Merge remote branch 'upstream/master'

[mono.git] / mono / mini / cpu-amd64.md

diff --git a/mono/mini/cpu-amd64.md b/mono/mini/cpu-amd64.md
index d3a2d0fe1d574a3d910f958008b06539f1b4ef90..b095aa9428095c2cbc04ac521a4d3096bb7ac608 100644 (file)
--- a/mono/mini/cpu-amd64.md
+++ b/mono/mini/cpu-amd64.md
@@ -53,14 +53,22 @@
 #
 # See the code in mini-x86.c for more details on how the specifiers are used.
 #
+#
+# Native Client Note: NaCl call sequences do not really reach > 32 bytes but 
+# the maximum length can be high, so if we get unlucky and wind up trying to 
+# emit a call sequence such that we are one or two bytes too long, we need to
+# pad out almost an entire 32 bytes.
+#
+
 break: len:2
 jmp: len:120
 tailcall: len:120 clob:c
 br: len:6
 label: len:0
+seq_point: len:25
 
-long_add: dest:i src1:i src2:i len:3 clob:1
-long_sub: dest:i src1:i src2:i len:3 clob:1
+long_add: dest:i src1:i src2:i len:3 clob:1 nacl:6
+long_sub: dest:i src1:i src2:i len:3 clob:1 nacl:6
 long_mul: dest:i src1:i src2:i len:4 clob:1
 long_div: dest:a src1:a src2:i len:16 clob:d
 long_div_un: dest:a src1:a src2:i len:16 clob:d
@@ -95,11 +103,11 @@ long_min_un: dest:i src1:i src2:i len:16 clob:1
 long_max: dest:i src1:i src2:i len:16 clob:1
 long_max_un: dest:i src1:i src2:i len:16 clob:1
 
-throw: src1:i len:18
-rethrow: src1:i len:18
-start_handler: len:9
-endfinally: len:9
-endfilter: src1:a len:9
+throw: src1:i len:18 nacl:50
+rethrow: src1:i len:18 nacl:50
+start_handler: len:16
+endfinally: len:9 nacl:22
+endfilter: src1:a len:9 nacl:19
 ckfinite: dest:f src1:f len:43
 ceq: dest:c len:8
 cgt: dest:c len:8
@@ -114,12 +122,11 @@ compare_imm: src1:i len:13
 icompare_imm: src1:i len:8
 fcompare: src1:f src2:f clob:a len:13
 oparglist: src1:b len:11
-setlret: dest:i src1:i src2:i len:5
-checkthis: src1:b len:5
-call: dest:a clob:c len:32
-voidcall: clob:c len:32
-voidcall_reg: src1:i clob:c len:32
-voidcall_membase: src1:b clob:c len:32
+checkthis: src1:b len:5 nacl:8
+call: dest:a clob:c len:32 nacl:64
+voidcall: clob:c len:32 nacl:64
+voidcall_reg: src1:i clob:c len:32 nacl:64
+voidcall_membase: src1:b clob:c len:32 nacl:64
 fcall: dest:f len:64 clob:c
 fcall_reg: dest:f src1:i len:64 clob:c
 fcall_membase: dest:f src1:b len:64 clob:c
@@ -129,48 +136,40 @@ lcall_membase: dest:a src1:b len:64 clob:c
 vcall: len:64 clob:c
 vcall_reg: src1:i len:64 clob:c
 vcall_membase: src1:b len:64 clob:c
-call_reg: dest:a src1:i len:32 clob:c
-call_membase: dest:a src1:b len:32 clob:c
+call_reg: dest:a src1:i len:32 clob:c nacl:64
+call_membase: dest:a src1:b len:32 clob:c nacl:64
 iconst: dest:i len:10
 i8const: dest:i len:10
 r4const: dest:f len:14
 r8const: dest:f len:9
 store_membase_imm: dest:b len:15
-store_membase_reg: dest:b src1:i len:9
-storei8_membase_reg: dest:b src1:i len:9
-storei1_membase_imm: dest:b len:11
-storei1_membase_reg: dest:b src1:c len:9
-storei2_membase_imm: dest:b len:13
-storei2_membase_reg: dest:b src1:i len:9
-storei4_membase_imm: dest:b len:13
-storei4_membase_reg: dest:b src1:i len:9
+store_membase_reg: dest:b src1:i len:9 nacl:11
+storei8_membase_reg: dest:b src1:i len:9 nacl:11
+storei1_membase_imm: dest:b len:11 nacl:15
+storei1_membase_reg: dest:b src1:c len:9 nacl:11
+storei2_membase_imm: dest:b len:13 nacl:15
+storei2_membase_reg: dest:b src1:i len:9 nacl:11
+storei4_membase_imm: dest:b len:13 nacl:15
+storei4_membase_reg: dest:b src1:i len:9 nacl:11
 storei8_membase_imm: dest:b len:18
 storer4_membase_reg: dest:b src1:f len:15
 storer8_membase_reg: dest:b src1:f len:10
-load_membase: dest:i src1:b len:8
-loadi1_membase: dest:c src1:b len:9
-loadu1_membase: dest:c src1:b len:9
-loadi2_membase: dest:i src1:b len:9
-loadu2_membase: dest:i src1:b len:9
-loadi4_membase: dest:i src1:b len:9
-loadu4_membase: dest:i src1:b len:9
-loadi8_membase: dest:i src1:b len:18
+load_membase: dest:i src1:b len:8 nacl:12
+loadi1_membase: dest:c src1:b len:9 nacl:12
+loadu1_membase: dest:c src1:b len:9 nacl:12
+loadi2_membase: dest:i src1:b len:9 nacl:12
+loadu2_membase: dest:i src1:b len:9 nacl:12
+loadi4_membase: dest:i src1:b len:9 nacl:12
+loadu4_membase: dest:i src1:b len:9 nacl:12
+loadi8_membase: dest:i src1:b len:18 nacl:14
 loadr4_membase: dest:f src1:b len:16
 loadr8_membase: dest:f src1:b len:16
-loadr8_spill_membase: src1:b len:9
 loadu4_mem: dest:i len:10
 amd64_loadi8_memindex: dest:i src1:i src2:i len:10
 move: dest:i src1:i len:3
-add_imm: dest:i src1:i len:8 clob:1
-sub_imm: dest:i src1:i len:8 clob:1
+add_imm: dest:i src1:i len:8 clob:1 nacl:11
+sub_imm: dest:i src1:i len:8 clob:1 nacl:11
 mul_imm: dest:i src1:i len:11
-# there is no actual support for division or reminder by immediate
-# we simulate them, though (but we need to change the burg rules 
-# to allocate a symbolic reg for src2)
-div_imm: dest:a src1:i src2:i len:16 clob:d
-div_un_imm: dest:a src1:i src2:i len:16 clob:d
-rem_imm: dest:d src1:i src2:i len:16 clob:a
-rem_un_imm: dest:d src1:i src2:i len:16 clob:a
 and_imm: dest:i src1:i len:8 clob:1
 or_imm: dest:i src1:i len:8 clob:1
 xor_imm: dest:i src1:i len:8 clob:1
@@ -194,7 +193,6 @@ cond_exc_nc: len:8
 cond_exc_iov: len:8
 cond_exc_ic: len:8
 
