Skip to content

[mono][jit] Fuse SIMD extract and insert on arm64 #92714

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 6 commits into from
Oct 4, 2023
Merged

[mono][jit] Fuse SIMD extract and insert on arm64 #92714

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
b409943
Fuse extract and insert into arm64 ins.
jandupej Sep 27, 2023
c8c340e
Extending to vector4.
jandupej Sep 27, 2023
0b73b3f
Ins index is now a function. Removed newline.
jandupej Sep 29, 2023
5a0bb3b
Refactoring.
jandupej Sep 29, 2023
828103a
SIMD extract ops have no side offects (to facilitate their eliminatio…
jandupej Sep 29, 2023
b388c46
Fixed arm32 build.
jandupej Sep 29, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 5 additions & 3 deletions src/mono/mono/mini/mini-arm64.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4065,18 +4065,20 @@ mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
break;
}

int idx_to = GTMREG_TO_UINT32 (ins->inst_c0) & 0xff;
int idx_from = GTMREG_TO_UINT32 (ins->inst_c0) >> 8;
if (dreg != sreg1) {
if (dreg != sreg2) {
arm_neon_mov (code, dreg, sreg1);
arm_neon_ins_e(code, t, dreg, sreg2, GTMREG_TO_UINT32 (ins->inst_c0), 0);
arm_neon_ins_e(code, t, dreg, sreg2, idx_to, idx_from);
} else {
arm_neon_mov (code, NEON_TMP_REG, sreg1);
arm_neon_ins_e(code, t, NEON_TMP_REG, sreg2, GTMREG_TO_UINT32 (ins->inst_c0), 0);
arm_neon_ins_e(code, t, NEON_TMP_REG, sreg2, idx_to, idx_from);
arm_neon_mov (code, dreg, NEON_TMP_REG);
}
} else {
g_assert (dreg != sreg2);
arm_neon_ins_e(code, t, dreg, sreg2, GTMREG_TO_UINT32 (ins->inst_c0), 0);
arm_neon_ins_e(code, t, dreg, sreg2, idx_to, idx_from);
}
break;
}
Expand Down
37 changes: 32 additions & 5 deletions src/mono/mono/mini/simd-intrinsics.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1077,7 +1077,18 @@ emit_vector_create_elementwise (
MonoClass *vklass = mono_class_from_mono_type_internal (vtype);
MonoInst *ins = emit_xzero (cfg, vklass);
for (int i = 0; i < fsig->param_count; ++i) {
if (!is_zero_const (args [i])) {
if (is_zero_const (args [i])) {
// element already set to zero
#ifdef TARGET_ARM64
} else if (!COMPILE_LLVM (cfg) && args [i]->opcode == type_to_extract_op (type) &&
(type == MONO_TYPE_R4 || type == MONO_TYPE_R8)) {
// OP_INSERT_Ix inserts from GP reg, not SIMD. Cannot optimize for int types.
int srcidx = args [i]->inst_c0;
ins = emit_simd_ins (cfg, vklass, op, ins->dreg, args [i]->sreg1);
ins->inst_c0 = i | (srcidx << 8);
ins->inst_c1 = type;
#endif
} else {
ins = emit_simd_ins (cfg, vklass, op, ins->dreg, args [i]->dreg);
ins->inst_c0 = i;
ins->inst_c1 = type;
Expand All @@ -1086,6 +1097,7 @@ emit_vector_create_elementwise (
return ins;
}


#if defined(TARGET_AMD64) || defined(TARGET_ARM64) || defined(TARGET_WASM)

static int
Expand Down Expand Up @@ -2287,10 +2299,25 @@ emit_sri_vector (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsi
}

int insert_op = type_to_insert_op (arg0_type);
MonoInst *ins = emit_simd_ins (cfg, klass, insert_op, args [0]->dreg, args [2]->dreg);
ins->inst_c0 = index;
ins->inst_c1 = arg0_type;
return ins;

#ifdef TARGET_ARM64
if (!COMPILE_LLVM (cfg) && args [2]->opcode == type_to_extract_op (arg0_type) && (arg0_type == MONO_TYPE_R4 || arg0_type == MONO_TYPE_R8)) {
// Optimize WithElement(GetElement(x, const_1), const_2) into one ins instruction on arm64
// OP_INSERT_Ix inserts from GP reg, not SIMD. Cannot optimize for int types.
int srcidx = args [2]->inst_c0;
MonoInst* ins = emit_simd_ins (cfg, klass, insert_op, args [0]->dreg, args [2]->sreg1);
ins->inst_c0 = index | (srcidx << 8);
ins->inst_c1 = arg0_type;
return ins;
}
else
#endif
{
MonoInst *ins = emit_simd_ins (cfg, klass, insert_op, args [0]->dreg, args [2]->dreg);
ins->inst_c0 = index;
ins->inst_c1 = arg0_type;
return ins;
}
}

if (!COMPILE_LLVM (cfg) && fsig->params [0]->type != MONO_TYPE_GENERICINST)
Expand Down