*/
void
ac_llvm_context_init(struct ac_llvm_context *ctx,
- enum chip_class chip_class, enum radeon_family family)
+ struct ac_llvm_compiler *compiler,
+ enum chip_class chip_class, enum radeon_family family,
+ enum ac_float_mode float_mode, unsigned wave_size,
+ unsigned ballot_mask_bits)
{
LLVMValueRef args[1];
ctx->chip_class = chip_class;
ctx->family = family;
- ctx->module = NULL;
- ctx->builder = NULL;
+ ctx->wave_size = wave_size;
+ ctx->ballot_mask_bits = ballot_mask_bits;
+ ctx->module = ac_create_module(wave_size == 32 ? compiler->tm_wave32
+ : compiler->tm,
+ ctx->context);
+ ctx->builder = ac_create_builder(ctx->context, float_mode);
ctx->voidt = LLVMVoidTypeInContext(ctx->context);
ctx->i1 = LLVMInt1TypeInContext(ctx->context);
ctx->v3f32 = LLVMVectorType(ctx->f32, 3);
ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
+ ctx->iN_wavemask = LLVMIntTypeInContext(ctx->context, ctx->wave_size);
+ ctx->iN_ballotmask = LLVMIntTypeInContext(ctx->context, ballot_mask_bits);
ctx->i8_0 = LLVMConstInt(ctx->i8, 0, false);
ctx->i8_1 = LLVMConstInt(ctx->i8, 1, false);
"amdgpu.uniform", 14);
ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
+ ctx->flow = calloc(1, sizeof(*ctx->flow));
}
void
ac_llvm_context_dispose(struct ac_llvm_context *ctx)
{
+ free(ctx->flow->stack);
free(ctx->flow);
ctx->flow = NULL;
- ctx->flow_depth_max = 0;
}
int
char *type_name = LLVMPrintTypeToString(type);
fprintf(stderr, "Error building type name for: %s\n",
type_name);
+ LLVMDisposeMessage(type_name);
return;
}
elem_type = LLVMGetElementType(type);
ac_build_ballot(struct ac_llvm_context *ctx,
LLVMValueRef value)
{
- const char *name = HAVE_LLVM >= 0x900 ? "llvm.amdgcn.icmp.i64.i32" : "llvm.amdgcn.icmp.i32";
+ const char *name;
+
+ if (HAVE_LLVM >= 0x900) {
+ if (ctx->wave_size == 64)
+ name = "llvm.amdgcn.icmp.i64.i32";
+ else
+ name = "llvm.amdgcn.icmp.i32.i32";
+ } else {
+ name = "llvm.amdgcn.icmp.i32";
+ }
LLVMValueRef args[3] = {
value,
ctx->i32_0,
args[0] = ac_to_integer(ctx, args[0]);
- return ac_build_intrinsic(ctx, name,
- ctx->i64, args, 3,
+ return ac_build_intrinsic(ctx, name, ctx->iN_wavemask, args, 3,
AC_FUNC_ATTR_NOUNWIND |
AC_FUNC_ATTR_READNONE |
AC_FUNC_ATTR_CONVERGENT);
LLVMConstInt(ctx->i32, LLVMIntNE, 0),
};
- assert(HAVE_LLVM >= 0x0800);
return ac_build_intrinsic(ctx, name, ctx->i64, args, 3,
AC_FUNC_ATTR_NOUNWIND |
AC_FUNC_ATTR_READNONE |
{
LLVMValueRef vote_set = ac_build_ballot(ctx, value);
return LLVMBuildICmp(ctx->builder, LLVMIntNE, vote_set,
- LLVMConstInt(ctx->i64, 0, 0), "");
+ LLVMConstInt(ctx->iN_wavemask, 0, 0), "");
}
LLVMValueRef
vote_set, active_set, "");
LLVMValueRef none = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
vote_set,
- LLVMConstInt(ctx->i64, 0, 0), "");
+ LLVMConstInt(ctx->iN_wavemask, 0, 0), "");
return LLVMBuildOr(ctx->builder, all, none, "");
}
return ac_build_gather_values(ctx, chan, dst_channels);
}
+/* Extract components [start, start + channels) from a vector.
+ */
+LLVMValueRef
+ac_extract_components(struct ac_llvm_context *ctx,
+ LLVMValueRef value,
+ unsigned start,
+ unsigned channels)
+{
+ LLVMValueRef chan[channels];
+
+ for (unsigned i = 0; i < channels; i++)
+ chan[i] = ac_llvm_extract_elem(ctx, value, i + start);
+
+ return ac_build_gather_values(ctx, chan, channels);
+}
+
/* Expand a scalar or vector to <4 x type> by filling the remaining channels
* with undef. Extract at most num_channels components from the input.
*/
}
static void
-ac_build_llvm7_buffer_store_common(struct ac_llvm_context *ctx,
- LLVMValueRef rsrc,
- LLVMValueRef data,
- LLVMValueRef vindex,
- LLVMValueRef voffset,
- unsigned num_channels,
- unsigned cache_policy,
- bool use_format)
-{
- LLVMValueRef args[] = {
- data,
- LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
- vindex ? vindex : ctx->i32_0,
- voffset,
- LLVMConstInt(ctx->i1, !!(cache_policy & ac_glc), 0),
- LLVMConstInt(ctx->i1, !!(cache_policy & ac_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_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
-}
-
-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,
- LLVMTypeRef return_channel_type,
- unsigned cache_policy,
- bool use_format,
- bool structurized)
+ac_build_buffer_store_common(struct ac_llvm_context *ctx,
+ LLVMValueRef rsrc,
+ LLVMValueRef data,
+ LLVMValueRef vindex,
+ LLVMValueRef voffset,
+ LLVMValueRef soffset,
+ unsigned num_channels,
+ LLVMTypeRef return_channel_type,
+ unsigned cache_policy,
+ bool use_format,
+ bool structurized)
{
LLVMValueRef args[6];
int idx = 0;
unsigned num_channels,
unsigned cache_policy)
{
- if (HAVE_LLVM >= 0x800) {
- ac_build_llvm8_buffer_store_common(ctx, rsrc, data, vindex,
- voffset, NULL, num_channels,
- ctx->f32, cache_policy,
- true, true);
- } else {
- ac_build_llvm7_buffer_store_common(ctx, rsrc, data, vindex, voffset,
- num_channels, cache_policy,
- true);
- }
+ ac_build_buffer_store_common(ctx, rsrc, data, vindex,
+ voffset, NULL, num_channels,
+ ctx->f32, cache_policy,
+ true, true);
}
/* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
offset = LLVMBuildAdd(ctx->builder, offset,
LLVMConstInt(ctx->i32, inst_offset, 0), "");
- if (HAVE_LLVM >= 0x800) {
- ac_build_llvm8_buffer_store_common(ctx, rsrc,
- ac_to_float(ctx, vdata),
- ctx->i32_0,
- voffset, offset,
- num_channels,
- ctx->f32,
- cache_policy,
- false, false);
- } else {
- if (voffset)
- offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
-
- ac_build_llvm7_buffer_store_common(ctx, rsrc,
- ac_to_float(ctx, vdata),
- ctx->i32_0, offset,
- num_channels, cache_policy,
- false);
- }
+ ac_build_buffer_store_common(ctx, rsrc, ac_to_float(ctx, vdata),
+ ctx->i32_0, voffset, offset,
+ num_channels, ctx->f32,
+ cache_policy, false, false);
return;
}
}
static LLVMValueRef
-ac_build_llvm7_buffer_load_common(struct ac_llvm_context *ctx,
- LLVMValueRef rsrc,
- LLVMValueRef vindex,
- LLVMValueRef voffset,
- unsigned num_channels,
- unsigned cache_policy,
- bool can_speculate,
- bool use_format)
-{
- LLVMValueRef args[] = {
- LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
- vindex ? vindex : ctx->i32_0,
- voffset,
- LLVMConstInt(ctx->i1, !!(cache_policy & ac_glc), 0),
