release/19.x: [llvm][CodeGen] Resolve issues when updating live intervals in window scheduler (#101945) #107338
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@arsenm What do you think about merging this PR to the release branch? |
|
@llvm/pr-subscribers-backend-hexagon Author: None (llvmbot) ChangesBackport 43ba109 Requested by: @huaatian Full diff: https://github.com/llvm/llvm-project/pull/107338.diff 2 Files Affected:
diff --git a/llvm/lib/CodeGen/ModuloSchedule.cpp b/llvm/lib/CodeGen/ModuloSchedule.cpp
index 0f29ebe3ee798d..b1a2bfaf789577 100644
--- a/llvm/lib/CodeGen/ModuloSchedule.cpp
+++ b/llvm/lib/CodeGen/ModuloSchedule.cpp
@@ -130,6 +130,7 @@ void ModuloScheduleExpander::generatePipelinedLoop() {
// Generate the prolog instructions that set up the pipeline.
generateProlog(MaxStageCount, KernelBB, VRMap, PrologBBs);
MF.insert(BB->getIterator(), KernelBB);
+ LIS.insertMBBInMaps(KernelBB);
// Rearrange the instructions to generate the new, pipelined loop,
// and update register names as needed.
@@ -210,6 +211,7 @@ void ModuloScheduleExpander::generateProlog(unsigned LastStage,
NewBB->transferSuccessors(PredBB);
PredBB->addSuccessor(NewBB);
PredBB = NewBB;
+ LIS.insertMBBInMaps(NewBB);
// Generate instructions for each appropriate stage. Process instructions
// in original program order.
@@ -283,6 +285,7 @@ void ModuloScheduleExpander::generateEpilog(
PredBB->replaceSuccessor(LoopExitBB, NewBB);
NewBB->addSuccessor(LoopExitBB);
+ LIS.insertMBBInMaps(NewBB);
if (EpilogStart == LoopExitBB)
EpilogStart = NewBB;
diff --git a/llvm/test/CodeGen/Hexagon/swp-ws-live-intervals.mir b/llvm/test/CodeGen/Hexagon/swp-ws-live-intervals.mir
new file mode 100644
index 00000000000000..7fa3cdf62d0902
--- /dev/null
+++ b/llvm/test/CodeGen/Hexagon/swp-ws-live-intervals.mir
@@ -0,0 +1,217 @@
+# REQUIRES: asserts
+#
+# RUN: llc --march=hexagon %s -run-pass=pipeliner -debug-only=pipeliner \
+# RUN: -window-sched=force -filetype=null -window-search-num=100 \
+# RUN: -window-search-ratio=100 -window-diff-limit=0 -verify-machineinstrs \
+# RUN: 2>&1 | FileCheck %s
+
+# The bug was reported at https://github.com/llvm/llvm-project/pull/99454.
+# It is caused by the corruption of live intervals in certain scenarios.
+#
+# We check the newly generated MBBs after successful scheduling here.
+# CHECK: Best window offset is {{[0-9]+}} and Best II is {{[0-9]+}}.
+# CHECK: prolog:
+# CHECK: bb.5:
+# CHECK: New block
+# CHECK: bb.6:
+# CHECK: epilog:
+# CHECK: bb.7:
+# CHECK: Best window offset is {{[0-9]+}} and Best II is {{[0-9]+}}.
+# CHECK: prolog:
+# CHECK: bb.8:
+# CHECK: New block
+# CHECK: bb.9:
+# CHECK: epilog:
+# CHECK: bb.10:
+
+--- |
+ target triple = "hexagon"
+
+ @_dp_ctrl_calc_tu_temp2_fp = global i64 0
+ @_dp_ctrl_calc_tu_temp1_fp = global i32 0
+ @dp_panel_update_tu_timings___trans_tmp_5 = global i64 0
+ @_dp_ctrl_calc_tu___trans_tmp_8 = global i64 0
+
+ declare void @llvm.lifetime.start.p0(i64 immarg, ptr nocapture)
+ declare i8 @div64_u64_rem(i32, ptr)
+ declare void @llvm.lifetime.end.p0(i64 immarg, ptr nocapture)
+
+ define void @dp_ctrl_calc_tu_parameters() {
+ if.end.i:
+ %rem.i11.i = alloca i64, align 8
+ %rem.i.i = alloca i64, align 8
+ call void @llvm.lifetime.start.p0(i64 8, ptr nonnull %rem.i11.i)
+ %call.i.i = call i8 @div64_u64_rem(i32 0, ptr nonnull %rem.i11.i)
+ %conv1.i.i = zext i8 %call.i.i to i64
+ %rem.promoted.i.i = load i64, ptr %rem.i11.i, align 8
+ br label %do.body.i.i
+
+ do.body.i.i:
+ %lsr.iv1 = phi i32 [ %lsr.iv.next2, %do.body.i.i ], [ -32, %if.end.i ]
+ %sub9.i.i = phi i64 [ %rem.promoted.i.i, %if.end.i ], [ %sub8.i.i.7, %do.body.i.i ]
+ %res_abs.0.i.i = phi i64 [ %conv1.i.i, %if.end.i ], [ %res_abs.1.i.i.7, %do.body.i.i ]
+ %cmp.not.i.i = icmp ne i64 %sub9.i.i, 0
+ %sub.i.neg.i = sext i1 %cmp.not.i.i to i64
+ %sub8.i.i = add i64 %sub9.i.i, %sub.i.neg.i
+ %0 = shl i64 %res_abs.0.i.i, 2
+ %1 = select i1 %cmp.not.i.i, i64 2, i64 0
+ %shl.i.i.5 = or disjoint i64 %0, %1
+ %cmp.not.i.i.5 = icmp ne i64 %sub8.i.i, 0
+ %sub.i.neg.i.5 = sext i1 %cmp.not.i.i.5 to i64
+ %sub8.i.i.5 = add i64 %sub8.i.i, %sub.i.neg.i.5
+ %or.i.i.5 = zext i1 %cmp.not.i.i.5 to i64
+ %res_abs.1.i.i.5 = or disjoint i64 %shl.i.i.5, %or.i.i.5
+ %cmp.not.i.i.6 = icmp ne i64 %sub8.i.i.5, 0
+ %sub.i.neg.i.6 = sext i1 %cmp.not.i.i.6 to i64
+ %sub8.i.i.6 = add i64 %sub8.i.i.5, %sub.i.neg.i.6
+ %2 = shl i64 %res_abs.1.i.i.5, 2
+ %3 = select i1 %cmp.not.i.i.6, i64 2, i64 0
+ %shl.i.i.7 = or disjoint i64 %2, %3
+ %cmp.not.i.i.7 = icmp ne i64 %sub8.i.i.6, 0
+ %sub.i.neg.i.7 = sext i1 %cmp.not.i.i.7 to i64
+ %sub8.i.i.7 = add i64 %sub8.i.i.6, %sub.i.neg.i.7
+ %or.i.i.7 = zext i1 %cmp.not.i.i.7 to i64
+ %res_abs.1.i.i.7 = or disjoint i64 %shl.i.i.7, %or.i.i.7
+ %lsr.iv.next2 = add nsw i32 %lsr.iv1, 8
+ %tobool.not.i.i.7 = icmp eq i32 %lsr.iv.next2, 0
+ br i1 %tobool.not.i.i.7, label %fec_check.i, label %do.body.i.i
+
+ fec_check.i:
+ call void @llvm.lifetime.end.p0(i64 8, ptr nonnull %rem.i11.i)
+ store i64 %res_abs.1.i.i.7, ptr @_dp_ctrl_calc_tu_temp2_fp, align 8
+ call void @llvm.lifetime.start.p0(i64 8, ptr nonnull %rem.i11.i)
+ %call.i12.i = call i8 @div64_u64_rem(i32 0, ptr nonnull %rem.i11.i)
+ %conv1.i13.i = zext i8 %call.i12.i to i64
+ %rem.promoted.i14.i = load i64, ptr %rem.i11.i, align 8
+ br label %do.body.i15.i
+
+ do.body.i15.i:
+ %lsr.iv = phi i32 [ %lsr.iv.next, %do.body.i15.i ], [ -32, %fec_check.i ]
+ %sub9.i16.i = phi i64 [ %rem.promoted.i14.i, %fec_check.i ], [ %sub8.i22.i.7, %do.body.i15.i ]
+ %res_abs.0.i17.i = phi i64 [ %conv1.i13.i, %fec_check.i ], [ %res_abs.1.i24.i.7, %do.body.i15.i ]
+ %cmp.not.i20.i = icmp ugt i64 %sub9.i16.i, 999
