qbit/go
1
0
Fork 0
mirror of https://github.com/golang/go synced 2024-11-24 05:30:24 -07:00
Code Releases Activity

cmd/compile: better propagation of desired registers

Browse Source 
This fixes two independent problems:

We normally propagate desired registers backwards through opcodes that
are marked resultInArg0. Unfortunately for the desired register
computation, ADDQconst is not marked as resultInArg0. This is because
the amd64 backend can write it out as LEAQ instead if the input and
output registers don't match. For desired register purposes, we want
to treat ADDQconst as resultInArg0, so that we get an ADDQ instead of
a LEAQ if we can.

Desired registers don't currently work for tuple-generating opcodes.
Declare that the desired register applies to the first element of the
tuple, and propagate the desired register back through Select0.

Noticed when fixing #51964

Change-Id: I83346b988882cd58c2d7e7e5b419a2b9a244ab66
Reviewed-on: https://go-review.googlesource.com/c/go/+/396035
Trust: Keith Randall <khr@golang.org>
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: David Chase <drchase@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
This commit is contained in:
Keith Randall 2022-03-26 10:05:04 -07:00
parent 23756207fb
commit bdd0f0f780
1 changed files with 16 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
src/cmd/compile/internal/ssa/regalloc.go
View File

@ -1237,7 +1237,7 @@ func (s *regAllocState) regalloc(f *Func) {
desired.clobber(j.regs)
desired.add(v.Args[j.idx].ID, pickReg(j.regs))
}
if opcodeTable[v.Op].resultInArg0 {
if opcodeTable[v.Op].resultInArg0 || v.Op == OpAMD64ADDQconst || v.Op == OpAMD64ADDLconst || v.Op == OpSelect0 {
if opcodeTable[v.Op].commutative {
desired.addList(v.Args[1].ID, prefs)
}
@ -1598,11 +1598,13 @@ func (s *regAllocState) regalloc(f *Func) {
}
}
}
for _, r := range dinfo[idx].out {
if r != noRegister && (mask&^s.used)>>r&1 != 0 {
// Desired register is allowed and unused.
mask = regMask(1) << r
break
if out.idx == 0 { // desired registers only apply to the first element of a tuple result
for _, r := range dinfo[idx].out {
if r != noRegister && (mask&^s.used)>>r&1 != 0 {
// Desired register is allowed and unused.
mask = regMask(1) << r
break
}
}
}
// Avoid registers we're saving for other values.
@ -2581,7 +2583,12 @@ func (s *regAllocState) computeLive() {
desired.add(v.Args[j.idx].ID, pickReg(j.regs))
}
// Set desired register of input 0 if this is a 2-operand instruction.
if opcodeTable[v.Op].resultInArg0 {
if opcodeTable[v.Op].resultInArg0 || v.Op == OpAMD64ADDQconst || v.Op == OpAMD64ADDLconst || v.Op == OpSelect0 {
// ADDQconst is added here because we want to treat it as resultInArg0 for
// the purposes of desired registers, even though it is not an absolute requirement.
// This is because we'd rather implement it as ADDQ instead of LEAQ.
// Same for ADDLconst
// Select0 is added here to propagate the desired register to the tuple-generating instruction.
if opcodeTable[v.Op].commutative {
desired.addList(v.Args[1].ID, prefs)
}
@ -2706,6 +2713,8 @@ type desiredStateEntry struct {
ID ID
// Registers it would like to be in, in priority order.
// Unused slots are filled with noRegister.
// For opcodes that return tuples, we track desired registers only
// for the first element of the tuple.
regs [4]register
}