packages/gcc
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
gcc/0164-LLC-Allocation-Fix-some-bugs-and-remove-variable-pre.patch

925 lines
33 KiB
Diff
Raw Blame History

From 1722afc51311a6bb0b892df50602f660c706162f Mon Sep 17 00:00:00 2001
From: liuf9 <liufeiyang6@huawei.com>
Date: Fri, 15 Dec 2023 11:25:48 +0800
Subject: [PATCH 2/2] [LLC Allocation] Fix some bugs and remove variable
prefetch tool. After outer loop analysis, it is possible to get nested loops
for kernel candidates and this situation has conflicts with the early exiting
criterion for kernel filtering process and we restrict this criterion for
innermost loops only. We also fix some pass configuration bugs in common.opt
and params.opt. We remove variable prefetch tool due to the consideration of
unsafe inputs from users.
---
gcc/common.opt | 2 +-
gcc/params.opt | 24 +-
gcc/testsuite/gcc.dg/llc-allocate/llc-2.c | 2 +-
.../llc-allocate/llc-issue-builtin-prefetch.c | 48 ----
.../llc-allocate/llc-tool-insertion-1.c | 48 ----
.../llc-allocate/llc-tool-insertion-2.c | 48 ----
.../llc-allocate/llc-tool-insertion-3.c | 48 ----
.../llc-allocate/llc-tool-insertion-4.c | 47 ---
.../llc-allocate/llc-tool-insertion-5.c | 48 ----
.../llc-allocate/llc-tool-insertion-6.c | 47 ---
.../llc-tool-insertion-7-null-var-name.c | 52 ----
.../llc-tool-insertion-8-tmp-var-name.c | 54 ----
gcc/tree-ssa-llc-allocate.c | 267 +-----------------
13 files changed, 11 insertions(+), 724 deletions(-)
delete mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-issue-builtin-prefetch.c
delete mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-1.c
delete mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-2.c
delete mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-3.c
delete mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-4.c
delete mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-5.c
delete mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-6.c
delete mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-7-null-var-name.c
delete mode 100644 gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-8-tmp-var-name.c
diff --git a/gcc/common.opt b/gcc/common.opt
index 56ad9a378..a8a2264ee 100644
--- a/gcc/common.opt
+++ b/gcc/common.opt
@@ -2255,7 +2255,7 @@ Generate prefetch instructions, if available, for arrays in loops. The prefetch
level can control the optimize level to array prefetch.
fllc-allocate
-Common Report Var(flag_llc_allocate) Init(-1) Optimization
+Common Report Var(flag_llc_allocate) Optimization
Generate LLC hint instructions.
fprofile
diff --git a/gcc/params.opt b/gcc/params.opt
index 792ca5c35..ef7bea311 100644
--- a/gcc/params.opt
+++ b/gcc/params.opt
@@ -1030,13 +1030,11 @@ Common Joined UInteger Var(param_mem_access_num) Init(3) Param Optimization
Memory access num.
-param=prefetch-offset=
-Common Joined UInteger Var(param_prefetch_offset) Init(1024)
-IntegerRange(1, 999999) Param Optimization
+Common Joined UInteger Var(param_prefetch_offset) Init(1024) IntegerRange(1, 999999) Param Optimization
Prefetch Offset, which is usually a power of two due to cache line size.
-param=branch-prob-threshold=
-Common Joined UInteger Var(param_branch_prob_threshold) Init(80) IntegerRange(50, 100)
-Param Optimization
+Common Joined UInteger Var(param_branch_prob_threshold) Init(80) IntegerRange(50, 100) Param Optimization
High Execution Rate Branch Threshold.
-param=issue-topn=
@@ -1051,24 +1049,6 @@ Force issue the topn LLC mem_ref hint, without generating dynamic multi-branches
Common Joined UInteger Var(param_llc_capacity_per_core) Init(114) IntegerRange(0, 999999) Param
LLC capacity per core.
