mirrors/mesa
0
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2024-11-23 18:24:13 +08:00
Code Issues Packages Projects Releases Wiki Activity

ac,radv,radeonsi: remove LLVM 7 support

Browse Source 
Now that LLVM 9 will be released soon, we will only support
LLVM 8, 9 and master (10).

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This commit is contained in:
Samuel Pitoiset 2019-08-01 11:18:43 +02:00
parent 3e03a3fc53
commit 1fd60db4a1
14 changed files with 66 additions and 321 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
meson.build
View File

@ -1243,7 +1243,7 @@ if with_gallium_opencl
endif endif
if with_amd_vk or with_gallium_radeonsi if with_amd_vk or with_gallium_radeonsi
_llvm_version = '>= 7.0.0' _llvm_version = '>= 8.0.0'
elif with_gallium_swr elif with_gallium_swr
_llvm_version = '>= 6.0.0' _llvm_version = '>= 6.0.0'
elif with_gallium_opencl or with_gallium_r600 elif with_gallium_opencl or with_gallium_r600

3
src/amd/common/ac_gpu_info.c
View File

@ -487,8 +487,7 @@ bool ac_query_gpu_info(int fd, void *dev_p,
} }
info->has_gds_ordered_append = info->chip_class >= GFX7 && info->has_gds_ordered_append = info->chip_class >= GFX7 &&
info->drm_minor >= 29 && info->drm_minor >= 29;
HAVE_LLVM >= 0x0800;
return true; return true;
} }

305
src/amd/common/ac_llvm_build.c
View File

@ -492,7 +492,6 @@ LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
LLVMConstInt(ctx->i32, LLVMIntNE, 0), LLVMConstInt(ctx->i32, LLVMIntNE, 0),
}; };
assert(HAVE_LLVM >= 0x0800);
return ac_build_intrinsic(ctx, name, ctx->i64, args, 3, return ac_build_intrinsic(ctx, name, ctx->i64, args, 3,
AC_FUNC_ATTR_NOUNWIND | AC_FUNC_ATTR_NOUNWIND |
AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_READNONE |
@ -1150,41 +1149,6 @@ static unsigned get_load_cache_policy(struct ac_llvm_context *ctx,
(ctx->chip_class >= GFX10 && cache_policy & ac_glc ? ac_dlc : 0); (ctx->chip_class >= GFX10 && cache_policy & ac_glc ? ac_dlc : 0);
} }
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 static void
ac_build_llvm8_buffer_store_common(struct ac_llvm_context *ctx, ac_build_llvm8_buffer_store_common(struct ac_llvm_context *ctx,
LLVMValueRef rsrc, LLVMValueRef rsrc,
@ -1235,16 +1199,10 @@ ac_build_buffer_store_format(struct ac_llvm_context *ctx,
unsigned num_channels, unsigned num_channels,
unsigned cache_policy) unsigned cache_policy)
{ {
if (HAVE_LLVM >= 0x800) { ac_build_llvm8_buffer_store_common(ctx, rsrc, data, vindex,
ac_build_llvm8_buffer_store_common(ctx, rsrc, data, vindex, voffset, NULL, num_channels,
voffset, NULL, num_channels, ctx->f32, cache_policy,
ctx->f32, cache_policy, true, true);
true, true);
} else {
ac_build_llvm7_buffer_store_common(ctx, rsrc, data, vindex, voffset,
num_channels, cache_policy,
true);
}
} }
/* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4. /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
@ -1294,25 +1252,14 @@ ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
offset = LLVMBuildAdd(ctx->builder, offset, offset = LLVMBuildAdd(ctx->builder, offset,
LLVMConstInt(ctx->i32, inst_offset, 0), ""); LLVMConstInt(ctx->i32, inst_offset, 0), "");
if (HAVE_LLVM >= 0x800) { ac_build_llvm8_buffer_store_common(ctx, rsrc,
ac_build_llvm8_buffer_store_common(ctx, rsrc, ac_to_float(ctx, vdata),
ac_to_float(ctx, vdata), ctx->i32_0,
ctx->i32_0, voffset, offset,
voffset, offset, num_channels,
num_channels, ctx->f32,
ctx->f32, cache_policy,
cache_policy, false, false);
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);
}
return; return;
} }
@ -1330,42 +1277,6 @@ ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
immoffset, num_channels, dfmt, nfmt, cache_policy); immoffset, num_channels, dfmt, nfmt, cache_policy);
} }
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 static LLVMValueRef
ac_build_llvm8_buffer_load_common(struct ac_llvm_context *ctx, ac_build_llvm8_buffer_load_common(struct ac_llvm_context *ctx,
LLVMValueRef rsrc, LLVMValueRef rsrc,
@ -1425,7 +1336,7 @@ ac_build_buffer_load(struct ac_llvm_context *ctx,
offset = LLVMBuildAdd(ctx->builder, offset, soffset, ""); offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
if (allow_smem && !(cache_policy & ac_slc) && 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); assert(vindex == NULL);
LLVMValueRef result[8]; LLVMValueRef result[8];
@ -1435,19 +1346,15 @@ ac_build_buffer_load(struct ac_llvm_context *ctx,
offset = LLVMBuildAdd(ctx->builder, offset, offset = LLVMBuildAdd(ctx->builder, offset,
LLVMConstInt(ctx->i32, 4, 0), ""); 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] = { LLVMValueRef args[3] = {
rsrc, rsrc,
offset, offset,
LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0), LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0),
}; };
result[i] = ac_build_intrinsic(ctx, intrname, result[i] = ac_build_intrinsic(ctx,
ctx->f32, args, num_args, "llvm.amdgcn.s.buffer.load.f32",
AC_FUNC_ATTR_READNONE | ctx->f32, args, 3,
(HAVE_LLVM < 0x0800 ? AC_FUNC_ATTR_LEGACY : 0)); AC_FUNC_ATTR_READNONE);
} }
if (num_channels == 1) if (num_channels == 1)
return result[0]; return result[0];
@ -1457,18 +1364,11 @@ ac_build_buffer_load(struct ac_llvm_context *ctx,
return ac_build_gather_values(ctx, result, num_channels); return ac_build_gather_values(ctx, result, num_channels);
} }
if (HAVE_LLVM >= 0x0800) { return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex,
return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, offset, ctx->i32_0,
offset, ctx->i32_0, num_channels, ctx->f32,
num_channels, ctx->f32, cache_policy,
cache_policy, can_speculate, false, false);
can_speculate, false,
false);
}
return ac_build_llvm7_buffer_load_common(ctx, rsrc, vindex, offset,
num_channels, cache_policy,
