ac: use new LLVM 8 intrinsic when storing 16-bit values
[mesa.git] / src / amd / common / ac_llvm_build.c
index 2d873d7910248643fb395035d311fef782164c54..73dd0230785b11a23cb3b2fb31454e5e89fe77e0 100644 (file)
@@ -1082,6 +1082,106 @@ LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
        return ac_build_load_custom(ctx, base_ptr, index, true, true, false);
 }
 
+static void
+ac_build_buffer_store_common(struct ac_llvm_context *ctx,
+                            LLVMValueRef rsrc,
+                            LLVMValueRef data,
+                            LLVMValueRef vindex,
+                            LLVMValueRef voffset,
+                            unsigned num_channels,
+                            bool glc,
+                            bool slc,
+                            bool writeonly_memory,
+                            bool use_format)
+{
+       LLVMValueRef args[] = {
+               data,
+               LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
+               vindex ? vindex : ctx->i32_0,
+               voffset,
+               LLVMConstInt(ctx->i1, glc, 0),
+               LLVMConstInt(ctx->i1, slc, 0)
+       };
+       unsigned func = CLAMP(num_channels, 1, 3) - 1;
+
+       const char *type_names[] = {"f32", "v2f32", "v4f32"};
+       char name[256];
+
+       if (use_format) {
+               snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.format.%s",
+                        type_names[func]);
+       } else {
+               snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.%s",
+                        type_names[func]);
+       }
+
+       ac_build_intrinsic(ctx, name, ctx->voidt, args, ARRAY_SIZE(args),
+                          ac_get_store_intr_attribs(writeonly_memory));
+}
+
+static void
+ac_build_llvm8_buffer_store_common(struct ac_llvm_context *ctx,
+                                  LLVMValueRef rsrc,
+                                  LLVMValueRef data,
+                                  LLVMValueRef vindex,
+                                  LLVMValueRef voffset,
+                                  LLVMValueRef soffset,
+                                  unsigned num_channels,
+                                  bool glc,
+                                  bool slc,
+                                  bool writeonly_memory,
+                                  bool use_format,
+                                  bool structurized)
+{
+       LLVMValueRef args[6];
+       int idx = 0;
+       args[idx++] = data;
+       args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
+       if (structurized)
+               args[idx++] = vindex ? vindex : ctx->i32_0;
+       args[idx++] = voffset ? voffset : ctx->i32_0;
+       args[idx++] = soffset ? soffset : ctx->i32_0;
+       args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
+       unsigned func = CLAMP(num_channels, 1, 3) - 1;
+
+       const char *type_names[] = {"f32", "v2f32", "v4f32"};
+       const char *indexing_kind = structurized ? "struct" : "raw";
+       char name[256];
+
+       if (use_format) {
+               snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.format.%s",
+                        indexing_kind, type_names[func]);
+       } else {
+               snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.%s",
+                        indexing_kind, type_names[func]);
+       }
+
+       ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
+                          ac_get_store_intr_attribs(writeonly_memory));
+}
+
+void
+ac_build_buffer_store_format(struct ac_llvm_context *ctx,
+                            LLVMValueRef rsrc,
+                            LLVMValueRef data,
+                            LLVMValueRef vindex,
+                            LLVMValueRef voffset,
+                            unsigned num_channels,
+                            bool glc,
+                            bool writeonly_memory)
+{
+       if (HAVE_LLVM >= 0x800) {
+               ac_build_llvm8_buffer_store_common(ctx, rsrc, data, vindex,
+                                                  voffset, NULL, num_channels,
+                                                  glc, false, writeonly_memory,
+                                                  true, true);
+       } else {
+               ac_build_buffer_store_common(ctx, rsrc, data, vindex, voffset,
+                                            num_channels, glc, false,
+                                            writeonly_memory, true);
+       }
+}
+
 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
  * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
  * or v4i32 (num_channels=3,4).
@@ -1150,9 +1250,7 @@ ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 
                ac_build_intrinsic(ctx, name, ctx->voidt,
                                   args, ARRAY_SIZE(args),
-                                  writeonly_memory ?
-                                  AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY :
-                                  AC_FUNC_ATTR_WRITEONLY);
+                                  ac_get_store_intr_attribs(writeonly_memory));
                return;
        }
 
