[inline asm] Add -print-schedule scheduling comments to inline asm

[RKSimon]

Bugzilla Link	35162
Resolution	FIXED
Resolved on	Nov 19, 2017 06:13
Version	unspecified
OS	Windows NT
Blocks	#31672
CC	@adibiagio,@topperc,@efriedma-quic,@rotateright,@ZviRackover
Fixed by commit(s)	317782

Extended Description

As part of increasing scheduler model test coverage it would be very useful to be get scheduling comments on ir instructions from inline asm:

define i64 @​test_bsr64(i64 %a0, i64 *%a1) nounwind {
%1 = tail call i64 asm "bsr $1, $0", "=r,r"(i64 %a0) nounwind
%2 = tail call i64 asm "bsr $1, $0", "=r,*m"(i64 *%a1) nounwind
%3 = or i64 %1, %2
ret i64 %3
}

define i64 @​test_bswap64(i64 %a0) nounwind {
%1 = tail call i64 asm "bswap $0", "=r,0"(i64 %a0) nounwind
ret i64 %1
}

Interestingly the bswap case succeeds but the bsr cases fail:

llc -mtriple=x86_64-unknown -mcpu=btver2 -print-schedule

test_bsr64: # @​test_bsr64

BB#0:

#APP
bsrq	(%rsi), %rax
#NO_APP
#APP
bsrq	%rdi, %rcx
#NO_APP
orq	%rcx, %rax              #  sched: [1:0.50]
retq                            #  sched: [4:1.00]

test_bswap64: # @​test_bswap64

BB#0:

bswapq	%rdi                    #  sched: [1:0.50]
movq	%rdi, %rax              #  sched: [1:0.17]
retq                            #  sched: [4:1.00]

[efriedma-quic]

bswap "works" because we specifically have a hack in isel to turn it into an ISD::BSWAP (see X86TargetLowering::ExpandInlineAsm).

[RKSimon]

bswap "works" because we specifically have a hack in isel to turn it into an
ISD::BSWAP (see X86TargetLowering::ExpandInlineAsm).

Huh, I've never paid attention to that code before. Is it something we should be encouraging or is bswap just an aberation?

[efriedma-quic]

I don't think that's something we should be copying in general. Assembly is complicated enough without the compiler introducing unexpected transforms into user code.

The hack for bswap was added a very long time ago; see https://reviews.llvm.org/rL31999.

[llvmbot]

I understood the root case of the issue: inline asm uses a rather specific way to emit instructions that's why we don't see the scheduler numbers. I've already found the possible way to fix the issue and I'm going to publish it asap.

[RKSimon]

rL317782 - thanks Andrew