[SYCL] Enhance device code split call graph analysis (#8589)

This patch introduces significant changes to how device code split
detects functions and global variables which should be included into a
cloned module.

There are two main changes done to that:
1. analysis algorithm now traces uses of global variables to allow
adding all globals into every split module
2. analysis algorithm now traces indirect calls, trying to define a list
of all functions which are potentially called indirectly to avoid the
need to disabled device code split completely in presence of indirect
calls

Both things are implemented through new `DependencyGraph` entity, which
replaces `CallGraph` entity we used.
Instead of calls, that new graph is built over _uses_ of functions and
variables to understand which functions and global variables are used by
which functions and global variables.

The most tricky part here is indirect calls: we can't understand which
exact function is being called by an indirect call. However, we can
compile a list of _potentially_-called function by comparing function
signatures with signature of an indirect call.

On top of that, ESIMD handling is refactored by this patch:
- outlined ESIMD-specific handling into a separate function
- created new ESIMD-specific device code split helper

New ESIMD-specific device code split helper is needed, because we should
use different rules for ESIMD and non-ESIMD parts of a module when
splitting it to two. For ESIMD part we want to grab all ESIMD-functions
even if they were not considered as entry points in the original module.
For non-ESIMD part we **don't want** to grab _any_ ESIMD-functions, even
if they are referenced/used by non-ESIMD functions.

Both of those special rules come from `invoke_simd` feature support:
non-ESIMD kernel can indirectly reference a ESIMD function. Since those
different kind of functions require different processing, we have to
completely separate them before processing step. Non-ESIMD module could
be incomplete as a result of such split, but it will be merged back with
ESIMD module after ESIMD lowering. That merge step is required for
`invoke_simd` functionality.

---------

Co-authored-by: Cai, Justin <[email protected]>

Author: AlexeySachkov

llvm/test/tools/sycl-post-link/assert/indirect-with-split-2.ll

@@ -9,8 +9,16 @@
 ; marked as using asserts.
 
 ; RUN: sycl-post-link -split=auto -symbols -S < %s -o %t.table
-; RUN: FileCheck %s -input-file=%t_0.prop -check-prefix=PRESENCE-CHECK
-; RUN: FileCheck %s -input-file=%t_0.prop -check-prefix=ABSENCE-CHECK
+; RUN: FileCheck %s -input-file=%t_0.prop -check-prefixes=CHECK,CHECK0 \
+; RUN:     --implicit-check-not TU1
+; RUN: FileCheck %s -input-file=%t_1.prop -check-prefixes=CHECK,CHECK1 \
+; RUN:     --implicit-check-not TU0
+;
+; CHECK: [SYCL/assert used]
+; CHECK0-DAG: main_TU1_kernel0
+; CHECK0-DAG: main_TU1_kernel1
+;
+; CHECK1: main_TU0_kernel0
 
 target datalayout = "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024"
 target triple = "spir64-unknown-linux"
@@ -40,7 +48,7 @@ entry:
 }
 
 ; ABSENCE-CHECK-NOT: empty_kernel
-define dso_local spir_kernel void @empty_kernel() {
+define dso_local spir_kernel void @empty_kernel() #2 {
   %1 = ptrtoint void ()* @bar to i64
   ret void
 }

llvm/test/tools/sycl-post-link/assert/indirect-with-split.ll

@@ -7,7 +7,21 @@
 ; marked as using asserts.
 
 ; RUN: sycl-post-link -split=auto -symbols -S < %s -o %t.table
-; RUN: FileCheck %s -input-file=%t_0.prop
+; RUN: FileCheck %s -input-file=%t_0.prop --check-prefixes=CHECK,CHECK1 \
+; RUN:     --implicit-check-not TU0
+; RUN: FileCheck %s -input-file=%t_1.prop --check-prefixes=CHECK,CHECK0 \
+; RUN:     --implicit-check-not TU1 --implicit-check-not kernel1
+;
+; With recent improvements to device code split, this file is actually being
+; split to two modules and one of them does not contain "indirectly-referenced"
+; function, meaning that only direct users of 'assert' will be mentioned in
+; device image properties.
+;
+; CHECK: [SYCL/assert used]
+; CHECK0: main_TU0_kernel0
+;
+; CHECK1-DAG: main_TU1_kernel0
+; CHECK1-DAG: main_TU1_kernel1
 
 target datalayout = "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024"
 target triple = "spir64-unknown-linux"
@@ -20,9 +34,6 @@ target triple = "spir64-unknown-linux"
 @__spirv_BuiltInLocalInvocationId = external dso_local local_unnamed_addr addrspace(1) constant <3 x i64>, align 32
 @_ZL10assert_fmt = internal addrspace(2) constant [85 x i8] c"%s:%d: %s: global id: [%lu,%lu,%lu], local id: [%lu,%lu,%lu] Assertion `%s` failed.\0A\00", align 1
 
