[llvm] 6b856ab - [PseudoProbe] Use probe id as the base dwarf discriminator for callsites (#65685)

[via llvm-commits]

Author: Hongtao Yu
Date: 2023-09-08T09:49:54-07:00
New Revision: 6b856abc6fec04b523c17aab96f9877f40e7b7ab

URL: https://github.com/llvm/llvm-project/commit/6b856abc6fec04b523c17aab96f9877f40e7b7ab
DIFF: https://github.com/llvm/llvm-project/commit/6b856abc6fec04b523c17aab96f9877f40e7b7ab.diff

LOG: [PseudoProbe] Use probe id as the base dwarf discriminator for callsites (#65685)

With `-fpseudo-probe-for-profiling`, the dwarf discriminator for a
callsite will be overwritten to pseudo probe related information for
that callsite. The probe information is encoded in a special format
(i.e., with all lowest three digits be one) in order to be distinguished
from regular dwarf discriminator. The special encoding format will be
decoded to zero by the regular discriminator logic. This means all
callsites would have a zero discriminator in both the sample profile and
the compiler, for classic AutoFDO. This is inconvenient in that no
decent classic AutoFDO can be generated from a pseudo probe build. I'm
mitigating the issue by allowing callsite probe id to be used as the
base dwarf discriminator for classic AutoFDO, since probe id is also
unique and can be used to differentiate callsites on the same source
line.

Added: 
    llvm/test/tools/llvm-profgen/inline-probe-afdo.test

Modified: 
    llvm/include/llvm/IR/DebugInfoMetadata.h
    llvm/test/Transforms/SampleProfile/pseudo-probe-inline.ll

Removed: 
    


################################################################################
diff  --git a/llvm/include/llvm/IR/DebugInfoMetadata.h b/llvm/include/llvm/IR/DebugInfoMetadata.h
index b9f6d39a7491566..9beb514b87125af 100644
--- a/llvm/include/llvm/IR/DebugInfoMetadata.h
+++ b/llvm/include/llvm/IR/DebugInfoMetadata.h
@@ -22,6 +22,7 @@
 #include "llvm/ADT/iterator_range.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/Metadata.h"
+#include "llvm/IR/PseudoProbe.h"
 #include "llvm/Support/Casting.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Discriminator.h"
@@ -2075,6 +2076,14 @@ class DILocation : public MDNode {
   static unsigned
   getBaseDiscriminatorFromDiscriminator(unsigned D,
                                         bool IsFSDiscriminator = false) {
+    // Return the probe id instead of zero for a pseudo probe discriminator.
+    // This should help 
diff erenciate callsites with same line numbers to
+    // achieve a decent AutoFDO profile under -fpseudo-probe-for-profiling,
+    // where the original callsite dwarf discriminator is overwritten by
+    // callsite probe information.
+    if (isPseudoProbeDiscriminator(D))
+      return PseudoProbeDwarfDiscriminator::extractProbeIndex(D);
+
     if (IsFSDiscriminator)
       return getMaskedDiscriminator(D, getBaseDiscriminatorBits());
     return getUnsignedFromPrefixEncoding(D);

diff  --git a/llvm/test/Transforms/SampleProfile/pseudo-probe-inline.ll b/llvm/test/Transforms/SampleProfile/pseudo-probe-inline.ll
index 8157bebdc3777be..18cbd857d97bb2c 100644
--- a/llvm/test/Transforms/SampleProfile/pseudo-probe-inline.ll
+++ b/llvm/test/Transforms/SampleProfile/pseudo-probe-inline.ll
@@ -106,6 +106,8 @@ if.end:
 ;YAML-NEXT:    - Line:            '1'
 ;YAML-NEXT:    - String:          ':'
 ;YAML-NEXT:    - Column:          '11'
+;YAML-NEXT:    - String:          .
+;YAML-NEXT:    - Disc:            '2'
 ;YAML-NEXT:    - String:          ';'
 ;YAML-NEXT:  ...
 ;YAML:  --- !Analysis

diff  --git a/llvm/test/tools/llvm-profgen/inline-probe-afdo.test b/llvm/test/tools/llvm-profgen/inline-probe-afdo.test
new file mode 100644
index 000000000000000..763bf3b0f8c394a
--- /dev/null
+++ b/llvm/test/tools/llvm-profgen/inline-probe-afdo.test
@@ -0,0 +1,82 @@
+; RUN: llvm-profgen --format=text --use-dwarf-correlation --ignore-stack-samples --perfscript=%S/Inputs/cs-preinline-cost.perfscript --binary=%S/Inputs/cs-preinline-cost.perfbin --output %t
+; RUN: FileCheck %s --input-file %t
+
+; CHECK:     main:947937:0
+; CHECK-NEXT:  2: 545
+; CHECK-NEXT:  3: 545
+; CHECK-NEXT:  5: 545
+; CHECK-NEXT:  7: 0
+; CHECK-NEXT:  65496: 545
+; CHECK-NEXT:  3.7: _Z3fooi:915794
+; CHECK-NEXT:   1: 545
+; CHECK-NEXT:   5: 545
+; CHECK-NEXT:   6: 272
+; CHECK-NEXT:   10: 273
+; CHECK-NEXT:   11: 180
+; CHECK-NEXT:   12: 6965
+; CHECK-NEXT:   13: 6965
+; CHECK-NEXT:   14: 6965
+; CHECK-NEXT:   15: 6965
+; CHECK-NEXT:   20: 182
+; CHECK-NEXT:   21: 6958
+; CHECK-NEXT:   22: 6958
+; CHECK-NEXT:   23: 6958
+; CHECK-NEXT:   24: 6958
+; CHECK-NEXT:   29: 272
+; CHECK-NEXT:   65529: 182
+; CHECK-NEXT:  4.8: _Z3fooi:16338
+; CHECK-NEXT:   1: 272
+; CHECK-NEXT:   6: 545
+
+
+
+
+; binary is built with the source below using the following command line:
+;   clang -O3 -g -fpseudo-probe-for-profiling test.cpp
+;
+;#include <stdio.h>
+;
+;volatile int state = 9000;
+;
+;int foo(int x) {
+;    if (x == 0) {
+;        return 7;
+;    }
+;
+;    if ((x & 1) == 0) {
+;        state--;
+;        return 9;
+;    }
+;
+;    if (state > 5000) {
+;        while (state > 5000) {
+;               for (int i = 50; i >= 0; i--) {
+;                state *= 6;
+;                state /= 7;
+;                state -= 1;
+;            }
+;        }
+;    }
+;    else {
+;        while (state < 5000) {
+;            for (int i = 50; i >= 0; i--) {
+;                state *= 6;
+;                state /= 5;
+;                state += 1;
+;            }
+;        }
+;    }
+;
+;    return state;
+;}
+;
+;volatile int cnt = 10000000;//10000000;
+;int main() {
+;    int r = 0;
+;    for (int i = 0; i < cnt; i++) {
+;      r += foo(i);
+;      r -= foo(i & (~1));
+;      r += foo(0);
+;    }
+;    return r;
+;}