-long_conv_to_ovf_i: dest:i src1:i src2:i len:40
 long_mul_ovf: dest:i src1:i src2:i clob:1 len:16
 long_mul_ovf_un: dest:i src1:i src2:i len:22
 long_shr_imm: dest:i src1:i clob:1 len:11
@@ -255,8 +253,9 @@ float_clt_membase: dest:i src1:f src2:b len:35
 float_clt_un_membase: dest:i src1:f src2:b len:42
 float_conv_to_u: dest:i src1:f len:46
 fmove: dest:f src1:f len:8
-call_handler: len:14
+call_handler: len:14 clob:c nacl:52
 aot_const: dest:i len:10
+nacl_gc_safe_point: clob:c
 x86_test_null: src1:i len:5
 x86_compare_membase_reg: src1:b src2:i len:9
 x86_compare_membase_imm: src1:b len:13
@@ -272,7 +271,7 @@ x86_push_imm: len:6
 x86_push_membase: src1:b len:8
 x86_push_obj: src1:b len:40
 x86_lea: dest:i src1:i src2:i len:8
-x86_lea_membase: dest:i src1:i len:11
+x86_lea_membase: dest:i src1:i len:11 nacl:14
 x86_xchg: src1:i src2:i clob:x len:2
 x86_fpop: src1:f len:3
 x86_seteq_membase: src1:b len:9
@@ -298,8 +297,8 @@ atomic_exchange_i4: src1:b src2:i dest:a len:32
 atomic_add_i8: src1:b src2:i dest:i len:32
 atomic_add_new_i8: src1:b src2:i dest:i len:32
 atomic_exchange_i8: src1:b src2:i dest:a len:32
-atomic_cas_imm_i4: src1:b src2:i dest:a len:32
-atomic_cas_imm_i8: src1:b src2:i dest:a len:32
+atomic_cas_i4: src1:b src2:i src3:a dest:a len:24
+atomic_cas_i8: src1:b src2:i src3:a dest:a len:24
 memory_barrier: len:16
 adc: dest:i src1:i src2:i len:3 clob:1
 addcc: dest:i src1:i src2:i len:3 clob:1
@@ -307,22 +306,20 @@ subcc: dest:i src1:i src2:i len:3 clob:1
 adc_imm: dest:i src1:i len:8 clob:1
 sbb: dest:i src1:i src2:i len:3 clob:1
 sbb_imm: dest:i src1:i len:8 clob:1
-br_reg: src1:i len:3
+br_reg: src1:i len:3 nacl:8
 sin: dest:f src1:f len:32
 cos: dest:f src1:f len:32
 abs: dest:f src1:f clob:1 len:32
 tan: dest:f src1:f len:59
 atan: dest:f src1:f len:9
 sqrt: dest:f src1:f len:32
-bigmul: len:3 dest:i src1:a src2:i
-bigmul_un: len:3 dest:i src1:a src2:i
 sext_i1: dest:i src1:i len:4
 sext_i2: dest:i src1:i len:4
 sext_i4: dest:i src1:i len:8
 
 # 32 bit opcodes
-int_add: dest:i src1:i src2:i clob:1 len:4
-int_sub: dest:i src1:i src2:i clob:1 len:4
+int_add: dest:i src1:i src2:i clob:1 len:4 nacl:7
+int_sub: dest:i src1:i src2:i clob:1 len:4 nacl:7
 int_mul: dest:i src1:i src2:i clob:1 len:4
 int_mul_ovf: dest:i src1:i src2:i clob:1 len:32
 int_mul_ovf_un: dest:i src1:i src2:i clob:1 len:32
@@ -342,8 +339,8 @@ int_sbb: dest:i src1:i src2:i clob:1 len:4
 int_sbb_imm: dest:i src1:i clob:1 len:8
 int_addcc: dest:i src1:i src2:i clob:1 len:16
 int_subcc: dest:i src1:i src2:i clob:1 len:16
-int_add_imm: dest:i src1:i clob:1 len:8
-int_sub_imm: dest:i src1:i clob:1 len:8
+int_add_imm: dest:i src1:i clob:1 len:8 nacl:10
+int_sub_imm: dest:i src1:i clob:1 len:8 nacl:10
 int_mul_imm: dest:i src1:i clob:1 len:32
 int_div_imm: dest:a src1:i clob:d len:32
 int_div_un_imm: dest:a src1:i clob:d len:32
@@ -380,12 +377,14 @@ int_bge_un: len:8
 int_ble: len:8
 int_ble_un: len:8
 