can_speculate, false);
} }
LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx, LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
@ -1479,44 +1379,9 @@ LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
unsigned cache_policy, unsigned cache_policy,
bool can_speculate) bool can_speculate)
{ {
if (HAVE_LLVM >= 0x800) { return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0, num_channels, ctx->f32,
num_channels, ctx->f32, cache_policy, can_speculate, true, true);
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);
} }
/// Translate a (dfmt, nfmt) pair into a chip-appropriate combined format /// Translate a (dfmt, nfmt) pair into a chip-appropriate combined format
@ -1615,36 +1480,12 @@ ac_build_tbuffer_load(struct ac_llvm_context *ctx,
bool can_speculate, bool can_speculate,
bool structurized) /* only matters for LLVM 8+ */ bool structurized) /* only matters for LLVM 8+ */
{ {
if (HAVE_LLVM >= 0x800) { voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
return ac_build_llvm8_tbuffer_load(ctx, rsrc, vindex, voffset, return ac_build_llvm8_tbuffer_load(ctx, rsrc, vindex, voffset,
soffset, num_channels, soffset, num_channels,
dfmt, nfmt, cache_policy, dfmt, nfmt, cache_policy,
can_speculate, structurized); 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 LLVMValueRef
@ -1872,20 +1713,13 @@ ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
for (unsigned i = 0; i < load_num_channels; ++i) { for (unsigned i = 0; i < load_num_channels; ++i) {
tmp = LLVMBuildAdd(ctx->builder, soffset, tmp = LLVMBuildAdd(ctx->builder, soffset,
LLVMConstInt(ctx->i32, i << load_log_size, false), ""); LLVMConstInt(ctx->i32, i << load_log_size, false), "");
if (HAVE_LLVM >= 0x0800) { LLVMTypeRef channel_type = load_log_size == 0 ? ctx->i8 :
LLVMTypeRef channel_type = load_log_size == 0 ? ctx->i8 : load_log_size == 1 ? ctx->i16 : ctx->i32;
load_log_size == 1 ? ctx->i16 : ctx->i32; unsigned num_channels = 1 << (MAX2(load_log_size, 2) - 2);
unsigned num_channels = 1 << (MAX2(load_log_size, 2) - 2); loads[i] = ac_build_llvm8_buffer_load_common(
loads[i] = ac_build_llvm8_buffer_load_common( ctx, rsrc, vindex, voffset, tmp,
ctx, rsrc, vindex, voffset, tmp, num_channels, channel_type, cache_policy,
num_channels, channel_type, cache_policy, can_speculate, false, true);
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);
}
if (load_log_size >= 2) if (load_log_size >= 2)
loads[i] = ac_to_integer(ctx, loads[i]); loads[i] = ac_to_integer(ctx, loads[i]);
} }
@ -2108,37 +1942,12 @@ ac_build_tbuffer_store(struct ac_llvm_context *ctx,
unsigned cache_policy, unsigned cache_policy,
bool structurized) /* only matters for LLVM 8+ */ bool structurized) /* only matters for LLVM 8+ */
{ {
if (HAVE_LLVM >= 0x800) { voffset = LLVMBuildAdd(ctx->builder, voffset ? voffset : ctx->i32_0,
voffset = LLVMBuildAdd(ctx->builder, immoffset, "");
voffset ? voffset : ctx->i32_0,
immoffset, "");
ac_build_llvm8_tbuffer_store(ctx, rsrc, vdata, vindex, voffset, ac_build_llvm8_tbuffer_store(ctx, rsrc, vdata, vindex, voffset,
soffset, num_channels, dfmt, nfmt, soffset, num_channels, dfmt, nfmt,
cache_policy, structurized); 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 void
@ -2879,22 +2688,10 @@ LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
width, width,
}; };
LLVMValueRef result = ac_build_intrinsic(ctx, return ac_build_intrinsic(ctx, is_signed ? "llvm.amdgcn.sbfe.i32" :
is_signed ? "llvm.amdgcn.sbfe.i32" : "llvm.amdgcn.ubfe.i32",
"llvm.amdgcn.ubfe.i32", ctx->i32, args, 3, AC_FUNC_ATTR_READNONE);
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 result;
} }
LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0, LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
@ -3869,15 +3666,9 @@ ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef la
LLVMValueRef LLVMValueRef
ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane) ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane)
{ {
if (HAVE_LLVM >= 0x0800) { return ac_build_intrinsic(ctx, "llvm.amdgcn.writelane", ctx->i32,
return ac_build_intrinsic(ctx, "llvm.amdgcn.writelane", ctx->i32, (LLVMValueRef []) {value, lane, src}, 3,
(LLVMValueRef []) {value, lane, src}, 3, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
}
LLVMValueRef pred = LLVMBuildICmp(ctx->builder, LLVMIntEQ, lane,
ac_get_thread_id(ctx), "");
return LLVMBuildSelect(ctx->builder, pred, value, src, "");
} }
LLVMValueRef LLVMValueRef

