[Mesa-dev] [PATCH 08/10] amd/common: add ac_compile_module_to_elf

Nicolai Hähnle nhaehnle at gmail.com
Fri May 3 11:18:27 UTC 2019


From: Nicolai Hähnle <nicolai.haehnle at amd.com>

A new variant of ac_compile_module_to_binary that allows us to
keep the entire ELF around.
---
 src/amd/common/ac_llvm_helper.cpp | 88 ++++++++++++++++++++++++++++---
 src/amd/common/ac_llvm_util.h     |  2 +
 2 files changed, 83 insertions(+), 7 deletions(-)

diff --git a/src/amd/common/ac_llvm_helper.cpp b/src/amd/common/ac_llvm_helper.cpp
index dcfb8008546..834c5d7f94e 100644
--- a/src/amd/common/ac_llvm_helper.cpp
+++ b/src/amd/common/ac_llvm_helper.cpp
@@ -22,23 +22,27 @@
  * of the Software.
  *
  */
 
 /* based on Marek's patch to lp_bld_misc.cpp */
 
 // Workaround http://llvm.org/PR23628
 #pragma push_macro("DEBUG")
 #undef DEBUG
 
+#include <cstring>
+
 #include "ac_binary.h"
 #include "ac_llvm_util.h"
 
+#include "util/macros.h"
+
 #include <llvm-c/Core.h>
 #include <llvm/Target/TargetMachine.h>
 #include <llvm/IR/IRBuilder.h>
 #include <llvm/Analysis/TargetLibraryInfo.h>
 #include <llvm/Transforms/IPO.h>
 
 #include <llvm/IR/LegacyPassManager.h>
 
 void ac_add_attr_dereferenceable(LLVMValueRef val, uint64_t bytes)
 {
@@ -102,28 +106,90 @@ ac_create_target_library_info(const char *triple)
 {
 	return reinterpret_cast<LLVMTargetLibraryInfoRef>(new llvm::TargetLibraryInfoImpl(llvm::Triple(triple)));
 }
 
 void
 ac_dispose_target_library_info(LLVMTargetLibraryInfoRef library_info)
 {
 	delete reinterpret_cast<llvm::TargetLibraryInfoImpl *>(library_info);
 }
 
+/* Implementation of raw_pwrite_stream that works on malloc()ed memory for
+ * better compatibility with C code. */
+struct raw_memory_ostream : public llvm::raw_pwrite_stream {
+	char *buffer;
+	size_t written;
+	size_t bufsize;
+
+	raw_memory_ostream()
+	{
+		buffer = NULL;
+		written = 0;
+		bufsize = 0;
+		SetUnbuffered();
+	}
+
+	~raw_memory_ostream()
+	{
+		free(buffer);
+	}
+
+	void clear()
+	{
+		written = 0;
+	}
+
+	void take(char *&out_buffer, size_t &out_size)
+	{
+		out_buffer = buffer;
+		out_size = written;
+		buffer = NULL;
+		written = 0;
+		bufsize = 0;
+	}
+
+	void flush() = delete;
+
+	void write_impl(const char *ptr, size_t size) override
+	{
+		if (unlikely(written + size < written))
+			abort();
+		if (written + size > bufsize) {
+			bufsize = MAX3(1024, written + size, bufsize / 3 * 4);
+			buffer = (char *)realloc(buffer, bufsize);
+			if (!buffer) {
+				fprintf(stderr, "amd: out of memory allocating ELF buffer\n");
+				abort();
+			}
+		}
+		memcpy(buffer + written, ptr, size);
+		written += size;
+	}
+
+	void pwrite_impl(const char *ptr, size_t size, uint64_t offset) override
+	{
+		assert(offset == (size_t)offset &&
+		       offset + size >= offset && offset + size <= written);
+		memcpy(buffer + offset, ptr, size);
+	}
+
+	uint64_t current_pos() const override
+	{
+		return written;
+	}
+};
+
 /* The LLVM compiler is represented as a pass manager containing passes for
  * optimizations, instruction selection, and code generation.
  */
 struct ac_compiler_passes {
-	ac_compiler_passes(): ostream(code_string) {}
-
-	llvm::SmallString<0> code_string;  /* ELF shader binary */
-	llvm::raw_svector_ostream ostream; /* stream for appending data to the binary */
+	raw_memory_ostream ostream; /* ELF shader binary stream */
 	llvm::legacy::PassManager passmgr; /* list of passes */
 };
 
 struct ac_compiler_passes *ac_create_llvm_passes(LLVMTargetMachineRef tm)
 {
 	struct ac_compiler_passes *p = new ac_compiler_passes();
 	if (!p)
 		return NULL;
 
 	llvm::TargetMachine *TM = reinterpret_cast<llvm::TargetMachine*>(tm);
@@ -142,28 +208,36 @@ void ac_destroy_llvm_passes(struct ac_compiler_passes *p)
 {
 	delete p;
 }
 
 /* This returns false on failure. */
 bool ac_compile_module_to_binary(struct ac_compiler_passes *p, LLVMModuleRef module,
 				 struct ac_shader_binary *binary)
 {
 	p->passmgr.run(*llvm::unwrap(module));
 
-	llvm::StringRef data = p->ostream.str();
-	bool success = ac_elf_read(data.data(), data.size(), binary);
-	p->code_string = ""; /* release the ELF shader binary */
+	bool success = ac_elf_read(p->ostream.buffer, p->ostream.written, binary);
+	p->ostream.clear();
 
 	if (!success)
 		fprintf(stderr, "amd: cannot read an ELF shader binary\n");
 	return success;
 }
 
+/* This returns false on failure. */
+bool ac_compile_module_to_elf(struct ac_compiler_passes *p, LLVMModuleRef module,
+			      char **pelf_buffer, size_t *pelf_size)
+{
+	p->passmgr.run(*llvm::unwrap(module));
+	p->ostream.take(*pelf_buffer, *pelf_size);
+	return true;
+}
+
 void ac_llvm_add_barrier_noop_pass(LLVMPassManagerRef passmgr)
 {
 	llvm::unwrap(passmgr)->add(llvm::createBarrierNoopPass());
 }
 
 void ac_enable_global_isel(LLVMTargetMachineRef tm)
 {
   reinterpret_cast<llvm::TargetMachine*>(tm)->setGlobalISel(true);
 }
diff --git a/src/amd/common/ac_llvm_util.h b/src/amd/common/ac_llvm_util.h
index 6d961c06f8a..7447d349bd5 100644
--- a/src/amd/common/ac_llvm_util.h
+++ b/src/amd/common/ac_llvm_util.h
@@ -135,18 +135,20 @@ void ac_init_llvm_once(void);
 
 bool ac_init_llvm_compiler(struct ac_llvm_compiler *compiler,
 			   enum radeon_family family,
 			   enum ac_target_machine_options tm_options);
 void ac_destroy_llvm_compiler(struct ac_llvm_compiler *compiler);
 
 struct ac_compiler_passes *ac_create_llvm_passes(LLVMTargetMachineRef tm);
 void ac_destroy_llvm_passes(struct ac_compiler_passes *p);
 bool ac_compile_module_to_binary(struct ac_compiler_passes *p, LLVMModuleRef module,
 				 struct ac_shader_binary *binary);
+bool ac_compile_module_to_elf(struct ac_compiler_passes *p, LLVMModuleRef module,
+			      char **pelf_buffer, size_t *pelf_size);
 void ac_llvm_add_barrier_noop_pass(LLVMPassManagerRef passmgr);
 void ac_enable_global_isel(LLVMTargetMachineRef tm);
 
 #ifdef __cplusplus
 }
 #endif
 
 #endif /* AC_LLVM_UTIL_H */
-- 
2.20.1



More information about the mesa-dev mailing list