@@ -1181,9 +1279,7 @@ ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 
        ac_build_intrinsic(ctx, name, ctx->voidt,
                           args, ARRAY_SIZE(args),
-                          writeonly_memory ?
-                                  AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY :
-                                  AC_FUNC_ATTR_WRITEONLY);
+                          ac_get_store_intr_attribs(writeonly_memory));
 }
 
 static LLVMValueRef
@@ -1316,6 +1412,14 @@ ac_build_buffer_load(struct ac_llvm_context *ctx,
                return ac_build_gather_values(ctx, result, num_channels);
        }
 
+       if (HAVE_LLVM >= 0x0800) {
+               return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex,
+                                                        offset, ctx->i32_0,
+                                                        num_channels, glc, slc,
+                                                        can_speculate, false,
+                                                        false);
+       }
+
        return ac_build_buffer_load_common(ctx, rsrc, vindex, offset,
                                           num_channels, glc, slc,
                                           can_speculate, false);
@@ -1494,7 +1598,6 @@ ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
 LLVMValueRef
 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
                            LLVMValueRef rsrc,
-                           LLVMValueRef vindex,
                            LLVMValueRef voffset,
                            LLVMValueRef soffset,
                            LLVMValueRef immoffset,
@@ -1504,12 +1607,160 @@ ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
        unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
        LLVMValueRef res;
 
-       res = ac_build_tbuffer_load(ctx, rsrc, vindex, voffset, soffset,
-                                   immoffset, 1, dfmt, nfmt, glc, false, false, true);
+       res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
+                                       immoffset, 1, dfmt, nfmt, glc, false,
+                                       false);
 
