bugfixes and more features

This commit is contained in:
JanLJL
2022-09-27 18:04:59 +02:00
parent d5f1654aa8
commit 4f8e37d9fd
2 changed files with 437 additions and 30 deletions

View File

@@ -16,8 +16,11 @@ class EntryBuilder:
def classify(operands_types): def classify(operands_types):
load = "mem" in operands_types[:-1] load = "mem" in operands_types[:-1]
store = "mem" in operands_types[-1:] store = "mem" in operands_types[-1:]
vec = False
if any([vecr in operands_types for vecr in ["mm", "xmm", "ymm", "zmm"]]):
vec = True
assert not (load and store), "Can not process a combined load-store instruction." assert not (load and store), "Can not process a combined load-store instruction."
return load, store return load, store, vec
def build_description( def build_description(
self, instruction_name, operand_types, port_pressure=[], latency=0, comment=None self, instruction_name, operand_types, port_pressure=[], latency=0, comment=None
@@ -26,7 +29,7 @@ class EntryBuilder:
comment = " # " + comment comment = " # " + comment
else: else:
comment = "" comment = ""
description = "- name: {}{}\n operands:\n".format(instruction_name, comment) description = "- name: {}{}\n operands: {}\n".format(instruction_name, comment, "[]" if len(operand_types) == 0 else "")
for ot in operand_types: for ot in operand_types:
if ot == "imd": if ot == "imd":
@@ -64,14 +67,18 @@ class EntryBuilder:
def parse_port_pressure(self, port_pressure_str): def parse_port_pressure(self, port_pressure_str):
""" """
Example: Example:
1*p45+2*p0 -> [[1, '45'], [2, '0']] 1*p45+2*p0+2*p10,11 -> [[1, '45'], [2, '0'], [2, ['10', '11']]]
""" """
port_pressure = [] port_pressure = []
if port_pressure_str: if port_pressure_str:
for p in port_pressure_str.split("+"): for p in port_pressure_str.split("+"):
cycles, ports = p.split("*p") cycles, ports = p.split("*p")
if ports.startswith("(") and ports.endswith(")"): ports = ports.split(",")
ports = ports[1:-1].split(",") if len(ports) == 1:
ports = ports[0]
else:
ports = list(filter(lambda p: len(p) > 0, ports))
port_pressure.append([int(cycles), ports]) port_pressure.append([int(cycles), ports])
return port_pressure return port_pressure
@@ -92,23 +99,42 @@ class EntryBuilder:
return self.build_description(instruction_name, operand_types, port_pressure, latency) return self.build_description(instruction_name, operand_types, port_pressure, latency)
class EntryBuilderIntelPort9(EntryBuilder): class ArchEntryBuilder(EntryBuilder):
# for ICX
def build_description(self, instruction_name, operand_types, port_pressure=[], latency=0): def build_description(self, instruction_name, operand_types, port_pressure=[], latency=0):
load, store = self.classify(operand_types) # Intel ICX
# LD_pressure = [[1, "23"], [1, ["2D", "3D"]]]
# LD_pressure_vec = LD_pressure
# ST_pressure = [[1, "79"], [1, "48"]]
# ST_pressure_vec = ST_pressure
# LD_lat = 5
# ST_lat = 0
# Zen3
LD_pressure = [[1, ["11", "12", "13"]]]
LD_pressure_vec = [[1, ["11", "12"]]]
ST_pressure = [[1, ["12", "13"]]]
ST_pressure_vec = [[1, ["4"]], [1, ["13"]]]
LD_lat = 4
ST_lat = 0
load, store, vec = self.classify(operand_types)
if load: if load:
port_pressure += [[1, "23"], [1, ["2D", "3D"]]] if vec:
latency += 5 port_pressure += LD_pressure_vec
else:
port_pressure += LD_pressure
latency += LD_lat
comment = "with load" comment = "with load"
return EntryBuilder.build_description( return EntryBuilder.build_description(
self, instruction_name, operand_types, port_pressure, latency, comment self, instruction_name, operand_types, port_pressure, latency, comment
) )
if store: if store:
port_pressure = port_pressure + [[1, "79"], [1, "48"]] if vec:
port_pressure = port_pressure + ST_pressure_vec
else:
port_pressure = port_pressure + ST_pressure
operands = ["mem" if o == "mem" else o for o in operand_types] operands = ["mem" if o == "mem" else o for o in operand_types]
latency += 0 latency += ST_lat
return EntryBuilder.build_description( return EntryBuilder.build_description(
self, self,
instruction_name, instruction_name,
@@ -125,7 +151,7 @@ class EntryBuilderIntelPort9(EntryBuilder):
def get_description(instruction_form, port_pressure, latency, rhs_comment=None): def get_description(instruction_form, port_pressure, latency, rhs_comment=None):
entry = EntryBuilderIntelPort9().process_item(instruction_form, (port_pressure, latency)) entry = ArchEntryBuilder().process_item(instruction_form, (port_pressure, latency))
if rhs_comment is not None: if rhs_comment is not None:
max_length = max([len(line) for line in entry.split("\n")]) max_length = max([len(line) for line in entry.split("\n")])

View File

@@ -16,8 +16,11 @@ class MOVEntryBuilder:
def classify(operands_types): def classify(operands_types):
load = "mem" in operands_types[:-1] load = "mem" in operands_types[:-1]
store = "mem" in operands_types[-1:] store = "mem" in operands_types[-1:]
vec = False
if any([vecr in operands_types for vecr in ["mm", "xmm", "ymm", "zmm"]]):
vec = True
assert not (load and store), "Can not process a combined load-store instruction." assert not (load and store), "Can not process a combined load-store instruction."