- LLVMConstInt(ctx->i1, !!(cache_policy & ac_slc), 0)
- };
- unsigned func = CLAMP(num_channels, 1, 3) - 1;
-
- LLVMTypeRef types[] = {ctx->f32, ctx->v2f32, ctx->v4f32};
- const char *type_names[] = {"f32", "v2f32", "v4f32"};
- char name[256];
-
- if (use_format) {
- snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.format.%s",
- type_names[func]);
- } else {
- snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.%s",
- type_names[func]);
- }
-
- return ac_build_intrinsic(ctx, name, types[func], args,
- ARRAY_SIZE(args),
- ac_get_load_intr_attribs(can_speculate));
-}
-
-static LLVMValueRef
-ac_build_llvm8_buffer_load_common(struct ac_llvm_context *ctx,
- LLVMValueRef rsrc,
- LLVMValueRef vindex,
- LLVMValueRef voffset,
- LLVMValueRef soffset,
- unsigned num_channels,
- LLVMTypeRef channel_type,
- unsigned cache_policy,
- bool can_speculate,
- bool use_format,
- bool structurized)
+ac_build_buffer_load_common(struct ac_llvm_context *ctx,
+ LLVMValueRef rsrc,
+ LLVMValueRef vindex,
+ LLVMValueRef voffset,
+ LLVMValueRef soffset,
+ unsigned num_channels,
+ LLVMTypeRef channel_type,
+ unsigned cache_policy,
+ bool can_speculate,
+ bool use_format,
+ bool structurized)
{
LLVMValueRef args[5];
int idx = 0;
offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
if (allow_smem && !(cache_policy & ac_slc) &&
- (!(cache_policy & ac_glc) || (HAVE_LLVM >= 0x0800 && ctx->chip_class >= GFX8))) {
+ (!(cache_policy & ac_glc) || ctx->chip_class >= GFX8)) {
assert(vindex == NULL);
LLVMValueRef result[8];
offset = LLVMBuildAdd(ctx->builder, offset,
LLVMConstInt(ctx->i32, 4, 0), "");
}
- const char *intrname =
- HAVE_LLVM >= 0x0800 ? "llvm.amdgcn.s.buffer.load.f32"
- : "llvm.SI.load.const.v4i32";
- unsigned num_args = HAVE_LLVM >= 0x0800 ? 3 : 2;
LLVMValueRef args[3] = {
rsrc,
offset,
LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0),
};
- result[i] = ac_build_intrinsic(ctx, intrname,
- ctx->f32, args, num_args,
- AC_FUNC_ATTR_READNONE |
- (HAVE_LLVM < 0x0800 ? AC_FUNC_ATTR_LEGACY : 0));
+ result[i] = ac_build_intrinsic(ctx,
+ "llvm.amdgcn.s.buffer.load.f32",
+ ctx->f32, args, 3,
+ AC_FUNC_ATTR_READNONE);
}
if (num_channels == 1)
return result[0];
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, ctx->f32,
- cache_policy,
- can_speculate, false,
- false);
- }
-
- return ac_build_llvm7_buffer_load_common(ctx, rsrc, vindex, offset,
- num_channels, cache_policy,
- can_speculate, false);
+ return ac_build_buffer_load_common(ctx, rsrc, vindex,
+ offset, ctx->i32_0,
+ num_channels, ctx->f32,
+ cache_policy,
+ can_speculate, false, false);
}
LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
unsigned cache_policy,
bool can_speculate)
{
- if (HAVE_LLVM >= 0x800) {
- return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
- num_channels, ctx->f32,
- cache_policy, can_speculate, true, true);
- }
- return ac_build_llvm7_buffer_load_common(ctx, rsrc, vindex, voffset,
- num_channels, cache_policy,
- can_speculate, true);
-}
-
-LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
- LLVMValueRef rsrc,
- LLVMValueRef vindex,
- LLVMValueRef voffset,
- unsigned num_channels,
- unsigned cache_policy,
- bool can_speculate)
-{
- if (HAVE_LLVM >= 0x800) {
- return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