-; CHECK: [SYCL/assert used]
-
-; CHECK-DAG: main_TU0_kernel0
 define dso_local spir_kernel void @main_TU0_kernel0() #0 {
 entry:
   call spir_func void @_Z3foov()
@@ -40,7 +51,6 @@ entry:
   ret void
 }
 
-; CHECK-DAG: main_TU0_kernel1
 define dso_local spir_kernel void @main_TU0_kernel1() #0 {
 entry:
   call spir_func void @_Z4foo1v()
@@ -55,14 +65,12 @@ entry:
   ret void
 }
 
-; CHECK-DAG: main_TU1_kernel0
 define dso_local spir_kernel void @main_TU1_kernel0() #2 {
 entry:
   call spir_func void @_Z3foov()
   ret void
 }
 
-; CHECK-DAG: main_TU1_kernel1
 define dso_local spir_kernel void @main_TU1_kernel1() #2 {
 entry:
   call spir_func void @_Z4foo2v()

llvm/test/tools/sycl-post-link/device-code-split/auto-module-split-2.ll

@@ -1,24 +1,42 @@
 ; RUN: sycl-post-link -split=auto -symbols -S < %s -o %t.table
-; In precense of indirectly callable function auto mode is equal to no split,
-; which means that separate LLVM IR file for device is not generated and we only
-; need to check generated symbol table
-; RUN: FileCheck %s -input-file=%t_0.sym --check-prefixes CHECK
+;
+; This is the same as auto-module-split-1 test with the only difference is that
+; @_Z3foov is marked with "referenced-indirectly" attribute.
+; The purpose of this test is to make sure that we can still perform device code
+; split as usual, because that function is not a part of any indirect calls
+;
+; RUN: FileCheck %s -input-file=%t_0.ll --check-prefixes CHECK-TU0,CHECK
+; RUN: FileCheck %s -input-file=%t_1.ll --check-prefixes CHECK-TU1,CHECK
+; RUN: FileCheck %s -input-file=%t_0.sym --check-prefixes CHECK-TU0-TXT
+; RUN: FileCheck %s -input-file=%t_1.sym --check-prefixes CHECK-TU1-TXT
 
 target datalayout = "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024"
 target triple = "spir64-unknown-linux"
 
 $_Z3barIiET_S0_ = comdat any
 
+; CHECK-TU1-NOT: @{{.*}}GV{{.*}}
+; CHECK-TU0: @{{.*}}GV{{.*}} = internal addrspace(1) constant [1 x i32] [i32 42], align 4
 @_ZL2GV = internal addrspace(1) constant [1 x i32] [i32 42], align 4
 
-; CHECK: {{.*}}TU0_kernel0{{.*}}
+; CHECK-TU1: define dso_local spir_kernel void @{{.*}}TU0_kernel0{{.*}}
+; CHECK-TU1-TXT: {{.*}}TU0_kernel0{{.*}}
+; CHECK-TU0-NOT: define dso_local spir_kernel void @{{.*}}TU0_kernel0{{.*}}
+; CHECK-TU0-TXT-NOT: {{.*}}TU0_kernel0{{.*}}
+
+; CHECK-TU1: call spir_func void @{{.*}}foo{{.*}}()
 
 define dso_local spir_kernel void @_ZTSZ4mainE11TU0_kernel0() #0 {
 entry:
   call spir_func void @_Z3foov()
   ret void
 }
 
+; CHECK-TU1: define dso_local spir_func void @{{.*}}foo{{.*}}()
+; CHECK-TU0-NOT: define dso_local spir_func void @{{.*}}foo{{.*}}()
+
+; CHECK-TU1: call spir_func i32 @{{.*}}bar{{.*}}(i32 1)
+
 define dso_local spir_func void @_Z3foov() #2 {
 entry:
   %a = alloca i32, align 4
@@ -28,6 +46,9 @@ entry:
   ret void
 }
 
+; CHECK-TU1: define {{.*}} spir_func i32 @{{.*}}bar{{.*}}(i32 %arg)
+; CHECK-TU0-NOT: define {{.*}} spir_func i32 @{{.*}}bar{{.*}}(i32 %arg)
+
 ; Function Attrs: nounwind
 define linkonce_odr dso_local spir_func i32 @_Z3barIiET_S0_(i32 %arg) comdat {
 entry:
@@ -37,33 +58,51 @@ entry:
   ret i32 %0
 }
 