return load, store return load, store, vec
def build_description( def build_description(
self, instruction_name, operand_types, port_pressure=[], latency=0, comment=None self, instruction_name, operand_types, port_pressure=[], latency=0, comment=None
@@ -65,6 +68,10 @@ class MOVEntryBuilder:
if port_pressure_str: if port_pressure_str:
for p in port_pressure_str.split("+"): for p in port_pressure_str.split("+"):
cycles, ports = p.split("*p") cycles, ports = p.split("*p")
ports = ports.split(",")
if len(ports) == 1:
ports = ports[0]
port_pressure.append([int(cycles), ports]) port_pressure.append([int(cycles), ports])
return port_pressure return port_pressure
@@ -84,7 +91,7 @@ class MOVEntryBuilder:
class MOVEntryBuilderIntelNoPort7AGU(MOVEntryBuilder): class MOVEntryBuilderIntelNoPort7AGU(MOVEntryBuilder):
# for SNB and IVB # for SNB and IVB
def build_description(self, instruction_name, operand_types, port_pressure=[], latency=0): def build_description(self, instruction_name, operand_types, port_pressure=[], latency=0):
load, store = self.classify(operand_types) load, store, vec = self.classify(operand_types)
comment = None comment = None
if load: if load:
@@ -111,9 +118,8 @@ class MOVEntryBuilderIntelNoPort7AGU(MOVEntryBuilder):
class MOVEntryBuilderIntelPort9(MOVEntryBuilder): class MOVEntryBuilderIntelPort9(MOVEntryBuilder):
# for ICX # for ICX
def build_description(self, instruction_name, operand_types, port_pressure=[], latency=0): def build_description(self, instruction_name, operand_types, port_pressure=[], latency=0):
load, store = self.classify(operand_types) load, store, vec = self.classify(operand_types)
if load: if load:
port_pressure += [[1, "23"], [1, ["2D", "3D"]]] port_pressure += [[1, "23"], [1, ["2D", "3D"]]]
@@ -141,6 +147,381 @@ class MOVEntryBuilderIntelPort9(MOVEntryBuilder):
) )
class MOVEntryBuilderAMDZen3(MOVEntryBuilder):
# for Zen 3
def build_description(self, instruction_name, operand_types, port_pressure=[], latency=0):
load, store, vec = self.classify(operand_types)
if load and vec:
port_pressure += [[1, ["11", "12"]]]
latency += 4
comment = "with load"
return MOVEntryBuilder.build_description(
self, instruction_name, operand_types, port_pressure, latency, comment
)
elif load:
port_pressure += [[1, ["11", "12", "13"]]]
latency += 4
comment = "with load"
return MOVEntryBuilder.build_description(
self, instruction_name, operand_types, port_pressure, latency, comment
)
if store and vec:
port_pressure = port_pressure + [[1, ["4"]], [1, ["13"]]]
operands = ["mem" if o == "mem" else o for o in operand_types]
latency += 0
return MOVEntryBuilder.build_description(
self,
instruction_name,
operands,
port_pressure,
latency,
"with store",
)
elif store:
port_pressure = port_pressure + [[1, ["12", "13"]]]
operands = ["mem" if o == "mem" else o for o in operand_types]