+ %sub.i21.neg.i = select i1 %cmp.not.i20.i, i64 -1000, i64 0
+ %sub8.i22.i = add i64 %sub.i21.neg.i, %sub9.i16.i
+ %4 = shl i64 %res_abs.0.i17.i, 2
+ %5 = select i1 %cmp.not.i20.i, i64 2, i64 0
+ %shl.i19.i.7 = or disjoint i64 %4, %5
+ %cmp.not.i20.i.7 = icmp ugt i64 %sub8.i22.i, 999
+ %sub.i21.neg.i.7 = select i1 %cmp.not.i20.i.7, i64 -1000, i64 0
+ %sub8.i22.i.7 = add i64 %sub.i21.neg.i.7, %sub8.i22.i
+ %or.i23.i.7 = zext i1 %cmp.not.i20.i.7 to i64
+ %res_abs.1.i24.i.7 = or disjoint i64 %shl.i19.i.7, %or.i23.i.7
+ %lsr.iv.next = add nsw i32 %lsr.iv, 8
+ %tobool.not.i26.i.7 = icmp eq i32 %lsr.iv.next, 0
+ br i1 %tobool.not.i26.i.7, label %_dp_ctrl_calc_tu.exit, label %do.body.i15.i
+
+ _dp_ctrl_calc_tu.exit:
+ call void @llvm.lifetime.end.p0(i64 8, ptr nonnull %rem.i11.i)
+ %conv.i = trunc i64 %res_abs.1.i24.i.7 to i32
+ store i32 %conv.i, ptr @_dp_ctrl_calc_tu_temp1_fp, align 4
+ %conv5.i = and i64 %res_abs.1.i24.i.7, 4294967295
+ store i64 %conv5.i, ptr @dp_panel_update_tu_timings___trans_tmp_5, align 8
+ store i64 %res_abs.1.i.i.7, ptr @_dp_ctrl_calc_tu___trans_tmp_8, align 8
+ ret void
+ }
+
+...
+---
+name: dp_ctrl_calc_tu_parameters
+tracksRegLiveness: true
+stack:
+ - { id: 0, name: rem.i11.i, type: default, offset: 0, size: 8, alignment: 8}
+body: |
+ bb.0:
+ successors: %bb.1(0x80000000)
+
+ %0:intregs = A2_tfrsi 0
+ %1:intregs = PS_fi %stack.0.rem.i11.i, 0
+ %2:intregs = A2_tfrsi 0
+ %3:doubleregs = A4_combineir 0, %2
+ %4:doubleregs = L2_loadrd_io %stack.0.rem.i11.i, 0
+ %5:doubleregs = A2_tfrpi 0
+ J2_loop0i %bb.1, 4, implicit-def $lc0, implicit-def $sa0, implicit-def $usr
+
+ bb.1 (machine-block-address-taken):
+ successors: %bb.2(0x04000000), %bb.1(0x7c000000)
+
+ %6:doubleregs = PHI %4, %bb.0, %7, %bb.1
+ %8:doubleregs = PHI %3, %bb.0, %9, %bb.1
+ %10:predregs = C2_cmpeqp %6, %5
+ %11:intregs = C2_muxii %10, 0, -1
+ %12:doubleregs = A2_addsp %11, %6
+ %13:doubleregs = S2_asl_i_p %8, 2
+ %14:intregs = S2_setbit_i %13.isub_lo, 1
+ %15:intregs = C2_mux %10, %13.isub_lo, %14
+ %16:predregs = C2_cmpeqp %12, %5
+ %17:intregs = C2_muxii %16, 0, -1
+ %18:doubleregs = A2_addsp %17, %12
+ %19:intregs = S2_setbit_i %15, 0
+ %20:intregs = C2_mux %16, %15, %19
+ %21:predregs = C2_cmpeqp %18, %5
+ %22:intregs = C2_muxii %21, 0, -1
+ %23:doubleregs = A2_addsp %22, %18
+ %24:intregs = S2_asl_i_r %20, 2
+ %25:intregs = S2_extractu %8.isub_lo, 2, 28
+ %26:intregs = S2_asl_i_r_or %25, %13.isub_hi, 2
+ %27:intregs = S2_setbit_i %24, 1
+ %28:intregs = C2_mux %21, %24, %27
+ %29:predregs = C2_cmpeqp %23, %5
+ %30:intregs = C2_muxii %29, 0, -1
+ %7:doubleregs = A2_addsp %30, %23
+ %31:intregs = S2_setbit_i %28, 0
+ %32:intregs = C2_mux %29, %28, %31
+ %9:doubleregs = REG_SEQUENCE %26, %subreg.isub_hi, %32, %subreg.isub_lo
+ ENDLOOP0 %bb.1, implicit-def $pc, implicit-def $lc0, implicit $sa0, implicit $lc0
+ J2_jump %bb.2, implicit-def dead $pc
+
+ bb.2:
+ successors: %bb.3(0x80000000)
+
+ S2_storerdgp @_dp_ctrl_calc_tu_temp2_fp, %9, implicit $gp
+ %33:intregs = A2_tfrsi 0
+ %34:intregs = PS_fi %stack.0.rem.i11.i, 0
+ %35:intregs = A2_tfrsi 0
+ %36:doubleregs = L2_loadrd_io %stack.0.rem.i11.i, 0
+ %37:doubleregs = A2_tfrpi 124
+ %38:intregs = A2_tfrsi -1000
+ %39:intregs = A2_tfrsi -1
+ J2_loop0i %bb.3, 4, implicit-def $lc0, implicit-def $sa0, implicit-def $usr
+
+ bb.3 (machine-block-address-taken):
+ successors: %bb.4(0x04000000), %bb.3(0x7c000000)
+
+ %40:doubleregs = PHI %36, %bb.2, %41, %bb.3
+ %42:intregs = PHI %35, %bb.2, %43, %bb.3
+ %44:intregs = PHI %33, %bb.2, %45, %bb.3
+ %46:doubleregs = S2_lsr_i_p %40, 3
+ %47:predregs = C2_cmpgtup %46, %37
+ %48:intregs = C2_mux %47, %38, %33
+ %49:intregs = C2_mux %47, %39, %33
+ %50:doubleregs = REG_SEQUENCE %49, %subreg.isub_hi, %48, %subreg.isub_lo
+ %51:doubleregs = A2_addp %50, %40
+ %52:intregs = S2_asl_i_r %42, 2
+ %53:intregs = S2_extractu %42, 2, 30
+ %45:intregs = S2_asl_i_r_or %53, %44, 2
+ %54:intregs = S2_setbit_i %52, 1
+ %55:intregs = C2_mux %47, %54, %52
+ %56:doubleregs = S2_lsr_i_p %51, 3
+ %57:predregs = C2_cmpgtup %56, %37
+ %58:intregs = C2_mux %57, %38, %33
+ %59:intregs = C2_mux %57, %39, %33
+ %60:doubleregs = REG_SEQUENCE %59, %subreg.isub_hi, %58, %subreg.isub_lo
+ %41:doubleregs = A2_addp %60, %51
+ %61:intregs = S2_setbit_i %55, 0
+ %43:intregs = C2_mux %57, %61, %55
+ ENDLOOP0 %bb.3, implicit-def $pc, implicit-def $lc0, implicit $sa0, implicit $lc0
+ J2_jump %bb.4, implicit-def dead $pc
+
+ bb.4:
+ S2_storerigp @_dp_ctrl_calc_tu_temp1_fp, %43, implicit $gp
+ %62:intregs = A2_tfrsi 0
+ %63:doubleregs = REG_SEQUENCE %43, %subreg.isub_lo, %62, %subreg.isub_hi
+ S2_storerdgp @dp_panel_update_tu_timings___trans_tmp_5, %63, implicit $gp
+ S2_storerdgp @_dp_ctrl_calc_tu___trans_tmp_8, %9, implicit $gp
+ PS_jmpret $r31, implicit-def dead $pc
+
+...
|
|
Hi, since we are wrapping up LLVM 19.1.0 we are very strict with the fixes we pick at this point. Can you please respond to the following questions to help me understand if this has to be included in the final release or not. Is this PR a fix for a regression or a critical issue? What is the risk of accepting this into the release branch? What is the risk of NOT accepting this into the release branch? |
Okay~
1.This PR fixes an assertion failure issue introduced by our recently proposed window scheduling algorithm. This issue was reported by nathanchance when building the Linux kernel for ARCH=hexagon (reported in #99454).
2.This PR inserts three newly generated MBBs into LiveIntervals. I think the risk of this modification is low.
3.Without this PR, there is a low probability of an assertion failure occurring in this scenario: In a function containing more than two loops, a value generated by a previous loop is used by the subsequent loop in the epilogue, and the subsequent loop successfully performs software pipelining. The issue may occur depending on the actual live interval of the value. |
|
this was merged - not sure why github shows it as closed instead. |
Okay, I see that this PR has already been merged. Thank you！ |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Backport 43ba109
Requested by: @huaatian