+card_table_wbarrier: src1:a src2:i clob:d len:42
+
 relaxed_nop: len:2
 hard_nop: len:1
 
 # Linear IR opcodes
 nop: len:0
-dummy_use: len:0
+dummy_use: src1:i len:0
 dummy_store: len:0
 not_reached: len:0
 not_null: src1:i len:0
@@ -447,8 +446,8 @@ cmov_lgt_un: dest:i src1:i src2:i len:16 clob:1
 cmov_lle_un: dest:i src1:i src2:i len:16 clob:1
 cmov_llt_un: dest:i src1:i src2:i len:16 clob:1
 
-long_add_imm: dest:i src1:i clob:1 len:12
-long_sub_imm: dest:i src1:i clob:1 len:12
+long_add_imm: dest:i src1:i clob:1 len:12 nacl:15
+long_sub_imm: dest:i src1:i clob:1 len:12 nacl:15
 long_and_imm: dest:i src1:i clob:1 len:12
 long_or_imm: dest:i src1:i clob:1 len:12
 long_xor_imm: dest:i src1:i clob:1 len:12
@@ -495,6 +494,8 @@ vcall2: len:64 clob:c
 vcall2_reg: src1:i len:64 clob:c
 vcall2_membase: src1:b len:64 clob:c
 