latency += 0
return MOVEntryBuilder.build_description(
self,
instruction_name,
operands,
port_pressure,
latency,
"with store",
)
# Register only:
return MOVEntryBuilder.build_description(
self, instruction_name, operand_types, port_pressure, latency
)
#############################################################################
z3 = MOVEntryBuilderAMDZen3()
zen3_mov_instructions = [
# https://www.felixcloutier.com/x86/mov
("mov gpr gpr", ("1*p6789", 1)),
("mov gpr mem", ("", 0)),
("mov mem gpr", ("", 0)),
("mov imd gpr", ("1*p6789", 1)),
("mov imd mem", ("", 0)),
("movabs imd gpr", ("1*p6789", 1)), # AT&T version, port util to be verified
# https://www.felixcloutier.com/x86/movapd
("movapd xmm xmm", ("1*p0123", 1)),
("movapd xmm mem", ("", 0)),
("movapd mem xmm", ("", 0)),
("vmovapd xmm xmm", ("1*p0123", 1)),
("vmovapd xmm mem", ("", 0)),
("vmovapd mem xmm", ("", 0)),
("vmovapd ymm ymm", ("1*p0123", 1)),
("vmovapd ymm mem", ("", 0)),
("vmovapd mem ymm", ("", 0)),
# https://www.felixcloutier.com/x86/movaps
("movaps xmm xmm", ("1*p0123", 1)),
("movaps xmm mem", ("", 0)),
("movaps mem xmm", ("", 0)),
("vmovaps xmm xmm", ("1*p0123", 1)),
("vmovaps xmm mem", ("", 0)),
("vmovaps mem xmm", ("", 0)),
("vmovaps ymm ymm", ("1*p0123", 1)),
("vmovaps ymm mem", ("", 0)),
("vmovaps mem ymm", ("", 0)),
# https://www.felixcloutier.com/x86/movd:movq
("movd gpr mm", ("1*p0123", 1)),
("movd mem mm", ("", 0)),
("movq gpr mm", ("1*p0123", 1)),
("movq mem mm", ("", 0)),
("movd mm gpr", ("1*p0123", 1)),
("movd mm mem", ("", 0)),
("movq mm gpr", ("1*p0123", 1)),
("movq mm mem", ("", 0)),
("movd gpr xmm", ("1*p0123", 1)),
("movd mem xmm", ("", 0)),
("movq gpr xmm", ("1*p0123", 1)),
("movq mem xmm", ("", 0)),
("movd xmm gpr", ("1*p0123", 1)),
("movd xmm mem", ("", 0)),
("movq xmm gpr", ("1*p0123", 1)),
("movq xmm mem", ("", 0)),
("vmovd gpr xmm", ("1*p0123", 1)),
("vmovd mem xmm", ("", 0)),
("vmovq gpr xmm", ("1*p0123", 1)),
("vmovq mem xmm", ("", 0)),
("vmovd xmm gpr", ("1*p0123", 1)),
("vmovd xmm mem", ("", 0)),
("vmovq xmm gpr", ("1*p0123", 1)),
("vmovq xmm mem", ("", 0)),
# https://www.felixcloutier.com/x86/movddup
("movddup xmm xmm", ("1*p12", 1)),
("movddup mem xmm", ("", 0)),
("vmovddup xmm xmm", ("1*p12", 1)),
("vmovddup mem xmm", ("", 0)),
("vmovddup ymm ymm", ("1*p12", 1)),
("vmovddup mem ymm", ("", 0)),
# https://www.felixcloutier.com/x86/movdq2q
("movdq2q xmm mm", ("1*p0123", 1)),
# https://www.felixcloutier.com/x86/movdqa:vmovdqa32:vmovdqa64
("movdqa xmm xmm", ("1*p0123", 1)),
("movdqa mem xmm", ("", 0)),
("movdqa xmm mem", ("", 0)),
("vmovdqa xmm xmm", ("1*p0123", 1)),
("vmovdqa mem xmm", ("", 0)),
("vmovdqa xmm mem", ("", 0)),
("vmovdqa ymm ymm", ("1*p0123", 1)),
("vmovdqa mem ymm", ("", 0)),
("vmovdqa ymm mem", ("", 0)),
# https://www.felixcloutier.com/x86/movdqu:vmovdqu8:vmovdqu16:vmovdqu32:vmovdqu64
("movdqu xmm xmm", ("1*p0123", 1)),
("movdqu mem xmm", ("", 0)),
("movdqu xmm mem", ("", 0)),
("vmovdqu xmm xmm", ("1*p0123", 1)),
("vmovdqu mem xmm", ("", 0)),
("vmovdqu xmm mem", ("", 0)),
("vmovdqu ymm ymm", ("1*p0123", 1)),
("vmovdqu mem ymm", ("", 0)),
("vmovdqu ymm mem", ("", 0)),
# https://www.felixcloutier.com/x86/movhlps
("movhlps xmm xmm", ("1*p12", 1)),
("vmovhlps xmm xmm xmm", ("1*p12", 1)),
# https://www.felixcloutier.com/x86/movhpd
("movhpd mem xmm", ("1*p12", 1)),
("vmovhpd mem xmm xmm", ("1*p12", 1)),
("movhpd xmm mem", ("", 0)),
("vmovhpd mem xmm", ("", 0)),
# https://www.felixcloutier.com/x86/movhps
("movhps mem xmm", ("1*p12", 1)),
("vmovhps mem xmm xmm", ("1*p12", 1)),
("movhps xmm mem", ("", 0)),
("vmovhps mem xmm", ("", 0)),
# https://www.felixcloutier.com/x86/movlhps
("movlhps xmm xmm", ("1*p12", 1)),
("vmovlhps xmm xmm xmm", ("1*p12", 1)),
# https://www.felixcloutier.com/x86/movlpd
("movlpd mem xmm", ("1*p12", 1)),
("vmovlpd mem xmm xmm", ("1*p12", 1)),
("movlpd xmm mem", ("1*p12", 0)),
("vmovlpd mem xmm", ("1*p12", 1)),
# https://www.felixcloutier.com/x86/movlps
("movlps mem xmm", ("1*p12", 1)),
("vmovlps mem xmm xmm", ("1*p12", 1)),
("movlps xmm mem", ("1*p12", 0)),
("vmovlps mem xmm", ("1*p12", 1)),
# https://www.felixcloutier.com/x86/movmskpd
("movmskpd xmm gpr", ("1*p0123", 1)),
("vmovmskpd xmm gpr", ("1*p0123", 1)),
("vmovmskpd ymm gpr", ("1*p0123", 1)),
# https://www.felixcloutier.com/x86/movmskps
("movmskps xmm gpr", ("1*p0123", 1)),
("vmovmskps xmm gpr", ("1*p0123", 1)),
("vmovmskps ymm gpr", ("1*p0123", 1)),
# https://www.felixcloutier.com/x86/movntdq
("movntdq xmm mem", ("", 0)), # TODO NT-store: what latency to use?
("vmovntdq xmm mem", ("", 0)), # TODO NT-store: what latency to use?
("vmovntdq ymm mem", ("", 0)), # TODO NT-store: what latency to use?
# https://www.felixcloutier.com/x86/movntdqa
("movntdqa mem xmm", ("", 0)), # TODO NT-store: what latency to use?
("vmovntdqa mem xmm", ("", 0)), # TODO NT-store: what latency to use?
("vmovntdqa mem ymm", ("", 0)), # TODO NT-store: what latency to use?
# https://www.felixcloutier.com/x86/movnti
("movnti gpr mem", ("", 0)), # TODO NT-store: what latency to use?
# https://www.felixcloutier.com/x86/movntpd
("movntpd xmm mem", ("", 0)), # TODO NT-store: what latency to use?
("vmovntpd xmm mem", ("", 0)), # TODO NT-store: what latency to use?
("vmovntpd ymm mem", ("", 0)), # TODO NT-store: what latency to use?
# https://www.felixcloutier.com/x86/movntps
("movntps xmm mem", ("", 0)), # TODO NT-store: what latency to use?
("vmovntps xmm mem", ("", 0)), # TODO NT-store: what latency to use?
("vmovntps ymm mem", ("", 0)), # TODO NT-store: what latency to use?
# https://www.felixcloutier.com/x86/movntq
("movntq mm mem", ("", 0)), # TODO NT-store: what latency to use?
# https://www.felixcloutier.com/x86/movq
("movq mm mm", ("", 0)),
("movq mem mm", ("", 0)),
("movq mm mem", ("", 0)),
("movq xmm xmm", ("1*p0123", 1)),
("movq mem xmm", ("", 0)),
("movq xmm mem", ("", 0)),
("vmovq xmm xmm", ("1*p0123", 1)),
("vmovq mem xmm", ("", 0)),
("vmovq xmm mem", ("", 0)),
# https://www.felixcloutier.com/x86/movs:movsb:movsw:movsd:movsq
# TODO combined load-store is currently not supported
# ('movs mem mem', ()),
# https://www.felixcloutier.com/x86/movsd
("movsd xmm xmm", ("1*p0123", 1)),
("movsd mem xmm", ("", 0)),
("movsd xmm mem", ("", 0)),
("vmovsd xmm xmm xmm", ("1*p0123", 1)),
("vmovsd mem xmm", ("", 0)),
("vmovsd xmm mem", ("", 0)),
# https://www.felixcloutier.com/x86/movshdup
("movshdup xmm xmm", ("1*p12", 1)),
("movshdup mem xmm", ("", 0)),
("vmovshdup xmm xmm", ("1*p12", 1)),
("vmovshdup mem xmm", ("", 0)),
("vmovshdup ymm ymm", ("1*p12", 1)),
("vmovshdup mem ymm", ("", 0)),
# https://www.felixcloutier.com/x86/movsldup
("movsldup xmm xmm", ("1*p12", 1)),
("movsldup mem xmm", ("", 0)),
("vmovsldup xmm xmm", ("1*p12", 1)),
("vmovsldup mem xmm", ("", 0)),
("vmovsldup ymm ymm", ("1*p12", 1)),
("vmovsldup mem ymm", ("", 0)),
# https://www.felixcloutier.com/x86/movss
("movss xmm xmm", ("1*p0123", 1)),
("movss mem xmm", ("", 0)),
("vmovss xmm xmm xmm", ("1*p0123", 1)),
("vmovss mem xmm", ("", 0)),
("vmovss xmm xmm", ("1*p0123", 1)),
("vmovss xmm mem", ("", 0)),
("movss mem xmm", ("", 0)),
# https://www.felixcloutier.com/x86/movsx:movsxd
("movsx gpr gpr", ("1*p6789", 1)),
("movsx mem gpr", ("", 0)),
("movsxd gpr gpr", ("", 0)),
("movsxd mem gpr", ("", 0)),
("movsb gpr gpr", ("1*p6789", 1)), # AT&T version
("movsb mem gpr", ("", 0)), # AT&T version
("movsw gpr gpr", ("1*p6789", 1)), # AT&T version
("movsw mem gpr", ("", 0)), # AT&T version
("movsl gpr gpr", ("1*p6789", 1)), # AT&T version
("movsl mem gpr", ("", 0)), # AT&T version
("movsq gpr gpr", ("1*p6789", 1)), # AT&T version
("movsq mem gpr", ("", 0)), # AT&T version
# https://www.felixcloutier.com/x86/movupd
("movupd xmm xmm", ("1*p0123", 1)),
("movupd mem xmm", ("", 0)),
("movupd xmm mem", ("", 0)),
("vmovupd xmm xmm", ("1*p0123", 1)),
("vmovupd mem xmm", ("", 0)),
("vmovupd xmm mem", ("", 0)),
("vmovupd ymm ymm", ("1*p0123", 1)),
("vmovupd mem ymm", ("", 0)),
("vmovupd ymm mem", ("", 0)),
# https://www.felixcloutier.com/x86/movups
("movups xmm xmm", ("1*p0123", 1)),
("movups mem xmm", ("", 0)),
("movups xmm mem", ("", 0)),
("vmovups xmm xmm", ("1*p0123", 1)),
("vmovups mem xmm", ("", 0)),
("vmovups xmm mem", ("", 0)),
("vmovups ymm ymm", ("1*p0123", 1)),
("vmovups mem ymm", ("", 0)),
("vmovups ymm mem", ("", 0)),
# https://www.felixcloutier.com/x86/movzx
("movzx gpr gpr", ("1*p6789", 1)),
("movzx mem gpr", ("", 0)),
