[netcore][x64] Implement lowering of new SIMD OPs into SSE opcodes (m…

…ono/mono#16672) * [netcore][x64] Implement lowering of new SIMD OPs into SSE2 or SSE4.1 opcodes. * Implement LZCNT/POPCNT in mini JIT * Fix C++ build Commit migrated from mono/mono@cf2c857
dotnet · Sep 8, 2019 · 1eab0fe · 1eab0fe
1 parent 9a04875
commit 1eab0fe
Show file tree

Hide file tree

Showing 11 changed files with 513 additions and 27 deletions.
diff --git a/src/mono/mono/arch/amd64/amd64-codegen.h b/src/mono/mono/arch/amd64/amd64-codegen.h
@@ -1165,6 +1165,9 @@ typedef union {
 
 #define amd64_sse_prefetch_reg_membase(inst, arg, basereg, disp) emit_sse_reg_membase_op2((inst), (arg), (basereg), (disp), 0x0f, 0x18)
 
+#define amd64_sse_lzcnt_reg_reg_size(inst, dreg, reg, size) emit_sse_reg_reg_size((inst), (dreg), (reg), 0xf3, 0x0f, 0xbd, (size))
+#define amd64_sse_popcnt_reg_reg_size(inst, dreg, reg, size) emit_sse_reg_reg_size((inst), (dreg), (reg), 0xf3, 0x0f, 0xb8, (size))
+
 /* Generated from x86-codegen.h */
 
 #define amd64_breakpoint_size(inst,size) do { x86_breakpoint(inst); } while (0)

diff --git a/src/mono/mono/mini/cpu-amd64.md b/src/mono/mono/mini/cpu-amd64.md
@@ -827,3 +827,8 @@ generic_class_init: src1:A len:32 clob:c
 get_last_error: dest:i len:32
 
 fill_prof_call_ctx: src1:i len:128
+
+lzcnt32: dest:i src1:i len:16
+lzcnt64: dest:i src1:i len:16
+popcnt32: dest:i src1:i len:16
+popcnt64: dest:i src1:i len:16
diff --git a/src/mono/mono/mini/decompose.c b/src/mono/mono/mini/decompose.c
@@ -1557,7 +1557,7 @@ mono_decompose_array_access_opts (MonoCompile *cfg)
 						MONO_EMIT_NEW_UNALU (cfg, OP_SEXT_I4, index2_reg, ins->sreg2);
 						MONO_EMIT_DEFAULT_BOUNDS_CHECK (cfg, ins->sreg1, ins->inst_imm, index2_reg, ins->flags & MONO_INST_FAULT, ins->inst_p0);
 					} else {
-						MONO_ARCH_EMIT_BOUNDS_CHECK (cfg, ins->sreg1, ins->inst_imm, ins->sreg2);
+						MONO_ARCH_EMIT_BOUNDS_CHECK (cfg, ins->sreg1, ins->inst_imm, ins->sreg2, ins->inst_p0);
 					}
 					break;
 				case OP_NEWARR:

diff --git a/src/mono/mono/mini/ir-emit.h b/src/mono/mono/mini/ir-emit.h
@@ -955,23 +955,24 @@ static int ccount = 0;
 			else \
 				MONO_EMIT_NEW_LOAD_MEMBASE_OP_FLAGS (cfg, OP_LOADI4_MEMBASE, _length_reg, array_reg, offset, MONO_INST_INVARIANT_LOAD); \
 			MONO_EMIT_NEW_BIALU (cfg, OP_COMPARE, -1, _length_reg, index_reg); \
-			MONO_EMIT_NEW_COND_EXC (cfg, LE_UN, (const char*)(ex_name ? ex_name : "IndexOutOfRangeException")); \
+			MONO_EMIT_NEW_COND_EXC (cfg, LE_UN, ex_name); \
 	} while (0)
 
 #ifndef MONO_ARCH_EMIT_BOUNDS_CHECK
-#define MONO_ARCH_EMIT_BOUNDS_CHECK(cfg, array_reg, offset, index_reg) MONO_EMIT_DEFAULT_BOUNDS_CHECK ((cfg), (array_reg), (offset), (index_reg), TRUE, NULL)
+#define MONO_ARCH_EMIT_BOUNDS_CHECK(cfg, array_reg, offset, index_reg, ex_name) MONO_EMIT_DEFAULT_BOUNDS_CHECK ((cfg), (array_reg), (offset), (index_reg), TRUE, ex_name)
 #endif
 
 static inline void
 mini_emit_bounds_check_offset (MonoCompile *cfg, int array_reg, int array_length_offset, int index_reg, const char *ex_name)
 {
 	if (!(cfg->opt & MONO_OPT_UNSAFE)) {
+		ex_name = ex_name ? ex_name : "IndexOutOfRangeException";
 		if (!(cfg->opt & MONO_OPT_ABCREM)) {
 			MONO_EMIT_NULL_CHECK (cfg, array_reg, FALSE);
 			if (COMPILE_LLVM (cfg))
 				MONO_EMIT_DEFAULT_BOUNDS_CHECK ((cfg), (array_reg), (array_length_offset), (index_reg), TRUE, ex_name);
 			else
-				MONO_ARCH_EMIT_BOUNDS_CHECK ((cfg), (array_reg), (array_length_offset), (index_reg));
+				MONO_ARCH_EMIT_BOUNDS_CHECK ((cfg), (array_reg), (array_length_offset), (index_reg), ex_name);
 		} else {
 			MonoInst *ins;
 			MONO_INST_NEW ((cfg), ins, OP_BOUNDS_CHECK);