        return LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
 }
 
+static void
+ac_build_llvm8_tbuffer_store(struct ac_llvm_context *ctx,
+                            LLVMValueRef rsrc,
+                            LLVMValueRef vdata,
+                            LLVMValueRef vindex,
+                            LLVMValueRef voffset,
+                            LLVMValueRef soffset,
+                            unsigned num_channels,
+                            unsigned dfmt,
+                            unsigned nfmt,
+                            bool glc,
+                            bool slc,
+                            bool writeonly_memory,
+                            bool structurized)
+{
+       LLVMValueRef args[7];
+       int idx = 0;
+       args[idx++] = vdata;
+       args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
+       if (structurized)
+               args[idx++] = vindex ? vindex : ctx->i32_0;
+       args[idx++] = voffset ? voffset : ctx->i32_0;
+       args[idx++] = soffset ? soffset : ctx->i32_0;
+       args[idx++] = LLVMConstInt(ctx->i32, dfmt | (nfmt << 4), 0);
+       args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
+       unsigned func = CLAMP(num_channels, 1, 3) - 1;
+
+       const char *type_names[] = {"i32", "v2i32", "v4i32"};
+       const char *indexing_kind = structurized ? "struct" : "raw";
+       char name[256];
+
+       snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.store.%s",
+                indexing_kind, type_names[func]);
+
+       ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
+                          ac_get_store_intr_attribs(writeonly_memory));
+}
+
+static void
+ac_build_tbuffer_store(struct ac_llvm_context *ctx,
+                      LLVMValueRef rsrc,
+                      LLVMValueRef vdata,
+                      LLVMValueRef vindex,
+                      LLVMValueRef voffset,
+                      LLVMValueRef soffset,
+                      LLVMValueRef immoffset,
+                      unsigned num_channels,
+                      unsigned dfmt,
+                      unsigned nfmt,
+                      bool glc,
+                      bool slc,
+                      bool writeonly_memory,
+                      bool structurized) /* only matters for LLVM 8+ */
+{
+       if (HAVE_LLVM >= 0x800) {
+               voffset = LLVMBuildAdd(ctx->builder,
+                                      voffset ? voffset : ctx->i32_0,
+                                      immoffset, "");
+
+               ac_build_llvm8_tbuffer_store(ctx, rsrc, vdata, vindex, voffset,
+                                            soffset, num_channels, dfmt, nfmt,
+                                            glc, slc, writeonly_memory,
+                                            structurized);
+       } else {
+               LLVMValueRef params[] = {
+                       vdata,
+                       rsrc,
+                       vindex ? vindex : ctx->i32_0,
+                       voffset ? voffset : ctx->i32_0,
+                       soffset ? soffset : ctx->i32_0,
+                       immoffset,
+                       LLVMConstInt(ctx->i32, dfmt, false),
+                       LLVMConstInt(ctx->i32, nfmt, false),
+                       LLVMConstInt(ctx->i32, glc, false),
+                       LLVMConstInt(ctx->i32, slc, false),
+               };
+               unsigned func = CLAMP(num_channels, 1, 3) - 1;
+               const char *type_names[] = {"i32", "v2i32", "v4i32"};
+               char name[256];
+
+               snprintf(name, sizeof(name), "llvm.amdgcn.tbuffer.store.%s",
+                        type_names[func]);
+
+               ac_build_intrinsic(ctx, name, ctx->voidt, params, 10,
+                                  ac_get_store_intr_attribs(writeonly_memory));
+       }
+}
+
+void
+ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
+                             LLVMValueRef rsrc,
+                             LLVMValueRef vdata,
+                             LLVMValueRef vindex,
+                             LLVMValueRef voffset,
+                             LLVMValueRef soffset,
+                             LLVMValueRef immoffset,
+                             unsigned num_channels,
+                             unsigned dfmt,
+                             unsigned nfmt,
+                             bool glc,
+                             bool slc,
+                             bool writeonly_memory)
+{
+       ac_build_tbuffer_store(ctx, rsrc, vdata, vindex, voffset, soffset,
+                              immoffset, num_channels, dfmt, nfmt, glc, slc,
+                              writeonly_memory, true);
+}
+
+void
+ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
+                          LLVMValueRef rsrc,
+                          LLVMValueRef vdata,
+                          LLVMValueRef voffset,
+                          LLVMValueRef soffset,
+                          LLVMValueRef immoffset,
+                          unsigned num_channels,
+                          unsigned dfmt,
+                          unsigned nfmt,
+                          bool glc,
+                          bool slc,
+                          bool writeonly_memory)
+{
+       ac_build_tbuffer_store(ctx, rsrc, vdata, NULL, voffset, soffset,
+                              immoffset, num_channels, dfmt, nfmt, glc, slc,
+                              writeonly_memory, false);
+}
+
+void
+ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
+                            LLVMValueRef rsrc,
+                            LLVMValueRef vdata,
+                            LLVMValueRef voffset,
+                            LLVMValueRef soffset,
+                            bool glc,
+                            bool writeonly_memory)
+{
+       unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
+       unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
+
+       vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i16, "");
+       vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
+
+       ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
+                                  ctx->i32_0, 1, dfmt, nfmt, glc, false,
+                                  writeonly_memory);
+}
+
 /**
  * Set range metadata on an instruction.  This can only be used on load and
  * call instructions.  If you know an instruction can only produce the values
@@ -2155,19 +2406,18 @@ LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
        char *intr;
 
        if (bitsize == 32) {
-               intr = "llvm.floor.f32";
+               intr = "llvm.amdgcn.fract.f32";
                type = ctx->f32;
        } else {
-               intr = "llvm.floor.f64";
+               intr = "llvm.amdgcn.fract.f64";
                type = ctx->f64;
        }
 
        LLVMValueRef params[] = {
                src0,
        };
-       LLVMValueRef floor = ac_build_intrinsic(ctx, intr, type, params, 1,
-                                               AC_FUNC_ATTR_READNONE);
-       return LLVMBuildFSub(ctx->builder, src0, floor, "");
+       return ac_build_intrinsic(ctx, intr, type, params, 1,
+                                 AC_FUNC_ATTR_READNONE);
 }
 
 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,