("movzb gpr gpr", ("1*p6789", 1)), # AT&T version
("movzb mem gpr", ("", 0)), # AT&T version
("movzw gpr gpr", ("1*p6789", 1)), # AT&T version
("movzw mem gpr", ("", 0)), # AT&T version
("movzl gpr gpr", ("1*p6789", 1)), # AT&T version
("movzl mem gpr", ("", 0)), # AT&T version
("movzq gpr gpr", ("1*p6789", 1)), # AT&T version
("movzq mem gpr", ("", 0)), # AT&T version
# https://www.felixcloutier.com/x86/cmovcc
("cmova gpr gpr", ("1*p69", 1)),
("cmova mem gpr", ("", 0)),
("cmovae gpr gpr", ("1*p69", 1)),
("cmovae mem gpr", ("", 0)),
("cmovb gpr gpr", ("1*p69", 1)),
("cmovb mem gpr", ("", 0)),
("cmovbe gpr gpr", ("1*p69", 1)),
("cmovbe mem gpr", ("", 0)),
("cmovc gpr gpr", ("1*p69", 1)),
("cmovc mem gpr", ("", 0)),
("cmove gpr gpr", ("1*p69", 1)),
("cmove mem gpr", ("", 0)),
("cmovg gpr gpr", ("1*p69", 1)),
("cmovg mem gpr", ("", 0)),
("cmovge gpr gpr", ("1*p69", 1)),
("cmovge mem gpr", ("", 0)),
("cmovl gpr gpr", ("1*p69", 1)),
("cmovl mem gpr", ("", 0)),
("cmovle gpr gpr", ("1*p69", 1)),
("cmovle mem gpr", ("", 0)),
("cmovna gpr gpr", ("1*p69", 1)),
("cmovna mem gpr", ("", 0)),
("cmovnae gpr gpr", ("1*p69", 1)),
("cmovnae mem gpr", ("", 0)),
("cmovnb gpr gpr", ("1*p69", 1)),
("cmovnb mem gpr", ("", 0)),
("cmovnbe gpr gpr", ("1*p69", 1)),
("cmovnbe mem gpr", ("", 0)),
("cmovnc gpr gpr", ("1*p69", 1)),
("cmovnc mem gpr", ("", 0)),
("cmovne gpr gpr", ("1*p69", 1)),
("cmovne mem gpr", ("", 0)),
("cmovng gpr gpr", ("1*p69", 1)),
("cmovng mem gpr", ("", 0)),
("cmovnge gpr gpr", ("1*p69", 1)),
("cmovnge mem gpr", ("", 0)),
("cmovnl gpr gpr", ("1*p69", 1)),
("cmovnl mem gpr", ("", 0)),
("cmovno gpr gpr", ("1*p69", 1)),
("cmovno mem gpr", ("", 0)),
("cmovnp gpr gpr", ("1*p69", 1)),
("cmovnp mem gpr", ("", 0)),
("cmovns gpr gpr", ("1*p69", 1)),
("cmovns mem gpr", ("", 0)),
("cmovnz gpr gpr", ("1*p69", 1)),
("cmovnz mem gpr", ("", 0)),
("cmovo gpr gpr", ("1*p69", 1)),
("cmovo mem gpr", ("", 0)),
("cmovp gpr gpr", ("1*p69", 1)),
("cmovp mem gpr", ("", 0)),
("cmovpe gpr gpr", ("1*p69", 1)),
("cmovpe mem gpr", ("", 0)),
("cmovpo gpr gpr", ("1*p69", 1)),
("cmovpo mem gpr", ("", 0)),
("cmovs gpr gpr", ("1*p69", 1)),
("cmovs mem gpr", ("", 0)),
("cmovz gpr gpr", ("1*p69", 1)),
("cmovz mem gpr", ("", 0)),
# https://www.felixcloutier.com/x86/pmovmskb
("pmovmskb mm gpr", ("1*p0123", 1)),
("pmovmskb xmm gpr", ("1*p0123", 1)),
("vpmovmskb xmm gpr", ("1*p0123", 1)),
# https://www.felixcloutier.com/x86/pmovsx
("pmovsxbw xmm xmm", ("1*p12", 1)),
("pmovsxbw mem xmm", ("1*p12", 1)),
("pmovsxbd xmm xmm", ("1*p12", 1)),
("pmovsxbd mem xmm", ("1*p12", 1)),
("pmovsxbq xmm xmm", ("1*p12", 1)),
("pmovsxbq mem xmm", ("1*p12", 1)),
("vpmovsxbw xmm xmm", ("1*p12", 1)),
("vpmovsxbw mem xmm", ("1*p12", 1)),
("vpmovsxbd xmm xmm", ("1*p12", 1)),
("vpmovsxbd mem xmm", ("1*p12", 1)),
