[llvm-bugs] [Bug 31274] New: cost models should allow something more than an instruction as an input

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=31274

            Bug ID: 31274
           Summary: cost models should allow something more than an
                    instruction as an input
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: normal
          Priority: P
         Component: Transformation Utilities
          Assignee: unassignedbugs at nondot.org
          Reporter: spatel+llvm at rotateright.com
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

Filing a bug to keep track of a suggestion that has come up a few times
recently:
http://lists.llvm.org/pipermail/llvm-dev/2016-November/107489.html
http://lists.llvm.org/pipermail/llvm-dev/2016-November/106879.html

Here's a umax example to illustrate:

$ cat costmodel_patterns.ll
target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-apple-macosx10.12.0"

define i32 @max(i32* nocapture readonly %x, i32 %N) #0 {
entry:
  %cmp11 = icmp eq i32 %N, 0
  br i1 %cmp11, label %for.cond.cleanup, label %for.body.preheader

for.body.preheader:   
  %wide.trip.count = zext i32 %N to i64
  br label %for.body

for.cond.cleanup.loopexit:
  br label %for.cond.cleanup

for.cond.cleanup:        
  %ret.0.lcssa = phi i32 [ 0, %entry ], [ %.ret.0, %for.cond.cleanup.loopexit ]
  ret i32 %ret.0.lcssa

for.body:               
  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0,
%for.body.preheader ]
  %ret.012 = phi i32 [ %.ret.0, %for.body ], [ 0, %for.body.preheader ]
  %arrayidx = getelementptr inbounds i32, i32* %x, i64 %indvars.iv
  %0 = load i32, i32* %arrayidx, align 4
  %cmp1 = icmp ugt i32 %0, %ret.012
  %.ret.0 = select i1 %cmp1, i32 %0, i32 %ret.012
  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
  %exitcond = icmp eq i64 %indvars.iv.next, %wide.trip.count
  br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body
}

attributes #0 = { "target-features"="+avx" }

-----------------------------------------------------------------------------

This is IR for a target that has AVX, therefore, 'umax' is a single and simple 
instruction with expected throughput of 1 inst / cycle:
$ ./opt -loop-vectorize costmodel_patterns.ll -S | ./llc -o -  |grep max
...    
    vpmaxud    %xmm4, %xmm1, %xmm1
...

The cost model interface, however, is limited to providing costs for individual
IR instructions. For example:
  /// \returns The expected cost of compare and select instructions.
  int getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
                         Type *CondTy = nullptr) const;


That means we see something like this:

$ ./opt -cost-model -analyze costmodel_patterns.ll -S
Printing analysis 'Cost Model Analysis' for function 'max':
...
Cost Model: Found an estimated cost of 1 for instruction:   %cmp1 = icmp ugt
i32 %0, %ret.012
Cost Model: Found an estimated cost of 1 for instruction:   %.ret.0 = select i1
%cmp1, i32 %0, i32 %ret.012

...so we calculate a cost of '2' for a max idiom that should have a cost of '1'
(both in terms of machine instruction count and throughput).

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20161205/e9e18468/attachment.html>