This is an archive of the discontinued LLVM Phabricator instance.

Differential D20852

[AMDGPU] Remove exit-on-error in test (PR27761)
ClosedPublic

Authored by rovka on Jun 1 2016, 1:15 AM.

Details

Reviewers
• tstellarAMD
arsenm
Commits
rGe440f99913f9: [AMDGPU] Remove exit-on-error in test (PR27761)
rL273550: [AMDGPU] Remove exit-on-error in test (PR27761)
Summary

The exit-on-error flag was necessary in order to avoid an assertion when handling DYNAMIC_STACKALLOC nodes in SelectionDAGLegalize.

We can avoid the assertion by creating some dummy nodes. This enables us to remove the exit-on-error flag on the first 2 run lines (SI), but on the third run line (R600) we would run into another assertion when trying to reserve indirect registers. This patch also replaces that assertion with an early exit from the function.

Fixes PR27761.

Last exit-on-error patch for AMDGPU

Diff Detail

Event Timeline

rovka updated this revision to Diff 59177.Jun 1 2016, 1:15 AM
rovka retitled this revision from to [AMDGPU] Remove exit-on-error in test (PR27761).
rovka updated this object.
rovka added reviewers: arsenm, • tstellarAMD.
rovka added subscribers: llvm-commits, rengolin.
Herald added subscribers: kzhuravl, arsenm. · View Herald TranscriptJun 1 2016, 1:15 AM
rovka updated this revision to Diff 59710.Jun 6 2016, 6:10 AM

Update some tests that I initially forgot about.

rovka updated this revision to Diff 59711.Jun 6 2016, 6:13 AM

...with the initial changes too, ofc.

arsenm added inline comments.Jun 6 2016, 1:37 PM
lib/Target/AMDGPU/AMDGPUISelLowering.cpp
692 ↗(On Diff #59711)

I would prefer not using getIntPtrConstant since I would like to remove it in general from LLVM. This would more preferrably be getConstant(-1, Op.getValueType())

test/CodeGen/AMDGPU/promote-alloca-bitcast-function.ll
6

I think this can now have a check for the second function's error

rovka updated this revision to Diff 59850.Jun 7 2016, 3:34 AM

Using the -1 constant causes some assertions later on, but 0 works like a charm.
Updated the test, too.
Thanks.

rovka added a child revision: D21368: [llc] Remove exit-on-error flag (PR27759).Jun 15 2016, 3:50 AM
rovka mentioned this in D21368: [llc] Remove exit-on-error flag (PR27759).
rovka marked 2 inline comments as done.Jun 15 2016, 3:55 AM

Ping?

arsenm accepted this revision.Jun 22 2016, 9:58 AM
arsenm edited edge metadata.

LGTM

test/CodeGen/AMDGPU/private-memory-broken.ll
1–2

Can you remove the -mcpu=SI? I've been removing them from tests as they get touched

This revision is now accepted and ready to land.Jun 22 2016, 9:58 AM
rovka added inline comments.Jun 22 2016, 10:33 AM
test/CodeGen/AMDGPU/private-memory-broken.ll
1–2

Will do, thanks.

Closed by commit rL273550: [AMDGPU] Remove exit-on-error in test (PR27761) (authored by rovka). · Explain WhyJun 23 2016, 2:26 AM
This revision was automatically updated to reflect the committed changes.

Revision Contents

PathSize
test/
CodeGen/
AMDGPU/
4 lines
2 lines

Diff 59710

test/CodeGen/AMDGPU/private-memory-broken.ll

; RUN: not llc -verify-machineinstrs -march=amdgcn -mcpu=SI -exit-on-error %s -o /dev/null 2>&1 | FileCheck %s ; RUN: not llc -verify-machineinstrs -march=amdgcn -mcpu=SI %s -o /dev/null 2>&1 | FileCheck %s
; RUN: not llc -verify-machineinstrs -march=amdgcn -mcpu=tonga -exit-on-error %s -o /dev/null 2>&1 | FileCheck %s ; RUN: not llc -verify-machineinstrs -march=amdgcn -mcpu=tonga %s -o /dev/null 2>&1 | FileCheck %s
arsenmUnsubmitted
Not Done
ReplyInline Actions

Can you remove the -mcpu=SI? I've been removing them from tests as they get touched

arsenm: Can you remove the -mcpu=SI? I've been removing them from tests as they get touched
rovkaAuthorUnsubmitted
Not Done
ReplyInline Actions

Will do, thanks.

rovka: Will do, thanks.
; Make sure promote alloca pass doesn't crash ; Make sure promote alloca pass doesn't crash
; CHECK: unsupported call ; CHECK: unsupported call
declare i32 @foo(i32*) nounwind declare i32 @foo(i32*) nounwind
define void @call_private(i32 addrspace(1)* %out, i32 %in) nounwind { define void @call_private(i32 addrspace(1)* %out, i32 %in) nounwind {
Show All 11 Lines

test/CodeGen/AMDGPU/promote-alloca-bitcast-function.ll

; RUN: not llc -march=amdgcn -exit-on-error < %s 2>&1 | FileCheck %s ; RUN: not llc -march=amdgcn < %s 2>&1 | FileCheck %s
; Make sure that AMDGPUPromoteAlloca doesn't crash if the called ; Make sure that AMDGPUPromoteAlloca doesn't crash if the called
; function is a constantexpr cast of a function. ; function is a constantexpr cast of a function.
declare void @foo(float*) #0 declare void @foo(float*) #0
arsenmUnsubmitted
Done
ReplyInline Actions

I think this can now have a check for the second function's error

arsenm: I think this can now have a check for the second function's error
declare void @foo.varargs(...) #0 declare void @foo.varargs(...) #0
; CHECK: in function crash_call_constexpr_cast{{.*}}: unsupported call to function foo ; CHECK: in function crash_call_constexpr_cast{{.*}}: unsupported call to function foo
define void @crash_call_constexpr_cast() #0 { define void @crash_call_constexpr_cast() #0 {
%alloca = alloca i32 %alloca = alloca i32
call void bitcast (void (float*)* @foo to void (i32*)*)(i32* %alloca) #0 call void bitcast (void (float*)* @foo to void (i32*)*)(i32* %alloca) #0
ret void ret void
} }
define void @crash_call_constexpr_cast_varargs() #0 { define void @crash_call_constexpr_cast_varargs() #0 {
%alloca = alloca i32 %alloca = alloca i32
call void bitcast (void (...)* @foo.varargs to void (i32*)*)(i32* %alloca) #0 call void bitcast (void (...)* @foo.varargs to void (i32*)*)(i32* %alloca) #0
ret void ret void
} }
attributes #0 = { nounwind } attributes #0 = { nounwind }