[llvm] r228973 - Change max interleave factor to 12 for POWER7 and POWER8.

[Hal Finkel]

----- Original Message -----
> From: "Olivier Sallenave" <ohsallen at us.ibm.com>
> To: llvm-commits at cs.uiuc.edu
> Sent: Thursday, February 12, 2015 4:57:58 PM
> Subject: [llvm] r228973 - Change max interleave factor to 12 for POWER7 and	POWER8.
> 
> Author: ohsallen
> Date: Thu Feb 12 16:57:58 2015
> New Revision: 228973
> 
> URL: http://llvm.org/viewvc/llvm-project?rev=228973&view=rev
> Log:
> Change max interleave factor to 12 for POWER7 and POWER8.
> 
> Added:
>     llvm/trunk/test/Transforms/LoopVectorize/PowerPC/small-loop-rdx.ll
> Modified:
>     llvm/trunk/lib/Target/PowerPC/PPCTargetTransformInfo.cpp
> 
> Modified: llvm/trunk/lib/Target/PowerPC/PPCTargetTransformInfo.cpp
> URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/PowerPC/PPCTargetTransformInfo.cpp?rev=228973&r1=228972&r2=228973&view=diff
> ==============================================================================
> --- llvm/trunk/lib/Target/PowerPC/PPCTargetTransformInfo.cpp
> (original)
> +++ llvm/trunk/lib/Target/PowerPC/PPCTargetTransformInfo.cpp Thu Feb
> 12 16:57:58 2015
> @@ -226,6 +226,12 @@ unsigned PPCTTIImpl::getMaxInterleaveFac
>    if (Directive == PPC::DIR_E500mc || Directive == PPC::DIR_E5500)
>      return 1;
>  
> +  // For P7 and P8, floating-point instructions have a 6-cycle
> latency and
> +  // there are two execution units, so unroll by 12x for latency
> hiding.
> +  if (Directive == PPC::DIR_PWR7 ||
> +      Directive == PPC::DIR_PWR8)
> +    return 12;
> +
>    // For most things, modern systems have two execution units (and
>    // out-of-order execution).
>    return 2;
> 
> Added:
> llvm/trunk/test/Transforms/LoopVectorize/PowerPC/small-loop-rdx.ll
> URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/LoopVectorize/PowerPC/small-loop-rdx.ll?rev=228973&view=auto
> ==============================================================================
> ---
> llvm/trunk/test/Transforms/LoopVectorize/PowerPC/small-loop-rdx.ll
> (added)
> +++
> llvm/trunk/test/Transforms/LoopVectorize/PowerPC/small-loop-rdx.ll
> Thu Feb 12 16:57:58 2015
> @@ -0,0 +1,35 @@
> +; RUN: opt < %s -loop-vectorize -S -debug < %s 2>&1 | FileCheck %s
> +
> +; CHECK: LV: Unroll Factor is 12

You'll need to add:
; REQUIRES: asserts

to this test because it checks debug output (you'll break the -Asserts builds otherwise). However, I'd rather you not do that, just run the optimization and add a sufficient number of CHECK lines to make sure that we have the desired number of interleave iterations.

Something like this should do it:

CHECK: fadd
CHECK: fadd
CHECK: fadd
CHECK: fadd
CHECK: fadd
CHECK: fadd
CHECK: fadd
CHECK: fadd
CHECK: fadd
CHECK: fadd
CHECK: fadd
CHECK: fadd
CHECK-NOT: fadd

 -Hal

> +
> +target datalayout = "e-m:e-i64:64-n32:64"
> +target triple = "powerpc64le-ibm-linux-gnu"
> +
> +define void @test(double* nocapture readonly %arr, i32 signext %len)
> #0 {
> +entry:
> +  %cmp4 = icmp sgt i32 %len, 0
> +  br i1 %cmp4, label %for.body.lr.ph, label %for.end
> +
> +for.body.lr.ph:                                   ; preds = %entry
> +  %0 = add i32 %len, -1
> +  br label %for.body
> +
> +for.body:                                         ; preds =
> %for.body, %for.body.lr.ph
> +  %indvars.iv = phi i64 [ 0, %for.body.lr.ph ], [ %indvars.iv.next,
> %for.body ]
> +  %redx.05 = phi double [ 0.000000e+00, %for.body.lr.ph ], [ %add,
> %for.body ]
> +  %arrayidx = getelementptr inbounds double* %arr, i64 %indvars.iv
> +  %1 = load double* %arrayidx, align 8
> +  %add = fadd fast double %1, %redx.05
> +  %indvars.iv.next = add i64 %indvars.iv, 1
> +  %lftr.wideiv = trunc i64 %indvars.iv to i32
> +  %exitcond = icmp eq i32 %lftr.wideiv, %0
> +  br i1 %exitcond, label %for.end.loopexit, label %for.body
> +
> +for.end.loopexit:                                 ; preds =
> %for.body
> +  %add.lcssa = phi double [ %add, %for.body ]
> +  br label %for.end
> +
> +for.end:                                          ; preds =
> %for.end.loopexit, %entry
> +  %redx.0.lcssa = phi double [ 0.000000e+00, %entry ], [ %add.lcssa,
> %for.end.loopexit ]
> +  ret void
> +}
> 
> 
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> 

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory