[NFC] [TEST] Fix the threads.ll for Windows

[llitchev]

Summary

The codegen for Windows and Unix is different and 2 of the threads.ll were failing on Windows.

[ftynse]

Added a relevant code owner as reviewer. This test has been failing for several weeks now.

[nigelp-xmos]

The specified XCore backend code owner has moved on from XMOS and from working on XCore. I will have a look at this patch.

[ftynse]

In D93625#2477200, @nigelp-xmos wrote:

The specified XCore backend code owner has moved on from XMOS and from working on XCore. I will have a look at this patch.

Thanks! Note that they are still listed as a code owner - https://github.com/llvm/llvm-project/blob/main/llvm/CODE_OWNERS.TXT#L155.

[nigelp-xmos]

I've reproduced the failure. The test passes when I change -march=xcore to -mtriple=xcore-unknown-unknown, as in main branch commit https://github.com/llvm/llvm-project/commit/15ca54525d6c2927b2a51b871a9e343c7ce1c2ea. Does that commit solve this problem and unblock D91556 ? (I will continue to investigate to understand the reason for the register swap.)

[nigelp-xmos]

I think this patch can be cancelled, because the issue is already fixed in the main branch. But please say if I have missed something.

Test threads.ll started failing on Windows in commit https://github.com/llvm/llvm-project/commit/961f31d8ad14c66829991522d73e14b5a96ff6d4 . That commit changed function TargetMachine::shouldAssumeDSOLocal(). Previously, for both Windows and non-Windows, in f_tle in this test, in the DAG combine phase, for global tle:

TargetMachine::shouldAssumeDSOLocal returned true to:
TargetLowering::isOffsetFoldingLegal which returned true to:
SelectionDAG::isConstantIntBuildVectorOrConstantInt which returned global address node to:
DAGCombiner::visitADDLike which put the operands in a certain order.

With the new code, for non-Windows, in f_tle, shouldAssumeDSOLocal returns false, and the operands are not reordered in that way. But for Windows, the function still returns true before the new code is reached, because of special case:

if (TT.isOSBinFormatCOFF() || TT.isOSWindows()) return true;

So the new code produces different output for Windows and non-Windows.

Since we do not need to target Windows on XCore, we do not need to check for specific Windows results, and can use -mtriple to set OS unknown. So I think the existing solution in branch main already resolves the issue of the difference on Windows which this patch is addressing.

[llitchev]

That makes sense. If it works and there is no need to verify the tests passes for Windows (with Windows configuration) what we have in master is fine.

llvm/test/CodeGen/XCore/threads.ll

 ; CHECK: ldaw [[R0:r[0-9]]], dp[tl]
 ; CHECK: ldc [[R1:r[0-9]]], 8
 ; CHECK: ldc [[R2:r[0-9]]], 12
 ; r0 = id*12 + 8 + &tl
 ; CHECK: lmul {{r[0-9]}}, r0, r11, [[R2]], [[R0]], [[R1]]
   ret i32* getelementptr inbounds ([3 x i32], [3 x i32]* @tl, i32 0, i32 2)
 }
 
+; This test fails on Windows because the second and third
+; register of the add operations are swapped.
+; Windows test is below.
+; REQUIRES: !windows
 define i32* @f_tle() {
 ; CHECK-LABEL: f_tle:
 ; CHECK: get r11, id
 ; CHECK: shl [[R0:r[0-9]]], r11, 3
 ; CHECK: ldaw [[R1:r[0-9]]], dp[tle]
 ; r0 = &tl + id*8
 ; CHECK: add r0, [[R1]], [[R0]]
   ret i32* getelementptr inbounds ([2 x i32], [2 x i32]* @tle, i32 0, i32 0)
 }
 
+; Windows version of the above test.
+; REQUIRES: windows
+define i32* @f_tle_win() {
+; CHECK-LABEL: f_tle_win:
+; CHECK: get r11, id
+; CHECK: shl [[R0:r[0-9]]], r11, 3
+; CHECK: ldaw [[R1:r[0-9]]], dp[tle]
+; r0 = &tl + id*8
+; CHECK: add r0, [[R0]], [[R1]]
+  ret i32* getelementptr inbounds ([2 x i32], [2 x i32]* @tle, i32 0, i32 0)
+}
+
+; This test fails on Windows because the second and third
+; register of the first add operations are swapped.
+; Windows test is below.
+; REQUIRES: !windows
 define i32 @f_tlExpr () {
 ; CHECK-LABEL: f_tlExpr:
 ; CHECK: get r11, id
 ; CHECK: shl [[R0:r[0-9]]], r11, 3
 ; CHECK: ldaw [[R1:r[0-9]]], dp[tle]
 ; CHECK: add [[R2:r[0-9]]], [[R1]], [[R0]]
 ; CHECK: add r0, [[R2]], [[R2]]
   ret i32 add(
       i32 ptrtoint( i32* getelementptr inbounds ([2 x i32], [2 x i32]* @tle, i32 0, i32 0) to i32),
       i32 ptrtoint( i32* getelementptr inbounds ([2 x i32], [2 x i32]* @tle, i32 0, i32 0) to i32))
 }
 
+; Windows version of the above test.
+; REQUIRES: windows
+define i32 @f_tlExpr_win () {
+; CHECK-LABEL: f_tlExpr_win:
+; CHECK: get r11, id
+; CHECK: shl [[R0:r[0-9]]], r11, 3
+; CHECK: ldaw [[R1:r[0-9]]], dp[tle]
+; CHECK: add [[R2:r[0-9]]], [[R0]], [[R1]]
+; CHECK: add r0, [[R2]], [[R2]]
+  ret i32 add(
+      i32 ptrtoint( i32* getelementptr inbounds ([2 x i32], [2 x i32]* @tle, i32 0, i32 0) to i32),
+      i32 ptrtoint( i32* getelementptr inbounds ([2 x i32], [2 x i32]* @tle, i32 0, i32 0) to i32))
+}
+
 define void @phiNode1() {
 ; N.B. lowering of duplicate constexpr in a PHI node requires -O=0
 ; PHINODE-LABEL: phiNode1:
 ; PHINODE: get r11, id
 ; PHINODE-LABEL: .LBB11_1:
 ; PHINODE: get r11, id
 ; PHINODE: bu .LBB11_1
 entry: