2021-01-12 01:20:45 +08:00
|
|
|
; RUN: llc -O0 < %s | FileCheck %s
|
[FastISel] Sink local value materializations to first use
Summary:
Local values are constants, global addresses, and stack addresses that
can't be folded into the instruction that uses them. For example, when
storing the address of a global variable into memory, we need to
materialize that address into a register.
FastISel doesn't want to materialize any given local value more than
once, so it generates all local value materialization code at
EmitStartPt, which always dominates the current insertion point. This
allows it to maintain a map of local value registers, and it knows that
the local value area will always dominate the current insertion point.
The downside is that local value instructions are always emitted without
a source location. This is done to prevent jumpy line tables, but it
means that the local value area will be considered part of the previous
statement. Consider this C code:
call1(); // line 1
++global; // line 2
++global; // line 3
call2(&global, &local); // line 4
Today we end up with assembly and line tables like this:
.loc 1 1
callq call1
leaq global(%rip), %rdi
leaq local(%rsp), %rsi
.loc 1 2
addq $1, global(%rip)
.loc 1 3
addq $1, global(%rip)
.loc 1 4
callq call2
The LEA instructions in the local value area have no source location and
are treated as being on line 1. Stepping through the code in a debugger
and correlating it with the assembly won't make much sense, because
these materializations are only required for line 4.
This is actually problematic for the VS debugger "set next statement"
feature, which effectively assumes that there are no registers live
across statement boundaries. By sinking the local value code into the
statement and fixing up the source location, we can make that feature
work. This was filed as https://bugs.llvm.org/show_bug.cgi?id=35975 and
https://crbug.com/793819.
This change is obviously not enough to make this feature work reliably
in all cases, but I felt that it was worth doing anyway because it
usually generates smaller, more comprehensible -O0 code. I measured a
0.12% regression in code generation time with LLC on the sqlite3
amalgamation, so I think this is worth doing.
There are some special cases worth calling out in the commit message:
1. local values materialized for phis
2. local values used by no-op casts
3. dead local value code
Local values can be materialized for phis, and this does not show up as
a vreg use in MachineRegisterInfo. In this case, if there are no other
uses, this patch sinks the value to the first terminator, EH label, or
the end of the BB if nothing else exists.
Local values may also be used by no-op casts, which adds the register to
the RegFixups table. Without reversing the RegFixups map direction, we
don't have enough information to sink these instructions.
Lastly, if the local value register has no other uses, we can delete it.
This comes up when fastisel tries two instruction selection approaches
and the first materializes the value but fails and the second succeeds
without using the local value.
Reviewers: aprantl, dblaikie, qcolombet, MatzeB, vsk, echristo
Subscribers: dotdash, chandlerc, hans, sdardis, amccarth, javed.absar, zturner, llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D43093
llvm-svn: 327581
2018-03-15 05:54:21 +08:00
|
|
|
|
|
|
|
|
target datalayout = "e-m:x-p:32:32-i64:64-f80:32-n8:16:32-a:0:32-S32"
|
|
|
|
|
target triple = "i386-linux-gnu"
|
|
|
|
|
|
|
|
|
|
; Try some simple cases that show how local value sinking improves line tables.
|
|
|
|
|
|
2020-12-05 05:51:01 +08:00
|
|
|
@sink_across = external dso_local global i32
|
[FastISel] Sink local value materializations to first use
Summary:
Local values are constants, global addresses, and stack addresses that
can't be folded into the instruction that uses them. For example, when
storing the address of a global variable into memory, we need to
materialize that address into a register.
FastISel doesn't want to materialize any given local value more than
once, so it generates all local value materialization code at
EmitStartPt, which always dominates the current insertion point. This
allows it to maintain a map of local value registers, and it knows that
the local value area will always dominate the current insertion point.
The downside is that local value instructions are always emitted without
a source location. This is done to prevent jumpy line tables, but it
means that the local value area will be considered part of the previous
statement. Consider this C code:
call1(); // line 1
++global; // line 2
++global; // line 3
call2(&global, &local); // line 4
Today we end up with assembly and line tables like this:
.loc 1 1
callq call1
leaq global(%rip), %rdi
leaq local(%rsp), %rsi
.loc 1 2
addq $1, global(%rip)
.loc 1 3
addq $1, global(%rip)
.loc 1 4
callq call2
The LEA instructions in the local value area have no source location and
are treated as being on line 1. Stepping through the code in a debugger
and correlating it with the assembly won't make much sense, because
these materializations are only required for line 4.
This is actually problematic for the VS debugger "set next statement"
feature, which effectively assumes that there are no registers live
across statement boundaries. By sinking the local value code into the
statement and fixing up the source location, we can make that feature
work. This was filed as https://bugs.llvm.org/show_bug.cgi?id=35975 and
https://crbug.com/793819.
This change is obviously not enough to make this feature work reliably
in all cases, but I felt that it was worth doing anyway because it
usually generates smaller, more comprehensible -O0 code. I measured a
0.12% regression in code generation time with LLC on the sqlite3
amalgamation, so I think this is worth doing.
There are some special cases worth calling out in the commit message:
1. local values materialized for phis
2. local values used by no-op casts
3. dead local value code
Local values can be materialized for phis, and this does not show up as
a vreg use in MachineRegisterInfo. In this case, if there are no other
uses, this patch sinks the value to the first terminator, EH label, or
the end of the BB if nothing else exists.
Local values may also be used by no-op casts, which adds the register to
the RegFixups table. Without reversing the RegFixups map direction, we
don't have enough information to sink these instructions.
Lastly, if the local value register has no other uses, we can delete it.
This comes up when fastisel tries two instruction selection approaches
and the first materializes the value but fails and the second succeeds
without using the local value.
