cmd/ld: race enabled binaries are too large

[minux]

on linux/amd64, go test -c fmt produces a binary of 3847664 bytes, however,
go test -c -race fmt produces a binary of 14653432 bytes, almost 10MB larger!

This issue also affects the 1.4 release.

The problem is that this symbol:
0000000000c09bc0 0000000000982648 d runtime/race(.bss)
which should be put into the bss section, actually consumes space in the binary.

You can also see the problem in the program headers:

Program Headers:
  Type           Offset             VirtAddr           PhysAddr
                 FileSiz            MemSiz              Flags  Align
  PHDR           0x0000000000000040 0x0000000000400040 0x0000000000400040
                 0x00000000000001f8 0x00000000000001f8  R      1000
  INTERP         0x0000000000000be4 0x0000000000400be4 0x0000000000400be4
                 0x000000000000001c 0x000000000000001c  R      1
      [Requesting program interpreter: /lib64/ld-linux-x86-64.so.2]
  LOAD           0x0000000000000040 0x0000000000400040 0x0000000000400040
                 0x0000000000458630 0x0000000000458630  R E    1000
  LOAD           0x0000000000459000 0x0000000000859000 0x0000000000859000
                 0x00000000003991d8 0x00000000003991d8  R      1000
  LOAD           0x00000000007f3000 0x0000000000bf3000 0x0000000000bf3000
                 0x00000000009a2fa0 0x00000000009bffe8  RW     1000
  DYNAMIC        0x00000000007f3140 0x0000000000bf3140 0x0000000000bf3140
                 0x0000000000000140 0x0000000000000140  RW     8
  TLS            0x0000000000000000 0x0000000000000000 0x0000000000000000
                 0x0000000000000000 0x0000000000000010  R      8
  GNU_STACK      0x0000000000000000 0x0000000000000000 0x0000000000000000
                 0x0000000000000000 0x0000000000000000  RW     8
  LOOS+5041580   0x0000000000000000 0x0000000000000000 0x0000000000000000
                 0x0000000000000000 0x0000000000000000         8

The file size and mem size for the RW section is almost identical.

The race runtime syso file is correct, our linker somehow failed to put the bss in real bss.

   text    data     bss     dec     hex filename
 159066     484 9971272 10130822         9a9586 src/runtime/race/race_linux_amd64.syso

Should we also fix this for 1.4.1? Every race enabled binary is 10MB larger than it should be right now.

[minux]

Fix is easy:

diff --git a/src/cmd/ld/ldelf.c b/src/cmd/ld/ldelf.c
index 243c8d8..4e87943 100644
--- a/src/cmd/ld/ldelf.c
+++ b/src/cmd/ld/ldelf.c
@@ -545,7 +545,10 @@ ldelf(Biobuf *f, char *pkg, int64 len, char *pn)
                        s->type = SRODATA;
                        break;
                case ElfSectFlagAlloc + ElfSectFlagWrite:
-                       s->type = SNOPTRDATA;
+                       if(sect->type == ElfSectNobits)
+                               s->type = SNOPTRBSS;
+                       else
+                               s->type = SNOPTRDATA;
                        break;
                case ElfSectFlagAlloc + ElfSectFlagExec:
                        s->type = STEXT;

Candidate for 1.4.1?

[minux]

Turns out this bug is present ever since Go supports loading ELF files.
We didn't notice that because before 1.4, the race runtime doesn't
use that much BSS (up to 1.3.3, it only uses ~72KB of BSS).

Anyway, I believe this still is a 1.4 regression without any meaningful
workaround.

[davecheney]

That would explain why the release $GOROOT/bin/go is double the size of a
locally compiled version

lucky(/tmp/go/bin) % ls -l go ~/go/bin/go
-rwxr-xr-x 1 dfc dfc 21435888 Dec 11 12:20 go         // from golang.org
-rwxr-xr-x 1 dfc dfc  9746184 Dec 16 21:05 /home/dfc/go/bin/go  // compiled
locally

[minux]

mailed https://golang.org/cl/1889