- num_channels, ctx->f32,
- cache_policy, can_speculate, true, true);
- }
-
- LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 2, 0), "");
- LLVMValueRef stride = LLVMBuildExtractElement(ctx->builder, rsrc, ctx->i32_1, "");
- stride = LLVMBuildLShr(ctx->builder, stride, LLVMConstInt(ctx->i32, 16, 0), "");
-
- LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->builder,
- LLVMBuildICmp(ctx->builder, LLVMIntUGT, elem_count, stride, ""),
- elem_count, stride, "");
-
- LLVMValueRef new_rsrc = LLVMBuildInsertElement(ctx->builder, rsrc, new_elem_count,
- LLVMConstInt(ctx->i32, 2, 0), "");
-
- return ac_build_llvm7_buffer_load_common(ctx, new_rsrc, vindex, voffset,
- num_channels, cache_policy,
- can_speculate, true);
+ return ac_build_buffer_load_common(ctx, rsrc, vindex, voffset,
+ ctx->i32_0, num_channels, ctx->f32,
+ cache_policy, can_speculate,
+ true, true);
}
/// Translate a (dfmt, nfmt) pair into a chip-appropriate combined format
unsigned format;
switch (dfmt) {
default: unreachable("bad dfmt");
+ case V_008F0C_BUF_DATA_FORMAT_INVALID: format = V_008F0C_IMG_FORMAT_INVALID; break;
case V_008F0C_BUF_DATA_FORMAT_8: format = V_008F0C_IMG_FORMAT_8_UINT; break;
case V_008F0C_BUF_DATA_FORMAT_8_8: format = V_008F0C_IMG_FORMAT_8_8_UINT; break;
case V_008F0C_BUF_DATA_FORMAT_8_8_8_8: format = V_008F0C_IMG_FORMAT_8_8_8_8_UINT; break;
}
static LLVMValueRef
-ac_build_llvm8_tbuffer_load(struct ac_llvm_context *ctx,
+ac_build_tbuffer_load(struct ac_llvm_context *ctx,
LLVMValueRef rsrc,
LLVMValueRef vindex,
LLVMValueRef voffset,
LLVMValueRef soffset,
+ LLVMValueRef immoffset,
unsigned num_channels,
unsigned dfmt,
unsigned nfmt,
bool can_speculate,
bool structurized)
{
+ voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
+
LLVMValueRef args[6];
int idx = 0;
args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
ac_get_load_intr_attribs(can_speculate));
}
-static LLVMValueRef
-ac_build_tbuffer_load(struct ac_llvm_context *ctx,
- LLVMValueRef rsrc,
- LLVMValueRef vindex,
- LLVMValueRef voffset,
- LLVMValueRef soffset,
- LLVMValueRef immoffset,
- unsigned num_channels,
- unsigned dfmt,
- unsigned nfmt,
- unsigned cache_policy,
- bool can_speculate,
- bool structurized) /* only matters for LLVM 8+ */
-{
- if (HAVE_LLVM >= 0x800) {
- voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
-
- return ac_build_llvm8_tbuffer_load(ctx, rsrc, vindex, voffset,
- soffset, num_channels,
- dfmt, nfmt, cache_policy,
- can_speculate, structurized);
- }
-
- LLVMValueRef args[] = {
- rsrc,
- vindex ? vindex : ctx->i32_0,
- voffset,
- soffset,
- immoffset,
- LLVMConstInt(ctx->i32, dfmt, false),
- LLVMConstInt(ctx->i32, nfmt, false),
- LLVMConstInt(ctx->i1, !!(cache_policy & ac_glc), false),
- LLVMConstInt(ctx->i1, !!(cache_policy & ac_slc), false),
- };
- unsigned func = CLAMP(num_channels, 1, 3) - 1;
- LLVMTypeRef types[] = {ctx->i32, ctx->v2i32, ctx->v4i32};
- const char *type_names[] = {"i32", "v2i32", "v4i32"};
- char name[256];
-
- snprintf(name, sizeof(name), "llvm.amdgcn.tbuffer.load.%s",
- type_names[func]);
-
- return ac_build_intrinsic(ctx, name, types[func], args, 9,
- ac_get_load_intr_attribs(can_speculate));
-}
-
LLVMValueRef
ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
LLVMValueRef rsrc,
voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
/* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
- res = ac_build_llvm8_buffer_load_common(ctx, rsrc, NULL,
- voffset, soffset,
- 1, ctx->i16, cache_policy,
- false, false, false);
+ res = ac_build_buffer_load_common(ctx, rsrc, NULL,
+ voffset, soffset,
+ 1, ctx->i16, cache_policy,
+ false, false, false);
} else {
unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
/* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
- res = ac_build_llvm8_buffer_load_common(ctx, rsrc, NULL,
- voffset, soffset,
- 1, ctx->i8, cache_policy,
- false, false, false);
+ res = ac_build_buffer_load_common(ctx, rsrc, NULL,
+ voffset, soffset,
+ 1, ctx->i8, cache_policy,
+ false, false, false);
} else {
unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
for (unsigned i = 0; i < load_num_channels; ++i) {
tmp = LLVMBuildAdd(ctx->builder, soffset,
LLVMConstInt(ctx->i32, i << load_log_size, false), "");
- if (HAVE_LLVM >= 0x0800) {
- LLVMTypeRef channel_type = load_log_size == 0 ? ctx->i8 :
- load_log_size == 1 ? ctx->i16 : ctx->i32;
- unsigned num_channels = 1 << (MAX2(load_log_size, 2) - 2);
- loads[i] = ac_build_llvm8_buffer_load_common(
- ctx, rsrc, vindex, voffset, tmp,
- num_channels, channel_type, cache_policy,
- can_speculate, false, true);
- } else {
- tmp = LLVMBuildAdd(ctx->builder, voffset, tmp, "");
- loads[i] = ac_build_llvm7_buffer_load_common(
- ctx, rsrc, vindex, tmp,
- 1 << (load_log_size - 2), cache_policy, can_speculate, false);
- }
+ LLVMTypeRef channel_type = load_log_size == 0 ? ctx->i8 :
+ load_log_size == 1 ? ctx->i16 : ctx->i32;
+ unsigned num_channels = 1 << (MAX2(load_log_size, 2) - 2);
+ loads[i] = ac_build_buffer_load_common(
+ ctx, rsrc, vindex, voffset, tmp,
+ num_channels, channel_type, cache_policy,
+ can_speculate, false, true);
if (load_log_size >= 2)
loads[i] = ac_to_integer(ctx, loads[i]);
}
}
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,
- unsigned cache_policy,
- bool structurized)
+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,
+ unsigned cache_policy,
+ bool structurized)
{
+ voffset = LLVMBuildAdd(ctx->builder, voffset ? voffset : ctx->i32_0,
+ immoffset, "");
+
LLVMValueRef args[7];
int idx = 0;
args[idx++] = vdata;
AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
}
-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,
- unsigned cache_policy,
- 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,
- cache_policy, 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->i1, !!(cache_policy & ac_glc), false),
- LLVMConstInt(ctx->i1, !!(cache_policy & ac_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_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
- }
-}
-
void
ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
LLVMValueRef rsrc,
if (HAVE_LLVM >= 0x900) {
/* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
- ac_build_llvm8_buffer_store_common(ctx, rsrc, vdata, NULL,
- voffset, soffset, 1,
- ctx->i16, cache_policy,
- false, false);
+ ac_build_buffer_store_common(ctx, rsrc, vdata, NULL,
+ voffset, soffset, 1,
+ ctx->i16, cache_policy,
+ false, false);
} else {
unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
if (HAVE_LLVM >= 0x900) {
/* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
- ac_build_llvm8_buffer_store_common(ctx, rsrc, vdata, NULL,
- voffset, soffset, 1,
- ctx->i8, cache_policy,
- false, false);
+ ac_build_buffer_store_common(ctx, rsrc, vdata, NULL,
+ voffset, soffset, 1,
+ ctx->i8, cache_policy,
+ false, false);
} else {
unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
"llvm.amdgcn.mbcnt.lo", ctx->i32,
tid_args, 2, AC_FUNC_ATTR_READNONE);
- tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
- ctx->i32, tid_args,
- 2, AC_FUNC_ATTR_READNONE);
- set_range_metadata(ctx, tid, 0, 64);
+ if (ctx->wave_size == 32) {
+ tid = tid_args[1];
+ } else {
+ tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
+ ctx->i32, tid_args,
+ 2, AC_FUNC_ATTR_READNONE);
+ }
+ set_range_metadata(ctx, tid, 0, ctx->wave_size);
return tid;
}
case ac_atomic_and: return "and";
case ac_atomic_or: return "or";
case ac_atomic_xor: return "xor";
+ case ac_atomic_inc_wrap: return "inc";
+ case ac_atomic_dec_wrap: return "dec";
}
unreachable("bad atomic op");
}
width,
};
- LLVMValueRef result = ac_build_intrinsic(ctx,
- is_signed ? "llvm.amdgcn.sbfe.i32" :
- "llvm.amdgcn.ubfe.i32",
- ctx->i32, args, 3,
- AC_FUNC_ATTR_READNONE);
-
- if (HAVE_LLVM < 0x0800) {
- /* FIXME: LLVM 7+ returns incorrect result when count is 0.
- * https://bugs.freedesktop.org/show_bug.cgi?id=107276
- */
- LLVMValueRef zero = ctx->i32_0;
- LLVMValueRef icond = LLVMBuildICmp(ctx->builder, LLVMIntEQ, width, zero, "");
- result = LLVMBuildSelect(ctx->builder, icond, zero, result, "");
- }
+ return ac_build_intrinsic(ctx, is_signed ? "llvm.amdgcn.sbfe.i32" :
+ "llvm.amdgcn.ubfe.i32",
+ ctx->i32, args, 3, AC_FUNC_ATTR_READNONE);
- return result;
}
LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
static struct ac_llvm_flow *
get_current_flow(struct ac_llvm_context *ctx)
{
- if (ctx->flow_depth > 0)
- return &ctx->flow[ctx->flow_depth - 1];
+ if (ctx->flow->depth > 0)
+ return &ctx->flow->stack[ctx->flow->depth - 1];
return NULL;
}
static struct ac_llvm_flow *
get_innermost_loop(struct ac_llvm_context *ctx)
{
- for (unsigned i = ctx->flow_depth; i > 0; --i) {
- if (ctx->flow[i - 1].loop_entry_block)
- return &ctx->flow[i - 1];
+ for (unsigned i = ctx->flow->depth; i > 0; --i) {
+ if (ctx->flow->stack[i - 1].loop_entry_block)
+ return &ctx->flow->stack[i - 1];
}
return NULL;
}
{
struct ac_llvm_flow *flow;
- if (ctx->flow_depth >= ctx->flow_depth_max) {
- unsigned new_max = MAX2(ctx->flow_depth << 1,
+ if (ctx->flow->depth >= ctx->flow->depth_max) {
+ unsigned new_max = MAX2(ctx->flow->depth << 1,
AC_LLVM_INITIAL_CF_DEPTH);
- ctx->flow = realloc(ctx->flow, new_max * sizeof(*ctx->flow));
- ctx->flow_depth_max = new_max;
+ ctx->flow->stack = realloc(ctx->flow->stack, new_max * sizeof(*ctx->flow->stack));
+ ctx->flow->depth_max = new_max;
}
- flow = &ctx->flow[ctx->flow_depth];
- ctx->flow_depth++;
+ flow = &ctx->flow->stack[ctx->flow->depth];
+ ctx->flow->depth++;
flow->next_block = NULL;
flow->loop_entry_block = NULL;
static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx,
const char *name)
{
- assert(ctx->flow_depth >= 1);