Reviewers: aprantl, dblaikie, qcolombet, MatzeB, vsk, echristo
Subscribers: dotdash, chandlerc, hans, sdardis, amccarth, javed.absar, zturner, llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D43093
llvm-svn: 327581
2018-03-15 05:54:21 +08:00
|
|
|
|
|
|
|
|
declare void @simple_callee(i32, i32)
|
|
|
|
|
|
|
|
|
|
define void @simple() !dbg !5 {
|
|
|
|
|
store i32 44, i32* @sink_across, !dbg !7
|
|
|
|
|
call void @simple_callee(i32 13, i32 55), !dbg !8
|
|
|
|
|
ret void, !dbg !9
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; CHECK-LABEL: simple:
|
|
|
|
|
; CHECK-NOT: movl $13,
|
|
|
|
|
; CHECK: .loc 1 1 1 prologue_end
|
|
|
|
|
; CHECK: movl $44, sink_across
|
|
|
|
|
; CHECK: .loc 1 2 1
|
|
|
|
|
; CHECK: movl $13,
|
|
|
|
|
; CHECK: movl $55,
|
|
|
|
|
; CHECK: calll simple_callee
|
|
|
|
|
|
|
|
|
|
declare void @simple_reg_callee(i32 inreg, i32 inreg)
|
|
|
|
|
|
|
|
|
|
define void @simple_reg() !dbg !10 {
|
|
|
|
|
store i32 44, i32* @sink_across, !dbg !11
|
|
|
|
|
call void @simple_reg_callee(i32 inreg 13, i32 inreg 55), !dbg !12
|
|
|
|
|
ret void, !dbg !13
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; CHECK-LABEL: simple_reg:
|
|
|
|
|
; CHECK: .loc 1 4 1 prologue_end
|
|
|
|
|
; CHECK: movl $44, sink_across
|
|
|
|
|
; CHECK: .loc 1 5 1
|
|
|
|
|
; CHECK: movl $13,
|
|
|
|
|
; CHECK: movl $55,
|
|
|
|
|
; CHECK: calll simple_reg_callee
|
|
|
|
|
|
|
|
|
|
; There are two interesting cases where local values have no uses but are not
|
|
|
|
|
; dead: when the local value is directly used by a phi, and when the local
|
|
|
|
|
; value is used by a no-op cast instruction. In these cases, we get side tables
|
|
|
|
|
; referring to the local value vreg that we need to check.
|
|
|
|
|
|
|
|
|
|
define i8* @phi_const(i32 %c) !dbg !14 {
|
|
|
|
|
entry:
|
|
|
|
|
%tobool = icmp eq i32 %c, 0, !dbg !20
|
|
|
|
|
call void @llvm.dbg.value(metadata i1 %tobool, metadata !16, metadata !DIExpression()), !dbg !20
|
|
|
|
|
br i1 %tobool, label %if.else, label %if.then, !dbg !21
|
|
|
|
|
|
|
|
|
|
if.then: ; preds = %entry
|
|
|
|
|
br label %if.end, !dbg !22
|
|
|
|
|
|
|
|
|
|
if.else: ; preds = %entry
|
|
|
|
|
br label %if.end, !dbg !23
|
|
|
|
|
|
|
|
|
|
if.end: ; preds = %if.else, %if.then
|
|
|
|
|
%r.0 = phi i8* [ inttoptr (i32 42 to i8*), %if.then ], [ inttoptr (i32 1 to i8*), %if.else ], !dbg !24
|
|
|
|
|
call void @llvm.dbg.value(metadata i8* %r.0, metadata !18, metadata !DIExpression()), !dbg !24
|
|
|
|
|
ret i8* %r.0, !dbg !25
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; CHECK-LABEL: phi_const:
|
|
|
|
|
; CHECK: # %entry
|
|
|
|
|
; CHECK: cmpl $0,
|
|
|
|
|
; CHECK: # %if.then
|
|
|
|
|
; CHECK: movl $42,
|
|
|
|
|
; CHECK: jmp
|
|
|
|
|
; CHECK: # %if.else
|
|
|
|
|
; CHECK: movl $1,
|
|
|
|
|
; CHECK: # %if.end
|
|
|
|
|
|
|
|
|
|
define i8* @phi_const_cast(i32 %c) !dbg !26 {
|
|
|
|
|
entry:
|
|
|
|
|
%tobool = icmp eq i32 %c, 0, !dbg !32
|
|
|
|
|
call void @llvm.dbg.value(metadata i1 %tobool, metadata !28, metadata !DIExpression()), !dbg !32
|
|
|
|
|
br i1 %tobool, label %if.else, label %if.then, !dbg !33
|
|
|
|
|
|
|
|
|
|
if.then: ; preds = %entry
|
|
|
|
|
%v42 = inttoptr i32 42 to i8*, !dbg !34
|
|
|
|
|
call void @llvm.dbg.value(metadata i8* %v42, metadata !29, metadata !DIExpression()), !dbg !34
|
|
|
|
|
br label %if.end, !dbg !35
|
|
|
|
|
|
|
|
|
|
if.else: ; preds = %entry
|
|
|
|
|
%v1 = inttoptr i32 1 to i8*, !dbg !36
|
|
|
|
|
call void @llvm.dbg.value(metadata i8* %v1, metadata !30, metadata !DIExpression()), !dbg !36
|
|
|
|
|
br label %if.end, !dbg !37
|
|
|
|
|
|
|
|
|
|
if.end: ; preds = %if.else, %if.then
|
|
|
|
|
%r.0 = phi i8* [ %v42, %if.then ], [ %v1, %if.else ], !dbg !38
|
|
|
|
|
call void @llvm.dbg.value(metadata i8* %r.0, metadata !31, metadata !DIExpression()), !dbg !38
|
|
|
|
|
ret i8* %r.0, !dbg !39
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; CHECK-LABEL: phi_const_cast:
|
|
|
|
|
; CHECK: # %entry
|
|
|
|
|
; CHECK: cmpl $0,
|
|
|
|
|
; CHECK: # %if.then
|
|
|
|
|
; CHECK: movl $42, %[[REG:[a-z]+]]
|
|
|
|
|
; CHECK: #DEBUG_VALUE: phi_const_cast:4 <- $[[REG]]
|
|
|
|
|
; CHECK: jmp
|
|
|
|
|
; CHECK: # %if.else
|
|
|
|
|
; CHECK: movl $1, %[[REG:[a-z]+]]
|
|
|
|
|
; CHECK: #DEBUG_VALUE: phi_const_cast:5 <- $[[REG]]
|
|
|
|
|
; CHECK: # %if.end
|
|
|
|
|
|
|
|
|
|
declare void @may_throw() local_unnamed_addr #1
|
|
|
|
|
|
|
|
|
|
declare i32 @__gxx_personality_v0(...)
|
|
|
|
|
|
|
|
|
|
define i32 @invoke_phi() personality i32 (...)* @__gxx_personality_v0 {
|
|
|
|
|
entry:
|
|
|
|
|
store i32 42, i32* @sink_across
|
|
|
|
|
invoke void @may_throw()
|
|
|
|
|
to label %try.cont unwind label %lpad
|
|
|
|
|
|
|
|
|
|
lpad: ; preds = %entry
|
|
|
|
|
%0 = landingpad { i8*, i32 }
|
|
|
|
|
catch i8* null
|
|
|
|
|
store i32 42, i32* @sink_across
|
|
|
|
|
br label %try.cont
|
|
|
|
|
|
|
|
|
|
try.cont: ; preds = %entry, %lpad
|
|
|
|
|
%r.0 = phi i32 [ 13, %entry ], [ 55, %lpad ]
|
|
|
|
|
ret i32 %r.0
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; The constant materialization should be *after* the stores to sink_across, but
|
|
|
|
|
; before any EH_LABEL.
|
|
|
|
|
|
|
|
|
|
; CHECK-LABEL: invoke_phi:
|
|
|
|
|
; CHECK: movl $42, sink_across
|
|
|
|
|
; CHECK: movl $13, %{{[a-z]*}}
|
|
|
|
|
; CHECK: .Ltmp{{.*}}:
|
|
|
|
|
; CHECK: calll may_throw
|
|
|
|
|
; CHECK: .Ltmp{{.*}}:
|
|
|
|
|
; CHECK: jmp .LBB{{.*}}
|
|
|
|
|
; CHECK: .LBB{{.*}}: # %lpad
|
|
|
|
|
; CHECK: movl $42, sink_across
|
|
|
|
|
; CHECK: movl $55, %{{[a-z]*}}
|
2020-03-29 02:03:14 +08:00
|
|
|
; CHECK: .LBB{{.*}}: # %try.cont
|
|
|
|
|
; CHECK: retl
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
define i32 @lpad_phi() personality i32 (...)* @__gxx_personality_v0 {
|
|
|
|
|
entry:
|
|
|
|
|
store i32 42, i32* @sink_across
|
|
|
|
|
invoke void @may_throw()
|
|
|
|
|
to label %try.cont unwind label %lpad
|
|
|
|
|
|
|
|
|
|
lpad: ; preds = %entry
|
|
|
|
|
%p = phi i32 [ 11, %entry ] ; Trivial, but -O0 keeps it
|
|
|
|
|
%0 = landingpad { i8*, i32 }
|
|
|
|
|
catch i8* null
|
|
|
|
|
store i32 %p, i32* @sink_across
|
|
|
|
|
br label %try.cont
|
|
|
|
|
|
|
|
|
|
try.cont: ; preds = %entry, %lpad
|
|
|
|
|
%r.0 = phi i32 [ 13, %entry ], [ 55, %lpad ]
|
|
|
|
|
ret i32 %r.0
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; The constant materialization should be *after* the stores to sink_across, but
|
|
|
|
|
; before any EH_LABEL.
|
|
|
|
|
|
|
|
|
|
; CHECK-LABEL: lpad_phi:
|
|
|
|
|
; CHECK: movl $42, sink_across
|
|
|
|
|
; CHECK: movl $13, %{{[a-z]*}}
|
|
|
|
|
; CHECK: .Ltmp{{.*}}:
|
|
|
|
|
; CHECK: calll may_throw
|
|
|
|
|
; CHECK: .Ltmp{{.*}}:
|
|
|
|
|
; CHECK: jmp .LBB{{.*}}
|
|
|
|
|
; CHECK: .LBB{{.*}}: # %lpad
|
|
|
|
|
; CHECK-NEXT: .Ltmp{{.*}}:
|
|
|
|
|
; CHECK: movl {{.*}}, sink_across
|
|
|
|
|
; CHECK: movl $55, %{{[a-z]*}}
|
[FastISel] Sink local value materializations to first use
Summary:
Local values are constants, global addresses, and stack addresses that
can't be folded into the instruction that uses them. For example, when
storing the address of a global variable into memory, we need to
materialize that address into a register.
FastISel doesn't want to materialize any given local value more than
once, so it generates all local value materialization code at
EmitStartPt, which always dominates the current insertion point. This
allows it to maintain a map of local value registers, and it knows that
the local value area will always dominate the current insertion point.
The downside is that local value instructions are always emitted without
a source location. This is done to prevent jumpy line tables, but it
means that the local value area will be considered part of the previous
statement. Consider this C code:
call1(); // line 1
++global; // line 2
++global; // line 3
call2(&global, &local); // line 4
Today we end up with assembly and line tables like this:
.loc 1 1
callq call1
leaq global(%rip), %rdi
leaq local(%rsp), %rsi
.loc 1 2
addq $1, global(%rip)
.loc 1 3
addq $1, global(%rip)
.loc 1 4
callq call2
The LEA instructions in the local value area have no source location and
are treated as being on line 1. Stepping through the code in a debugger
and correlating it with the assembly won't make much sense, because
these materializations are only required for line 4.
This is actually problematic for the VS debugger "set next statement"
feature, which effectively assumes that there are no registers live
across statement boundaries. By sinking the local value code into the
statement and fixing up the source location, we can make that feature
work. This was filed as https://bugs.llvm.org/show_bug.cgi?id=35975 and
https://crbug.com/793819.
This change is obviously not enough to make this feature work reliably
in all cases, but I felt that it was worth doing anyway because it
usually generates smaller, more comprehensible -O0 code. I measured a
0.12% regression in code generation time with LLC on the sqlite3
amalgamation, so I think this is worth doing.
There are some special cases worth calling out in the commit message:
1. local values materialized for phis
2. local values used by no-op casts
3. dead local value code
Local values can be materialized for phis, and this does not show up as
a vreg use in MachineRegisterInfo. In this case, if there are no other
uses, this patch sinks the value to the first terminator, EH label, or
the end of the BB if nothing else exists.
Local values may also be used by no-op casts, which adds the register to
the RegFixups table. Without reversing the RegFixups map direction, we
don't have enough information to sink these instructions.
Lastly, if the local value register has no other uses, we can delete it.
This comes up when fastisel tries two instruction selection approaches
and the first materializes the value but fails and the second succeeds
without using the local value.
Reviewers: aprantl, dblaikie, qcolombet, MatzeB, vsk, echristo
Subscribers: dotdash, chandlerc, hans, sdardis, amccarth, javed.absar, zturner, llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D43093
llvm-svn: 327581
2018-03-15 05:54:21 +08:00
|
|
|
; CHECK: .LBB{{.*}}: # %try.cont
|
|
|
|
|
; CHECK: retl
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
; Function Attrs: nounwind readnone speculatable
|
|
|
|
|
declare void @llvm.dbg.value(metadata, metadata, metadata) #0
|
|
|
|
|
|
|
|
|
|
attributes #0 = { nounwind readnone speculatable }
|
|
|
|
|
|
|
|
|
|
!llvm.dbg.cu = !{!0}
|
|
|
|
|
!llvm.debugify = !{!3, !4}
|
|
|
|
|
!llvm.module.flags = !{!52, !53}
|
|
|
|
|
|
|
|
|
|
!0 = distinct !DICompileUnit(language: DW_LANG_C, file: !1, producer: "debugify", isOptimized: true, runtimeVersion: 0, emissionKind: FullDebug, enums: !2)
|
|
|
|
|
!1 = !DIFile(filename: "../llvm/test/CodeGen/X86/sink-local-value.ll", directory: "/")
|
|
|
|
|
!2 = !{}
|
|
|
|
|
!3 = !{i32 27}
|
|
|
|
|
!4 = !{i32 8}
|
[DebugInfo] Add DILabel metadata and intrinsic llvm.dbg.label.
In order to set breakpoints on labels and list source code around
labels, we need collect debug information for labels, i.e., label
name, the function label belong, line number in the file, and the
address label located. In order to keep these information in LLVM
IR and to allow backend to generate debug information correctly.
We create a new kind of metadata for labels, DILabel. The format
of DILabel is
!DILabel(scope: !1, name: "foo", file: !2, line: 3)
We hope to keep debug information as much as possible even the
code is optimized. So, we create a new kind of intrinsic for label
metadata to avoid the metadata is eliminated with basic block.
The intrinsic will keep existing if we keep it from optimized out.
The format of the intrinsic is
llvm.dbg.label(metadata !1)
It has only one argument, that is the DILabel metadata. The
intrinsic will follow the label immediately. Backend could get the
label metadata through the intrinsic's parameter.
We also create DIBuilder API for labels to be used by Frontend.
Frontend could use createLabel() to allocate DILabel objects, and use
insertLabel() to insert llvm.dbg.label intrinsic in LLVM IR.
Differential Revision: https://reviews.llvm.org/D45024
Patch by Hsiangkai Wang.
llvm-svn: 331841
2018-05-09 10:40:45 +08:00
|
|
|
!5 = distinct !DISubprogram(name: "simple", linkageName: "simple", scope: null, file: !1, line: 1, type: !6, isLocal: false, isDefinition: true, scopeLine: 1, isOptimized: true, unit: !0, retainedNodes: !2)
|
[FastISel] Sink local value materializations to first use
Summary:
Local values are constants, global addresses, and stack addresses that
can't be folded into the instruction that uses them. For example, when
storing the address of a global variable into memory, we need to
materialize that address into a register.
FastISel doesn't want to materialize any given local value more than
once, so it generates all local value materialization code at
EmitStartPt, which always dominates the current insertion point. This
allows it to maintain a map of local value registers, and it knows that
the local value area will always dominate the current insertion point.
The downside is that local value instructions are always emitted without
a source location. This is done to prevent jumpy line tables, but it
means that the local value area will be considered part of the previous
statement. Consider this C code:
call1(); // line 1
++global; // line 2
++global; // line 3
call2(&global, &local); // line 4
Today we end up with assembly and line tables like this:
.loc 1 1
callq call1
leaq global(%rip), %rdi
leaq local(%rsp), %rsi
.loc 1 2
addq $1, global(%rip)
.loc 1 3
addq $1, global(%rip)
.loc 1 4
callq call2
The LEA instructions in the local value area have no source location and
are treated as being on line 1. Stepping through the code in a debugger
and correlating it with the assembly won't make much sense, because
these materializations are only required for line 4.
This is actually problematic for the VS debugger "set next statement"
feature, which effectively assumes that there are no registers live
across statement boundaries. By sinking the local value code into the
statement and fixing up the source location, we can make that feature
work. This was filed as https://bugs.llvm.org/show_bug.cgi?id=35975 and
https://crbug.com/793819.
This change is obviously not enough to make this feature work reliably
in all cases, but I felt that it was worth doing anyway because it
usually generates smaller, more comprehensible -O0 code. I measured a
0.12% regression in code generation time with LLC on the sqlite3
amalgamation, so I think this is worth doing.
There are some special cases worth calling out in the commit message:
1. local values materialized for phis
2. local values used by no-op casts
3. dead local value code
Local values can be materialized for phis, and this does not show up as
a vreg use in MachineRegisterInfo. In this case, if there are no other
uses, this patch sinks the value to the first terminator, EH label, or
the end of the BB if nothing else exists.
Local values may also be used by no-op casts, which adds the register to
the RegFixups table. Without reversing the RegFixups map direction, we
don't have enough information to sink these instructions.
Lastly, if the local value register has no other uses, we can delete it.
This comes up when fastisel tries two instruction selection approaches
and the first materializes the value but fails and the second succeeds
without using the local value.
Reviewers: aprantl, dblaikie, qcolombet, MatzeB, vsk, echristo
Subscribers: dotdash, chandlerc, hans, sdardis, amccarth, javed.absar, zturner, llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D43093
llvm-svn: 327581
2018-03-15 05:54:21 +08:00
|
|
|
!6 = !DISubroutineType(types: !2)
|
|
|
|
|
!7 = !DILocation(line: 1, column: 1, scope: !5)
|
|
|
|
|
!8 = !DILocation(line: 2, column: 1, scope: !5)
|
|
|
|
|
!9 = !DILocation(line: 3, column: 1, scope: !5)
|
[DebugInfo] Add DILabel metadata and intrinsic llvm.dbg.label.
In order to set breakpoints on labels and list source code around
labels, we need collect debug information for labels, i.e., label
name, the function label belong, line number in the file, and the
address label located. In order to keep these information in LLVM
IR and to allow backend to generate debug information correctly.
We create a new kind of metadata for labels, DILabel. The format
of DILabel is
!DILabel(scope: !1, name: "foo", file: !2, line: 3)
We hope to keep debug information as much as possible even the
code is optimized. So, we create a new kind of intrinsic for label
metadata to avoid the metadata is eliminated with basic block.
The intrinsic will keep existing if we keep it from optimized out.
The format of the intrinsic is
llvm.dbg.label(metadata !1)
It has only one argument, that is the DILabel metadata. The
intrinsic will follow the label immediately. Backend could get the
label metadata through the intrinsic's parameter.
We also create DIBuilder API for labels to be used by Frontend.
Frontend could use createLabel() to allocate DILabel objects, and use
insertLabel() to insert llvm.dbg.label intrinsic in LLVM IR.
Differential Revision: https://reviews.llvm.org/D45024
Patch by Hsiangkai Wang.
llvm-svn: 331841
2018-05-09 10:40:45 +08:00
|
|
|
!10 = distinct !DISubprogram(name: "simple_reg", linkageName: "simple_reg", scope: null, file: !1, line: 4, type: !6, isLocal: false, isDefinition: true, scopeLine: 4, isOptimized: true, unit: !0, retainedNodes: !2)
|
[FastISel] Sink local value materializations to first use
Summary:
Local values are constants, global addresses, and stack addresses that
can't be folded into the instruction that uses them. For example, when
storing the address of a global variable into memory, we need to
materialize that address into a register.
FastISel doesn't want to materialize any given local value more than
once, so it generates all local value materialization code at
EmitStartPt, which always dominates the current insertion point. This
allows it to maintain a map of local value registers, and it knows that
the local value area will always dominate the current insertion point.
The downside is that local value instructions are always emitted without
a source location. This is done to prevent jumpy line tables, but it
means that the local value area will be considered part of the previous
statement. Consider this C code:
call1(); // line 1
++global; // line 2
++global; // line 3
call2(&global, &local); // line 4
Today we end up with assembly and line tables like this:
.loc 1 1
callq call1
leaq global(%rip), %rdi
leaq local(%rsp), %rsi
.loc 1 2
addq $1, global(%rip)
.loc 1 3
addq $1, global(%rip)
.loc 1 4
callq call2
The LEA instructions in the local value area have no source location and
are treated as being on line 1. Stepping through the code in a debugger
and correlating it with the assembly won't make much sense, because
these materializations are only required for line 4.
This is actually problematic for the VS debugger "set next statement"
feature, which effectively assumes that there are no registers live
across statement boundaries. By sinking the local value code into the
statement and fixing up the source location, we can make that feature
work. This was filed as https://bugs.llvm.org/show_bug.cgi?id=35975 and
https://crbug.com/793819.
This change is obviously not enough to make this feature work reliably
in all cases, but I felt that it was worth doing anyway because it
usually generates smaller, more comprehensible -O0 code. I measured a
0.12% regression in code generation time with LLC on the sqlite3
amalgamation, so I think this is worth doing.
There are some special cases worth calling out in the commit message:
1. local values materialized for phis
2. local values used by no-op casts
3. dead local value code
Local values can be materialized for phis, and this does not show up as
a vreg use in MachineRegisterInfo. In this case, if there are no other
uses, this patch sinks the value to the first terminator, EH label, or
the end of the BB if nothing else exists.
Local values may also be used by no-op casts, which adds the register to
the RegFixups table. Without reversing the RegFixups map direction, we
don't have enough information to sink these instructions.
Lastly, if the local value register has no other uses, we can delete it.
This comes up when fastisel tries two instruction selection approaches
and the first materializes the value but fails and the second succeeds
without using the local value.
Reviewers: aprantl, dblaikie, qcolombet, MatzeB, vsk, echristo
Subscribers: dotdash, chandlerc, hans, sdardis, amccarth, javed.absar, zturner, llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D43093
llvm-svn: 327581
2018-03-15 05:54:21 +08:00
|
|
|
!11 = !DILocation(line: 4, column: 1, scope: !10)
|
|
|
|
|
!12 = !DILocation(line: 5, column: 1, scope: !10)
|
|
|
|
|
!13 = !DILocation(line: 6, column: 1, scope: !10)
|
[DebugInfo] Add DILabel metadata and intrinsic llvm.dbg.label.
In order to set breakpoints on labels and list source code around
labels, we need collect debug information for labels, i.e., label
name, the function label belong, line number in the file, and the
address label located. In order to keep these information in LLVM
IR and to allow backend to generate debug information correctly.
We create a new kind of metadata for labels, DILabel. The format
of DILabel is
!DILabel(scope: !1, name: "foo", file: !2, line: 3)
We hope to keep debug information as much as possible even the
code is optimized. So, we create a new kind of intrinsic for label
metadata to avoid the metadata is eliminated with basic block.
The intrinsic will keep existing if we keep it from optimized out.
The format of the intrinsic is
llvm.dbg.label(metadata !1)
It has only one argument, that is the DILabel metadata. The
intrinsic will follow the label immediately. Backend could get the
label metadata through the intrinsic's parameter.
We also create DIBuilder API for labels to be used by Frontend.
Frontend could use createLabel() to allocate DILabel objects, and use
insertLabel() to insert llvm.dbg.label intrinsic in LLVM IR.
Differential Revision: https://reviews.llvm.org/D45024
Patch by Hsiangkai Wang.
llvm-svn: 331841
2018-05-09 10:40:45 +08:00
|
|
|
!14 = distinct !DISubprogram(name: "phi_const", linkageName: "phi_const", scope: null, file: !1, line: 7, type: !6, isLocal: false, isDefinition: true, scopeLine: 7, isOptimized: true, unit: !0, retainedNodes: !15)
|
[FastISel] Sink local value materializations to first use
Summary:
Local values are constants, global addresses, and stack addresses that
can't be folded into the instruction that uses them. For example, when
storing the address of a global variable into memory, we need to
materialize that address into a register.
FastISel doesn't want to materialize any given local value more than
once, so it generates all local value materialization code at
EmitStartPt, which always dominates the current insertion point. This
allows it to maintain a map of local value registers, and it knows that
the local value area will always dominate the current insertion point.
The downside is that local value instructions are always emitted without
a source location. This is done to prevent jumpy line tables, but it
means that the local value area will be considered part of the previous
statement. Consider this C code:
call1(); // line 1
++global; // line 2
++global; // line 3
call2(&global, &local); // line 4
Today we end up with assembly and line tables like this:
.loc 1 1
callq call1
leaq global(%rip), %rdi
leaq local(%rsp), %rsi
.loc 1 2
addq $1, global(%rip)
.loc 1 3
addq $1, global(%rip)
.loc 1 4
callq call2
The LEA instructions in the local value area have no source location and
are treated as being on line 1. Stepping through the code in a debugger
and correlating it with the assembly won't make much sense, because
these materializations are only required for line 4.
This is actually problematic for the VS debugger "set next statement"
feature, which effectively assumes that there are no registers live
across statement boundaries. By sinking the local value code into the
statement and fixing up the source location, we can make that feature
work. This was filed as https://bugs.llvm.org/show_bug.cgi?id=35975 and
https://crbug.com/793819.
This change is obviously not enough to make this feature work reliably
in all cases, but I felt that it was worth doing anyway because it
usually generates smaller, more comprehensible -O0 code. I measured a
0.12% regression in code generation time with LLC on the sqlite3
amalgamation, so I think this is worth doing.
There are some special cases worth calling out in the commit message:
1. local values materialized for phis
2. local values used by no-op casts
3. dead local value code
Local values can be materialized for phis, and this does not show up as
a vreg use in MachineRegisterInfo. In this case, if there are no other
uses, this patch sinks the value to the first terminator, EH label, or
the end of the BB if nothing else exists.
Local values may also be used by no-op casts, which adds the register to
the RegFixups table. Without reversing the RegFixups map direction, we
don't have enough information to sink these instructions.
Lastly, if the local value register has no other uses, we can delete it.
This comes up when fastisel tries two instruction selection approaches
and the first materializes the value but fails and the second succeeds
without using the local value.
Reviewers: aprantl, dblaikie, qcolombet, MatzeB, vsk, echristo
Subscribers: dotdash, chandlerc, hans, sdardis, amccarth, javed.absar, zturner, llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D43093
llvm-svn: 327581
2018-03-15 05:54:21 +08:00
|
|
|
!15 = !{!16, !18}
|
|
|
|
|
!16 = !DILocalVariable(name: "1", scope: !14, file: !1, line: 7, type: !17)
|
|
|
|
|
!17 = !DIBasicType(name: "ty8", size: 8, encoding: DW_ATE_unsigned)
|
|
|
|
|
!18 = !DILocalVariable(name: "2", scope: !14, file: !1, line: 11, type: !19)
|
|
|
|
|
!19 = !DIBasicType(name: "ty32", size: 32, encoding: DW_ATE_unsigned)
|
|
|
|
|
!20 = !DILocation(line: 7, column: 1, scope: !14)
|
|
|
|
|
!21 = !DILocation(line: 8, column: 1, scope: !14)
|
|
|
|
|
!22 = !DILocation(line: 9, column: 1, scope: !14)
|
|
|
|
|
!23 = !DILocation(line: 10, column: 1, scope: !14)
|
|
|
|
|
!24 = !DILocation(line: 11, column: 1, scope: !14)
|
|
|
|
|
!25 = !DILocation(line: 12, column: 1, scope: !14)
|
[DebugInfo] Add DILabel metadata and intrinsic llvm.dbg.label.
In order to set breakpoints on labels and list source code around
labels, we need collect debug information for labels, i.e., label
name, the function label belong, line number in the file, and the
address label located. In order to keep these information in LLVM
IR and to allow backend to generate debug information correctly.
We create a new kind of metadata for labels, DILabel. The format
of DILabel is
!DILabel(scope: !1, name: "foo", file: !2, line: 3)
We hope to keep debug information as much as possible even the
code is optimized. So, we create a new kind of intrinsic for label
metadata to avoid the metadata is eliminated with basic block.
The intrinsic will keep existing if we keep it from optimized out.
The format of the intrinsic is
llvm.dbg.label(metadata !1)
It has only one argument, that is the DILabel metadata. The
intrinsic will follow the label immediately. Backend could get the
label metadata through the intrinsic's parameter.
We also create DIBuilder API for labels to be used by Frontend.
Frontend could use createLabel() to allocate DILabel objects, and use
insertLabel() to insert llvm.dbg.label intrinsic in LLVM IR.
Differential Revision: https://reviews.llvm.org/D45024
Patch by Hsiangkai Wang.
llvm-svn: 331841
2018-05-09 10:40:45 +08:00
|
|
|
!26 = distinct !DISubprogram(name: "phi_const_cast", linkageName: "phi_const_cast", scope: null, file: !1, line: 13, type: !6, isLocal: false, isDefinition: true, scopeLine: 13, isOptimized: true, unit: !0, retainedNodes: !27)
|
[FastISel] Sink local value materializations to first use
Summary:
Local values are constants, global addresses, and stack addresses that
can't be folded into the instruction that uses them. For example, when
storing the address of a global variable into memory, we need to
materialize that address into a register.
FastISel doesn't want to materialize any given local value more than
once, so it generates all local value materialization code at
EmitStartPt, which always dominates the current insertion point. This
allows it to maintain a map of local value registers, and it knows that
the local value area will always dominate the current insertion point.
The downside is that local value instructions are always emitted without
a source location. This is done to prevent jumpy line tables, but it
means that the local value area will be considered part of the previous
statement. Consider this C code:
call1(); // line 1
++global; // line 2
++global; // line 3
call2(&global, &local); // line 4
Today we end up with assembly and line tables like this:
.loc 1 1
callq call1
leaq global(%rip), %rdi
leaq local(%rsp), %rsi
.loc 1 2
addq $1, global(%rip)
.loc 1 3
addq $1, global(%rip)
.loc 1 4
callq call2
The LEA instructions in the local value area have no source location and
are treated as being on line 1. Stepping through the code in a debugger
and correlating it with the assembly won't make much sense, because
these materializations are only required for line 4.
This is actually problematic for the VS debugger "set next statement"
feature, which effectively assumes that there are no registers live
across statement boundaries. By sinking the local value code into the
statement and fixing up the source location, we can make that feature
work. This was filed as https://bugs.llvm.org/show_bug.cgi?id=35975 and
https://crbug.com/793819.
This change is obviously not enough to make this feature work reliably
in all cases, but I felt that it was worth doing anyway because it
usually generates smaller, more comprehensible -O0 code. I measured a
0.12% regression in code generation time with LLC on the sqlite3
amalgamation, so I think this is worth doing.
There are some special cases worth calling out in the commit message:
1. local values materialized for phis
2. local values used by no-op casts
3. dead local value code
Local values can be materialized for phis, and this does not show up as
a vreg use in MachineRegisterInfo. In this case, if there are no other
uses, this patch sinks the value to the first terminator, EH label, or
the end of the BB if nothing else exists.
Local values may also be used by no-op casts, which adds the register to
the RegFixups table. Without reversing the RegFixups map direction, we
don't have enough information to sink these instructions.
Lastly, if the local value register has no other uses, we can delete it.
This comes up when fastisel tries two instruction selection approaches
and the first materializes the value but fails and the second succeeds
without using the local value.
Reviewers: aprantl, dblaikie, qcolombet, MatzeB, vsk, echristo
Subscribers: dotdash, chandlerc, hans, sdardis, amccarth, javed.absar, zturner, llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D43093
llvm-svn: 327581
2018-03-15 05:54:21 +08:00
|
|
|
!27 = !{!28, !29, !30, !31}
|
|
|
|
|
!28 = !DILocalVariable(name: "3", scope: !26, file: !1, line: 13, type: !17)
|
|
|
|
|
!29 = !DILocalVariable(name: "4", scope: !26, file: !1, line: 15, type: !19)
|
|
|
|
|
!30 = !DILocalVariable(name: "5", scope: !26, file: !1, line: 17, type: !19)
|
|
|
|
|
!31 = !DILocalVariable(name: "6", scope: !26, file: !1, line: 19, type: !19)
|
|
|
|
|
!32 = !DILocation(line: 13, column: 1, scope: !26)
|
|
|
|
|
!33 = !DILocation(line: 14, column: 1, scope: !26)
|
|
|
|
|
!34 = !DILocation(line: 15, column: 1, scope: !26)
|
|
|
|
|
!35 = !DILocation(line: 16, column: 1, scope: !26)
|
|
|
|
|
!36 = !DILocation(line: 17, column: 1, scope: !26)
|
|
|
|
|
!37 = !DILocation(line: 18, column: 1, scope: !26)
|
|
|
|
|
!38 = !DILocation(line: 19, column: 1, scope: !26)
|
|
|
|
|
!39 = !DILocation(line: 20, column: 1, scope: !26)
|
[DebugInfo] Add DILabel metadata and intrinsic llvm.dbg.label.
In order to set breakpoints on labels and list source code around
labels, we need collect debug information for labels, i.e., label
name, the function label belong, line number in the file, and the
address label located. In order to keep these information in LLVM
IR and to allow backend to generate debug information correctly.
We create a new kind of metadata for labels, DILabel. The format
of DILabel is
!DILabel(scope: !1, name: "foo", file: !2, line: 3)
We hope to keep debug information as much as possible even the
code is optimized. So, we create a new kind of intrinsic for label
metadata to avoid the metadata is eliminated with basic block.
The intrinsic will keep existing if we keep it from optimized out.
The format of the intrinsic is
llvm.dbg.label(metadata !1)
It has only one argument, that is the DILabel metadata. The
intrinsic will follow the label immediately. Backend could get the
label metadata through the intrinsic's parameter.
We also create DIBuilder API for labels to be used by Frontend.
Frontend could use createLabel() to allocate DILabel objects, and use
insertLabel() to insert llvm.dbg.label intrinsic in LLVM IR.
Differential Revision: https://reviews.llvm.org/D45024
Patch by Hsiangkai Wang.
llvm-svn: 331841
2018-05-09 10:40:45 +08:00
|
|
|
!40 = distinct !DISubprogram(name: "invoke_phi", linkageName: "invoke_phi", scope: null, file: !1, line: 21, type: !6, isLocal: false, isDefinition: true, scopeLine: 21, isOptimized: true, unit: !0, retainedNodes: !41)
|
[FastISel] Sink local value materializations to first use
Summary:
Local values are constants, global addresses, and stack addresses that
can't be folded into the instruction that uses them. For example, when
storing the address of a global variable into memory, we need to
materialize that address into a register.
FastISel doesn't want to materialize any given local value more than
once, so it generates all local value materialization code at
EmitStartPt, which always dominates the current insertion point. This
allows it to maintain a map of local value registers, and it knows that
the local value area will always dominate the current insertion point.
The downside is that local value instructions are always emitted without
a source location. This is done to prevent jumpy line tables, but it
means that the local value area will be considered part of the previous
statement. Consider this C code:
call1(); // line 1
++global; // line 2
++global; // line 3
call2(&global, &local); // line 4
Today we end up with assembly and line tables like this:
.loc 1 1
callq call1
leaq global(%rip), %rdi
leaq local(%rsp), %rsi
.loc 1 2
addq $1, global(%rip)
.loc 1 3
addq $1, global(%rip)
.loc 1 4
callq call2
The LEA instructions in the local value area have no source location and
are treated as being on line 1. Stepping through the code in a debugger
and correlating it with the assembly won't make much sense, because
these materializations are only required for line 4.
This is actually problematic for the VS debugger "set next statement"
feature, which effectively assumes that there are no registers live
across statement boundaries. By sinking the local value code into the
statement and fixing up the source location, we can make that feature
work. This was filed as https://bugs.llvm.org/show_bug.cgi?id=35975 and
https://crbug.com/793819.
This change is obviously not enough to make this feature work reliably
in all cases, but I felt that it was worth doing anyway because it
usually generates smaller, more comprehensible -O0 code. I measured a
0.12% regression in code generation time with LLC on the sqlite3
amalgamation, so I think this is worth doing.
There are some special cases worth calling out in the commit message:
1. local values materialized for phis
2. local values used by no-op casts
3. dead local value code
Local values can be materialized for phis, and this does not show up as
a vreg use in MachineRegisterInfo. In this case, if there are no other
uses, this patch sinks the value to the first terminator, EH label, or
the end of the BB if nothing else exists.
Local values may also be used by no-op casts, which adds the register to
the RegFixups table. Without reversing the RegFixups map direction, we
don't have enough information to sink these instructions.
Lastly, if the local value register has no other uses, we can delete it.
This comes up when fastisel tries two instruction selection approaches
and the first materializes the value but fails and the second succeeds
without using the local value.
Reviewers: aprantl, dblaikie, qcolombet, MatzeB, vsk, echristo
Subscribers: dotdash, chandlerc, hans, sdardis, amccarth, javed.absar, zturner, llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D43093
llvm-svn: 327581
2018-03-15 05:54:21 +08:00
|
|
|
!41 = !{!42, !44}
|
|
|
|
|
!42 = !DILocalVariable(name: "7", scope: !40, file: !1, line: 23, type: !43)
|
|
|
|
|
!43 = !DIBasicType(name: "ty64", size: 64, encoding: DW_ATE_unsigned)
|
|
|
|
|
!44 = !DILocalVariable(name: "8", scope: !40, file: !1, line: 26, type: !19)
|
|
|
|
|
!45 = !DILocation(line: 21, column: 1, scope: !40)
|
|
|
|
|
!46 = !DILocation(line: 22, column: 1, scope: !40)
|
|
|
|
|
!47 = !DILocation(line: 23, column: 1, scope: !40)
|
|
|
|
|
!48 = !DILocation(line: 24, column: 1, scope: !40)
|
|
|
|
|
!49 = !DILocation(line: 25, column: 1, scope: !40)
|
|
|
|
|
!50 = !DILocation(line: 26, column: 1, scope: !40)
|
|
|
|
|
!51 = !DILocation(line: 27, column: 1, scope: !40)
|
|
|
|
|
!52 = !{i32 2, !"Dwarf Version", i32 4}
|
|
|
|
|
!53 = !{i32 2, !"Debug Info Version", i32 3}
|