10
src/amd/common/ac_llvm_build.h
View File

@ -324,16 +324,6 @@ LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
unsigned cache_policy, unsigned cache_policy,
bool can_speculate); bool can_speculate);
/* load_format that handles the stride & element count better if idxen is
* disabled by LLVM. */
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);
LLVMValueRef LLVMValueRef
ac_build_tbuffer_load_short(struct ac_llvm_context *ctx, ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
LLVMValueRef rsrc, LLVMValueRef rsrc,

5
src/amd/common/ac_llvm_util.c
View File

@ -133,7 +133,7 @@ const char *ac_get_llvm_processor_name(enum radeon_family family)
return "gfx906"; return "gfx906";
case CHIP_RAVEN2: case CHIP_RAVEN2:
case CHIP_RENOIR: case CHIP_RENOIR:
return HAVE_LLVM >= 0x0800 ? "gfx909" : "gfx902"; return "gfx909";
case CHIP_ARCTURUS: case CHIP_ARCTURUS:
return "gfx908"; return "gfx908";
case CHIP_NAVI10: case CHIP_NAVI10:
@ -158,8 +158,7 @@ static LLVMTargetMachineRef ac_create_target_machine(enum radeon_family family,
LLVMTargetRef target = ac_get_llvm_target(triple); LLVMTargetRef target = ac_get_llvm_target(triple);
snprintf(features, sizeof(features), snprintf(features, sizeof(features),
"+DumpCode,-fp32-denormals,+fp64-denormals%s%s%s%s%s%s%s", "+DumpCode,-fp32-denormals,+fp64-denormals%s%s%s%s%s%s",
HAVE_LLVM >= 0x0800 ? "" : ",+vgpr-spilling",
family >= CHIP_NAVI10 && !(tm_options & AC_TM_WAVE32) ? family >= CHIP_NAVI10 && !(tm_options & AC_TM_WAVE32) ?
",+wavefrontsize64,-wavefrontsize32" : "", ",+wavefrontsize64,-wavefrontsize32" : "",
tm_options & AC_TM_SISCHED ? ",+si-scheduler" : "", tm_options & AC_TM_SISCHED ? ",+si-scheduler" : "",

24
src/amd/common/ac_nir_to_llvm.c
View File

@ -1308,21 +1308,12 @@ static LLVMValueRef build_tex_intrinsic(struct ac_nir_context *ctx,
if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) { if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa); unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
if (ctx->abi->gfx9_stride_size_workaround) { return ac_build_buffer_load_format(&ctx->ac,
return ac_build_buffer_load_format_gfx9_safe(&ctx->ac, args->resource,
args->resource, args->coords[0],
args->coords[0], ctx->ac.i32_0,
ctx->ac.i32_0, util_last_bit(mask),
util_last_bit(mask), 0, true);
0, true);
} else {
return ac_build_buffer_load_format(&ctx->ac,
args->resource,
args->coords[0],
ctx->ac.i32_0,
util_last_bit(mask),
0, true);
}
} }
args->opcode = ac_image_sample; args->opcode = ac_image_sample;
@ -2477,8 +2468,7 @@ static LLVMValueRef get_image_buffer_descriptor(struct ac_nir_context *ctx,
bool write, bool atomic) bool write, bool atomic)
{ {
LLVMValueRef rsrc = get_image_descriptor(ctx, instr, AC_DESC_BUFFER, write); LLVMValueRef rsrc = get_image_descriptor(ctx, instr, AC_DESC_BUFFER, write);
if (ctx->abi->gfx9_stride_size_workaround || if (ctx->abi->gfx9_stride_size_workaround_for_atomic && atomic) {
(ctx->abi->gfx9_stride_size_workaround_for_atomic && atomic)) {
LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 2, 0), ""); LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 2, 0), "");
LLVMValueRef stride = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 1, 0), ""); LLVMValueRef stride = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 1, 0), "");
stride = LLVMBuildLShr(ctx->ac.builder, stride, LLVMConstInt(ctx->ac.i32, 16, 0), ""); stride = LLVMBuildLShr(ctx->ac.builder, stride, LLVMConstInt(ctx->ac.i32, 16, 0), "");

