[X86][SchedModel] Add missing ReadAdvance for some arithmetic ops (PR51318 and PR51322).

This fixes a bug where implicit uses of EFLAGS were not marked as ReadAdvance in
the RM/MR variants of ADC/SBB (PR51318)

This also fixes the absence of ReadAdvance for the register operand of
RMW arithmetic instructions (PR51322).

Differential Revision: https://reviews.llvm.org/D107367
This commit is contained in:
Andrea Di Biagio 2021-08-03 17:10:42 +01:00
parent 680c71b127
commit 7a1a35a1d1
4 changed files with 49 additions and 37 deletions

View File

@ -708,6 +708,19 @@ class BinOpRM<bits<8> opcode, string mnemonic, X86TypeInfo typeinfo,
mnemonic, "{$src2, $src1|$src1, $src2}", pattern>,
Sched<[sched.Folded, sched.ReadAfterFold]>;
// BinOpRM - Instructions like "adc reg, reg, [mem]".
// There is an implicit register read at the end of the operand sequence.
class BinOpRM_ImplicitUse<bits<8> opcode, string mnemonic, X86TypeInfo typeinfo,
dag outlist, X86FoldableSchedWrite sched, list<dag> pattern>
: ITy<opcode, MRMSrcMem, typeinfo, outlist,
(ins typeinfo.RegClass:$src1, typeinfo.MemOperand:$src2),
mnemonic, "{$src2, $src1|$src1, $src2}", pattern>,
Sched<[sched.Folded, sched.ReadAfterFold,
// base, scale, index, offset, segment.
ReadDefault, ReadDefault, ReadDefault, ReadDefault, ReadDefault,
// implicit register read.
sched.ReadAfterFold]>;
// BinOpRM_F - Instructions like "cmp reg, [mem]".
class BinOpRM_F<bits<8> opcode, string mnemonic, X86TypeInfo typeinfo,
SDNode opnode>
@ -725,7 +738,7 @@ class BinOpRM_RF<bits<8> opcode, string mnemonic, X86TypeInfo typeinfo,
// BinOpRM_RFF - Instructions like "adc reg, reg, [mem]".
class BinOpRM_RFF<bits<8> opcode, string mnemonic, X86TypeInfo typeinfo,
SDNode opnode>
: BinOpRM<opcode, mnemonic, typeinfo, (outs typeinfo.RegClass:$dst), WriteADC,
: BinOpRM_ImplicitUse<opcode, mnemonic, typeinfo, (outs typeinfo.RegClass:$dst), WriteADC,
[(set typeinfo.RegClass:$dst, EFLAGS,
(opnode typeinfo.RegClass:$src1, (typeinfo.LoadNode addr:$src2),
EFLAGS))]>;
@ -805,7 +818,11 @@ class BinOpMR_RMW<bits<8> opcode, string mnemonic, X86TypeInfo typeinfo,
SDNode opnode>
: BinOpMR<opcode, mnemonic, typeinfo,
[(store (opnode (load addr:$dst), typeinfo.RegClass:$src), addr:$dst),
(implicit EFLAGS)]>, Sched<[WriteALURMW]>;
(implicit EFLAGS)]>, Sched<[WriteALURMW,
// base, scale, index, offset, segment
ReadDefault, ReadDefault, ReadDefault,
ReadDefault, ReadDefault,
WriteALU.ReadAfterFold]>; // reg
// BinOpMR_RMW_FF - Instructions like "adc [mem], reg".
class BinOpMR_RMW_FF<bits<8> opcode, string mnemonic, X86TypeInfo typeinfo,
@ -813,7 +830,12 @@ class BinOpMR_RMW_FF<bits<8> opcode, string mnemonic, X86TypeInfo typeinfo,
: BinOpMR<opcode, mnemonic, typeinfo,
[(store (opnode (load addr:$dst), typeinfo.RegClass:$src, EFLAGS),
addr:$dst),
(implicit EFLAGS)]>, Sched<[WriteADCRMW]>;
(implicit EFLAGS)]>, Sched<[WriteADCRMW,
// base, scale, index, offset, segment
ReadDefault, ReadDefault, ReadDefault,
ReadDefault, ReadDefault,
WriteALU.ReadAfterFold, // reg
WriteALU.ReadAfterFold]>; // EFLAGS
// BinOpMR_F - Instructions like "cmp [mem], reg".
class BinOpMR_F<bits<8> opcode, string mnemonic, X86TypeInfo typeinfo,

