
Is transformed into something equivalent to this:
Kernel void cf_test(global int* a, int b, int c, int e)
{
  Int d;
  If (b) {
    d = 0;
  } else {
    if (c < e) {
      d = a + b;
    } else {
      d = 0;
    }
  }
 *a = d;
}

by the visitBr code found in SelectionDAGBuilder::visitBr():1188.

However, if jumps are expensive or jumps are not supported and high level flow
control needs to be reconstructed. This is extremely inefficient. For example
on AMD GPU's, a single flow control instruction can take 40 cycles to execute,
but an bit instruction, can be executed every cycle. So obviously the
assumptions made by this block of code are inefficient on AMD hardware.
Increasing control flow has a direct impact on performance and removing the
extra 'and' or 'or' in order to short circuit the conditional evaluation does
not work for our target.

So in order to make this type of instruction rely more on target specific
information. I've added a new Boolean to the TargetLoweringInfo class called
JumpIsExpensive along with accessor functions.

Please review the patch and apply if acceptable.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20101129/159758b1/attachment.html>


More information about the llvm-commits mailing list