Frank Winter via llvm-dev
2015-Aug-19 18:43 UTC
[llvm-dev] SLP vectorizer on simple function not working
The attached function should be vectorizable with a SIMD width of 4.
However, the SLP vectorizer tells me it cannot schedule the bundle.
llvm-3.6/bin/opt -datalayout -basicaa -slp-vectorizer -debug-only SLP -S
< func_simple_scalar.ll -o tmp.ll &> out
SLP: Analyzing blocks in main.
SLP: Found 4 stores to vectorize.
SLP: Analyzing a store chain of length 4.
SLP: Analyzing a store chain of length 4
SLP: Analyzing 4 stores at offset 0
SLP: bundle: store float %4, float* %5
SLP: initialize schedule region to store float %4, float* %5
SLP: extend schedule region end to store float %10, float* %11
SLP: extend schedule region end to store float %16, float* %17
SLP: extend schedule region end to store float %22, float* %23
SLP: try schedule bundle [ store float %4, float* %5; store float %10,
float* %11; store float %16, float* %17; store float %22, float* %23]
in block entrypoint
SLP: update deps of [ store float %4, float* %5; store float
%10, float* %11; store float %16, float* %17; store float %22, float* %23]
SLP: update deps of / store float %10, float* %11
SLP: update deps of / store float %16, float* %17
SLP: update deps of / store float %22, float* %23
SLP: update deps of %21 = load float* %20
SLP: update deps of %22 = fmul float %21, %19
SLP: update deps of %19 = load float* %18
SLP: update deps of %15 = load float* %14
SLP: update deps of %16 = fmul float %15, %13
SLP: update deps of %13 = load float* %12
SLP: update deps of %9 = load float* %8
SLP: update deps of %10 = fmul float %9, %7
SLP: update deps of %7 = load float* %6
SLP: We are not able to schedule this bundle!
SLP: cancel scheduling of [ store float %4, float* %5; store float
%10, float* %11; store float %16, float* %17; store float %22, float* %23]
SLP: Calculating cost for tree of size 1.
SLP: Check whether the tree with height 1 is fully vectorizable .
SLP: Found cost=2147483647 for VF=4
What am I missing?
Thanks,
Frank
-------------- next part --------------
;; ModuleID = 'module'
target triple = "x86_64-unknown-linux-gnu"
define void @main(i64 %lo, i64 %hi, float* %arg0, float* %arg1, float* %arg2) {
entrypoint:
%0 = getelementptr float* %arg1, i32 0
%1 = load float* %0
%2 = getelementptr float* %arg2, i32 0
%3 = load float* %2
%4 = fmul float %3, %1
%5 = getelementptr float* %arg0, i32 0
store float %4, float* %5
%6 = getelementptr float* %arg1, i32 1
%7 = load float* %6
%8 = getelementptr float* %arg2, i32 1
%9 = load float* %8
%10 = fmul float %9, %7
%11 = getelementptr float* %arg0, i32 1
store float %10, float* %11
%12 = getelementptr float* %arg1, i32 2
%13 = load float* %12
%14 = getelementptr float* %arg2, i32 2
%15 = load float* %14
%16 = fmul float %15, %13
%17 = getelementptr float* %arg0, i32 2
store float %16, float* %17
%18 = getelementptr float* %arg1, i32 3
%19 = load float* %18
%20 = getelementptr float* %arg2, i32 3
%21 = load float* %20
%22 = fmul float %21, %19
%23 = getelementptr float* %arg0, i32 3
store float %22, float* %23
ret void
}
Frank Winter via llvm-dev
2015-Aug-19 19:22 UTC
[llvm-dev] SLP vectorizer on simple function not working
Okay, it was the missing 'noalias' attribute on the arguments. Damn, I thought I had that down. Sorry for the noise. Thanks, Frank On 08/19/2015 02:43 PM, Frank Winter via llvm-dev wrote:> The attached function should be vectorizable with a SIMD width of 4. > However, the SLP vectorizer tells me it cannot schedule the bundle. > > llvm-3.6/bin/opt -datalayout -basicaa -slp-vectorizer -debug-only SLP -S > < func_simple_scalar.ll -o tmp.ll &> out > > SLP: Analyzing blocks in main. > SLP: Found 4 stores to vectorize. > SLP: Analyzing a store chain of length 4. > SLP: Analyzing a store chain of length 4 > SLP: Analyzing 4 stores at offset 0 > SLP: bundle: store float %4, float* %5 > SLP: initialize schedule region to store float %4, float* %5 > SLP: extend schedule region end to store float %10, float* %11 > SLP: extend schedule region end to store float %16, float* %17 > SLP: extend schedule region end to store float %22, float* %23 > SLP: try schedule bundle [ store float %4, float* %5; store float %10, > float* %11; store float %16, float* %17; store float %22, float* %23] > in block entrypoint > SLP: update deps of [ store float %4, float* %5; store float > %10, float* %11; store float %16, float* %17; store float %22, float* > %23] > SLP: update deps of / store float %10, float* %11 > SLP: update deps of / store float %16, float* %17 > SLP: update deps of / store float %22, float* %23 > SLP: update deps of %21 = load float* %20 > SLP: update deps of %22 = fmul float %21, %19 > SLP: update deps of %19 = load float* %18 > SLP: update deps of %15 = load float* %14 > SLP: update deps of %16 = fmul float %15, %13 > SLP: update deps of %13 = load float* %12 > SLP: update deps of %9 = load float* %8 > SLP: update deps of %10 = fmul float %9, %7 > SLP: update deps of %7 = load float* %6 > SLP: We are not able to schedule this bundle! > SLP: cancel scheduling of [ store float %4, float* %5; store float > %10, float* %11; store float %16, float* %17; store float %22, float* > %23] > SLP: Calculating cost for tree of size 1. > SLP: Check whether the tree with height 1 is fully vectorizable . > SLP: Found cost=2147483647 for VF=4 > > What am I missing? > > Thanks, > Frank > > > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev >