radeonsi: implement 32-bit pointers in user data SGPRs (v2)
[mesa.git] / src / amd / common / ac_llvm_build.h
index 65a9a055a72978c0af8f218ae96e7f913b7dff47..0a49ad8ca13554ed8baccef94389513e708d01ef 100644 (file)
 #include <stdbool.h>
 #include <llvm-c/TargetMachine.h>
 
+#include "amd_family.h"
+
 #ifdef __cplusplus
 extern "C" {
 #endif
 
+#define HAVE_32BIT_POINTERS (HAVE_LLVM >= 0x0700)
+
+enum {
+       /* CONST is the only address space that selects SMEM loads */
+       AC_CONST_ADDR_SPACE = HAVE_LLVM >= 0x700 ? 4 : 2,
+       AC_LOCAL_ADDR_SPACE = 3,
+       AC_CONST_32BIT_ADDR_SPACE = 6, /* same as CONST, but the pointer type has 32 bits */
+};
+
 struct ac_llvm_context {
        LLVMContextRef context;
        LLVMModuleRef module;
@@ -40,11 +51,31 @@ struct ac_llvm_context {
        LLVMTypeRef voidt;
        LLVMTypeRef i1;
        LLVMTypeRef i8;
+       LLVMTypeRef i16;
        LLVMTypeRef i32;
+       LLVMTypeRef i64;
+       LLVMTypeRef intptr;
+       LLVMTypeRef f16;
        LLVMTypeRef f32;
+       LLVMTypeRef f64;
+       LLVMTypeRef v2i16;
+       LLVMTypeRef v2i32;
+       LLVMTypeRef v3i32;
        LLVMTypeRef v4i32;
+       LLVMTypeRef v2f32;
        LLVMTypeRef v4f32;
-       LLVMTypeRef v16i8;
+       LLVMTypeRef v8i32;
+
+       LLVMValueRef i32_0;
+       LLVMValueRef i32_1;
+       LLVMValueRef i64_0;
+       LLVMValueRef i64_1;
+       LLVMValueRef f32_0;
+       LLVMValueRef f32_1;
+       LLVMValueRef f64_0;
+       LLVMValueRef f64_1;
+       LLVMValueRef i1true;
+       LLVMValueRef i1false;
 
        unsigned range_md_kind;
        unsigned invariant_load_md_kind;
@@ -52,37 +83,87 @@ struct ac_llvm_context {
        unsigned fpmath_md_kind;
        LLVMValueRef fpmath_md_2p5_ulp;
        LLVMValueRef empty_md;
+
+       enum chip_class chip_class;
+       enum radeon_family family;
+
+       LLVMValueRef lds;
 };
 
 void
-ac_llvm_context_init(struct ac_llvm_context *ctx, LLVMContextRef context);
+ac_llvm_context_init(struct ac_llvm_context *ctx, LLVMContextRef context,
+                    enum chip_class chip_class, enum radeon_family family);
+
+int
+ac_get_llvm_num_components(LLVMValueRef value);
+
+int
+ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type);
 
 LLVMValueRef
-ac_emit_llvm_intrinsic(struct ac_llvm_context *ctx, const char *name,
-                      LLVMTypeRef return_type, LLVMValueRef *params,
-                      unsigned param_count, unsigned attrib_mask);
+ac_llvm_extract_elem(struct ac_llvm_context *ac,
+                    LLVMValueRef value,
+                    int index);
+
+unsigned ac_get_type_size(LLVMTypeRef type);
+
+LLVMTypeRef ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
+LLVMValueRef ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v);
+LLVMTypeRef ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
+LLVMValueRef ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v);
+
+LLVMValueRef
+ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
+                  LLVMTypeRef return_type, LLVMValueRef *params,
+                  unsigned param_count, unsigned attrib_mask);
 
 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize);
 
+LLVMValueRef
+ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
+            unsigned count_incoming, LLVMValueRef *values,
+            LLVMBasicBlockRef *blocks);
+
+void ac_build_optimization_barrier(struct ac_llvm_context *ctx,
+                                  LLVMValueRef *pvgpr);
+
+LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx);
+
+LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value);
+
+LLVMValueRef ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value);
+
+LLVMValueRef ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value);
+
+LLVMValueRef ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value);
+
+LLVMValueRef
+ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
+                              unsigned value_count, unsigned component);
+
 LLVMValueRef
 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
                                LLVMValueRef *values,
                                unsigned value_count,
                                unsigned value_stride,
-                               bool load);
+                               bool load,
+                               bool always_vector);
 LLVMValueRef
 ac_build_gather_values(struct ac_llvm_context *ctx,
                       LLVMValueRef *values,
                       unsigned value_count);
+LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
+                                    LLVMValueRef value,
+                                    unsigned num_channels);
 
 LLVMValueRef
-ac_emit_fdiv(struct ac_llvm_context *ctx,
-            LLVMValueRef num,
-            LLVMValueRef den);
+ac_build_fdiv(struct ac_llvm_context *ctx,
+             LLVMValueRef num,
+             LLVMValueRef den);
 
 void
 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
-                      bool is_deriv, bool is_array,
+                      bool is_deriv, bool is_array, bool is_lod,
                       LLVMValueRef *coords_arg,
                       LLVMValueRef *derivs_arg);
 
@@ -112,26 +193,25 @@ ac_build_indexed_store(struct ac_llvm_context *ctx,
                       LLVMValueRef base_ptr, LLVMValueRef index,
                       LLVMValueRef value);
 
-LLVMValueRef
-ac_build_indexed_load(struct ac_llvm_context *ctx,
-                     LLVMValueRef base_ptr, LLVMValueRef index,
-                     bool uniform);
-
-LLVMValueRef
-ac_build_indexed_load_const(struct ac_llvm_context *ctx,
-                           LLVMValueRef base_ptr, LLVMValueRef index);
+LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
+                          LLVMValueRef index);
+LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
+                                    LLVMValueRef base_ptr, LLVMValueRef index);
+LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
+                                  LLVMValueRef base_ptr, LLVMValueRef index);
 
 void
 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
                            LLVMValueRef rsrc,
                            LLVMValueRef vdata,
                            unsigned num_channels,
-                           LLVMValueRef vaddr,
+                           LLVMValueRef voffset,
                            LLVMValueRef soffset,
                            unsigned inst_offset,
-                           bool offen,
                            bool glc,
-                           bool slc);
+                           bool slc,
+                           bool writeonly_memory,
+                           bool swizzle_enable_hint);
 LLVMValueRef
 ac_build_buffer_load(struct ac_llvm_context *ctx,
                     LLVMValueRef rsrc,
@@ -142,7 +222,16 @@ ac_build_buffer_load(struct ac_llvm_context *ctx,
                     unsigned inst_offset,
                     unsigned glc,
                     unsigned slc,
-                    bool readonly_memory);
+                    bool can_speculate,
+                    bool allow_smem);
+
+LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
+                                        LLVMValueRef rsrc,
+                                        LLVMValueRef vindex,
+                                        LLVMValueRef voffset,
+                                        unsigned num_channels,
+                                        bool glc,
+                                        bool can_speculate);
 
 LLVMValueRef
 ac_get_thread_id(struct ac_llvm_context *ctx);
@@ -152,12 +241,10 @@ ac_get_thread_id(struct ac_llvm_context *ctx);
 #define AC_TID_MASK_LEFT     0xfffffffe
 
 LLVMValueRef
-ac_emit_ddxy(struct ac_llvm_context *ctx,
-            bool has_ds_bpermute,
-            uint32_t mask,
-            int idx,
-            LLVMValueRef lds,
-            LLVMValueRef val);
+ac_build_ddxy(struct ac_llvm_context *ctx,
+             uint32_t mask,
+             int idx,
+             LLVMValueRef val);
 
 #define AC_SENDMSG_GS 2
 #define AC_SENDMSG_GS_DONE 3
@@ -167,19 +254,27 @@ ac_emit_ddxy(struct ac_llvm_context *ctx,
 #define AC_SENDMSG_GS_OP_EMIT     (2 << 4)
 #define AC_SENDMSG_GS_OP_EMIT_CUT (3 << 4)
 
-void ac_emit_sendmsg(struct ac_llvm_context *ctx,
-                    uint32_t msg,
-                    LLVMValueRef wave_id);
+void ac_build_sendmsg(struct ac_llvm_context *ctx,
+                     uint32_t msg,
+                     LLVMValueRef wave_id);
 
-LLVMValueRef ac_emit_imsb(struct ac_llvm_context *ctx,
-                         LLVMValueRef arg,
-                         LLVMTypeRef dst_type);
+LLVMValueRef ac_build_imsb(struct ac_llvm_context *ctx,
+                          LLVMValueRef arg,
+                          LLVMTypeRef dst_type);
 
-LLVMValueRef ac_emit_umsb(struct ac_llvm_context *ctx,
+LLVMValueRef ac_build_umsb(struct ac_llvm_context *ctx,
                          LLVMValueRef arg,
                          LLVMTypeRef dst_type);
-
-LLVMValueRef ac_emit_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
+LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
+                          LLVMValueRef b);
+LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
+                          LLVMValueRef b);
+LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
+                          LLVMValueRef b);
+LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
+                          LLVMValueRef b);
+LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
+LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
 
 struct ac_export_args {
        LLVMValueRef out[4];
@@ -190,7 +285,9 @@ struct ac_export_args {
         bool valid_mask;
 };
 
-void ac_emit_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
+void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a);
+
+void ac_build_export_null(struct ac_llvm_context *ctx);
 
 enum ac_image_opcode {
        ac_image_sample,
@@ -218,14 +315,51 @@ struct ac_image_args {
        bool da;
 };
 
-LLVMValueRef ac_emit_image_opcode(struct ac_llvm_context *ctx,
-                                 struct ac_image_args *a);
-LLVMValueRef ac_emit_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
-                                  LLVMValueRef args[2]);
-void ac_emit_kill(struct ac_llvm_context *ctx, LLVMValueRef value);
-LLVMValueRef ac_emit_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
-                        LLVMValueRef offset, LLVMValueRef width,
-                        bool is_signed);
+LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
+                                  struct ac_image_args *a);
+LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
+                                   LLVMValueRef args[2]);
+LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
+                                    LLVMValueRef args[2]);
+LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
+                                    LLVMValueRef args[2]);
+LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
+                                LLVMValueRef args[2], unsigned bits, bool hi);
+LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
+                                LLVMValueRef args[2], unsigned bits, bool hi);
+LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1);
+void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1);
+LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
+                         LLVMValueRef offset, LLVMValueRef width,
+                         bool is_signed);
+
+void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16);
+
+void ac_get_image_intr_name(const char *base_name,
+                           LLVMTypeRef data_type,
+                           LLVMTypeRef coords_type,
+                           LLVMTypeRef rsrc_type,
+                           char *out_name, unsigned out_len);
+
+void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
+                           LLVMValueRef main_fn,
+                           uint8_t *vs_output_param_offset,
+                           uint32_t num_outputs,
+                           uint8_t *num_param_exports);
+void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
+
+void ac_declare_lds_as_pointer(struct ac_llvm_context *ac);
+LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
+                        LLVMValueRef dw_addr);
+void ac_lds_store(struct ac_llvm_context *ctx,
+                 LLVMValueRef dw_addr, LLVMValueRef value);
+
+LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
+                        LLVMTypeRef dst_type,
+                        LLVMValueRef src0);
+
+LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type);
+LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type);
 
 #ifdef __cplusplus
 }