EABI requires to pass structurs larger than 32 bits by reference

[chbessonova]

p. 3.5 MSP430-EABI

Structures or Unions Passed and Returned by Reference
Structures (including classes) and unions larger than 32 bits are passed and returned by reference.
To pass a structure or union by reference, the caller places its address in the appropriate location: either in a register or on the stack, according to its position in the argument list...

Example:

struct t { long i; long j; };
void boo(struct t a1) {}

int test() {
  struct t a1 = { 1000, 10001 };
  boo(a1);
  return 0;
}

clang:

test:                                   ; @test
; %bb.0:
	push	r4
	mov	r1, r4
	sub	#16, r1
	mov	&.Ltest.a1+6, r12
	mov	r12, -2(r4)
	mov	&.Ltest.a1+4, r12
	mov	r12, -4(r4)
	mov	&.Ltest.a1+2, r12
	mov	r12, -6(r4)
	mov	&.Ltest.a1, r12
	mov	r12, -8(r4)
	mov	-2(r4), r12
	mov	r1, r13
	mov	r12, 6(r13)
	mov	-4(r4), r12
	mov	r12, 4(r13)
	mov	-6(r4), r12
	mov	r12, 2(r13)
	mov	-8(r4), r12
	mov	r12, 0(r13)
	call	#boo
	clr	r12
	add	#16, r1
	pop	r4
	ret

clang -02:

test:                                   ; @test
; %bb.0:
	push	r4
	mov	r1, r4
	sub	#8, r1
	mov	&.Ltest.a1+6, 6(r1)
	mov	&.Ltest.a1+4, 4(r1)
	mov	&.Ltest.a1+2, 2(r1)
	mov	&.Ltest.a1, 0(r1)
	call	#boo
	clr	r12
	add	#8, r1
	pop	r4
	ret

gcc:

test:
	SUB.W	#8, R1
	MOV.W	#1000, @R1
	MOV.W	#0, 2(R1)
	MOV.W	#10001, 4(R1)
	MOV.W	#0, 6(R1)
	MOV.W	R1, R12
	CALL	#boo
	MOV.B	#0, R12
	ADD.W	#8, R1
	RET

[mskvortsov]

So in clang case we have a garbage in r12. This is incorrect.

[asl]

How does IR look like?

…

On Tue, Oct 23, 2018, 08:40 mskvortsov ***@***.***> wrote: So in clang case we have a garbage in r12. This is incorrect. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#18 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEHQFcTeJ9E8I0TWdlTiCaxCoi0nIPRks5unzhfgaJpZM4X1yLv> .

[chbessonova]

For -O2 version:

target datalayout = "e-m:e-p:16:16-i32:16-i64:16-f32:16-f64:16-a:8-n8:16-S16"
target triple = "msp430"

%struct.t = type { i32, i32 }

@test.a1 = private unnamed_addr constant %struct.t { i32 1000, i32 10001 }, align 2

; Function Attrs: nounwind
define dso_local i16 @test() local_unnamed_addr #0 {
  tail call void @boo(%struct.t* byval nonnull align 2 @test.a1) #2
  ret i16 0
}

declare dso_local void @boo(%struct.t* byval align 2) local_unnamed_addr #1

[asl]

Ok, so it seems that our byval lowering is broken.

…

On Tue, Oct 23, 2018, 08:59 chbessonova ***@***.***> wrote: For -O2 version: target datalayout = "e-m:e-p:16:16-i32:16-i64:16-f32:16-f64:16-a:8-n8:16-S16" target triple = "msp430" %struct.t = type { i32, i32 } @test.a1 = private unnamed_addr constant %struct.t { i32 1000, i32 10001 }, align 2 ; Function Attrs: nounwind define dso_local i16 @test() local_unnamed_addr #0 { tail call void @boo(%struct.t* byval nonnull align 2 @test.a1) #2 ret i16 0 } declare dso_local void @boo(%struct.t* byval align 2) local_unnamed_addr #1 — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#18 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEHQMcOVszRlpYfvbeLL_GdXkntNc3Cks5unzzrgaJpZM4X1yLv> .

[chbessonova]

Small note:
MSP430 EABI requires to pass/return structures larger than 32 bit by reference, but it says nothing specific about smaller structures (i.e. no requirements to pass/return them directly through registers/stack). Both TI GCC and TI compiler pass any structs (even smaller that 32 bit) by reference currently.