GHC creates untagged closure references in strict field of static closures.

Pretty defines a handful of shorthands of the form: lparen = char '('

These get compiled to static closures as expected:

STG:

Pretty.lparen1 :: Pretty.TextDetails
[GblId, Caf=NoCafRefs, Str=m1, Unf=OtherCon []] =
    CCS_DONT_CARE Pretty.Chr! ['('#];

Pretty.lparen :: Pretty.Doc
[GblId, Caf=NoCafRefs, Str=m3, Unf=OtherCon []] =
    CCS_DONT_CARE Pretty.TextBeside! [Pretty.lparen1 1# Pretty.Empty];

==================== Asm code ====================
2019-04-08 21:17:52.2702151 UTC

.section .data
.align 8
.align 1
.globl Pretty.lparen1_closure
Pretty.lparen1_closure:
	.quad	Pretty.Chr_con_info
	.quad	40


==================== Asm code ====================
2019-04-08 21:17:52.2891641 UTC

.section .data
.align 8
.align 1
.globl Pretty.lparen_closure
Pretty.lparen_closure:
	.quad	Pretty.TextBeside_con_info
	.quad	Pretty.lparen1_closure+1
	.quad	Pretty.Empty_closure+1
	.quad	1
	.quad	3

So far so good, however in the X86 pretty printer we have this code
which is essentially a redefinition of lparen:

stg

lvl67_rvZv :: Pretty.Doc
[GblId, Caf=NoCafRefs, Str=m3, Unf=OtherCon []] =
    CCS_DONT_CARE Pretty.TextBeside! [Pretty.lparen1 1# Pretty.Empty];

==================== Asm code ====================
2019-04-08 23:28:52.8210512 UTC

.section .data
.align 8
.align 1
lvl67_rvZv_closure:
	.quad	Pretty.TextBeside_con_info
	.quad	Pretty.lparen1_closure
	.quad	Pretty.Empty_closure+1
	.quad	1
	.quad	3

Where lparen doesn't get tagged. It seems related to #15155 (closed) however there is no casting going on in this case.

This happens with all the char docs as far as I can tell.

So far I wasn't able to reproduce the issue outside of GHC. But I'Ve reproduced it on both Linux and Windows with HEAD.

Sidenote: It also seems we inline all the lparens, rparens, ... bindings just to then then float them out again, which causes them to be duplicated. Once in the pretty module, once in Ppr. Maybe also something we could avoid.