ctx->i16 = LLVMIntTypeInContext(ctx->context, 16);
ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
ctx->i64 = LLVMIntTypeInContext(ctx->context, 64);
+ ctx->i128 = LLVMIntTypeInContext(ctx->context, 128);
ctx->intptr = ctx->i32;
ctx->f16 = LLVMHalfTypeInContext(ctx->context);
ctx->f32 = LLVMFloatTypeInContext(ctx->context);
ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
ctx->i64_0 = LLVMConstInt(ctx->i64, 0, false);
ctx->i64_1 = LLVMConstInt(ctx->i64, 1, false);
+ ctx->i128_0 = LLVMConstInt(ctx->i128, 0, false);
+ ctx->i128_1 = LLVMConstInt(ctx->i128, 1, false);
ctx->f16_0 = LLVMConstReal(ctx->f16, 0.0);
ctx->f16_1 = LLVMConstReal(ctx->f16, 1.0);
ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
} else {
LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
+ LLVMTypeRef type = LLVMTypeOf(*pvgpr);
+ unsigned bitsize = ac_get_elem_bits(ctx, type);
LLVMValueRef vgpr = *pvgpr;
- LLVMTypeRef vgpr_type = LLVMTypeOf(vgpr);
- unsigned vgpr_size = ac_get_type_size(vgpr_type);
+ LLVMTypeRef vgpr_type;
+ unsigned vgpr_size;
LLVMValueRef vgpr0;
+ if (bitsize < 32)
+ vgpr = LLVMBuildZExt(ctx->builder, vgpr, ctx->i32, "");
+
+ vgpr_type = LLVMTypeOf(vgpr);
+ vgpr_size = ac_get_type_size(vgpr_type);
+
assert(vgpr_size % 4 == 0);
vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
+ if (bitsize < 32)
+ vgpr = LLVMBuildTrunc(builder, vgpr, type, "");
+
*pvgpr = vgpr;
}
}
LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
LLVMValueRef value)
{
- const char *name = LLVM_VERSION_MAJOR >= 9 ? "llvm.amdgcn.icmp.i64.i1" : "llvm.amdgcn.icmp.i1";
+ const char *name;
+
+ if (LLVM_VERSION_MAJOR >= 9) {
+ if (ctx->wave_size == 64)
+ name = "llvm.amdgcn.icmp.i64.i1";
+ else
+ name = "llvm.amdgcn.icmp.i32.i1";
+ } else {
+ name = "llvm.amdgcn.icmp.i1";
+ }
LLVMValueRef args[3] = {
value,
ctx->i1false,
LLVMConstInt(ctx->i32, LLVMIntNE, 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);
LLVMValueRef voffset,
LLVMValueRef soffset,
unsigned inst_offset,
- unsigned cache_policy,
- bool swizzle_enable_hint)
+ unsigned cache_policy)
{
/* Split 3 channel stores, because only LLVM 9+ support 3-channel
* intrinsics. */
v01 = ac_build_gather_values(ctx, v, 2);
ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
- soffset, inst_offset, cache_policy,
- swizzle_enable_hint);
+ soffset, inst_offset, cache_policy);
ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
soffset, inst_offset + 8,
- cache_policy,
- swizzle_enable_hint);
+ cache_policy);
return;
}
* (voffset is swizzled, but soffset isn't swizzled).
* llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
*/
- if (!swizzle_enable_hint) {
+ if (!(cache_policy & ac_swizzled)) {
LLVMValueRef offset = soffset;
if (inst_offset)
bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
switch (bitsize) {
+ case 128:
+ result = ac_build_intrinsic(ctx, "llvm.ctpop.i128", ctx->i128,
+ (LLVMValueRef []) { src0 }, 1,
+ AC_FUNC_ATTR_READNONE);
+ result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
+ break;
case 64:
result = ac_build_intrinsic(ctx, "llvm.ctpop.i64", ctx->i64,
(LLVMValueRef []) { src0 }, 1,
_ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
bool exchange_rows, bool bound_ctrl)
{
+ LLVMTypeRef type = LLVMTypeOf(src);
+ LLVMValueRef result;
+
+ src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
+
LLVMValueRef args[6] = {
src,
src,
ctx->i1true, /* fi */
bound_ctrl ? ctx->i1true : ctx->i1false,
};
- return ac_build_intrinsic(ctx, exchange_rows ? "llvm.amdgcn.permlanex16"
- : "llvm.amdgcn.permlane16",
- ctx->i32, args, 6,
- AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
+
+ result = ac_build_intrinsic(ctx, exchange_rows ? "llvm.amdgcn.permlanex16"
+ : "llvm.amdgcn.permlane16",
+ ctx->i32, args, 6,
+ AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
+
+ return LLVMBuildTrunc(ctx->builder, result, type, "");
}
static LLVMValueRef
src = ac_to_integer(ctx, src);
unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
LLVMValueRef ret;
- if (bits == 32) {
- ret = _ac_build_permlane16(ctx, src, sel, exchange_rows,
- bound_ctrl);
- } else {
+ if (bits > 32) {
assert(bits % 32 == 0);
LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
LLVMValueRef src_vector =
LLVMConstInt(ctx->i32, i,
0), "");
}
+ } else {
+ ret = _ac_build_permlane16(ctx, src, sel, exchange_rows,
+ bound_ctrl);
}
return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
}
ac_build_alu_op(struct ac_llvm_context *ctx, LLVMValueRef lhs, LLVMValueRef rhs, nir_op op)
{
bool _64bit = ac_get_type_size(LLVMTypeOf(lhs)) == 8;
+ bool _32bit = ac_get_type_size(LLVMTypeOf(lhs)) == 4;
switch (op) {
case nir_op_iadd: return LLVMBuildAdd(ctx->builder, lhs, rhs, "");
case nir_op_fadd: return LLVMBuildFAdd(ctx->builder, lhs, rhs, "");
LLVMBuildICmp(ctx->builder, LLVMIntULT, lhs, rhs, ""),
lhs, rhs, "");
case nir_op_fmin: return ac_build_intrinsic(ctx,
- _64bit ? "llvm.minnum.f64" : "llvm.minnum.f32",
- _64bit ? ctx->f64 : ctx->f32,
+ _64bit ? "llvm.minnum.f64" : _32bit ? "llvm.minnum.f32" : "llvm.minnum.f16",
+ _64bit ? ctx->f64 : _32bit ? ctx->f32 : ctx->f16,
(LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
case nir_op_imax: return LLVMBuildSelect(ctx->builder,
LLVMBuildICmp(ctx->builder, LLVMIntSGT, lhs, rhs, ""),
LLVMBuildICmp(ctx->builder, LLVMIntUGT, lhs, rhs, ""),
lhs, rhs, "");
case nir_op_fmax: return ac_build_intrinsic(ctx,
- _64bit ? "llvm.maxnum.f64" : "llvm.maxnum.f32",
- _64bit ? ctx->f64 : ctx->f32,
+ _64bit ? "llvm.maxnum.f64" : _32bit ? "llvm.maxnum.f32" : "llvm.maxnum.f16",
+ _64bit ? ctx->f64 : _32bit ? ctx->f32 : ctx->f16,
(LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
case nir_op_iand: return LLVMBuildAnd(ctx->builder, lhs, rhs, "");
case nir_op_ior: return LLVMBuildOr(ctx->builder, lhs, rhs, "");
}
}
+/**
+ * \param src The value to shift.
+ * \param identity The value to use the first lane.
+ * \param maxprefix specifies that the result only needs to be correct for a
+ * prefix of this many threads
+ * \return src, shifted 1 lane up, and identity shifted into lane 0.
