radeonsi: Cache LLVMTargetMachineRef in context instead of in screen

Fixes a crash in genymotion with several threads compiling shaders
concurrently.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89746

Cc: 10.5 <[email protected]>
Reviewed-by: Tom Stellard <[email protected]>

6 files changed, 41 insertions, 30 deletions

diff --git a/src/gallium/drivers/radeonsi/si_compute.c b/src/gallium/drivers/radeonsi/si_compute.c
index 8609b894108..89bef2e7afd 100644
--- a/src/gallium/drivers/radeonsi/si_compute.c
+++ b/src/gallium/drivers/radeonsi/si_compute.c
@@ -130,7 +130,8 @@ static void *si_create_compute_state(
 	        for (i = 0; i < program->num_kernels; i++) {
 		        LLVMModuleRef mod = radeon_llvm_get_kernel_module(program->llvm_ctx, i,
                                                         code, header->num_bytes);
-			si_compile_llvm(sctx->screen, &program->kernels[i], mod);
+			si_compile_llvm(sctx->screen, &program->kernels[i], sctx->tm,
+					mod);
 			LLVMDisposeModule(mod);
 		}
 	}
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index d335bda44f4..0eada72ad5a 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -69,6 +69,11 @@ static void si_destroy_context(struct pipe_context *context)
 	si_pm4_cleanup(sctx);
 
 	r600_common_context_cleanup(&sctx->b);
+
+#if HAVE_LLVM >= 0x0306
+	LLVMDisposeTargetMachine(sctx->tm);
+#endif
+
 	FREE(sctx);
 }
 
@@ -77,6 +82,12 @@ static struct pipe_context *si_create_context(struct pipe_screen *screen, void *
 	struct si_context *sctx = CALLOC_STRUCT(si_context);
 	struct si_screen* sscreen = (struct si_screen *)screen;
 	struct radeon_winsys *ws = sscreen->b.ws;
+	LLVMTargetRef r600_target;
+#if HAVE_LLVM >= 0x0306
+	const char *triple = "amdgcn--";
+#else
+	const char *triple = "r600--";
+#endif
 	int shader, i;
 
 	if (sctx == NULL)
@@ -170,6 +181,17 @@ static struct pipe_context *si_create_context(struct pipe_screen *screen, void *
 	 */
 	sctx->scratch_waves = 32 * sscreen->b.info.max_compute_units;
 
+#if HAVE_LLVM >= 0x0306
+	/* Initialize LLVM TargetMachine */
+	r600_target = radeon_llvm_get_r600_target(triple);
+	sctx->tm = LLVMCreateTargetMachine(r600_target, triple,
+					   r600_get_llvm_processor_name(sscreen->b.family),
+					   "+DumpCode,+vgpr-spilling",
+					   LLVMCodeGenLevelDefault,
+					   LLVMRelocDefault,
+					   LLVMCodeModelDefault);
+#endif
+
 	return &sctx->b.b;
 fail:
 	si_destroy_context(&sctx->b.b);
@@ -445,12 +467,6 @@ static void si_destroy_screen(struct pipe_screen* pscreen)
 	if (!sscreen->b.ws->unref(sscreen->b.ws))
 		return;
 
-#if HAVE_LLVM >= 0x0306
-	// r600_destroy_common_screen() frees sscreen, so we need to make
-	// sure to dispose the TargetMachine before we call it.
-	LLVMDisposeTargetMachine(sscreen->tm);
-#endif
-
 	r600_destroy_common_screen(&sscreen->b);
 }
 
@@ -508,12 +524,7 @@ static bool si_initialize_pipe_config(struct si_screen *sscreen)
 struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws)
 {
 	struct si_screen *sscreen = CALLOC_STRUCT(si_screen);
-	LLVMTargetRef r600_target;
-#if HAVE_LLVM >= 0x0306
-	const char *triple = "amdgcn--";
-#else
-	const char *triple = "r600--";
-#endif
+
 	if (sscreen == NULL) {
 		return NULL;
 	}
@@ -541,13 +552,5 @@ struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws)
 	/* Create the auxiliary context. This must be done last. */
 	sscreen->b.aux_context = sscreen->b.b.context_create(&sscreen->b.b, NULL);
 
-#if HAVE_LLVM >= 0x0306
-	/* Initialize LLVM TargetMachine */
-	r600_target = radeon_llvm_get_r600_target(triple);
-	sscreen->tm = LLVMCreateTargetMachine(r600_target, triple,
-				r600_get_llvm_processor_name(sscreen->b.family),
-				"+DumpCode,+vgpr-spilling", LLVMCodeGenLevelDefault, LLVMRelocDefault,
-				LLVMCodeModelDefault);
-#endif
 	return &sscreen->b.b;
 }
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 422b873db04..f98c7a83744 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -85,7 +85,6 @@ struct si_compute;
 
 struct si_screen {
 	struct r600_common_screen	b;
-	LLVMTargetMachineRef		tm;
 };
 
 struct si_sampler_view {
@@ -203,6 +202,8 @@ struct si_context {
 	struct pipe_resource	*esgs_ring;
 	struct pipe_resource	*gsvs_ring;
 
+	LLVMTargetMachineRef		tm;
+
 	/* SI state handling */
 	union si_state	queued;
 	union si_state	emitted;
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 4dcf7566af0..b4709acc465 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -72,6 +72,7 @@ struct si_shader_context
 	int param_streamout_offset[4];
 	int param_vertex_id;
 	int param_instance_id;
+	LLVMTargetMachineRef tm;
 	LLVMValueRef const_md;
 	LLVMValueRef const_resource[SI_NUM_CONST_BUFFERS];
 	LLVMValueRef ddxy_lds;
@@ -2697,13 +2698,13 @@ int si_shader_binary_read(struct si_screen *sscreen,
 }
 
 int si_compile_llvm(struct si_screen *sscreen, struct si_shader *shader,
-							LLVMModuleRef mod)
+		    LLVMTargetMachineRef tm, LLVMModuleRef mod)
 {
 	int r = 0;
 	bool dump = r600_can_dump_shader(&sscreen->b,
 			shader->selector ? shader->selector->tokens : NULL);
 	r = radeon_llvm_compile(mod, &shader->binary,
-		r600_get_llvm_processor_name(sscreen->b.family), dump, sscreen->tm);
+		r600_get_llvm_processor_name(sscreen->b.family), dump, tm);
 
 	if (r) {
 		return r;
@@ -2791,7 +2792,7 @@ static int si_generate_gs_copy_shader(struct si_screen *sscreen,
 		fprintf(stderr, "Copy Vertex Shader for Geometry Shader:\n\n");
 
 	r = si_compile_llvm(sscreen, si_shader_ctx->shader,
-			    bld_base->base.gallivm->module);
+			    si_shader_ctx->tm, bld_base->base.gallivm->module);
 
 	radeon_llvm_dispose(&si_shader_ctx->radeon_bld);
 
@@ -2836,7 +2837,8 @@ static void si_dump_key(unsigned shader, union si_shader_key *key)
 	}
 }
 
-int si_shader_create(struct si_screen *sscreen, struct si_shader *shader)
+int si_shader_create(struct si_screen *sscreen, LLVMTargetMachineRef tm,
+		     struct si_shader *shader)
 {
 	struct si_shader_selector *sel = shader->selector;
 	struct tgsi_token *tokens = sel->tokens;
@@ -2909,6 +2911,7 @@ int si_shader_create(struct si_screen *sscreen, struct si_shader *shader)
 	si_shader_ctx.shader = shader;
 	si_shader_ctx.type = tgsi_get_processor_type(tokens);
 	si_shader_ctx.screen = sscreen;
+	si_shader_ctx.tm = tm;
 
 	switch (si_shader_ctx.type) {
 	case TGSI_PROCESSOR_VERTEX:
@@ -2964,7 +2967,7 @@ int si_shader_create(struct si_screen *sscreen, struct si_shader *shader)
 	radeon_llvm_finalize_module(&si_shader_ctx.radeon_bld);
 
 	mod = bld_base->base.gallivm->module;
-	r = si_compile_llvm(sscreen, shader, mod);
+	r = si_compile_llvm(sscreen, shader, tm, mod);
 	if (r) {
 		fprintf(stderr, "LLVM failed to compile shader\n");
 		goto out;
diff --git a/src/gallium/drivers/radeonsi/si_shader.h b/src/gallium/drivers/radeonsi/si_shader.h
index 5b602ac3428..51055afe36a 100644
--- a/src/gallium/drivers/radeonsi/si_shader.h
+++ b/src/gallium/drivers/radeonsi/si_shader.h
@@ -183,9 +183,10 @@ static inline struct si_shader* si_get_vs_state(struct si_context *sctx)
 }
 
 /* radeonsi_shader.c */
-int si_shader_create(struct si_screen *sscreen, struct si_shader *shader);
+int si_shader_create(struct si_screen *sscreen, LLVMTargetMachineRef tm,
+		     struct si_shader *shader);
 int si_compile_llvm(struct si_screen *sscreen, struct si_shader *shader,
-		    LLVMModuleRef mod);
+		    LLVMTargetMachineRef tm, LLVMModuleRef mod);
 void si_shader_destroy(struct pipe_context *ctx, struct si_shader *shader);
 unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index);
 int si_shader_binary_read(struct si_screen *sscreen, struct si_shader *shader,
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index 382738a2121..b0a6fb9fc84 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -405,6 +405,7 @@ static INLINE void si_shader_selector_key(struct pipe_context *ctx,
 static int si_shader_select(struct pipe_context *ctx,
 			    struct si_shader_selector *sel)
 {
+	struct si_context *sctx = (struct si_context *)ctx;
 	union si_shader_key key;
 	struct si_shader * shader = NULL;
 	int r;
@@ -444,7 +445,8 @@ static int si_shader_select(struct pipe_context *ctx,
 
 		shader->next_variant = sel->current;
 		sel->current = shader;
-		r = si_shader_create((struct si_screen*)ctx->screen, shader);
+		r = si_shader_create((struct si_screen*)ctx->screen, sctx->tm,
+				     shader);
 		if (unlikely(r)) {
 			R600_ERR("Failed to build shader variant (type=%u) %d\n",
 				 sel->type, r);