[CodeGenPrepare] Store splitting creates invalid alignments on big-endian targets, and under-aligned stores on all targets

[legrosbuffle]

Bugzilla Link	44877
Resolution	FIXED
Resolved on	Feb 19, 2020 09:46
Version	trunk
OS	Linux
Blocks	#43900
CC	@gchatelet,@zmodem,@LebedevRI,@smeenai,@rotateright

[legrosbuffle]

Full discussion: https://reviews.llvm.org/D74311

[legrosbuffle]

Fixed by rG15488ff24b4a

[zmodem]

From the code review:
"We probably want to backport the big-endian fix into 10.0 in some form?"

Clement, can you help with this? I tried cherry-picking the whole commit over but it doesn't apply cleanly.

[legrosbuffle]

Clement, can you help with this? I tried cherry-picking the whole commit
over but it doesn't apply cleanly.

I guess the migration to typed Align() had not made it to 10.0.

I'll need to do a manual merge. How can I get the patch to you afterwards (sorry, first time this happens to me) ? Thanks.

[legrosbuffle]

cherry-pick of rG15488ff24b4a on top of 10

[zmodem]

Created attachment 23122 [details]
cherry-pick of rG15488ff24b4a on top of 10

Thanks! Pushed that as b3cf704

[smeenai]

I’m seeing a failure related to the relevant commit on the release/10.x branch (b3cf704). Specifically, I see a failure in the test “Transforms/CodeGenPrepare/PowerPC/split-store-alignment.ll”. Here are the results from me locally running this:

apl@apl-mbp ~/code/llvm-project/build (release/10.x) $ ./bin/llvm-lit -sv test/Transforms/CodeGenPrepare/PowerPC
FAIL: LLVM :: Transforms/CodeGenPrepare/PowerPC/split-store-alignment.ll (1 of 1)
******************** TEST 'LLVM :: Transforms/CodeGenPrepare/PowerPC/split-store-alignment.ll' FAILED ********************
Script:

: 'RUN: at line 2'; /Users/apl/code/llvm-project/build/bin/opt -S -codegenprepare -mtriple=powerpc64-unknown-linux-gnu -data-layout="E-m:e-i64:64-n32:64" -force-split-store < /Users/apl/code/llvm-project/llvm/test/Transforms/CodeGenPrepare/PowerPC/split-store-alignment.ll | /Users/apl/code/llvm-project/build/bin/FileCheck --check-prefixes=ALL,BE /Users/apl/code/llvm-project/llvm/test/Transforms/CodeGenPrepare/PowerPC/split-store-alignment.ll
: 'RUN: at line 3'; /Users/apl/code/llvm-project/build/bin/opt -S -codegenprepare -mtriple=powerpc64le-unknown-linux-gnu -data-layout="e-m:e-i64:64-n32:64" -force-split-store < /Users/apl/code/llvm-project/llvm/test/Transforms/CodeGenPrepare/PowerPC/split-store-alignment.ll | /Users/apl/code/llvm-project/build/bin/FileCheck --check-prefixes=ALL,LE /Users/apl/code/llvm-project/llvm/test/Transforms/CodeGenPrepare/PowerPC/split-store-alignment.ll

Exit Code: 1

Command Output (stderr):

/Users/apl/code/llvm-project/llvm/test/Transforms/CodeGenPrepare/PowerPC/split-store-alignment.ll:12:12: error: BE-NEXT: expected string not found in input
; BE-NEXT: [[TMP1:%.]] = bitcast i64 [[P:%.]] to i32
^
:12:2: note: scanning from here
store i64 %o, i64* %p, align 1
^
/Users/apl/code/llvm-project/llvm/test/Transforms/CodeGenPrepare/PowerPC/split-store-alignment.ll:48:12: error: BE-NEXT: expected string not found in input
; BE-NEXT: [[TMP1:%.]] = bitcast i64 [[P:%.]] to i32
^
:22:2: note: scanning from here
store i64 %o, i64* %p, align 2
^
/Users/apl/code/llvm-project/llvm/test/Transforms/CodeGenPrepare/PowerPC/split-store-alignment.ll:84:12: error: BE-NEXT: expected string not found in input
; BE-NEXT: [[TMP1:%.]] = bitcast i64 [[P:%.]] to i32
^
:32:2: note: scanning from here
store i64 %o, i64* %p, align 8
^

--

---

It looks like the test is expecting to see several things between the or and the store but running the first invocation of opt yields this:
define void @​split_store_align1(float %x, i64* %p) {
%b = bitcast float %x to i32
%z = zext i32 0 to i64
%s = shl nuw nsw i64 %z, 32
%z2 = zext i32 %b to i64
%o = or i64 %s, %z2
store i64 %o, i64* %p, align 1
ret void
}

[smeenai]

(posting that on behalf of a coworker who doesn't have a Bugzilla account yet, but it's blocking our internal integration of 10.0)

[zmodem]

That's strange. The test passes on my machine, and the output of the first run line looks like this:

$ /work/llvm.monorepo.rel/build.release/bin/opt -S -codegenprepare -mtriple=powerpc64-unknown-linux-gnu -data-layout="E-m:e-i64:64-n32:64" -force-split-store < /work/llvm.monorepo.rel/llvm/test/Transforms/CodeGenPrepare/PowerPC/split-store-alignment.ll | head -19
; ModuleID = ''
source_filename = ""
target datalayout = "E-m:e-i64:64-n32:64"
target triple = "powerpc64-unknown-linux-gnu"

define void @​split_store_align1(float %x, i64* %p) {
%b = bitcast float %x to i32
%z = zext i32 0 to i64
%s = shl nuw nsw i64 %z, 32
%z2 = zext i32 %b to i64
%o = or i64 %s, %z2
%1 = bitcast i64* %p to i32*
%2 = getelementptr i32, i32* %1, i32 1
store i32 %b, i32* %2, align 1
%3 = bitcast i64* %p to i32*
store i32 0, i32* %3, align 1
ret void
}

Could it be due to a difference in host architecture or build configuration or something?

[legrosbuffle]

Hm. I think I know what's happening. Are you by any chance compiling without support for PowerPC ? I just noticed that I failed to backport the base commit for the change, which included the addition of the local lit config checking for PowerPC targets.

[legrosbuffle]

Indeed, I could reproduce by compiling only with X86 support.
I have committed be45a5a, which fixes the issue on my machine.

[zmodem]

Indeed, I could reproduce by compiling only with X86 support.
I have committed be45a5a, which fixes the
issue on my machine.

Thanks!

[smeenai]

Yup, we weren't enabling the PowerPC target. Thanks for the fix!