cmd/internal/obj/arm64: fix constant pool size calculation error

The current calculation method of constant pool size is:
  c.pool.size = -c.pool.size & (funcAlign - 1)
  c.pool.size += uint32(sz)
This doesn't make sense. This CL changes it as:
  if q.As == ADWORD {
    c.pool.size = roundUp(c.pool.size, 8)
  }
  c.pool.size += uint32(sz)
which takes into account the padding size generated by aligning DWORD to
8 bytes.

It's unnecessary to set the Pc field in addpool and addpool128 because
the Pc value will be reset in function span7, so remove the related lines.

Change-Id: I5eb8f259be55a6b97fc2c20958b4a602bffa4f88
Reviewed-on: https://go-review.googlesource.com/c/go/+/298609
Reviewed-by: eric fang <eric.fang@arm.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Trust: eric fang <eric.fang@arm.com>
Run-TryBot: eric fang <eric.fang@arm.com>
TryBot-Result: Go Bot <gobot@golang.org>

diff --git a/src/cmd/internal/obj/arm64/asm7.go b/src/cmd/internal/obj/arm64/asm7.go
index 20f1843951e6347510575c6ca84a0fc556a4ca18..b0e29c26e814497f0e0e4feccbc2584d58433918 100644 (file)
--- a/src/cmd/internal/obj/arm64/asm7.go
+++ b/src/cmd/internal/obj/arm64/asm7.go
@@ -1200,19 +1200,16 @@ func (c *ctxt7) flushpool(p *obj.Prog, skip int) {
 // addpool128 adds a 128-bit constant to literal pool by two consecutive DWORD
 // instructions, the 128-bit constant is formed by ah.Offset<<64+al.Offset.
 func (c *ctxt7) addpool128(p *obj.Prog, al, ah *obj.Addr) {
-       lit := al.Offset
        q := c.newprog()
        q.As = ADWORD
        q.To.Type = obj.TYPE_CONST
-       q.To.Offset = lit
-       q.Pc = int64(c.pool.size)
+       q.To.Offset = al.Offset
 
-       lit = ah.Offset
        t := c.newprog()
        t.As = ADWORD
        t.To.Type = obj.TYPE_CONST
-       t.To.Offset = lit
-       t.Pc = int64(c.pool.size + 8)
+       t.To.Offset = ah.Offset
+
        q.Link = t
 
        if c.blitrl == nil {
@@ -1223,6 +1220,7 @@ func (c *ctxt7) addpool128(p *obj.Prog, al, ah *obj.Addr) {
        }
 
        c.elitrl = t
+       c.pool.size = roundUp(c.pool.size, 16)
        c.pool.size += 16
        p.Pool = q
 }
@@ -1267,7 +1265,6 @@ func (c *ctxt7) addpool(p *obj.Prog, a *obj.Addr) {
 
        q := c.newprog()
        *q = *t
-       q.Pc = int64(c.pool.size)
        if c.blitrl == nil {
                c.blitrl = q
                c.pool.start = uint32(p.Pc)
@@ -1275,11 +1272,24 @@ func (c *ctxt7) addpool(p *obj.Prog, a *obj.Addr) {
                c.elitrl.Link = q
        }
        c.elitrl = q
-       c.pool.size = -c.pool.size & (funcAlign - 1)
+       if q.As == ADWORD {
+               // make DWORD 8-byte aligned, this is not required by ISA,
+               // just to avoid performance penalties when loading from
+               // the constant pool across a cache line.
+               c.pool.size = roundUp(c.pool.size, 8)
+       }
        c.pool.size += uint32(sz)
        p.Pool = q
 }
 
+// roundUp rounds up x to "to".
+func roundUp(x, to uint32) uint32 {
+       if to == 0 || to&(to-1) != 0 {
+               log.Fatalf("rounded up to a value that is not a power of 2: %d\n", to)
+       }
+       return (x + to - 1) &^ (to - 1)
+}
+
 func (c *ctxt7) regoff(a *obj.Addr) uint32 {
        c.instoffset = 0
        c.aclass(a)