cmd/compile/internal/amd64: improve fix up code for signed division

In order to avoid a CPU exception resulting from signed overflow, the signed
division code tests if the divisor is -1 and if it is, runs fix up code to
manually compute the quotient and remainder (thus avoiding IDIV and potential
signed overflow).

However, the way that this is currently structured means that the normal code
path for the case where the divisor is not -1 results in five instructions
and two branches (CMP, JEQ, followed by sign extension, IDIV and another JMP
to skip over the fix up code).

Rework the fix up code such that the final JMP is incurred by the less likely
divisor is -1 code path, rather than more likely code path (which is already
more expensive due to IDIV). This result in a four instruction sequence
(CMP, JNE, sign extension, IDIV), with only a single branch.

Updates #59089

Change-Id: Ie8d065750a178518d7397e194920b201afeb0530
Reviewed-on: https://go-review.googlesource.com/c/go/+/482658
Run-TryBot: Joel Sing <[email protected]>
Reviewed-by: Keith Randall <[email protected]>
TryBot-Result: Gopher Robot <[email protected]>
Reviewed-by: Keith Randall <[email protected]>
Reviewed-by: Michael Knyszek <[email protected]>

src/cmd/compile/internal/amd64/ssa.go

@@ -336,7 +336,6 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
  // Result[0] (the quotient) is in AX.
  // Result[1] (the remainder) is in DX.
  r := v.Args[1].Reg()
- var j1 *obj.Prog
 
  var opCMP, opNEG, opSXD obj.As
  switch v.Op {
@@ -350,28 +349,19 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 
  // CPU faults upon signed overflow, which occurs when the most
  // negative int is divided by -1. Handle divide by -1 as a special case.
+ var j1, j2 *obj.Prog
  if ssa.DivisionNeedsFixUp(v) {
  c := s.Prog(opCMP)
  c.From.Type = obj.TYPE_REG
  c.From.Reg = r
  c.To.Type = obj.TYPE_CONST
  c.To.Offset = -1
- j1 = s.Prog(x86.AJEQ)
- j1.To.Type = obj.TYPE_BRANCH
- }
-
- // Sign extend dividend and perform division.
- s.Prog(opSXD)
- p := s.Prog(v.Op.Asm())
- p.From.Type = obj.TYPE_REG
- p.From.Reg = r
 
- if j1 != nil {
- // Skip over -1 fixup code.
- j2 := s.Prog(obj.AJMP)
- j2.To.Type = obj.TYPE_BRANCH
+ // Divisor is not -1, proceed with normal division.
+ j1 = s.Prog(x86.AJNE)
+ j1.To.Type = obj.TYPE_BRANCH
 
- // Issue -1 fixup code.
+ // Divisor is -1, manually compute quotient and remainder via fixup code.
  // n / -1 = -n
  n1 := s.Prog(opNEG)
  n1.To.Type = obj.TYPE_REG
@@ -383,7 +373,21 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
  // TODO(khr): issue only the -1 fixup code we need.
  // For instance, if only the quotient is used, no point in zeroing the remainder.
 
- j1.To.SetTarget(n1)
+ // Skip over normal division.
+ j2 = s.Prog(obj.AJMP)
+ j2.To.Type = obj.TYPE_BRANCH
+ }
+
+ // Sign extend dividend and perform division.
+ p := s.Prog(opSXD)
+ if j1 != nil {
+ j1.To.SetTarget(p)
+ }
+ p = s.Prog(v.Op.Asm())
+ p.From.Type = obj.TYPE_REG
+ p.From.Reg = r
+
+ if j2 != nil {
  j2.To.SetTarget(s.Pc())
  }