View File

@ -1,19 +1,16 @@
# NOTE: Assertions have been autogenerated by utils/update_mca_test_checks.py
# RUN: llvm-mca -mtriple=x86_64-unknown-unknown -mcpu=btver2 -iterations=4 -timeline < %s | FileCheck %s
# FIXME: PR51318
# Missing read-advance for the implicit use of register EFLAGS.
adc 4(%rsp), %eax
# CHECK: Iterations: 4
# CHECK-NEXT: Instructions: 4
# CHECK-NEXT: Total Cycles: 19
# CHECK-NEXT: Total Cycles: 10
# CHECK-NEXT: Total uOps: 4
# CHECK: Dispatch Width: 2
# CHECK-NEXT: uOps Per Cycle: 0.21
# CHECK-NEXT: IPC: 0.21
# CHECK-NEXT: uOps Per Cycle: 0.40
# CHECK-NEXT: IPC: 0.40
# CHECK-NEXT: Block RThroughput: 1.0
# CHECK: Instruction Info:
@ -52,13 +49,12 @@ adc 4(%rsp), %eax
# CHECK-NEXT: 1.00 1.00 - - - - - 1.00 - - - - - - adcl 4(%rsp), %eax
# CHECK: Timeline view:
# CHECK-NEXT: 012345678
# CHECK-NEXT: Index 0123456789
# CHECK: [0,0] DeeeeER . . . adcl 4(%rsp), %eax
# CHECK-NEXT: [1,0] D====eeeeER . . adcl 4(%rsp), %eax
# CHECK-NEXT: [2,0] .D=======eeeeER. . adcl 4(%rsp), %eax
# CHECK-NEXT: [3,0] .D===========eeeeER adcl 4(%rsp), %eax
# CHECK: [0,0] DeeeeER . adcl 4(%rsp), %eax
# CHECK-NEXT: [1,0] D=eeeeER . adcl 4(%rsp), %eax
# CHECK-NEXT: [2,0] .D=eeeeER. adcl 4(%rsp), %eax
# CHECK-NEXT: [3,0] .D==eeeeER adcl 4(%rsp), %eax
# CHECK: Average Wait times (based on the timeline view):
# CHECK-NEXT: [0]: Executions
@ -67,4 +63,4 @@ adc 4(%rsp), %eax
# CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
# CHECK: [0] [1] [2] [3]
# CHECK-NEXT: 0. 4 6.5 0.3 0.0 adcl 4(%rsp), %eax
# CHECK-NEXT: 0. 4 2.0 0.3 0.0 adcl 4(%rsp), %eax

View File

@ -1,20 +1,17 @@
# NOTE: Assertions have been autogenerated by utils/update_mca_test_checks.py
# RUN: llvm-mca -mtriple=x86_64-unknown-unknown -mcpu=btver2 -iterations=1 -timeline < %s | FileCheck %s
# FIXME: PR51322
# Missing read-advance for register EAX.
add %eax, %eax
adc %eax, 4(%rsp)
# CHECK: Iterations: 1
# CHECK-NEXT: Instructions: 2
# CHECK-NEXT: Total Cycles: 10
# CHECK-NEXT: Total Cycles: 9
# CHECK-NEXT: Total uOps: 2
# CHECK: Dispatch Width: 2
# CHECK-NEXT: uOps Per Cycle: 0.20
# CHECK-NEXT: IPC: 0.20
# CHECK-NEXT: uOps Per Cycle: 0.22
# CHECK-NEXT: IPC: 0.22
# CHECK-NEXT: Block RThroughput: 1.5
# CHECK: Instruction Info:
@ -55,10 +52,10 @@ adc %eax, 4(%rsp)
# CHECK-NEXT: 2.00 - - - - - - 1.00 - 1.00 - - - - adcl %eax, 4(%rsp)
# CHECK: Timeline view:
# CHECK-NEXT: Index 0123456789
# CHECK-NEXT: Index 012345678
# CHECK: [0,0] DeER . . addl %eax, %eax
# CHECK-NEXT: [0,1] D=eeeeeeER adcl %eax, 4(%rsp)
# CHECK: [0,0] DeER . . addl %eax, %eax
# CHECK-NEXT: [0,1] DeeeeeeER adcl %eax, 4(%rsp)
# CHECK: Average Wait times (based on the timeline view):
# CHECK-NEXT: [0]: Executions
@ -68,5 +65,5 @@ adc %eax, 4(%rsp)
# CHECK: [0] [1] [2] [3]
# CHECK-NEXT: 0. 1 1.0 1.0 0.0 addl %eax, %eax
# CHECK-NEXT: 1. 1 2.0 0.0 0.0 adcl %eax, 4(%rsp)
# CHECK-NEXT: 1 1.5 0.5 0.0 <total>
# CHECK-NEXT: 1. 1 1.0 0.0 0.0 adcl %eax, 4(%rsp)
# CHECK-NEXT: 1 1.0 0.5 0.0 <total>

View File

@ -1,20 +1,17 @@
# NOTE: Assertions have been autogenerated by utils/update_mca_test_checks.py
# RUN: llvm-mca -mtriple=x86_64-unknown-unknown -mcpu=btver2 -iterations=1 -timeline < %s | FileCheck %s
# FIXME: PR51322
# Missing read-advance for register EAX.
add %eax, %eax
add %eax, 4(%rsp)
# CHECK: Iterations: 1
# CHECK-NEXT: Instructions: 2
# CHECK-NEXT: Total Cycles: 10
# CHECK-NEXT: Total Cycles: 9
# CHECK-NEXT: Total uOps: 2
# CHECK: Dispatch Width: 2
# CHECK-NEXT: uOps Per Cycle: 0.20
# CHECK-NEXT: IPC: 0.20
# CHECK-NEXT: uOps Per Cycle: 0.22
# CHECK-NEXT: IPC: 0.22
# CHECK-NEXT: Block RThroughput: 1.0
# CHECK: Instruction Info:
@ -55,10 +52,10 @@ add %eax, 4(%rsp)
# CHECK-NEXT: 1.00 - - - - - - 1.00 - 1.00 - - - - addl %eax, 4(%rsp)
# CHECK: Timeline view:
# CHECK-NEXT: Index 0123456789
# CHECK-NEXT: Index 012345678
# CHECK: [0,0] DeER . . addl %eax, %eax
# CHECK-NEXT: [0,1] D=eeeeeeER addl %eax, 4(%rsp)
# CHECK: [0,0] DeER . . addl %eax, %eax
# CHECK-NEXT: [0,1] DeeeeeeER addl %eax, 4(%rsp)
# CHECK: Average Wait times (based on the timeline view):
# CHECK-NEXT: [0]: Executions
@ -68,5 +65,5 @@ add %eax, 4(%rsp)
# CHECK: [0] [1] [2] [3]
# CHECK-NEXT: 0. 1 1.0 1.0 0.0 addl %eax, %eax
# CHECK-NEXT: 1. 1 2.0 0.0 0.0 addl %eax, 4(%rsp)
# CHECK-NEXT: 1 1.5 0.5 0.0 <total>
# CHECK-NEXT: 1. 1 1.0 0.0 0.0 addl %eax, 4(%rsp)
# CHECK-NEXT: 1 1.0 0.5 0.0 <total>