[AMDGPU] Graph-based Module Splitting Rewrite (#104763)

Major rewrite of the AMDGPUSplitModule pass in order to better support
it long-term.

Highlights:
- Removal of the "SML" logging system in favor of just using CL options
and LLVM_DEBUG, like any other pass in LLVM.
- The SML system started from good intentions, but it was too flawed and
messy to be of any real use. It was also a real pain to use and made the
code more annoying to maintain.
 - Graph-based module representation with DOTGraph printing support
- The graph represents the module accurately, with bidirectional, typed
edges between nodes (a node usually represents one function).
- Nodes are assigned IDs starting from 0, which allows us to represent a
set of nodes as a BitVector. This makes comparing 2 sets of nodes to
find common dependencies a trivial task. Merging two clusters of nodes
together is also really trivial.
 - No more defaulting to "P0" for external calls
- Roots that can reach non-copyable dependencies (such as external
calls) are now grouped together in a single "cluster" that can go into
any partition.
 - No more defaulting to "P0" for indirect calls
- New representation for module splitting proposals that can be graded
and compared.
- Graph-search algorithm that can explore multiple branches/assignments
for a cluster of functions, up to a maximum depth.
- With the default max depth of 8, we can create up to 256 propositions
to try and find the best one.
- We can still fall back to a greedy approach upon reaching max depth.
That greedy approach uses almost identical heuristics to the previous
version of the pass.

All of this gives us a lot of room to experiment with new heuristics or
even entirely different splitting strategies if we need to. For
instance, the graph representation has room for abstract nodes, e.g. if
we need to represent some global variables or external constraints. We
could also introduce more edge types to model other type of relations
between nodes, etc.

I also designed the graph representation & the splitting strategies to
be as fast as possible, and it seems to have paid off. Some quick tests
showed that we spend pretty much all of our time in the CloneModule
function, with the actual splitting logic being >1% of the runtime.

Author: Pierre-vh

llvm/lib/Target/AMDGPU/AMDGPUSplitModule.cpp

@@ -1,30 +1,24 @@
-; RUN: llvm-split -o %t %s -j 3 -mtriple amdgcn-amd-amdhsa -amdgpu-module-splitting-large-function-threshold=0
-; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
-; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
-; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+; RUN: llvm-split -o %t %s -j 3 -mtriple amdgcn-amd-amdhsa -amdgpu-module-splitting-large-threshold=0
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 --implicit-check-not=define %s
 
 ; 3 kernels:
 ;   - A does a direct call to HelperA
 ;   - B is storing @HelperA
 ;   - C does a direct call to HelperA
 ;
-; The helper functions will get externalized, which will force A and C into P0 as
-; external functions cannot be duplicated.
-
-; CHECK0: define hidden void @HelperA()
-; CHECK0: define amdgpu_kernel void @A()
-; CHECK0: declare amdgpu_kernel void @B(ptr)
-; CHECK0: define amdgpu_kernel void @C()
-
-; CHECK1: declare hidden void @HelperA()
-; CHECK1: declare amdgpu_kernel void @A()
-; CHECK1: declare amdgpu_kernel void @B(ptr)
-; CHECK1: declare amdgpu_kernel void @C()
-
-; CHECK2: declare hidden void @HelperA()
-; CHECK2: declare amdgpu_kernel void @A()
-; CHECK2: define amdgpu_kernel void @B(ptr %dst)
-; CHECK2: declare amdgpu_kernel void @C()
+; The helper functions will get externalized, so C/A will end up
+; in the same partition.
+
+; P0 is empty.
+; CHECK0: declare
+
+; CHECK1: define amdgpu_kernel void @B(ptr %dst)
+
+; CHECK2: define hidden void @HelperA()
+; CHECK2: define amdgpu_kernel void @A()
+; CHECK2: define amdgpu_kernel void @C()
 
 define internal void @HelperA() {
   ret void

llvm/test/tools/llvm-split/AMDGPU/address-taken-externalize-with-call.ll

@@ -1,4 +1,4 @@
-; RUN: llvm-split -o %t %s -j 2 -mtriple amdgcn-amd-amdhsa -amdgpu-module-splitting-large-function-threshold=0
+; RUN: llvm-split -o %t %s -j 2 -mtriple amdgcn-amd-amdhsa -amdgpu-module-splitting-large-threshold=0
 ; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
 ; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
 

llvm/test/tools/llvm-split/AMDGPU/address-taken-externalize.ll

@@ -1,15 +1,12 @@
 ; RUN: llvm-split -o %t %s -j 2 -mtriple amdgcn-amd-amdhsa
 ; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
-; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
+; RUN: not llvm-dis -o - %t1
 
-; Check that all declarations are put into each partition.
+; Empty module without any defs should result in a single output module that is
+; an exact copy of the input.
 
 ; CHECK0: declare void @A
 ; CHECK0: declare void @B
 
-; CHECK1: declare void @A
-; CHECK1: declare void @B
-
 declare void @A()
-
 declare void @B()

llvm/test/tools/llvm-split/AMDGPU/debug-name-hiding.ll

@@ -1,6 +1,6 @@
 ; RUN: llvm-split -o %t %s -j 2 -mtriple amdgcn-amd-amdhsa
-; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
-; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 --implicit-check-not=define %s
 
 ; 3 kernels:
 ;   - A calls nothing
@@ -13,16 +13,12 @@
 ; Additionally, @PerryThePlatypus gets externalized as
 ; the alias counts as taking its address.
 
-; CHECK0-NOT: define
-; CHECK0: @Perry = internal alias ptr (), ptr @PerryThePlatypus
-; CHECK0: define hidden void @PerryThePlatypus()
-; CHECK0: define amdgpu_kernel void @B
-; CHECK0: define amdgpu_kernel void @C
-; CHECK0-NOT: define
+; CHECK0: define amdgpu_kernel void @A
 
-; CHECK1-NOT: define
-; CHECK1: define amdgpu_kernel void @A
-; CHECK1-NOT: define
+; CHECK1: @Perry = internal alias ptr (), ptr @PerryThePlatypus
+; CHECK1: define hidden void @PerryThePlatypus()
+; CHECK1: define amdgpu_kernel void @B
+; CHECK1: define amdgpu_kernel void @C
 
 @Perry = internal alias ptr(), ptr @PerryThePlatypus
 

llvm/test/tools/llvm-split/AMDGPU/debug-non-kernel-root.ll

@@ -1,27 +1,21 @@
 ; RUN: llvm-split -o %t %s -j 3 -mtriple amdgcn-amd-amdhsa
-; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
-; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
-; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 --implicit-check-not=define %s
 
 ; 3 kernels with each their own dependencies should go into 3
 ; distinct partitions. The most expensive kernel should be
 ; seen first and go into the last partition.
 
-; CHECK0-NOT: define
 ; CHECK0: define amdgpu_kernel void @C
 ; CHECK0: define internal void @HelperC
 ; CHECK0-NOT: define
 
-; CHECK1-NOT: define
 ; CHECK1: define amdgpu_kernel void @A
 ; CHECK1: define internal void @HelperA
-; CHECK1-NOT: define
 
-; CHECK2-NOT: define
 ; CHECK2: define amdgpu_kernel void @B
 ; CHECK2: define internal void @HelperB
-; CHECK2-NOT: define
-
 
 define amdgpu_kernel void @A() {
   call void @HelperA()

llvm/test/tools/llvm-split/AMDGPU/declarations.ll

@@ -1,29 +1,20 @@
 ; RUN: llvm-split -o %t %s -j 4 -mtriple amdgcn-amd-amdhsa
-; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
-; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
-; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
-; RUN: llvm-dis -o - %t3 | FileCheck --check-prefix=CHECK3 %s
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t3 | FileCheck --check-prefix=CHECK3 --implicit-check-not=define %s
 
-; Both overridable helper should go in P0.
+; CHECK0: define internal void @PrivateHelper1()
+; CHECK0: define amdgpu_kernel void @D
 
-; CHECK0-NOT: define
-; CHECK0: define available_externally void @OverridableHelper0()
-; CHECK0: define internal void @OverridableHelper1()
-; CHECK0: define amdgpu_kernel void @A
-; CHECK0: define amdgpu_kernel void @B
-; CHECK0-NOT: define
+; CHECK1: define internal void @PrivateHelper0()
+; CHECK1: define amdgpu_kernel void @C
 
-; CHECK1-NOT: define
+; CHECK2: define internal void @OverridableHelper1()
+; CHECK2: define amdgpu_kernel void @B
 
-; CHECK2-NOT: define
-; CHECK2: define internal void @PrivateHelper1()
-; CHECK2: define amdgpu_kernel void @D
-; CHECK2-NOT: define
-
-; CHECK3-NOT: define
-; CHECK3: define internal void @PrivateHelper0()
-; CHECK3: define amdgpu_kernel void @C
-; CHECK3-NOT: define
+; CHECK3: define available_externally void @OverridableHelper0()
+; CHECK3: define amdgpu_kernel void @A
 
 define available_externally void @OverridableHelper0() {
   ret void

llvm/test/tools/llvm-split/AMDGPU/kernels-alias-dependencies.ll

@@ -1,7 +1,7 @@
 ; RUN: llvm-split -o %t %s -j 3 -mtriple amdgcn-amd-amdhsa
-; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
-; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
-; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 --implicit-check-not=define %s
 
 ; We have 4 kernels:
 ;   - Each kernel has an internal helper
@@ -15,25 +15,19 @@
 ; indirect call. HelperC/D should also end up in P0 as they
 ; are dependencies of HelperB.
 
-; CHECK0-NOT: define
-; CHECK0: define hidden void @HelperA
-; CHECK0: define hidden void @HelperB
-; CHECK0: define hidden void @CallCandidate
-; CHECK0: define internal void @HelperC
 ; CHECK0: define internal void @HelperD
-; CHECK0: define amdgpu_kernel void @A
-; CHECK0: define amdgpu_kernel void @B
-; CHECK0-NOT: define
+; CHECK0: define amdgpu_kernel void @D
 
-; CHECK1-NOT: define
-; CHECK1: define internal void @HelperD
-; CHECK1: define amdgpu_kernel void @D
-; CHECK1-NOT: define
+; CHECK1: define internal void @HelperC
+; CHECK1: define amdgpu_kernel void @C
 
-; CHECK2-NOT: define
+; CHECK2: define hidden void @HelperA
+; CHECK2: define hidden void @HelperB
+; CHECK2: define hidden void @CallCandidate
 ; CHECK2: define internal void @HelperC
-; CHECK2: define amdgpu_kernel void @C
-; CHECK2-NOT: define
+; CHECK2: define internal void @HelperD
+; CHECK2: define amdgpu_kernel void @A
+; CHECK2: define amdgpu_kernel void @B
 
 @addrthief = global [3 x ptr] [ptr @HelperA, ptr @HelperB, ptr @CallCandidate]
 

llvm/test/tools/llvm-split/AMDGPU/kernels-cost-ranking.ll

@@ -1,21 +1,15 @@
 ; RUN: llvm-split -o %t %s -j 3 -mtriple amdgcn-amd-amdhsa
-; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
-; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
-; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
-
-; CHECK0-NOT: define
-; CHECK0: define void @ExternalHelper
-; CHECK0: define amdgpu_kernel void @A
-; CHECK0: define amdgpu_kernel void @B
-; CHECK0-NOT: define
-
-; CHECK1-NOT: define
-; CHECK1: define amdgpu_kernel void @D
-; CHECK1-NOT: define
-
-; CHECK2-NOT: define
-; CHECK2: define amdgpu_kernel void @C
-; CHECK2-NOT: define
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 --implicit-check-not=define %s
+
+; CHECK0: define amdgpu_kernel void @D
+
+; CHECK1: define amdgpu_kernel void @C
+
+; CHECK2: define void @ExternalHelper
+; CHECK2: define amdgpu_kernel void @A
+; CHECK2: define amdgpu_kernel void @B
 
 define void @ExternalHelper() {
   ret void

llvm/test/tools/llvm-split/AMDGPU/kernels-dependency-external.ll

@@ -1,26 +1,20 @@
 ; RUN: llvm-split -o %t %s -j 3 -mtriple amdgcn-amd-amdhsa -amdgpu-module-splitting-no-externalize-globals
-; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
-; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
-; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 --implicit-check-not=define %s
 
 ; 3 kernels use private/internal global variables.
 ; The GVs should be copied in each partition as needed.
 
-; CHECK0-NOT: define
 ; CHECK0: @bar = internal constant ptr
 ; CHECK0: define amdgpu_kernel void @C
-; CHECK0-NOT: define
 
-; CHECK1-NOT: define
 ; CHECK1: @foo = private constant ptr
 ; CHECK1: define amdgpu_kernel void @A
-; CHECK1-NOT: define
 
-; CHECK2-NOT: define
 ; CHECK2: @foo = private constant ptr
 ; CHECK2: @bar = internal constant ptr
 ; CHECK2: define amdgpu_kernel void @B
-; CHECK2-NOT: define
 
 @foo = private constant ptr poison
 @bar = internal constant ptr poison

llvm/test/tools/llvm-split/AMDGPU/kernels-dependency-indirect.ll

@@ -1,28 +1,22 @@
 ; RUN: llvm-split -o %t %s -j 3 -mtriple amdgcn-amd-amdhsa
-; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
-; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
-; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 --implicit-check-not=define %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 --implicit-check-not=define %s
 
 ; 3 kernels use private/internal global variables.
 ; The GVs should be copied in each partition as needed.
 
-; CHECK0-NOT: define
 ; CHECK0: @foo = hidden constant ptr poison
 ; CHECK0: @bar = hidden constant ptr poison
 ; CHECK0: define amdgpu_kernel void @C
-; CHECK0-NOT: define
 
-; CHECK1-NOT: define
 ; CHECK1: @foo = external hidden constant ptr{{$}}
 ; CHECK1: @bar = external hidden constant ptr{{$}}
 ; CHECK1: define amdgpu_kernel void @A
-; CHECK1-NOT: define
 
-; CHECK2-NOT: define
 ; CHECK2: @foo = external hidden constant ptr{{$}}
 ; CHECK2: @bar = external hidden constant ptr{{$}}
 ; CHECK2: define amdgpu_kernel void @B
-; CHECK2-NOT: define
 
 @foo = private constant ptr poison
 @bar = internal constant ptr poison