+dyn_call: src1:i src2:i len:64 clob:c nacl:128
+
 localloc_imm: dest:i len:84
 
 load_mem: dest:i len:16
@@ -502,3 +503,213 @@ loadi8_mem: dest:i len:16
 loadi4_mem: dest:i len:16
 loadu1_mem: dest:i len:16
 loadu2_mem: dest:i len:16
+
+
+#SIMD
+
+addps: dest:x src1:x src2:x len:4 clob:1
+divps: dest:x src1:x src2:x len:4 clob:1
+mulps: dest:x src1:x src2:x len:4 clob:1
+subps: dest:x src1:x src2:x len:4 clob:1
+maxps: dest:x src1:x src2:x len:4 clob:1
+minps: dest:x src1:x src2:x len:4 clob:1
+compps: dest:x src1:x src2:x len:5 clob:1
+andps: dest:x src1:x src2:x len:4 clob:1
+andnps: dest:x src1:x src2:x len:4 clob:1
+orps: dest:x src1:x src2:x len:4 clob:1
+xorps: dest:x src1:x src2:x len:4 clob:1
+
+haddps: dest:x src1:x src2:x len:5 clob:1
+hsubps: dest:x src1:x src2:x len:5 clob:1
+addsubps: dest:x src1:x src2:x len:5 clob:1
+dupps_low: dest:x src1:x len:5
+dupps_high: dest:x src1:x len:5
+
+addpd: dest:x src1:x src2:x len:5 clob:1
+divpd: dest:x src1:x src2:x len:5 clob:1
+mulpd: dest:x src1:x src2:x len:5 clob:1
+subpd: dest:x src1:x src2:x len:5 clob:1
+maxpd: dest:x src1:x src2:x len:5 clob:1
+minpd: dest:x src1:x src2:x len:5 clob:1
+comppd: dest:x src1:x src2:x len:6 clob:1
+andpd: dest:x src1:x src2:x len:5 clob:1
+andnpd: dest:x src1:x src2:x len:5 clob:1
+orpd: dest:x src1:x src2:x len:5 clob:1
+xorpd: dest:x src1:x src2:x len:5 clob:1
+sqrtpd: dest:x src1:x len:5 clob:1
+
+haddpd: dest:x src1:x src2:x len:6 clob:1
+hsubpd: dest:x src1:x src2:x len:6 clob:1
+addsubpd: dest:x src1:x src2:x len:6 clob:1
+duppd: dest:x src1:x len:6
+
+pand: dest:x src1:x src2:x len:5 clob:1
+por: dest:x src1:x src2:x len:5 clob:1
+pxor: dest:x src1:x src2:x len:5 clob:1
+
+sqrtps: dest:x src1:x len:5
+rsqrtps: dest:x src1:x len:5
+rcpps: dest:x src1:x len:5
+
+pshufflew_high: dest:x src1:x len:6
+pshufflew_low: dest:x src1:x len:6
+pshuffled: dest:x src1:x len:6
+
+extract_mask: dest:i src1:x len:6
+
+paddb: dest:x src1:x src2:x len:5 clob:1
+paddw: dest:x src1:x src2:x len:5 clob:1
+paddd: dest:x src1:x src2:x len:5 clob:1
+paddq: dest:x src1:x src2:x len:5 clob:1
+
+psubb: dest:x src1:x src2:x len:5 clob:1
+psubw: dest:x src1:x src2:x len:5 clob:1
+psubd: dest:x src1:x src2:x len:5 clob:1
+psubq: dest:x src1:x src2:x len:5 clob:1
+
+pmaxb_un: dest:x src1:x src2:x len:5 clob:1
+pmaxw_un: dest:x src1:x src2:x len:6 clob:1
+pmaxd_un: dest:x src1:x src2:x len:6 clob:1
+
+pmaxb: dest:x src1:x src2:x len:6 clob:1
+pmaxw: dest:x src1:x src2:x len:5 clob:1
+pmaxd: dest:x src1:x src2:x len:6 clob:1
+
+pavgb_un: dest:x src1:x src2:x len:5 clob:1
+pavgw_un: dest:x src1:x src2:x len:5 clob:1
+
+pminb_un: dest:x src1:x src2:x len:5 clob:1
+pminw_un: dest:x src1:x src2:x len:6 clob:1
+pmind_un: dest:x src1:x src2:x len:6 clob:1
+
+pminb: dest:x src1:x src2:x len:6 clob:1
+pminw: dest:x src1:x src2:x len:5 clob:1
+pmind: dest:x src1:x src2:x len:6 clob:1
+
+pcmpeqb: dest:x src1:x src2:x len:5 clob:1
+pcmpeqw: dest:x src1:x src2:x len:5 clob:1
+pcmpeqd: dest:x src1:x src2:x len:5 clob:1
+pcmpeqq: dest:x src1:x src2:x len:6 clob:1
+
+pcmpgtb: dest:x src1:x src2:x len:5 clob:1
+pcmpgtw: dest:x src1:x src2:x len:5 clob:1
+pcmpgtd: dest:x src1:x src2:x len:5 clob:1
+pcmpgtq: dest:x src1:x src2:x len:6 clob:1
+
+psumabsdiff: dest:x src1:x src2:x len:5 clob:1
+
+unpack_lowb: dest:x src1:x src2:x len:5 clob:1
+unpack_loww: dest:x src1:x src2:x len:5 clob:1
+unpack_lowd: dest:x src1:x src2:x len:5 clob:1
+unpack_lowq: dest:x src1:x src2:x len:5 clob:1
+unpack_lowps: dest:x src1:x src2:x len:5 clob:1
+unpack_lowpd: dest:x src1:x src2:x len:5 clob:1
+
+unpack_highb: dest:x src1:x src2:x len:5 clob:1
+unpack_highw: dest:x src1:x src2:x len:5 clob:1
+unpack_highd: dest:x src1:x src2:x len:5 clob:1
+unpack_highq: dest:x src1:x src2:x len:5 clob:1
+unpack_highps: dest:x src1:x src2:x len:5 clob:1
+unpack_highpd: dest:x src1:x src2:x len:5 clob:1
+
+packw: dest:x src1:x src2:x len:5 clob:1 
+packd: dest:x src1:x src2:x len:5 clob:1 
+
+packw_un: dest:x src1:x src2:x len:5 clob:1 
+packd_un: dest:x src1:x src2:x len:6 clob:1 
+
+paddb_sat: dest:x src1:x src2:x len:5 clob:1
+paddb_sat_un: dest:x src1:x src2:x len:5 clob:1
+
+paddw_sat: dest:x src1:x src2:x len:5 clob:1
+paddw_sat_un: dest:x src1:x src2:x len:5 clob:1
+
+psubb_sat: dest:x src1:x src2:x len:5 clob:1
+psubb_sat_un: dest:x src1:x src2:x len:5 clob:1
+
+psubw_sat: dest:x src1:x src2:x len:5 clob:1
+psubw_sat_un: dest:x src1:x src2:x len:5 clob:1
+
+pmulw: dest:x src1:x src2:x len:5 clob:1
+pmuld: dest:x src1:x src2:x len:6 clob:1
+pmulq: dest:x src1:x src2:x len:5 clob:1
+
+pmul_high_un: dest:x src1:x src2:x len:5 clob:1
+pmul_high: dest:x src1:x src2:x len:5 clob:1
+
+pshrw: dest:x src1:x len:6 clob:1
+pshrw_reg: dest:x src1:x src2:x len:5 clob:1
+
+psarw: dest:x src1:x len:6 clob:1
+psarw_reg: dest:x src1:x src2:x len:5 clob:1
+
+pshlw: dest:x src1:x len:6 clob:1
+pshlw_reg: dest:x src1:x src2:x len:5 clob:1
+
+pshrd: dest:x src1:x len:6 clob:1
+pshrd_reg: dest:x src1:x src2:x len:5 clob:1
+
+psard: dest:x src1:x len:6 clob:1
+psard_reg: dest:x src1:x src2:x len:5 clob:1
+
+pshld: dest:x src1:x len:6 clob:1
+pshld_reg: dest:x src1:x src2:x len:5 clob:1
+
+pshrq: dest:x src1:x len:6 clob:1
+pshrq_reg: dest:x src1:x src2:x len:5 clob:1
+
+pshlq: dest:x src1:x len:6 clob:1
+pshlq_reg: dest:x src1:x src2:x len:5 clob:1
+
+xmove: dest:x src1:x len:5
+xzero: dest:x len:5
+
+iconv_to_x: dest:x src1:i len:5
+extract_i4: dest:i src1:x len:5
+
+extract_i8: dest:i src1:x len:9
+
+extract_i2: dest:i src1:x len:13
+extract_u2: dest:i src1:x len:13
+extract_i1: dest:i src1:x len:13
+extract_u1: dest:i src1:x len:13
+extract_r8: dest:f src1:x len:5 
+
+iconv_to_r8_raw: dest:f src1:i len:10
+
+insert_i2: dest:x src1:x src2:i len:6 clob:1
+
+extractx_u2: dest:i src1:x len:6
+insertx_u1_slow: dest:x src1:i src2:i len:18 clob:x
+
+insertx_i4_slow: dest:x src1:x src2:i len:16 clob:x
+insertx_i8_slow: dest:x src1:x src2:i len:13
+insertx_r4_slow: dest:x src1:x src2:f len:24
+insertx_r8_slow: dest:x src1:x src2:f len:24
+
+loadx_membase: dest:x src1:b len:9
+storex_membase: dest:b src1:x len:9
+storex_membase_reg: dest:b src1:x len:9
+
+loadx_aligned_membase: dest:x src1:b len:7
+storex_aligned_membase_reg: dest:b src1:x len:7
+storex_nta_membase_reg: dest:b src1:x len:7
+
+fconv_to_r8_x: dest:x src1:f len:4 
+xconv_r8_to_i4: dest:y src1:x len:7
+
+prefetch_membase: src1:b len:4
+
+expand_i2: dest:x src1:i len:18
+expand_i4: dest:x src1:i len:11
+expand_i8: dest:x src1:i len:11
+expand_r4: dest:x src1:f len:16
+expand_r8: dest:x src1:f len:13
+
+liverange_start: len:0
+liverange_end: len:0
+gc_liveness_def: len:0
+gc_liveness_use: len:0
+gc_spill_slot_liveness_def: len:0
+gc_param_slot_liveness_def: len:0
+