1
src/amd/common/ac_shader_abi.h
View File

@ -209,7 +209,6 @@ struct ac_shader_abi {
/* Whether to workaround GFX9 ignoring the stride for the buffer size if IDXEN=0 /* Whether to workaround GFX9 ignoring the stride for the buffer size if IDXEN=0
* and LLVM optimizes an indexed load with constant index to IDXEN=0. */ * and LLVM optimizes an indexed load with constant index to IDXEN=0. */
bool gfx9_stride_size_workaround;
bool gfx9_stride_size_workaround_for_atomic; bool gfx9_stride_size_workaround_for_atomic;
/* Whether bounds checks are required */ /* Whether bounds checks are required */

2
src/amd/vulkan/radv_device.c
View File

@ -979,7 +979,7 @@ void radv_GetPhysicalDeviceFeatures2(
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: { case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: {
VkPhysicalDeviceFloat16Int8FeaturesKHR *features = VkPhysicalDeviceFloat16Int8FeaturesKHR *features =
(VkPhysicalDeviceFloat16Int8FeaturesKHR*)ext; (VkPhysicalDeviceFloat16Int8FeaturesKHR*)ext;
features->shaderFloat16 = pdevice->rad_info.chip_class >= GFX8 && HAVE_LLVM >= 0x0800; features->shaderFloat16 = pdevice->rad_info.chip_class >= GFX8;
features->shaderInt8 = true; features->shaderInt8 = true;
break; break;
} }

2
src/amd/vulkan/radv_extensions.py
View File

@ -138,7 +138,7 @@ EXTENSIONS = [
Extension('VK_AMD_buffer_marker', 1, True), Extension('VK_AMD_buffer_marker', 1, True),
Extension('VK_AMD_draw_indirect_count', 1, True), Extension('VK_AMD_draw_indirect_count', 1, True),
Extension('VK_AMD_gcn_shader', 1, True), Extension('VK_AMD_gcn_shader', 1, True),
Extension('VK_AMD_gpu_shader_half_float', 1, 'device->rad_info.chip_class >= GFX9 && HAVE_LLVM >= 0x0800'), Extension('VK_AMD_gpu_shader_half_float', 1, 'device->rad_info.chip_class >= GFX9'),
Extension('VK_AMD_gpu_shader_int16', 1, 'device->rad_info.chip_class >= GFX9'), Extension('VK_AMD_gpu_shader_int16', 1, 'device->rad_info.chip_class >= GFX9'),
Extension('VK_AMD_rasterization_order', 1, 'device->has_out_of_order_rast'), Extension('VK_AMD_rasterization_order', 1, 'device->has_out_of_order_rast'),
Extension('VK_AMD_shader_ballot', 1, 'device->use_shader_ballot'), Extension('VK_AMD_shader_ballot', 1, 'device->use_shader_ballot'),

1
src/amd/vulkan/radv_nir_to_llvm.c
View File

@ -4360,7 +4360,6 @@ LLVMModuleRef ac_translate_nir_to_llvm(struct ac_llvm_compiler *ac_llvm,
ctx.abi.load_sampler_desc = radv_get_sampler_desc; ctx.abi.load_sampler_desc = radv_get_sampler_desc;
ctx.abi.load_resource = radv_load_resource; ctx.abi.load_resource = radv_load_resource;
ctx.abi.clamp_shadow_reference = false; ctx.abi.clamp_shadow_reference = false;
ctx.abi.gfx9_stride_size_workaround = ctx.ac.chip_class == GFX9 && HAVE_LLVM < 0x800;
ctx.abi.robust_buffer_access = options->robust_buffer_access; ctx.abi.robust_buffer_access = options->robust_buffer_access;
/* Because the new raw/struct atomic intrinsics are buggy with LLVM 8, /* Because the new raw/struct atomic intrinsics are buggy with LLVM 8,