("vpmovsxbq xmm xmm", ("1*p12", 1)),
("vpmovsxbq mem xmm", ("1*p12", 1)),
("vpmovsxbw xmm ymm", ("1*p0123", 1)),
("vpmovsxbw mem ymm", ("1*p12", 1)),
("vpmovsxbd xmm ymm", ("1*p0123", 1)),
("vpmovsxbd mem ymm", ("1*p12", 1)),
("vpmovsxbq xmm ymm", ("1*p0123", 1)),
("vpmovsxbq mem ymm", ("1*p12", 1)),
# https://www.felixcloutier.com/x86/pmovzx
("pmovzxbw xmm xmm", ("1*p12", 1)),
("pmovzxbw mem xmm", ("1*p12", 1)),
("vpmovzxbw xmm xmm", ("1*p12", 1)),
("vpmovzxbw mem xmm", ("1*p12", 1)),
("vpmovzxbw xmm ymm", ("1*p0123", 1)),
("vpmovzxbw mem ymm", ("1*p12", 1)),
#################################################################
# https://www.felixcloutier.com/x86/movbe
("movbe gpr mem", ("1*p67", 5)),
("movbe mem gpr", ("1*p67", 5)),
################################################
# https://www.felixcloutier.com/x86/movq2dq
("movq2dq mm xmm", ("2*p0123", 1)),
]
p9 = MOVEntryBuilderIntelPort9() p9 = MOVEntryBuilderIntelPort9()
icx_mov_instructions = [ icx_mov_instructions = [
@@ -517,11 +898,11 @@ icx_mov_instructions = [
("vpmovsxbd mem xmm", ("1*p15", 1)), ("vpmovsxbd mem xmm", ("1*p15", 1)),
("vpmovsxbq xmm xmm", ("1*p15", 1)), ("vpmovsxbq xmm xmm", ("1*p15", 1)),
("vpmovsxbq mem xmm", ("1*p15", 1)), ("vpmovsxbq mem xmm", ("1*p15", 1)),
("vpmovsxbw ymm ymm", ("1*p5", 1)), ("vpmovsxbw xmm ymm", ("1*p5", 1)),
("vpmovsxbw mem ymm", ("1*p5", 1)), ("vpmovsxbw mem ymm", ("1*p5", 1)),
("vpmovsxbd ymm ymm", ("1*p5", 1)), ("vpmovsxbd xmm ymm", ("1*p5", 1)),
("vpmovsxbd mem ymm", ("1*p5", 1)), ("vpmovsxbd mem ymm", ("1*p5", 1)),
("vpmovsxbq ymm ymm", ("1*p5", 1)), ("vpmovsxbq xmm ymm", ("1*p5", 1)),
("vpmovsxbq mem ymm", ("1*p5", 1)), ("vpmovsxbq mem ymm", ("1*p5", 1)),
("vpmovsxbw ymm zmm", ("1*p5", 3)), ("vpmovsxbw ymm zmm", ("1*p5", 3)),
("vpmovsxbw mem zmm", ("1*p5", 1)), ("vpmovsxbw mem zmm", ("1*p5", 1)),
@@ -572,7 +953,7 @@ class MOVEntryBuilderIntelWithPort7AGU(MOVEntryBuilder):
# for HSW, BDW, SKX and CSX # for HSW, BDW, SKX and CSX
def build_description(self, instruction_name, operand_types, port_pressure=[], latency=0): def build_description(self, instruction_name, operand_types, port_pressure=[], latency=0):
load, store = self.classify(operand_types) load, store, vec = self.classify(operand_types)
if load: if load:
port_pressure += [[1, "23"], [1, ["2D", "3D"]]] port_pressure += [[1, "23"], [1, ["2D", "3D"]]]
@@ -794,7 +1175,6 @@ snb_mov_instructions = [
("movss mem xmm", ("", 0)), ("movss mem xmm", ("", 0)),
("vmovss xmm xmm xmm", ("1*p5", 1)), ("vmovss xmm xmm xmm", ("1*p5", 1)),
("vmovss mem xmm", ("", 0)), ("vmovss mem xmm", ("", 0)),
("vmovss xmm xmm", ("1*p5", 1)),
("vmovss xmm mem", ("", 0)), ("vmovss xmm mem", ("", 0)),
("movss mem xmm", ("", 0)), ("movss mem xmm", ("", 0)),
# https://www.felixcloutier.com/x86/movsx:movsxd # https://www.felixcloutier.com/x86/movsx:movsxd
@@ -919,11 +1299,11 @@ snb_mov_instructions = [
("vpmovsxbd mem xmm", ("1*p15", 1)), ("vpmovsxbd mem xmm", ("1*p15", 1)),
("vpmovsxbq xmm xmm", ("1*p15", 1)), ("vpmovsxbq xmm xmm", ("1*p15", 1)),
("vpmovsxbq mem xmm", ("1*p15", 1)), ("vpmovsxbq mem xmm", ("1*p15", 1)),
("vpmovsxbw ymm ymm", ("1*p15", 1)), ("vpmovsxbw xmm ymm", ("1*p15", 1)),
("vpmovsxbw mem ymm", ("1*p15", 1)), ("vpmovsxbw mem ymm", ("1*p15", 1)),
("vpmovsxbd ymm ymm", ("1*p15", 1)), ("vpmovsxbd xmm ymm", ("1*p15", 1)),
("vpmovsxbd mem ymm", ("1*p15", 1)), ("vpmovsxbd mem ymm", ("1*p15", 1)),
("vpmovsxbq ymm ymm", ("1*p15", 1)), ("vpmovsxbq xmm ymm", ("1*p15", 1)),
("vpmovsxbq mem ymm", ("1*p15", 1)), ("vpmovsxbq mem ymm", ("1*p15", 1)),
# https://www.felixcloutier.com/x86/pmovzx # https://www.felixcloutier.com/x86/pmovzx
("pmovzxbw xmm xmm", ("1*p15", 1)), ("pmovzxbw xmm xmm", ("1*p15", 1)),
@@ -1079,11 +1459,11 @@ hsw_mov_instructions = list(
("vpmovsxbd mem xmm", ("1*p5", 1)), ("vpmovsxbd mem xmm", ("1*p5", 1)),
("vpmovsxbq xmm xmm", ("1*p5", 1)), ("vpmovsxbq xmm xmm", ("1*p5", 1)),
("vpmovsxbq mem xmm", ("1*p5", 1)), ("vpmovsxbq mem xmm", ("1*p5", 1)),
("vpmovsxbw ymm ymm", ("1*p5", 1)), ("vpmovsxbw xmm ymm", ("1*p5", 1)),
("vpmovsxbw mem ymm", ("1*p5", 1)), ("vpmovsxbw mem ymm", ("1*p5", 1)),
("vpmovsxbd ymm ymm", ("1*p5", 1)), ("vpmovsxbd xmm ymm", ("1*p5", 1)),
("vpmovsxbd mem ymm", ("1*p5", 1)), ("vpmovsxbd mem ymm", ("1*p5", 1)),
("vpmovsxbq ymm ymm", ("1*p5", 1)), ("vpmovsxbq xmm ymm", ("1*p5", 1)),
("vpmovsxbq mem ymm", ("1*p5", 1)), ("vpmovsxbq mem ymm", ("1*p5", 1)),
# https://www.felixcloutier.com/x86/pmovzx # https://www.felixcloutier.com/x86/pmovzx
("pmovzxbw xmm xmm", ("1*p5", 1)), ("pmovzxbw xmm xmm", ("1*p5", 1)),
@@ -1234,6 +1614,7 @@ def get_description(arch, rhs_comment=None):
"skx": "\n".join([p7.process_item(*item) for item in skx_mov_instructions]), "skx": "\n".join([p7.process_item(*item) for item in skx_mov_instructions]),
"csx": "\n".join([p7.process_item(*item) for item in csx_mov_instructions]), "csx": "\n".join([p7.process_item(*item) for item in csx_mov_instructions]),
"icx": "\n".join([p9.process_item(*item) for item in icx_mov_instructions]), "icx": "\n".join([p9.process_item(*item) for item in icx_mov_instructions]),
"zen3": "\n".join([z3.process_item(*item) for item in zen3_mov_instructions]),
} }
description = descriptions[arch] description = descriptions[arch]
@@ -1255,7 +1636,7 @@ if __name__ == "__main__":
import sys import sys
if len(sys.argv) != 2: if len(sys.argv) != 2:
print("Usage: {} (snb|ivb|hsw|bdw|skx|csx|icx)".format(sys.argv[0])) print("Usage: {} (snb|ivb|hsw|bdw|skx|csx|icx|zen3)".format(sys.argv[0]))
sys.exit(0) sys.exit(0)
try: try: