[llvm] r195265 - MachineBlockPlacement: Strengthen the source order bias when picking an exit block.

Benjamin Kramer benny.kra at googlemail.com
Wed Nov 20 11:08:44 PST 2013


Author: d0k
Date: Wed Nov 20 13:08:44 2013
New Revision: 195265

URL: http://llvm.org/viewvc/llvm-project?rev=195265&view=rev
Log:
MachineBlockPlacement: Strengthen the source order bias when picking an exit block.

We now only allow breaking source order if the exit block frequency is
significantly higher than the other exit block. The actual bias is
currently under a flag so the best cut-off can be found; the flag
defaults to the old behavior. The idea is to get some benchmark coverage
over different values for the flag and pick the best one.

When we require the new frequency to be at least 20% higher than the old
frequency I see a 5% speedup on zlib's deflate when compressing a random
file on x86_64/westmere. Hal reported a small speedup on Fhourstones on
a BG/Q and no regressions in the test suite.

The test case is the full long_match function from zlib's deflate. I was
reluctant to add it for previous tweaks to branch probabilities because
it's large and potentially fragile, but changed my mind since it's an
important use case and more likely to break with all the current work
going into the PGO infrastructure.

Differential Revision: http://llvm-reviews.chandlerc.com/D2202

Added:
    llvm/trunk/test/CodeGen/X86/zlib-longest-match.ll
Modified:
    llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp

Modified: llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp?rev=195265&r1=195264&r2=195265&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp (original)
+++ llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp Wed Nov 20 13:08:44 2013
@@ -58,6 +58,13 @@ static cl::opt<unsigned> AlignAllBlock("
                                                 "blocks in the function."),
                                        cl::init(0), cl::Hidden);
 
+// FIXME: Find a good default for this flag and remove the flag.
+static cl::opt<unsigned>
+ExitBlockBias("block-placement-exit-block-bias",
+              cl::desc("Block frequency percentage a loop exit block needs "
+                       "over the original exit to be considered the new exit."),
+              cl::init(0), cl::Hidden);
+
 namespace {
 class BlockChain;
 /// \brief Type for our function-wide basic block -> block chain mapping.
@@ -691,13 +698,15 @@ MachineBlockPlacement::findBestLoopExit(
       DEBUG(dbgs() << "    exiting: " << getBlockName(*I) << " -> "
                    << getBlockName(*SI) << " [L:" << SuccLoopDepth
                    << "] (" << ExitEdgeFreq << ")\n");
-      // Note that we slightly bias this toward an existing layout successor to
-      // retain incoming order in the absence of better information.
-      // FIXME: Should we bias this more strongly? It's pretty weak.
+      // Note that we bias this toward an existing layout successor to retain
+      // incoming order in the absence of better information. The exit must have
+      // a frequency higher than the current exit before we consider breaking
+      // the layout.
+      BranchProbability Bias(100 - ExitBlockBias, 100);
       if (!ExitingBB || BestExitLoopDepth < SuccLoopDepth ||
           ExitEdgeFreq > BestExitEdgeFreq ||
           ((*I)->isLayoutSuccessor(*SI) &&
-           !(ExitEdgeFreq < BestExitEdgeFreq))) {
+           !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) {
         BestExitEdgeFreq = ExitEdgeFreq;
         ExitingBB = *I;
       }

Added: llvm/trunk/test/CodeGen/X86/zlib-longest-match.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/zlib-longest-match.ll?rev=195265&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/zlib-longest-match.ll (added)
+++ llvm/trunk/test/CodeGen/X86/zlib-longest-match.ll Wed Nov 20 13:08:44 2013
@@ -0,0 +1,240 @@
+; RUN: llc -march=x86-64 < %s -block-placement-exit-block-bias=20 | FileCheck %s
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.9.0"
+
+; This is longest_match, the hot function from zlib's deflate implementation.
+
+%struct.internal_state = type { %struct.z_stream_s*, i32, i8*, i64, i8*, i32, i32, %struct.gz_header_s*, i32, i8, i32, i32, i32, i32, i8*, i64, i16*, i16*, i32, i32, i32, i32, i32, i64, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, [573 x %struct.ct_data_s], [61 x %struct.ct_data_s], [39 x %struct.ct_data_s], %struct.tree_desc_s, %struct.tree_desc_s, %struct.tree_desc_s, [16 x i16], [573 x i32], i32, i32, [573 x i8], i8*, i32, i32, i16*, i64, i64, i32, i32, i16, i32, i64 }
+%struct.z_stream_s = type { i8*, i32, i64, i8*, i32, i64, i8*, %struct.internal_state*, i8* (i8*, i32, i32)*, void (i8*, i8*)*, i8*, i32, i64, i64 }
+%struct.gz_header_s = type { i32, i64, i32, i32, i8*, i32, i32, i8*, i32, i8*, i32, i32, i32 }
+%struct.ct_data_s = type { %union.anon, %union.anon.0 }
+%union.anon = type { i16 }
+%union.anon.0 = type { i16 }
+%struct.tree_desc_s = type { %struct.ct_data_s*, i32, %struct.static_tree_desc_s* }
+%struct.static_tree_desc_s = type { i32 }
+
+; CHECK-LABEL: longest_match:
+
+; Verify that there are no spills or reloads in the loop exit block. This loop
+; is mostly cold, only %do.cond125 and %land.rhs131 are hot.
+; CHECK: %do.cond125
+; CHECK-NOT: {{Spill|Reload}}
+; CHECK: jbe
+
+; Verify that block placement doesn't destroy source order. It's important that
+; the two hot blocks are laid out close to each other.
+; CHECK-NEXT: %land.rhs131
+; CHECK: jne
+; CHECK: jmp
+define i32 @longest_match(%struct.internal_state* nocapture %s, i32 %cur_match) nounwind {
+entry:
+  %max_chain_length = getelementptr inbounds %struct.internal_state* %s, i64 0, i32 31
+  %0 = load i32* %max_chain_length, align 4
+  %window = getelementptr inbounds %struct.internal_state* %s, i64 0, i32 14
+  %1 = load i8** %window, align 8
+  %strstart = getelementptr inbounds %struct.internal_state* %s, i64 0, i32 27
+  %2 = load i32* %strstart, align 4
+  %idx.ext = zext i32 %2 to i64
+  %add.ptr = getelementptr inbounds i8* %1, i64 %idx.ext
+  %prev_length = getelementptr inbounds %struct.internal_state* %s, i64 0, i32 30
+  %3 = load i32* %prev_length, align 4
+  %nice_match1 = getelementptr inbounds %struct.internal_state* %s, i64 0, i32 36
+  %4 = load i32* %nice_match1, align 4
+  %w_size = getelementptr inbounds %struct.internal_state* %s, i64 0, i32 11
+  %5 = load i32* %w_size, align 4
+  %sub = add i32 %5, -262
+  %cmp = icmp ugt i32 %2, %sub
+  %sub6 = sub i32 %2, %sub
+  %sub6. = select i1 %cmp, i32 %sub6, i32 0
+  %prev7 = getelementptr inbounds %struct.internal_state* %s, i64 0, i32 16
+  %6 = load i16** %prev7, align 8
+  %w_mask = getelementptr inbounds %struct.internal_state* %s, i64 0, i32 13
+  %7 = load i32* %w_mask, align 4
+  %add.ptr11.sum = add i64 %idx.ext, 258
+  %add.ptr12 = getelementptr inbounds i8* %1, i64 %add.ptr11.sum
+  %sub13 = add nsw i32 %3, -1
+  %idxprom = sext i32 %sub13 to i64
+  %add.ptr.sum = add i64 %idxprom, %idx.ext
+  %arrayidx = getelementptr inbounds i8* %1, i64 %add.ptr.sum
+  %8 = load i8* %arrayidx, align 1
+  %idxprom14 = sext i32 %3 to i64
+  %add.ptr.sum213 = add i64 %idxprom14, %idx.ext
+  %arrayidx15 = getelementptr inbounds i8* %1, i64 %add.ptr.sum213
+  %9 = load i8* %arrayidx15, align 1
+  %good_match = getelementptr inbounds %struct.internal_state* %s, i64 0, i32 35
+  %10 = load i32* %good_match, align 4
+  %cmp17 = icmp ult i32 %3, %10
+  %shr = lshr i32 %0, 2
+  %chain_length.0 = select i1 %cmp17, i32 %0, i32 %shr
+  %lookahead = getelementptr inbounds %struct.internal_state* %s, i64 0, i32 29
+  %11 = load i32* %lookahead, align 4
+  %cmp18 = icmp ugt i32 %4, %11
+  %. = select i1 %cmp18, i32 %11, i32 %4
+  %match_start = getelementptr inbounds %struct.internal_state* %s, i64 0, i32 28
+  %add.ptr.sum217 = add i64 %idx.ext, 1
+  %arrayidx44 = getelementptr inbounds i8* %1, i64 %add.ptr.sum217
+  %add.ptr.sum218 = add i64 %idx.ext, 2
+  %add.ptr50 = getelementptr inbounds i8* %1, i64 %add.ptr.sum218
+  %sub.ptr.lhs.cast = ptrtoint i8* %add.ptr12 to i64
+  br label %do.body
+
+do.body:                                          ; preds = %land.rhs131, %entry
+  %best_len.0 = phi i32 [ %best_len.1, %land.rhs131 ], [ %3, %entry ]
+  %chain_length.1 = phi i32 [ %dec, %land.rhs131 ], [ %chain_length.0, %entry ]
+  %cur_match.addr.0 = phi i32 [ %conv128, %land.rhs131 ], [ %cur_match, %entry ]
+  %scan_end1.0 = phi i8 [ %scan_end1.1, %land.rhs131 ], [ %8, %entry ]
+  %scan_end.0 = phi i8 [ %scan_end.1, %land.rhs131 ], [ %9, %entry ]
+  %idx.ext23 = zext i32 %cur_match.addr.0 to i64
+  %add.ptr24 = getelementptr inbounds i8* %1, i64 %idx.ext23
+  %idxprom25 = sext i32 %best_len.0 to i64
+  %add.ptr24.sum = add i64 %idx.ext23, %idxprom25
+  %arrayidx26 = getelementptr inbounds i8* %1, i64 %add.ptr24.sum
+  %12 = load i8* %arrayidx26, align 1
+  %cmp28 = icmp eq i8 %12, %scan_end.0
+  br i1 %cmp28, label %lor.lhs.false, label %do.cond125
+
+lor.lhs.false:                                    ; preds = %do.body
+  %sub30 = add nsw i32 %best_len.0, -1
+  %idxprom31 = sext i32 %sub30 to i64
+  %add.ptr24.sum214 = add i64 %idx.ext23, %idxprom31
+  %arrayidx32 = getelementptr inbounds i8* %1, i64 %add.ptr24.sum214
+  %13 = load i8* %arrayidx32, align 1
+  %cmp35 = icmp eq i8 %13, %scan_end1.0
+  br i1 %cmp35, label %lor.lhs.false37, label %do.cond125
+
+lor.lhs.false37:                                  ; preds = %lor.lhs.false
+  %14 = load i8* %add.ptr24, align 1
+  %15 = load i8* %add.ptr, align 1
+  %cmp40 = icmp eq i8 %14, %15
+  br i1 %cmp40, label %lor.lhs.false42, label %do.cond125
+
+lor.lhs.false42:                                  ; preds = %lor.lhs.false37
+  %add.ptr24.sum215 = add i64 %idx.ext23, 1
+  %incdec.ptr = getelementptr inbounds i8* %1, i64 %add.ptr24.sum215
+  %16 = load i8* %incdec.ptr, align 1
+  %17 = load i8* %arrayidx44, align 1
+  %cmp46 = icmp eq i8 %16, %17
+  br i1 %cmp46, label %if.end49, label %do.cond125
+
+if.end49:                                         ; preds = %lor.lhs.false42
+  %incdec.ptr.sum = add i64 %idx.ext23, 2
+  %incdec.ptr51 = getelementptr inbounds i8* %1, i64 %incdec.ptr.sum
+  br label %do.cond
+
+do.cond:                                          ; preds = %land.lhs.true100, %if.end49
+  %match.0 = phi i8* [ %incdec.ptr51, %if.end49 ], [ %incdec.ptr103, %land.lhs.true100 ]
+  %scan.1 = phi i8* [ %add.ptr50, %if.end49 ], [ %incdec.ptr101, %land.lhs.true100 ]
+  %incdec.ptr53 = getelementptr inbounds i8* %scan.1, i64 1
+  %18 = load i8* %incdec.ptr53, align 1
+  %incdec.ptr55 = getelementptr inbounds i8* %match.0, i64 1
+  %19 = load i8* %incdec.ptr55, align 1
+  %cmp57 = icmp eq i8 %18, %19
+  br i1 %cmp57, label %land.lhs.true, label %do.end
+
+land.lhs.true:                                    ; preds = %do.cond
+  %incdec.ptr59 = getelementptr inbounds i8* %scan.1, i64 2
+  %20 = load i8* %incdec.ptr59, align 1
+  %incdec.ptr61 = getelementptr inbounds i8* %match.0, i64 2
+  %21 = load i8* %incdec.ptr61, align 1
+  %cmp63 = icmp eq i8 %20, %21
+  br i1 %cmp63, label %land.lhs.true65, label %do.end
+
+land.lhs.true65:                                  ; preds = %land.lhs.true
+  %incdec.ptr66 = getelementptr inbounds i8* %scan.1, i64 3
+  %22 = load i8* %incdec.ptr66, align 1
+  %incdec.ptr68 = getelementptr inbounds i8* %match.0, i64 3
+  %23 = load i8* %incdec.ptr68, align 1
+  %cmp70 = icmp eq i8 %22, %23
+  br i1 %cmp70, label %land.lhs.true72, label %do.end
+
+land.lhs.true72:                                  ; preds = %land.lhs.true65
+  %incdec.ptr73 = getelementptr inbounds i8* %scan.1, i64 4
+  %24 = load i8* %incdec.ptr73, align 1
+  %incdec.ptr75 = getelementptr inbounds i8* %match.0, i64 4
+  %25 = load i8* %incdec.ptr75, align 1
+  %cmp77 = icmp eq i8 %24, %25
+  br i1 %cmp77, label %land.lhs.true79, label %do.end
+
+land.lhs.true79:                                  ; preds = %land.lhs.true72
+  %incdec.ptr80 = getelementptr inbounds i8* %scan.1, i64 5
+  %26 = load i8* %incdec.ptr80, align 1
+  %incdec.ptr82 = getelementptr inbounds i8* %match.0, i64 5
+  %27 = load i8* %incdec.ptr82, align 1
+  %cmp84 = icmp eq i8 %26, %27
+  br i1 %cmp84, label %land.lhs.true86, label %do.end
+
+land.lhs.true86:                                  ; preds = %land.lhs.true79
+  %incdec.ptr87 = getelementptr inbounds i8* %scan.1, i64 6
+  %28 = load i8* %incdec.ptr87, align 1
+  %incdec.ptr89 = getelementptr inbounds i8* %match.0, i64 6
+  %29 = load i8* %incdec.ptr89, align 1
+  %cmp91 = icmp eq i8 %28, %29
+  br i1 %cmp91, label %land.lhs.true93, label %do.end
+
+land.lhs.true93:                                  ; preds = %land.lhs.true86
+  %incdec.ptr94 = getelementptr inbounds i8* %scan.1, i64 7
+  %30 = load i8* %incdec.ptr94, align 1
+  %incdec.ptr96 = getelementptr inbounds i8* %match.0, i64 7
+  %31 = load i8* %incdec.ptr96, align 1
+  %cmp98 = icmp eq i8 %30, %31
+  br i1 %cmp98, label %land.lhs.true100, label %do.end
+
+land.lhs.true100:                                 ; preds = %land.lhs.true93
+  %incdec.ptr101 = getelementptr inbounds i8* %scan.1, i64 8
+  %32 = load i8* %incdec.ptr101, align 1
+  %incdec.ptr103 = getelementptr inbounds i8* %match.0, i64 8
+  %33 = load i8* %incdec.ptr103, align 1
+  %cmp105 = icmp eq i8 %32, %33
+  %cmp107 = icmp ult i8* %incdec.ptr101, %add.ptr12
+  %or.cond = and i1 %cmp105, %cmp107
+  br i1 %or.cond, label %do.cond, label %do.end
+
+do.end:                                           ; preds = %land.lhs.true100, %land.lhs.true93, %land.lhs.true86, %land.lhs.true79, %land.lhs.true72, %land.lhs.true65, %land.lhs.true, %do.cond
+  %scan.2 = phi i8* [ %incdec.ptr101, %land.lhs.true100 ], [ %incdec.ptr94, %land.lhs.true93 ], [ %incdec.ptr87, %land.lhs.true86 ], [ %incdec.ptr80, %land.lhs.true79 ], [ %incdec.ptr73, %land.lhs.true72 ], [ %incdec.ptr66, %land.lhs.true65 ], [ %incdec.ptr59, %land.lhs.true ], [ %incdec.ptr53, %do.cond ]
+  %sub.ptr.rhs.cast = ptrtoint i8* %scan.2 to i64
+  %sub.ptr.sub = sub i64 %sub.ptr.lhs.cast, %sub.ptr.rhs.cast
+  %conv109 = trunc i64 %sub.ptr.sub to i32
+  %sub110 = sub nsw i32 258, %conv109
+  %cmp112 = icmp sgt i32 %sub110, %best_len.0
+  br i1 %cmp112, label %if.then114, label %do.cond125
+
+if.then114:                                       ; preds = %do.end
+  store i32 %cur_match.addr.0, i32* %match_start, align 4
+  %cmp115 = icmp slt i32 %sub110, %.
+  br i1 %cmp115, label %if.end118, label %do.end135
+
+if.end118:                                        ; preds = %if.then114
+  %sub119 = add nsw i32 %sub110, -1
+  %idxprom120 = sext i32 %sub119 to i64
+  %add.ptr111.sum = add i64 %idxprom120, %idx.ext
+  %arrayidx121 = getelementptr inbounds i8* %1, i64 %add.ptr111.sum
+  %34 = load i8* %arrayidx121, align 1
+  %idxprom122 = sext i32 %sub110 to i64
+  %add.ptr111.sum216 = add i64 %idxprom122, %idx.ext
+  %arrayidx123 = getelementptr inbounds i8* %1, i64 %add.ptr111.sum216
+  %35 = load i8* %arrayidx123, align 1
+  br label %do.cond125
+
+do.cond125:                                       ; preds = %if.end118, %do.end, %lor.lhs.false42, %lor.lhs.false37, %lor.lhs.false, %do.body
+  %best_len.1 = phi i32 [ %best_len.0, %do.body ], [ %best_len.0, %lor.lhs.false ], [ %best_len.0, %lor.lhs.false37 ], [ %best_len.0, %lor.lhs.false42 ], [ %sub110, %if.end118 ], [ %best_len.0, %do.end ]
+  %scan_end1.1 = phi i8 [ %scan_end1.0, %do.body ], [ %scan_end1.0, %lor.lhs.false ], [ %scan_end1.0, %lor.lhs.false37 ], [ %scan_end1.0, %lor.lhs.false42 ], [ %34, %if.end118 ], [ %scan_end1.0, %do.end ]
+  %scan_end.1 = phi i8 [ %scan_end.0, %do.body ], [ %scan_end.0, %lor.lhs.false ], [ %scan_end.0, %lor.lhs.false37 ], [ %scan_end.0, %lor.lhs.false42 ], [ %35, %if.end118 ], [ %scan_end.0, %do.end ]
+  %and = and i32 %cur_match.addr.0, %7
+  %idxprom126 = zext i32 %and to i64
+  %arrayidx127 = getelementptr inbounds i16* %6, i64 %idxprom126
+  %36 = load i16* %arrayidx127, align 2
+  %conv128 = zext i16 %36 to i32
+  %cmp129 = icmp ugt i32 %conv128, %sub6.
+  br i1 %cmp129, label %land.rhs131, label %do.end135
+
+land.rhs131:                                      ; preds = %do.cond125
+  %dec = add i32 %chain_length.1, -1
+  %cmp132 = icmp eq i32 %dec, 0
+  br i1 %cmp132, label %do.end135, label %do.body
+
+do.end135:                                        ; preds = %land.rhs131, %do.cond125, %if.then114
+  %best_len.2 = phi i32 [ %best_len.1, %land.rhs131 ], [ %best_len.1, %do.cond125 ], [ %sub110, %if.then114 ]
+  %cmp137 = icmp ugt i32 %best_len.2, %11
+  %.best_len.2 = select i1 %cmp137, i32 %11, i32 %best_len.2
+  ret i32 %.best_len.2
+}





More information about the llvm-commits mailing list