11
src/amd/vulkan/radv_shader.c
View File

@ -942,15 +942,8 @@ static void radv_init_llvm_target()
* *
* "mesa" is the prefix for error messages. * "mesa" is the prefix for error messages.
*/ */
if (HAVE_LLVM >= 0x0800) { const char *argv[2] = { "mesa", "-simplifycfg-sink-common=false" };
const char *argv[2] = { "mesa", "-simplifycfg-sink-common=false" }; LLVMParseCommandLineOptions(2, argv, NULL);
LLVMParseCommandLineOptions(2, argv, NULL);
} else {
const char *argv[3] = { "mesa", "-simplifycfg-sink-common=false",
"-amdgpu-skip-threshold=1" };
LLVMParseCommandLineOptions(3, argv, NULL);
}
} }
static once_flag radv_init_llvm_target_once_flag = ONCE_FLAG_INIT; static once_flag radv_init_llvm_target_once_flag = ONCE_FLAG_INIT;

5
src/gallium/drivers/radeonsi/si_shader.c
View File

@ -5305,11 +5305,6 @@ static bool si_get_external_symbol(void *data, const char *name, uint64_t *value
/* Enable scratch coalescing. */ /* Enable scratch coalescing. */
*value = S_008F04_BASE_ADDRESS_HI(*scratch_va >> 32) | *value = S_008F04_BASE_ADDRESS_HI(*scratch_va >> 32) |
S_008F04_SWIZZLE_ENABLE(1); S_008F04_SWIZZLE_ENABLE(1);
if (HAVE_LLVM < 0x0800) {
/* Old LLVM created an R_ABS32_HI relocation for
* this symbol. */
*value <<= 32;
}
return true; return true;
} }

7
src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
View File

@ -840,8 +840,7 @@ static void atomic_emit(
vindex = args.coords[0]; /* for buffers only */ vindex = args.coords[0]; /* for buffers only */
} }
if (HAVE_LLVM >= 0x0800 && if (inst->Src[0].Register.File != TGSI_FILE_BUFFER &&
inst->Src[0].Register.File != TGSI_FILE_BUFFER &&
inst->Memory.Texture == TGSI_TEXTURE_BUFFER) { inst->Memory.Texture == TGSI_TEXTURE_BUFFER) {
LLVMValueRef buf_args[7]; LLVMValueRef buf_args[7];
unsigned num_args = 0; unsigned num_args = 0;
@ -866,9 +865,7 @@ static void atomic_emit(
return; return;
} }
if (inst->Src[0].Register.File == TGSI_FILE_BUFFER || if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
(HAVE_LLVM < 0x0800 &&
inst->Memory.Texture == TGSI_TEXTURE_BUFFER)) {
LLVMValueRef buf_args[7]; LLVMValueRef buf_args[7];
unsigned num_args = 0; unsigned num_args = 0;

9
src/gallium/drivers/radeonsi/si_state.c
View File

@ -3801,14 +3801,7 @@ si_make_buffer_descriptor(struct si_screen *screen, struct si_resource *buf,
* - For VMEM and inst.IDXEN == 0 or STRIDE == 0, it's in byte units. * - For VMEM and inst.IDXEN == 0 or STRIDE == 0, it's in byte units.
* - For VMEM and inst.IDXEN == 1 and STRIDE != 0, it's in units of STRIDE. * - For VMEM and inst.IDXEN == 1 and STRIDE != 0, it's in units of STRIDE.
*/ */
if (screen->info.chip_class == GFX9 && HAVE_LLVM < 0x0800) if (screen->info.chip_class == GFX8)
/* When vindex == 0, LLVM < 8.0 sets IDXEN = 0, thus changing units
* from STRIDE to bytes. This works around it by setting
* NUM_RECORDS to at least the size of one element, so that
* the first element is readable when IDXEN == 0.
*/
num_records = num_records ? MAX2(num_records, stride) : 0;
else if (screen->info.chip_class == GFX8)
num_records *= stride; num_records *= stride;
state[4] = 0; state[4] = 0;