+ assert(ctx->flow->depth >= 1);
- if (ctx->flow_depth >= 2) {
- struct ac_llvm_flow *flow = &ctx->flow[ctx->flow_depth - 2];
+ if (ctx->flow->depth >= 2) {
+ struct ac_llvm_flow *flow = &ctx->flow->stack[ctx->flow->depth - 2];
return LLVMInsertBasicBlockInContext(ctx->context,
flow->next_block, name);
LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
set_basicblock_name(current_branch->next_block, "endif", label_id);
- ctx->flow_depth--;
+ ctx->flow->depth--;
}
void ac_build_endloop(struct ac_llvm_context *ctx, int label_id)
LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
set_basicblock_name(current_loop->next_block, "endloop", label_id);
- ctx->flow_depth--;
+ ctx->flow->depth--;
}
void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id)
LLVMConstInt(ctx->i32, i, 0), "");
}
}
+ if (LLVMGetTypeKind(src_type) == LLVMPointerTypeKind)
+ return LLVMBuildIntToPtr(ctx->builder, ret, src_type, "");
return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
}
LLVMValueRef
ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane)
{
- /* TODO: Use the actual instruction when LLVM adds an intrinsic for it.
- */
- LLVMValueRef pred = LLVMBuildICmp(ctx->builder, LLVMIntEQ, lane,
- ac_get_thread_id(ctx), "");
- return LLVMBuildSelect(ctx->builder, pred, value, src, "");
+ return ac_build_intrinsic(ctx, "llvm.amdgcn.writelane", ctx->i32,
+ (LLVMValueRef []) {value, lane, src}, 3,
+ AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
}
LLVMValueRef
ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask)
{
+ if (ctx->wave_size == 32) {
+ return ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
+ (LLVMValueRef []) { mask, ctx->i32_0 },
+ 2, AC_FUNC_ATTR_READNONE);
+ }
LLVMValueRef mask_vec = LLVMBuildBitCast(ctx->builder, mask,
LLVMVectorType(ctx->i32, 2),
"");
if (ctx->chip_class >= GFX10) {
result = inclusive ? src : identity;
} else {
- if (inclusive)
- result = src;
- else
- result = ac_build_dpp(ctx, identity, src, dpp_wf_sr1, 0xf, 0xf, false);
+ if (!inclusive)
+ src = ac_build_dpp(ctx, identity, src, dpp_wf_sr1, 0xf, 0xf, false);
+ result = src;
}
if (maxprefix <= 1)
return result;
get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
LLVMTypeOf(identity), "");
- result = ac_build_scan(ctx, op, result, identity, 64, true);
+ result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, true);
return ac_build_wwm(ctx, result);
}
get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
LLVMTypeOf(identity), "");
- result = ac_build_scan(ctx, op, result, identity, 64, false);
+ result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, false);
return ac_build_wwm(ctx, result);
}
if (ws->maxwaves <= 1)
return;
- const LLVMValueRef i32_63 = LLVMConstInt(ctx->i32, 63, false);
+ const LLVMValueRef last_lane = LLVMConstInt(ctx->i32, ctx->wave_size - 1, false);
LLVMBuilderRef builder = ctx->builder;
LLVMValueRef tid = ac_get_thread_id(ctx);
LLVMValueRef tmp;
- tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, i32_63, "");
+ tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, last_lane, "");
ac_build_ifcc(ctx, tmp, 1000);
LLVMBuildStore(builder, ws->src, LLVMBuildGEP(builder, ws->scratch, &ws->waveidx, 1, ""));
ac_build_endif(ctx, 1000);