Reorganize pass-pipeline for Polly

[MatthiasJReisinger]

We now move Polly after LoopRotate since at this point the LLVM IR is more amenable to Polly. For example, a simple gemm kernel like

@polly function gemm!(A,B,C)
    m,n = size(A)
    n,o = size(B)
    @inbounds for i=1:m, j=1:o, k=1:n
        C[i,j] += A[i,k] * B[k,j]
    end
end

could not be optimized before this change. However, this new position after LoopRotate also makes it necessary to rerun InstCombine before Polly to get rid of hindering phi instructions that are introduced by LCSSA. Additionally, we add Polly's CodePreparation and CodegenCleanUp passes. Parts of these changes have already been discussed in https://groups.google.com/forum/#!topic/polly-dev/P2RZUlKRo0I. The reason why they were held back was to see if further adaptions to the pass-pipeline will become apparent, but for now this configuration seems to be convenient.

@TobiG @vtjnash @timholy

[tobiasgrosser]

Why are makefile changes necessary to reorganize the pass pipeline?

[MatthiasJReisinger]

To use the CodePreparationPass in jitlayers.cpp it was necessary to include polly/LinkAllPasses.h which in turn includes polly/Config/config.h which is placed in $(shell $(LLVM_CONFIG_HOST) --obj-root)/tools/polly/include.

[tobiasgrosser]

On 07/29/2016 01:10 PM, Matthias J. Reisinger wrote:

In src/Makefile
#17696 (comment):

@@ -54,7 +54,7 @@ FLAGS += -I$(shell $(LLVM_CONFIG_HOST) --includedir)
LLVM_LIBS := all
ifeq ($(USE_POLLY),1)
LLVMLINK += -lPolly -lPollyISL
-FLAGS += -I$(shell $(LLVM_CONFIG_HOST) --src-root)/tools/polly/include
+FLAGS += -I$(shell $(LLVM_CONFIG_HOST) --src-root)/tools/polly/include -I$(shell $(LLVM_CONFIG_HOST) --obj-root)/tools/polly/include

To use the CodePreparationPass in jitlayers.cpp it was necessary to
include polly/LinkAllPasses.h which in turn includes
polly/Config/config.h which is placed in $(shell $(LLVM_CONFIG_HOST)
--obj-root)/tools/polly/include.

OK.

Tobias

[ViralBShah]

Let's merge this after branching for 0.5.

[tobiasgrosser]

Codegen cleanup and codegen prepare passes are partially independent changes. I would suggest to add them in separate commits (may be the still pull-request though).

[MatthiasJReisinger]

@TobiG thank you for the feedback! I will address this and update the patch. @ViralBShah I agree, it's not necessary to merge this before 0.5.

[vtjnash]

lgtm. Since this is all behind preprocessor macros anyways, I see no particular reason it matters if this merges before the v0.5 branch point or after.