[PATCH] R600/SI: Fix bug in TTI loop unrolling preferences

[Hal Finkel]

LGTM.

 -Hal

----- Original Message -----
> From: "Tom Stellard" <thomas.stellard at amd.com>
> To: llvm-commits at cs.uiuc.edu
> Cc: "Tom Stellard" <thomas.stellard at amd.com>
> Sent: Wednesday, February 4, 2015 12:32:08 PM
> Subject: [PATCH] R600/SI: Fix bug in TTI loop unrolling preferences
> 
> We should be setting UnrollingPreferences::MaxCount to MAX_UINT
> instead
> of UnrollingPreferences::Count.
> 
> Count is a 'forced unrolling factor', while MaxCount sets an upper
> limit to the unrolling factor.
> 
> Setting Count to MAX_UINT was causing the loop in the testcase to be
> unrolled 15 times, when it only had a maximum of 4 iterations.
> ---
>  lib/Target/R600/AMDGPUTargetTransformInfo.cpp |  2 +-
>  test/CodeGen/R600/tti-unroll-prefs.ll         | 58
>  +++++++++++++++++++++++++++
>  2 files changed, 59 insertions(+), 1 deletion(-)
>  create mode 100644 test/CodeGen/R600/tti-unroll-prefs.ll
> 
> diff --git a/lib/Target/R600/AMDGPUTargetTransformInfo.cpp
> b/lib/Target/R600/AMDGPUTargetTransformInfo.cpp
> index 4647ddf..68f4600 100644
> --- a/lib/Target/R600/AMDGPUTargetTransformInfo.cpp
> +++ b/lib/Target/R600/AMDGPUTargetTransformInfo.cpp
> @@ -30,7 +30,7 @@ using namespace llvm;
>  void AMDGPUTTIImpl::getUnrollingPreferences(Loop *L,
>                                              TTI::UnrollingPreferences
>                                              &UP) {
>    UP.Threshold = 300; // Twice the default.
> -  UP.Count = UINT_MAX;
> +  UP.MaxCount = UINT_MAX;
>    UP.Partial = true;
>  
>    // TODO: Do we want runtime unrolling?
> diff --git a/test/CodeGen/R600/tti-unroll-prefs.ll
> b/test/CodeGen/R600/tti-unroll-prefs.ll
> new file mode 100644
> index 0000000..0009c42
> --- /dev/null
> +++ b/test/CodeGen/R600/tti-unroll-prefs.ll
> @@ -0,0 +1,58 @@
> +; RUN: opt -loop-unroll -S -mtriple=amdgcn-- -mcpu=SI %s | FileCheck
> %s
> +
> +; This IR comes from this OpenCL C code:
> +;
> +; if (b + 4 > a) {
> +;   for (int i = 0; i < 4; i++, b++) {
> +;     if (b + 1 <= a)
> +;       *(dst + c + b) = 0;
> +;     else
> +;       break;
> +;   }
> +; }
> +;
> +; This test is meant to check that this loop isn't unrolled into
> more than
> +; four iterations.  The loop unrolling preferences we currently use
> cause this
> +; loop to not be unrolled at all, but that may change in the future.
> +
> +; CHECK-LABEL: @test
> +; CHECK: store i8 0, i8 addrspace(1)*
> +; CHECK-NOT: store i8 0, i8 addrspace(1)*
> +; CHECK: ret void
> +define void @test(i8 addrspace(1)* nocapture %dst, i32 %a, i32 %b,
> i32 %c) {
> +entry:
> +  %add = add nsw i32 %b, 4
> +  %cmp = icmp sgt i32 %add, %a
> +  br i1 %cmp, label %for.cond.preheader, label %if.end7
> +
> +for.cond.preheader:                               ; preds = %entry
> +  %cmp313 = icmp slt i32 %b, %a
> +  br i1 %cmp313, label %if.then4.lr.ph, label %if.end7.loopexit
> +
> +if.then4.lr.ph:                                   ; preds =
> %for.cond.preheader
> +  %0 = sext i32 %c to i64
> +  br label %if.then4
> +
> +if.then4:                                         ; preds =
> %if.then4.lr.ph, %if.then4
> +  %i.015 = phi i32 [ 0, %if.then4.lr.ph ], [ %inc, %if.then4 ]
> +  %b.addr.014 = phi i32 [ %b, %if.then4.lr.ph ], [ %add2, %if.then4
> ]
> +  %add2 = add nsw i32 %b.addr.014, 1
> +  %1 = sext i32 %b.addr.014 to i64
> +  %add.ptr.sum = add nsw i64 %1, %0
> +  %add.ptr5 = getelementptr inbounds i8 addrspace(1)* %dst, i64
> %add.ptr.sum
> +  store i8 0, i8 addrspace(1)* %add.ptr5, align 1
> +  %inc = add nsw i32 %i.015, 1
> +  %cmp1 = icmp slt i32 %inc, 4
> +  %cmp3 = icmp slt i32 %add2, %a
> +  %or.cond = and i1 %cmp3, %cmp1
> +  br i1 %or.cond, label %if.then4, label
> %for.cond.if.end7.loopexit_crit_edge
> +
> +for.cond.if.end7.loopexit_crit_edge:              ; preds =
> %if.then4
> +  br label %if.end7.loopexit
> +
> +if.end7.loopexit:                                 ; preds =
> %for.cond.if.end7.loopexit_crit_edge, %for.cond.preheader
> +  br label %if.end7
> +
> +if.end7:                                          ; preds =
> %if.end7.loopexit, %entry
> +  ret void
> +}
> --
> 1.8.5.5
> 
> _______________________________________________
> 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