+ */
+static LLVMValueRef
+ac_wavefront_shift_right_1(struct ac_llvm_context *ctx, LLVMValueRef src,
+ LLVMValueRef identity, unsigned maxprefix)
+{
+ if (ctx->chip_class >= GFX10) {
+ /* wavefront shift_right by 1 on GFX10 (emulate dpp_wf_sr1) */
+ LLVMValueRef active, tmp1, tmp2;
+ LLVMValueRef tid = ac_get_thread_id(ctx);
+
+ tmp1 = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
+
+ tmp2 = ac_build_permlane16(ctx, src, (uint64_t)~0, true, false);
+
+ if (maxprefix > 32) {
+ active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid,
+ LLVMConstInt(ctx->i32, 32, false), "");
+
+ tmp2 = LLVMBuildSelect(ctx->builder, active,
+ ac_build_readlane(ctx, src,
+ LLVMConstInt(ctx->i32, 31, false)),
+ tmp2, "");
+
+ active = LLVMBuildOr(ctx->builder, active,
+ LLVMBuildICmp(ctx->builder, LLVMIntEQ,
+ LLVMBuildAnd(ctx->builder, tid,
+ LLVMConstInt(ctx->i32, 0x1f, false), ""),
+ LLVMConstInt(ctx->i32, 0x10, false), ""), "");
+ return LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
+ } else if (maxprefix > 16) {
+ active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid,
+ LLVMConstInt(ctx->i32, 16, false), "");
+
+ return LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
+ }
+ } else if (ctx->chip_class >= GFX8) {
+ return ac_build_dpp(ctx, identity, src, dpp_wf_sr1, 0xf, 0xf, false);
+ }
+
+ /* wavefront shift_right by 1 on SI/CI */
+ LLVMValueRef active, tmp1, tmp2;
+ LLVMValueRef tid = ac_get_thread_id(ctx);
+ tmp1 = ac_build_ds_swizzle(ctx, src, (1 << 15) | dpp_quad_perm(0, 0, 1, 2));
+ tmp2 = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x18, 0x03, 0x00));
+ active = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
+ LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0x7, 0), ""),
+ LLVMConstInt(ctx->i32, 0x4, 0), "");
+ tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
+ tmp2 = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x10, 0x07, 0x00));
+ active = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
+ LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0xf, 0), ""),
+ LLVMConstInt(ctx->i32, 0x8, 0), "");
+ tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
+ tmp2 = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x00, 0x0f, 0x00));
+ active = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
+ LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0x1f, 0), ""),
+ LLVMConstInt(ctx->i32, 0x10, 0), "");
+ tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
+ tmp2 = ac_build_readlane(ctx, src, LLVMConstInt(ctx->i32, 31, 0));
+ active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid, LLVMConstInt(ctx->i32, 32, 0), "");
+ tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
+ active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid, LLVMConstInt(ctx->i32, 0, 0), "");
+ return LLVMBuildSelect(ctx->builder, active, identity, tmp1, "");
+}
+
/**
* \param maxprefix specifies that the result only needs to be correct for a
* prefix of this many threads
- *
- * TODO: add inclusive and excluse scan functions for GFX6.
*/
static LLVMValueRef
ac_build_scan(struct ac_llvm_context *ctx, nir_op op, LLVMValueRef src, LLVMValueRef identity,
{
LLVMValueRef result, tmp;
- if (ctx->chip_class >= GFX10) {
- result = inclusive ? src : identity;
- } else {
- if (!inclusive)
- src = ac_build_dpp(ctx, identity, src, dpp_wf_sr1, 0xf, 0xf, false);
- result = src;
+ if (!inclusive)
+ src = ac_wavefront_shift_right_1(ctx, src, identity, maxprefix);
+
+ result = src;
+
+ if (ctx->chip_class <= GFX7) {
+ assert(maxprefix == 64);
+ LLVMValueRef tid = ac_get_thread_id(ctx);
+ LLVMValueRef active;
+ tmp = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x1e, 0x00, 0x00));
+ active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
+ LLVMBuildAnd(ctx->builder, tid, ctx->i32_1, ""),
+ ctx->i32_0, "");
+ tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+ result = ac_build_alu_op(ctx, result, tmp, op);
+ tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1c, 0x01, 0x00));
+ active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
+ LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 2, 0), ""),
+ ctx->i32_0, "");
+ tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+ result = ac_build_alu_op(ctx, result, tmp, op);
+ tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x18, 0x03, 0x00));
+ active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
+ LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 4, 0), ""),
+ ctx->i32_0, "");
+ tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+ result = ac_build_alu_op(ctx, result, tmp, op);
+ tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x10, 0x07, 0x00));
+ active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
+ LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 8, 0), ""),
+ ctx->i32_0, "");
+ tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+ result = ac_build_alu_op(ctx, result, tmp, op);
+ tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x00, 0x0f, 0x00));
+ active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
+ LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 16, 0), ""),
+ ctx->i32_0, "");
+ tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+ result = ac_build_alu_op(ctx, result, tmp, op);
+ tmp = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, 0));
+ active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
+ LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 32, 0), ""),
+ ctx->i32_0, "");
+ tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+ result = ac_build_alu_op(ctx, result, tmp, op);
+ return result;
}
+
if (maxprefix <= 1)
return result;
tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
return result;
if (ctx->chip_class >= GFX10) {
- /* dpp_row_bcast{15,31} are not supported on gfx10. */
- LLVMBuilderRef builder = ctx->builder;
LLVMValueRef tid = ac_get_thread_id(ctx);
- LLVMValueRef cc;
- /* TODO-GFX10: Can we get better code-gen by putting this into
- * a branch so that LLVM generates EXEC mask manipulations? */
- if (inclusive)
- tmp = result;
- else
- tmp = ac_build_alu_op(ctx, result, src, op);
- tmp = ac_build_permlane16(ctx, tmp, ~(uint64_t)0, true, false);
- tmp = ac_build_alu_op(ctx, result, tmp, op);
- cc = LLVMBuildAnd(builder, tid, LLVMConstInt(ctx->i32, 16, false), "");
- cc = LLVMBuildICmp(builder, LLVMIntNE, cc, ctx->i32_0, "");
- result = LLVMBuildSelect(builder, cc, tmp, result, "");
+ LLVMValueRef active;
+
+ tmp = ac_build_permlane16(ctx, result, ~(uint64_t)0, true, false);
+
+ active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
+ LLVMBuildAnd(ctx->builder, tid,
+ LLVMConstInt(ctx->i32, 16, false), ""),
+ ctx->i32_0, "");
+
+ tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+
+ result = ac_build_alu_op(ctx, result, tmp, op);
+
if (maxprefix <= 32)
return result;
- if (inclusive)
- tmp = result;
- else
- tmp = ac_build_alu_op(ctx, result, src, op);
- tmp = ac_build_readlane(ctx, tmp, LLVMConstInt(ctx->i32, 31, false));
- tmp = ac_build_alu_op(ctx, result, tmp, op);
- cc = LLVMBuildICmp(builder, LLVMIntUGE, tid,
- LLVMConstInt(ctx->i32, 32, false), "");
- result = LLVMBuildSelect(builder, cc, tmp, result, "");
+ tmp = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, false));
+
+ active = LLVMBuildICmp(ctx->builder, LLVMIntUGE, tid,
+ LLVMConstInt(ctx->i32, 32, false), "");
+
+ tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
+
+ result = ac_build_alu_op(ctx, result, tmp, op);
return result;
}
if (cluster_size == 32) return ac_build_wwm(ctx, result);
if (ctx->chip_class >= GFX8) {
- if (ctx->chip_class >= GFX10)
- swap = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, false));
- else
- swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
- result = ac_build_alu_op(ctx, result, swap, op);
- result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 63, 0));
+ if (ctx->wave_size == 64) {