-; CHECK: {{.*}}TU0_kernel1{{.*}}
+; CHECK-TU1: define dso_local spir_kernel void @{{.*}}TU0_kernel1{{.*}}()
+; CHECK-TU1-TXT: {{.*}}TU0_kernel1{{.*}}
+; CHECK-TU0-NOT: define dso_local spir_kernel void @{{.*}}TU0_kernel1{{.*}}()
+; CHECK-TU0-TXT-NOT: {{.*}}TU0_kernel1{{.*}}
+
+; CHECK-TU1: call spir_func void @{{.*}}foo1{{.*}}()
 
 define dso_local spir_kernel void @_ZTSZ4mainE11TU0_kernel1() #0 {
 entry:
   call spir_func void @_Z4foo1v()
   ret void
 }
 
+; CHECK-TU1: define dso_local spir_func void @{{.*}}foo1{{.*}}()
+; CHECK-TU0-NOT: define dso_local spir_func void @{{.*}}foo1{{.*}}()
+
 ; Function Attrs: nounwind
 define dso_local spir_func void @_Z4foo1v() {
 entry:
   %a = alloca i32, align 4
   store i32 2, i32* %a, align 4
   ret void
 }
-; CHECK: {{.*}}TU1_kernel{{.*}}
+
+; CHECK-TU1-NOT: define dso_local spir_kernel void @{{.*}}TU1_kernel{{.*}}()
+; CHECK-TU1-TXT-NOT: {{.*}}TU1_kernel{{.*}}
+; CHECK-TU0: define dso_local spir_kernel void @{{.*}}TU1_kernel{{.*}}()
+; CHECK-TU0-TXT: {{.*}}TU1_kernel{{.*}}
+
+; CHECK-TU0: call spir_func void @{{.*}}foo2{{.*}}()
 
 define dso_local spir_kernel void @_ZTSZ4mainE10TU1_kernel() #1 {
 entry:
   call spir_func void @_Z4foo2v()
   ret void
 }
 
+; CHECK-TU1-NOT: define dso_local spir_func void @{{.*}}foo2{{.*}}()
+; CHECK-TU0: define dso_local spir_func void @{{.*}}foo2{{.*}}()
+
 ; Function Attrs: nounwind
 define dso_local spir_func void @_Z4foo2v() {
 entry:
   %a = alloca i32, align 4
+; CHECK-TU0: %0 = load i32, i32 addrspace(4)* getelementptr inbounds ([1 x i32], [1 x i32] addrspace(4)* addrspacecast ([1 x i32] addrspace(1)* @{{.*}}GV{{.*}} to [1 x i32] addrspace(4)*), i64 0, i64 0), align 4
   %0 = load i32, i32 addrspace(4)* getelementptr inbounds ([1 x i32], [1 x i32] addrspace(4)* addrspacecast ([1 x i32] addrspace(1)* @_ZL2GV to [1 x i32] addrspace(4)*), i64 0, i64 0), align 4
   %add = add nsw i32 4, %0
   store i32 %add, i32* %a, align 4
@@ -74,8 +113,15 @@ attributes #0 = { "sycl-module-id"="TU1.cpp" }
 attributes #1 = { "sycl-module-id"="TU2.cpp" }
 attributes #2 = { "referenced-indirectly" }
 
+; Metadata is saved in both modules.
+; CHECK: !opencl.spir.version = !{!0, !0}
+; CHECK: !spirv.Source = !{!1, !1}
+
 !opencl.spir.version = !{!0, !0}
 !spirv.Source = !{!1, !1}
 
+; CHECK: !0 = !{i32 1, i32 2}
+; CHECK: !1 = !{i32 4, i32 100000}
+
 !0 = !{i32 1, i32 2}
 !1 = !{i32 4, i32 100000}

llvm/test/tools/sycl-post-link/device-code-split/auto-module-split-3.ll

@@ -1,8 +1,31 @@
 ; RUN: sycl-post-link -split=auto -symbols -S < %s -o %t.table
-; In precense of indirect calls auto mode is equal to no split,
-; which means that separate LLVM IR file for device is not generated and we only
-; need to check generated symbol table
-; RUN: FileCheck %s -input-file=%t_0.sym --check-prefixes CHECK
+;
+; In precense of indirect calls we start matching functions using their
+; signatures, i.e. we have an indirect call to i32(i32) function within
+; @_Z3foov, which means that all functions with i32(i32) signature should be
+; placed in the same module as @_Z3foov.
+;
+; RUN: FileCheck %s -input-file=%t_0.ll --check-prefixes CHECK-TU0-IR \
+; RUN:     --implicit-check-not TU0_kernel --implicit-check-not _Z3foov
+; RUN: FileCheck %s -input-file=%t_1.ll --check-prefixes CHECK-TU1-IR \
+; RUN:     --implicit-check-not TU1_kernel --implicit-check-not _Z4foo2v
+; RUN: FileCheck %s -input-file=%t_0.sym --check-prefixes CHECK-TU0-SYM
+; RUN: FileCheck %s -input-file=%t_1.sym --check-prefixes CHECK-TU1-SYM
+;
+; CHECK-TU0-SYM: _ZTSZ4mainE11TU1_kernel0
+; CHECK-TU0-SYM: _ZTSZ4mainE11TU1_kernel1
+;
+; CHECK-TU1-SYM: _ZTSZ4mainE10TU0_kernel
+;
+; CHECK-TU0-IR: @_ZL2GV = internal addrspace(1) constant
+; CHECK-TU0-IR: define dso_local spir_kernel void @_ZTSZ4mainE11TU1_kernel0
+; CHECK-TU0-IR: define dso_local spir_func i32 @_Z4foo1v
+; CHECK-TU0-IR: define dso_local spir_kernel void @_ZTSZ4mainE11TU1_kernel1
+; CHECK-TU0-IR: define dso_local spir_func void @_Z4foo2v
+;
+; CHECK-TU1-IR: define dso_local spir_kernel void @_ZTSZ4mainE10TU0_kernel
+; CHECK-TU1-IR: define dso_local spir_func void @_Z3foov
+; CHECK-TU1-IR: define dso_local spir_func i32 @_Z4foo1v
 
 target datalayout = "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024"
 target triple = "spir64-unknown-linux"
@@ -11,9 +34,7 @@ $_Z3barIiET_S0_ = comdat any
 
 @_ZL2GV = internal addrspace(1) constant [1 x i32] [i32 42], align 4
 
-; CHECK: {{.*}}TU0_kernel0{{.*}}
-
-define dso_local spir_kernel void @_ZTSZ4mainE11TU0_kernel0() #0 {
+define dso_local spir_kernel void @_ZTSZ4mainE10TU0_kernel() #0 {
 entry:
   call spir_func void @_Z3foov()
   ret void
@@ -38,24 +59,23 @@ entry:
   ret i32 %0
 }
 
-; CHECK: {{.*}}TU0_kernel1{{.*}}
-
-define dso_local spir_kernel void @_ZTSZ4mainE11TU0_kernel1() #0 {
+define dso_local spir_kernel void @_ZTSZ4mainE11TU1_kernel0() #1 {
 entry:
-  call spir_func void @_Z4foo1v()
+  %a = alloca i32, align 4
+  %arg = load i32, i32* %a, align 4
+  %call = call spir_func i32 @_Z4foo1v(i32 %arg)
   ret void
 }
 
 ; Function Attrs: nounwind
-define dso_local spir_func void @_Z4foo1v() {
+define dso_local spir_func i32 @_Z4foo1v(i32 %arg) {
 entry:
   %a = alloca i32, align 4
-  store i32 2, i32* %a, align 4
-  ret void
+  store i32 %arg, i32* %a, align 4
+  ret i32 %arg
 }
-; CHECK: {{.*}}TU1_kernel{{.*}}
 
-define dso_local spir_kernel void @_ZTSZ4mainE10TU1_kernel() #1 {
+define dso_local spir_kernel void @_ZTSZ4mainE11TU1_kernel1() #1 {
 entry:
   call spir_func void @_Z4foo2v()
   ret void

llvm/test/tools/sycl-post-link/device-code-split/auto-module-split-func-ptr.ll

@@ -1,8 +1,20 @@
 ; RUN: sycl-post-link -split=auto -symbols -S < %s -o %t.table
-; RUN: FileCheck %s -input-file=%t_0.sym
-
-; This test checkes that module is not split if function pointer's user is not
-; CallInst.
+; RUN: FileCheck %s -input-file=%t_0.sym --check-prefix=CHECK-SYM0
+; RUN: FileCheck %s -input-file=%t_1.sym --check-prefix=CHECK-SYM1
+; RUN: FileCheck %s -input-file=%t_0.ll --check-prefix=CHECK-IR0
+; RUN: FileCheck %s -input-file=%t_1.ll --check-prefix=CHECK-IR1
+
+; This test checkes that we can properly perform device code split by tracking
+; all uses of functions (not only direct calls)
+
+; CHECK-SYM0: kernel2
+; CHECK-SYM1: kernel1
+;
+; CHECK-IR0: define dso_local spir_kernel void @kernel2
+;
+; CHECK-IR1: @_Z2f1iTable = weak global [1 x i32 (i32)*] [i32 (i32)* @_Z2f1i]
+; CHECK-IR1: define dso_local spir_func i32 @_Z2f1i
+; CHECK-IR1: define weak_odr dso_local spir_kernel void @kernel1
 
 target datalayout = "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-n8:16:32:64"
 target triple = "spir64_x86_64-unknown-unknown"