Fixes for top-of-tree LLVM (#6546)

Makefile

@@ -1064,9 +1064,25 @@ $(BUILD_DIR)/initmod.%_32_debug.ll: $(SRC_DIR)/runtime/%.cpp $(BUILD_DIR)/clang_
 	@mkdir -p $(@D)
 	$(CLANG) $(CXX_WARNING_FLAGS) -g -DDEBUG_RUNTIME -O3 $(RUNTIME_CXX_FLAGS) -fpic -m32 -target $(RUNTIME_TRIPLE_32) -DCOMPILING_HALIDE_RUNTIME -DBITS_32 -emit-llvm -S $(SRC_DIR)/runtime/$*.cpp -o $@ -MMD -MP -MF $(BUILD_DIR)/initmod.$*_32_debug.d
 
+ifneq (,$(findstring $(LLVM_VERSION_TIMES_10), 120 130))
+# For LLVM14+, we must add elementtype() annotations to some of our LLVM IR;
+# earlier versions either don't understand that keyword at all, or don't support
+# the uses we have for it. Rather than forking these sources, for now we'll just
+# edit the files at build time to remove the offending uses. Note that while we could use `sed`
+# here, that isn't an option for CMake builds (since they must support Windows environments without
+# such tooling); to ensure consistent transformations in all builds, we'll use the tool here, too.
+#
+# (This may well need attention in the future, depending on how the LLVM opaque-pointers work proceeeds;
+# see https://llvm.org/docs/OpaquePointers.html)
+$(BUILD_DIR)/initmod.%_ll.ll: $(SRC_DIR)/runtime/%.ll $(BIN_DIR)/regexp_replace
+	@mkdir -p $(@D)
+	$(BIN_DIR)/regexp_replace 'elementtype\(i[0-9]+\)' '' < $(SRC_DIR)/runtime/$*.ll > $(BUILD_DIR)/initmod.$*_ll.ll
+else
 $(BUILD_DIR)/initmod.%_ll.ll: $(SRC_DIR)/runtime/%.ll
 	@mkdir -p $(@D)
 	cp $(SRC_DIR)/runtime/$*.ll $(BUILD_DIR)/initmod.$*_ll.ll
+endif
+
 
 $(BUILD_DIR)/initmod.%.bc: $(BUILD_DIR)/initmod.%.ll $(BUILD_DIR)/llvm_ok
 	$(LLVM_AS) $(BUILD_DIR)/initmod.$*.ll -o $(BUILD_DIR)/initmod.$*.bc
@@ -1088,6 +1104,10 @@ $(BIN_DIR)/binary2cpp: $(ROOT_DIR)/tools/binary2cpp.cpp
 	@mkdir -p $(@D)
 	$(CXX) $< -o $@
 
+$(BIN_DIR)/regexp_replace: $(ROOT_DIR)/tools/regexp_replace.cpp
+	@mkdir -p $(@D)
+	$(CXX) -std=c++11 $< -o $@
+
 $(BUILD_DIR)/initmod_ptx.%_ll.o: $(BUILD_DIR)/initmod_ptx.%_ll.cpp
 	$(CXX) -c $< -o $@ -MMD -MP -MF $(BUILD_DIR)/$*.d -MT $(BUILD_DIR)/$*.o
 

src/runtime/CMakeLists.txt

@@ -293,9 +293,32 @@ foreach (i IN LISTS RUNTIME_LL)
     set(BC "initmod.${i}.bc")
     set(INITMOD "_initmod_${i}.cpp")
 
+    # For LLVM14+, we must add elementtype() annotations to some of our LLVM IR;
+    # earlier versions either don't understand that keyword at all, or don't support
+    # the uses we have for it. Rather than forking these sources, for now we'll just
+    # edit the files at build time to remove the offending uses. Note that `sed` and friends aren't
+    # an option for CMake builds (since they must support Windows environments without
+    # such tooling).
+    #
+    # (This may well need attention in the future, depending on how the LLVM opaque-pointers work proceeeds;
+    # see https://llvm.org/docs/OpaquePointers.html)
+
+    set(LL_TRANSFORMED "${LL}.transformed.ll")
+    if (Halide_LLVM_VERSION VERSION_LESS 14.0)
+        add_custom_command(OUTPUT "${LL_TRANSFORMED}"
+                           COMMAND regexp_replace "elementtype\\(i[0-9]+\\)" "" < "$<SHELL_PATH:${CMAKE_CURRENT_SOURCE_DIR}/${LL}>" > "${LL_TRANSFORMED}"
+                           DEPENDS "${LL}" regexp_replace
+                           VERBATIM)
+    else ()
+        add_custom_command(OUTPUT "${LL_TRANSFORMED}"
+                           COMMAND ${CMAKE_COMMAND} -E copy "$<SHELL_PATH:${CMAKE_CURRENT_SOURCE_DIR}/${LL}>" "${LL_TRANSFORMED}"
+                           DEPENDS "${LL}"
+                           VERBATIM)
+    endif ()
+
     add_custom_command(OUTPUT "${BC}"
-                       COMMAND llvm-as "$<SHELL_PATH:${CMAKE_CURRENT_SOURCE_DIR}/${LL}>" -o "${BC}"
-                       DEPENDS "${LL}"
+                       COMMAND llvm-as "${LL_TRANSFORMED}" -o "${BC}"
+                       DEPENDS "${LL_TRANSFORMED}"
                        VERBATIM)
     add_custom_command(OUTPUT "${INITMOD}"
                        COMMAND binary2cpp "halide_internal_initmod_${i}_ll" < "${BC}" > "${INITMOD}"

src/runtime/hvx_128.ll

@@ -423,7 +423,7 @@ define weak_odr void @halide.hexagon.vscatter_acc.w.w(i8* %buf_ptr, i32 %size, <
 }
 
 define weak_odr void @halide.hexagon.scatter.release(i8* %ptr) nounwind uwtable {
-  call void asm sideeffect "vmem($0 + #0):scatter_release\0A; v1 = vmem($0 + #0)\0A", "=*m,*m,~{v1}"(i8* %ptr, i8* %ptr)
+  call void asm sideeffect "vmem($0 + #0):scatter_release\0A; v1 = vmem($0 + #0)\0A", "=*m,*m,~{v1}"(i8* elementtype(i8) %ptr, i8* elementtype(i8) %ptr)
   ret void
 }
 

src/runtime/x86.ll

@@ -1,28 +1,24 @@
-; Note that this is only used for LLVM 8.0+
 define weak_odr <16 x i8>  @paddusbx16(<16 x i8> %a0, <16 x i8> %a1) nounwind alwaysinline {
   %1 = add <16 x i8> %a0, %a1
   %2 = icmp ugt <16 x i8> %a0, %1
   %3 = select <16 x i1> %2, <16 x i8> <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>, <16 x i8> %1
   ret <16 x i8> %3
 }
 
-; Note that this is only used for LLVM 8.0+
 define weak_odr <8 x i16> @padduswx8(<8 x i16> %a0, <8 x i16> %a1) nounwind alwaysinline {
   %1 = add <8 x i16> %a0, %a1
   %2 = icmp ugt <8 x i16> %a0, %1
   %3 = select <8 x i1> %2, <8 x i16> <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1>, <8 x i16> %1
   ret <8 x i16> %3
 }
 
-; Note that this is only used for LLVM 8.0+
 define weak_odr <16 x i8> @psubusbx16(<16 x i8> %a0, <16 x i8> %a1) nounwind alwaysinline {
   %1 = icmp ugt <16 x i8> %a0, %a1
   %2 = select <16 x i1> %1, <16 x i8> %a0, <16 x i8> %a1
   %3 = sub <16 x i8> %2, %a1
   ret <16 x i8> %3
 }
 
-; Note that this is only used for LLVM 8.0+
 define weak_odr <8 x i16> @psubuswx8(<8 x i16> %a0, <8 x i16> %a1) nounwind alwaysinline {
   %1 = icmp ugt <8 x i16> %a0, %a1
   %2 = select <8 x i1> %1, <8 x i16> %a0, <8 x i16> %a1
@@ -125,12 +121,12 @@ define weak_odr <4 x float> @fast_inverse_sqrt_f32x4(<4 x float> %x) nounwind uw
 ; -- A version without stack spills tends to confuse the x86-32 code generator
 ; and cause it to fail via running out of registers.
 define weak_odr void @x86_cpuid_halide(i32* %info) nounwind uwtable {
-  call void asm sideeffect inteldialect "xchg ebx, esi\0A\09mov eax, dword ptr $$0 $0\0A\09mov ecx, dword ptr $$4 $0\0A\09cpuid\0A\09mov dword ptr $$0 $0, eax\0A\09mov dword ptr $$4 $0, ebx\0A\09mov dword ptr $$8 $0, ecx\0A\09mov dword ptr $$12 $0, edx\0A\09xchg ebx, esi", "=*m,~{eax},~{ebx},~{ecx},~{edx},~{esi},~{dirflag},~{fpsr},~{flags}"(i32* %info)
+  call void asm sideeffect inteldialect "xchg ebx, esi\0A\09mov eax, dword ptr $$0 $0\0A\09mov ecx, dword ptr $$4 $0\0A\09cpuid\0A\09mov dword ptr $$0 $0, eax\0A\09mov dword ptr $$4 $0, ebx\0A\09mov dword ptr $$8 $0, ecx\0A\09mov dword ptr $$12 $0, edx\0A\09xchg ebx, esi", "=*m,~{eax},~{ebx},~{ecx},~{edx},~{esi},~{dirflag},~{fpsr},~{flags}"(i32* elementtype(i32) %info)
   ret void
 }
 
 ; Same, but ensure that we save/restore the full 64 bits of rbx/rsi.
 define weak_odr void @x64_cpuid_halide(i32* %info) nounwind uwtable {
-  call void asm sideeffect inteldialect "xchg rbx, rsi\0A\09mov eax, dword ptr $$0 $0\0A\09mov ecx, dword ptr $$4 $0\0A\09cpuid\0A\09mov dword ptr $$0 $0, eax\0A\09mov dword ptr $$4 $0, ebx\0A\09mov dword ptr $$8 $0, ecx\0A\09mov dword ptr $$12 $0, edx\0A\09xchg rbx, rsi", "=*m,~{eax},~{ebx},~{ecx},~{edx},~{esi},~{dirflag},~{fpsr},~{flags}"(i32* %info)
+  call void asm sideeffect inteldialect "xchg rbx, rsi\0A\09mov eax, dword ptr $$0 $0\0A\09mov ecx, dword ptr $$4 $0\0A\09cpuid\0A\09mov dword ptr $$0 $0, eax\0A\09mov dword ptr $$4 $0, ebx\0A\09mov dword ptr $$8 $0, ecx\0A\09mov dword ptr $$12 $0, edx\0A\09xchg rbx, rsi", "=*m,~{eax},~{ebx},~{ecx},~{edx},~{esi},~{dirflag},~{fpsr},~{flags}"(i32* elementtype(i32) %info)
   ret void
 }

tools/CMakeLists.txt

@@ -8,6 +8,9 @@ target_compile_options(build_halide_h PRIVATE $<$<CXX_COMPILER_ID:MSVC>:/wd4996>
 add_executable(binary2cpp binary2cpp.cpp)
 target_compile_options(binary2cpp PRIVATE $<$<CXX_COMPILER_ID:MSVC>:/wd4996>)
 
+add_executable(regexp_replace regexp_replace.cpp)
+target_compile_options(regexp_replace PRIVATE $<$<CXX_COMPILER_ID:MSVC>:/wd4996>)
+
 ##
 # Interface target for enabling PNG/JPEG support in Halide
 ##

tools/regexp_replace.cpp

@@ -0,0 +1,25 @@
+#include <cassert>
+#include <cstring>
+#include <iostream>
+#include <regex>
+
+// A utility that does a single regexp-based replace on stdin and dumps it to stdout.
+// Exists solely because we can't rely on (e.g.) `sed` being available in Windows
+// build environments. Usage is basically equivalent to `sed -e 's/regex/replacement/g'`
+// Note that if regex is an empty string, this becomes a simple line-by-line file copy.
+
+int main(int argc, const char **argv) {
+    if (argc != 3) {
+        fprintf(stderr, "Usage: %s regex replacement\n", argv[0]);
+        return -1;
+    }
+    std::regex re(argv[1]);
+
+    std::string line;
+    while (std::getline(std::cin, line)) {
+        std::regex_replace(std::ostreambuf_iterator<char>(std::cout),
+                           line.begin(), line.end(), re, argv[2]);
+        std::cout << std::endl;
+    }
+    return 0;
+}