[llvm-branch-commits] [llvm] release/19.x: [llvm][CodeGen] Resolve issues when updating live intervals in window scheduler (#101945) (PR #107338)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Wed Sep 4 17:59:29 PDT 2024
https://github.com/llvmbot created https://github.com/llvm/llvm-project/pull/107338
Backport 43ba1097ee747b4ec5e757762ed0c9df6255a292
Requested by: @huaatian
>From 5b44b9cfaa0d87901d029537532bcf8d21259117 Mon Sep 17 00:00:00 2001
From: Hua Tian <akiratian at tencent.com>
Date: Thu, 15 Aug 2024 19:03:27 +0800
Subject: [PATCH] [llvm][CodeGen] Resolve issues when updating live intervals
in window scheduler (#101945)
Corrupted live interval information can cause window scheduling to crash
in some cases. By adding the missing MBB's live interval information in the
ModuloScheduleExpander, the information can be correctly analyzed in
the window scheduler.
(cherry picked from commit 43ba1097ee747b4ec5e757762ed0c9df6255a292)
---
llvm/lib/CodeGen/ModuloSchedule.cpp | 3 +
.../CodeGen/Hexagon/swp-ws-live-intervals.mir | 217 ++++++++++++++++++
2 files changed, 220 insertions(+)
create mode 100644 llvm/test/CodeGen/Hexagon/swp-ws-live-intervals.mir
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
+
+...
More information about the llvm-branch-commits
mailing list