+ if (ctx->chip_class >= GFX10)
+ swap = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, false));
+ else
+ swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
+ result = ac_build_alu_op(ctx, result, swap, op);
+ result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 63, 0));
+ }
+
return ac_build_wwm(ctx, result);
} else {
swap = ac_build_readlane(ctx, result, ctx->i32_0);
} else if (bitsize == 32) {
intr = "llvm.canonicalize.f32";
type = ctx->f32;
- } else if (bitsize == 64) {
+ } else {
intr = "llvm.canonicalize.f64";
type = ctx->f64;
}
args->enabled_channels = mask;
}
+/* Send GS Alloc Req message from the first wave of the group to SPI.
+ * Message payload is:
+ * - bits 0..10: vertices in group
+ * - bits 12..22: primitives in group
+ */
+void ac_build_sendmsg_gs_alloc_req(struct ac_llvm_context *ctx, LLVMValueRef wave_id,
+ LLVMValueRef vtx_cnt, LLVMValueRef prim_cnt)
+{
+ LLVMBuilderRef builder = ctx->builder;
+ LLVMValueRef tmp;
+
+ ac_build_ifcc(ctx, LLVMBuildICmp(builder, LLVMIntEQ, wave_id, ctx->i32_0, ""), 5020);
+
+ tmp = LLVMBuildShl(builder, prim_cnt, LLVMConstInt(ctx->i32, 12, false),"");
+ tmp = LLVMBuildOr(builder, tmp, vtx_cnt, "");
+ ac_build_sendmsg(ctx, AC_SENDMSG_GS_ALLOC_REQ, tmp);
+
+ ac_build_endif(ctx, 5020);
+}
+
+LLVMValueRef ac_pack_prim_export(struct ac_llvm_context *ctx,
+ const struct ac_ngg_prim *prim)
+{
+ /* The prim export format is:
+ * - bits 0..8: index 0
+ * - bit 9: edge flag 0
+ * - bits 10..18: index 1
+ * - bit 19: edge flag 1
+ * - bits 20..28: index 2
+ * - bit 29: edge flag 2
+ * - bit 31: null primitive (skip)
+ */
+ LLVMBuilderRef builder = ctx->builder;
+ LLVMValueRef tmp = LLVMBuildZExt(builder, prim->isnull, ctx->i32, "");
+ LLVMValueRef result = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->i32, 31, false), "");
+
+ for (unsigned i = 0; i < prim->num_vertices; ++i) {
+ tmp = LLVMBuildShl(builder, prim->index[i],
+ LLVMConstInt(ctx->i32, 10 * i, false), "");
+ result = LLVMBuildOr(builder, result, tmp, "");
+ tmp = LLVMBuildZExt(builder, prim->edgeflag[i], ctx->i32, "");
+ tmp = LLVMBuildShl(builder, tmp,
+ LLVMConstInt(ctx->i32, 10 * i + 9, false), "");
+ result = LLVMBuildOr(builder, result, tmp, "");
+ }
+ return result;
+}
+
+void ac_build_export_prim(struct ac_llvm_context *ctx,
+ const struct ac_ngg_prim *prim)
+{
+ struct ac_export_args args;
+
+ if (prim->passthrough) {
+ args.out[0] = prim->passthrough;
+ } else {
+ args.out[0] = ac_pack_prim_export(ctx, prim);
+ }
+
+ args.out[0] = LLVMBuildBitCast(ctx->builder, args.out[0], ctx->f32, "");
+ args.out[1] = LLVMGetUndef(ctx->f32);
+ args.out[2] = LLVMGetUndef(ctx->f32);
+ args.out[3] = LLVMGetUndef(ctx->f32);
+
+ args.target = V_008DFC_SQ_EXP_PRIM;
+ args.enabled_channels = 1;
+ args.done = true;
+ args.valid_mask = false;
+ args.compr = false;
+
+ ac_build_export(ctx, &args);
+}
+
+static LLVMTypeRef
+arg_llvm_type(enum ac_arg_type type, unsigned size, struct ac_llvm_context *ctx)
+{
+ if (type == AC_ARG_FLOAT) {
+ return size == 1 ? ctx->f32 : LLVMVectorType(ctx->f32, size);
+ } else if (type == AC_ARG_INT) {
+ return size == 1 ? ctx->i32 : LLVMVectorType(ctx->i32, size);
+ } else {
+ LLVMTypeRef ptr_type;
+ switch (type) {
+ case AC_ARG_CONST_PTR:
+ ptr_type = ctx->i8;
+ break;
+ case AC_ARG_CONST_FLOAT_PTR:
+ ptr_type = ctx->f32;
+ break;
+ case AC_ARG_CONST_PTR_PTR:
+ ptr_type = ac_array_in_const32_addr_space(ctx->i8);
+ break;
+ case AC_ARG_CONST_DESC_PTR:
+ ptr_type = ctx->v4i32;
+ break;
+ case AC_ARG_CONST_IMAGE_PTR:
+ ptr_type = ctx->v8i32;
+ break;
+ default:
+ unreachable("unknown arg type");
+ }
+ if (size == 1) {
+ return ac_array_in_const32_addr_space(ptr_type);
+ } else {
+ assert(size == 2);
+ return ac_array_in_const_addr_space(ptr_type);
+ }
+ }
+}
+
+LLVMValueRef
+ac_build_main(const struct ac_shader_args *args,
+ struct ac_llvm_context *ctx,
+ enum ac_llvm_calling_convention convention,
+ const char *name, LLVMTypeRef ret_type,
+ LLVMModuleRef module)
+{
+ LLVMTypeRef arg_types[AC_MAX_ARGS];
+
+ for (unsigned i = 0; i < args->arg_count; i++) {
+ arg_types[i] = arg_llvm_type(args->args[i].type,
+ args->args[i].size, ctx);
+ }
+
+ LLVMTypeRef main_function_type =
+ LLVMFunctionType(ret_type, arg_types, args->arg_count, 0);
+
+ LLVMValueRef main_function =
+ LLVMAddFunction(module, name, main_function_type);
+ LLVMBasicBlockRef main_function_body =
+ LLVMAppendBasicBlockInContext(ctx->context, main_function, "main_body");
+ LLVMPositionBuilderAtEnd(ctx->builder, main_function_body);
+
+ LLVMSetFunctionCallConv(main_function, convention);
+ for (unsigned i = 0; i < args->arg_count; ++i) {
+ LLVMValueRef P = LLVMGetParam(main_function, i);
+
+ if (args->args[i].file != AC_ARG_SGPR)
+ continue;
+
+ ac_add_function_attr(ctx->context, main_function, i + 1, AC_FUNC_ATTR_INREG);
+
+ if (LLVMGetTypeKind(LLVMTypeOf(P)) == LLVMPointerTypeKind) {
+ ac_add_function_attr(ctx->context, main_function, i + 1, AC_FUNC_ATTR_NOALIAS);
+ ac_add_attr_dereferenceable(P, UINT64_MAX);
+ }
+ }
+
+ ctx->main_function = main_function;
+ return main_function;
+}
+
+void ac_build_s_endpgm(struct ac_llvm_context *ctx)
+{
+ LLVMTypeRef calltype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
+ LLVMValueRef code = LLVMConstInlineAsm(calltype, "s_endpgm", "", true, false);
+ LLVMBuildCall(ctx->builder, code, NULL, 0, "");
+}
+
+LLVMValueRef ac_prefix_bitcount(struct ac_llvm_context *ctx,
+ LLVMValueRef mask, LLVMValueRef index)
+{
+ LLVMBuilderRef builder = ctx->builder;
+ LLVMTypeRef type = LLVMTypeOf(mask);
+
+ LLVMValueRef bit = LLVMBuildShl(builder, LLVMConstInt(type, 1, 0),
+ LLVMBuildZExt(builder, index, type, ""), "");
+ LLVMValueRef prefix_bits = LLVMBuildSub(builder, bit, LLVMConstInt(type, 1, 0), "");
+ LLVMValueRef prefix_mask = LLVMBuildAnd(builder, mask, prefix_bits, "");
+ return ac_build_bit_count(ctx, prefix_mask);
+}
+
+/* Compute the prefix sum of the "mask" bit array with 128 elements (bits). */
+LLVMValueRef ac_prefix_bitcount_2x64(struct ac_llvm_context *ctx,
+ LLVMValueRef mask[2], LLVMValueRef index)
+{
+ LLVMBuilderRef builder = ctx->builder;
+#if 0
+ /* Reference version using i128. */
+ LLVMValueRef input_mask =
+ LLVMBuildBitCast(builder, ac_build_gather_values(ctx, mask, 2), ctx->i128, "");
+
+ return ac_prefix_bitcount(ctx, input_mask, index);
+#else
+ /* Optimized version using 2 64-bit masks. */
+ LLVMValueRef is_hi, is_0, c64, c128, all_bits;
+ LLVMValueRef prefix_mask[2], shift[2], mask_bcnt0, prefix_bcnt[2];
+
+ /* Compute the 128-bit prefix mask. */
+ c64 = LLVMConstInt(ctx->i32, 64, 0);
+ c128 = LLVMConstInt(ctx->i32, 128, 0);
+ all_bits = LLVMConstInt(ctx->i64, UINT64_MAX, 0);
+ /* The first index that can have non-zero high bits in the prefix mask is 65. */
+ is_hi = LLVMBuildICmp(builder, LLVMIntUGT, index, c64, "");
+ is_0 = LLVMBuildICmp(builder, LLVMIntEQ, index, ctx->i32_0, "");
+ mask_bcnt0 = ac_build_bit_count(ctx, mask[0]);
+
+ for (unsigned i = 0; i < 2; i++) {
+ shift[i] = LLVMBuildSub(builder, i ? c128 : c64, index, "");
+ /* For i==0, index==0, the right shift by 64 doesn't give the desired result,
+ * so we handle it by the is_0 select.
+ * For i==1, index==64, same story, so we handle it by the last is_hi select.
+ * For i==0, index==64, we shift by 0, which is what we want.
+ */
+ prefix_mask[i] = LLVMBuildLShr(builder, all_bits,
+ LLVMBuildZExt(builder, shift[i], ctx->i64, ""), "");
+ prefix_mask[i] = LLVMBuildAnd(builder, mask[i], prefix_mask[i], "");
+ prefix_bcnt[i] = ac_build_bit_count(ctx, prefix_mask[i]);
+ }
+
+ prefix_bcnt[0] = LLVMBuildSelect(builder, is_0, ctx->i32_0, prefix_bcnt[0], "");
+ prefix_bcnt[0] = LLVMBuildSelect(builder, is_hi, mask_bcnt0, prefix_bcnt[0], "");
+ prefix_bcnt[1] = LLVMBuildSelect(builder, is_hi, prefix_bcnt[1], ctx->i32_0, "");
+
+ return LLVMBuildAdd(builder, prefix_bcnt[0], prefix_bcnt[1], "");
+#endif
+}