--param=target-variables=
-Common Joined Var(param_target_variables) Init("") Param Optimization
---param=target-variables=<var>[,<var>,...] Target variables for prefetching, separated by comma,
-without space. The representation of a variable can be complex and containing space, please surround
-it by quotation marks and escape special characters in Linux. The input length should be no more
-than 512 characters.
-
--param=use-ref-group-index=
-Common Joined UInteger Var(param_use_ref_group_index) Init(0) IntegerRange(0, 1) Param Optimization
-Prefetch the target variables by their indices in sorted ref_groups, use together with parameter
-target-variables.
-
--param=mem-ref-index=
-Common Joined Var(param_mem_ref_index) Init("") Param Optimization
---param=mem-ref-index=<idx>[,<idx>,...] Prefetch the target variable at the memory reference
-location with the index of customized order, separated by comma, without space. The input length
-should be no more than 512 characters.
-
-param=filter-kernels=
Common Joined UInteger Var(param_filter_kernels) Init(1) IntegerRange(0, 1) Param
Allow LLC allocate pass to greedily filter kernels by traversing the corresponding basic blocks
diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-2.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-2.c
index 9f8a5c307..f8b1cc5c1 100644
--- a/gcc/testsuite/gcc.dg/llc-allocate/llc-2.c
+++ b/gcc/testsuite/gcc.dg/llc-allocate/llc-2.c
@@ -45,7 +45,7 @@ main (int argc, char *argv[])
/* { dg-final { scan-tree-dump-not "static_data_size:" "llc_allocate" } } */
/* { dg-final { scan-tree-dump-times "\{ (?:\\d+\\(\\d+\\) ){1}\}" 2 "llc_allocate" } } */
/* { dg-final { scan-tree-dump-not ", size: (?!(0\.000000))" "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-times ", size: 0\.000000" 8 "llc_allocate" } } */
+/* { dg-final { scan-tree-dump-times ", size: 0\.000000" 6 "llc_allocate" } } */
/* { dg-final { scan-tree-dump-times "\\d x_data \\(0.000000, 1, 0\\) : 3" 2 "llc_allocate" } } */
/* { dg-final { scan-tree-dump-times "\\d A_j \\(0.000000, 1, 0\\) : 2" 2 "llc_allocate" } } */
/* { dg-final { scan-tree-dump-times "\\d A_data \\(0.000000, 1, 0\\) : 2" 2 "llc_allocate" } } */
diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-issue-builtin-prefetch.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-issue-builtin-prefetch.c
deleted file mode 100644
index 2a58c501f..000000000
--- a/gcc/testsuite/gcc.dg/llc-allocate/llc-issue-builtin-prefetch.c
+++ /dev/null
@@ -1,48 +0,0 @@
-/* { dg-do compile { target { aarch64*-*-linux* } } } */
-/* { dg-options "-O3 -march=armv8.2-a+sve -funroll-loops -ffast-math -static -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param target-variables=uPtr" } */
-
-#include <stdio.h>
-
-#define N 131590
-#define F 384477
-
-double diagPtr[N];
-double psiPtr[N];
-double ApsiPtr[N];
-int lPtr[F];
-int uPtr[F];
-double lowerPtr[F];
-double upperPtr[F];
-
-void
-AMUL (double *diagPtr, double *psiPtr, double *ApsiPtr, int *lPtr,
- int *uPtr, double *lowerPtr, double *upperPtr, int nCells, int nFaces)
-{
- for (int cell=0; cell<nCells; cell++)
- ApsiPtr[cell] = diagPtr[cell]*psiPtr[cell];
-
- for (int face=0; face<nFaces; face++)
- {
- ApsiPtr[uPtr[face]] += lowerPtr[face]*psiPtr[lPtr[face]];
- ApsiPtr[lPtr[face]] += upperPtr[face]*psiPtr[uPtr[face]];
- }
-}
-
-int
-main (int argc, char *argv[])
-{
- int nCells = N;
- int nFaces = F;
- int testIter = 2;
-
- for (int i=0; i<testIter; i++)
- AMUL (diagPtr,psiPtr,ApsiPtr,lPtr,uPtr,lowerPtr,upperPtr,nCells,nFaces);
-
- return 0;
-}
-
-/* { dg-final { scan-tree-dump-not "runtime issue" "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-times "static issue" 2 "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-times "insert prfm" 2 "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-not "\[&\]?uPtr(?:_\\d+\\(D\\))? \\+ \\d{4};" "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-times "\[&\]?uPtr(?:_\\d+\\(D\\))? \\+ \[_\]\\d{1,4};" 2 "llc_allocate" } } */
diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-1.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-1.c
deleted file mode 100644
index 276781c4f..000000000
--- a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-1.c
+++ /dev/null
@@ -1,48 +0,0 @@
-/* { dg-do compile { target { aarch64*-*-linux* } } } */
-/* { dg-options "-O3 -march=armv8.2-a+sve -funroll-loops -ffast-math -static -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param target-variables=lPtr" } */
-
-#include <stdio.h>
-
-#define N 131590
-#define F 384477
-
-double diagPtr[N];
-double psiPtr[N];
-double ApsiPtr[N];
-int lPtr[F];
-int uPtr[F];
-double lowerPtr[F];
-double upperPtr[F];
-
-void
-AMUL (double *diagPtr, double *psiPtr, double *ApsiPtr, int *lPtr,
- int *uPtr, double *lowerPtr, double *upperPtr, int nCells, int nFaces)
-{
- for (int cell=0; cell<nCells; cell++)
- ApsiPtr[cell] = diagPtr[cell]*psiPtr[cell];
-
- for (int face=0; face<nFaces; face++)
- {
- ApsiPtr[uPtr[face]] += lowerPtr[face]*psiPtr[lPtr[face]];
- ApsiPtr[lPtr[face]] += upperPtr[face]*psiPtr[uPtr[face]];
- }
-}
-
-int
-main (int argc, char *argv[])
-{
- int nCells = N;
- int nFaces = F;
- int testIter = 2;
-
- for (int i=0; i<testIter; i++)
- AMUL (diagPtr,psiPtr,ApsiPtr,lPtr,uPtr,lowerPtr,upperPtr,nCells,nFaces);
-
- return 0;
-}
-
-/* { dg-final { scan-tree-dump-times "NOTICE: Prefetching target variable \""
- " lPtr \"" 2 "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-not "runtime issue" "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-times "static issue" 2 "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-times "insert prfm" 2 "llc_allocate" } } */
diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-2.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-2.c
deleted file mode 100644
index 57c76f4a6..000000000
--- a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-2.c
+++ /dev/null
@@ -1,48 +0,0 @@
-/* { dg-do compile { target { aarch64*-*-linux* } } } */
-/* { dg-options "-O3 -march=armv8.2-a+sve -funroll-loops -ffast-math -static -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param target-variables=,lPtr, --param mem-ref-index=5" } */
-
-#include <stdio.h>
-
-#define N 131590
-#define F 384477
-
-double diagPtr[N];
-double psiPtr[N];
-double ApsiPtr[N];
-int lPtr[F];
-int uPtr[F];
-double lowerPtr[F];
-double upperPtr[F];
-
-void
-AMUL (double *diagPtr, double *psiPtr, double *ApsiPtr, int *lPtr,
- int *uPtr, double *lowerPtr, double *upperPtr, int nCells, int nFaces)
-{
- for (int cell=0; cell<nCells; cell++)
- ApsiPtr[cell] = diagPtr[cell]*psiPtr[cell];
-
- for (int face=0; face<nFaces; face++)
- {
- ApsiPtr[uPtr[face]] += lowerPtr[face]*psiPtr[lPtr[face]];
- ApsiPtr[lPtr[face]] += upperPtr[face]*psiPtr[uPtr[face]];
- }
-}
-
-int
-main (int argc, char *argv[])
-{
- int nCells = N;
- int nFaces = F;
- int testIter = 2;
-
- for (int i=0; i<testIter; i++)
- AMUL (diagPtr,psiPtr,ApsiPtr,lPtr,uPtr,lowerPtr,upperPtr,nCells,nFaces);
-
- return 0;
-}
-
-/* { dg-final { scan-tree-dump-times "WARNING: The target data_ref index is "
- "out of range." 2 "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-not "runtime issue" "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-times "static issue" 2 "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-times "insert prfm" 2 "llc_allocate" } } */
diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-3.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-3.c
deleted file mode 100644
index d9c053566..000000000
--- a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-3.c
+++ /dev/null
@@ -1,48 +0,0 @@
-/* { dg-do compile { target { aarch64*-*-linux* } } } */
-/* { dg-options "-O3 -march=armv8.2-a+sve -funroll-loops -ffast-math -static -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param target-variables=lPtr,uPtr,, --param mem-ref-index=5" } */
-
-#include <stdio.h>
-
-#define N 131590
-#define F 384477
-
-double diagPtr[N];
-double psiPtr[N];
-double ApsiPtr[N];
-int lPtr[F];
-int uPtr[F];
-double lowerPtr[F];
-double upperPtr[F];
-
-void
-AMUL (double *diagPtr, double *psiPtr, double *ApsiPtr, int *lPtr,
- int *uPtr, double *lowerPtr, double *upperPtr, int nCells, int nFaces)
-{
- for (int cell=0; cell<nCells; cell++)
- ApsiPtr[cell] = diagPtr[cell]*psiPtr[cell];
-
- for (int face=0; face<nFaces; face++)
- {
- ApsiPtr[uPtr[face]] += lowerPtr[face]*psiPtr[lPtr[face]];
- ApsiPtr[lPtr[face]] += upperPtr[face]*psiPtr[uPtr[face]];
- }
-}
-
-int
-main (int argc, char *argv[])
-{
- int nCells = N;
- int nFaces = F;
- int testIter = 2;
-
- for (int i=0; i<testIter; i++)
- AMUL (diagPtr,psiPtr,ApsiPtr,lPtr,uPtr,lowerPtr,upperPtr,nCells,nFaces);
-
- return 0;
-}
-
-/* { dg-final { scan-tree-dump-not "WARNING: The number of provided memory "
- "reference indices is less" "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-not "runtime issue" "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-times "static issue" 2 "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-times "insert prfm" 4 "llc_allocate" } } */
diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-4.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-4.c
deleted file mode 100644
index b87f9903d..000000000
--- a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-4.c
+++ /dev/null
@@ -1,47 +0,0 @@
-/* { dg-do compile { target { aarch64*-*-linux* } } } */
-/* { dg-options "-O3 -march=armv8.2-a+sve -funroll-loops -ffast-math -static -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param target-variables=lPtr --param use-ref-group-index=1" } */
-
-#include <stdio.h>
-
-#define N 131590
-#define F 384477
-
-double diagPtr[N];
-double psiPtr[N];
-double ApsiPtr[N];
-int lPtr[F];
-int uPtr[F];
-double lowerPtr[F];
-double upperPtr[F];
-
-void
-AMUL (double *diagPtr, double *psiPtr, double *ApsiPtr, int *lPtr,
- int *uPtr, double *lowerPtr, double *upperPtr, int nCells, int nFaces)
-{
- for (int cell=0; cell<nCells; cell++)
- ApsiPtr[cell] = diagPtr[cell]*psiPtr[cell];
-
- for (int face=0; face<nFaces; face++)
- {
- ApsiPtr[uPtr[face]] += lowerPtr[face]*psiPtr[lPtr[face]];
- ApsiPtr[lPtr[face]] += upperPtr[face]*psiPtr[uPtr[face]];
- }
-}
-
-int
-main (int argc, char *argv[])
-{
- int nCells = N;
- int nFaces = F;
- int testIter = 2;
-
- for (int i=0; i<testIter; i++)
- AMUL (diagPtr,psiPtr,ApsiPtr,lPtr,uPtr,lowerPtr,upperPtr,nCells,nFaces);
-
- return 0;
-}
-
-/* { dg-final { scan-tree-dump-times "ERROR: not an unsigned integer" 1
- "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-not "runtime issue" "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-not "static issue" "llc_allocate" } } */
diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-5.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-5.c
deleted file mode 100644
index d07836765..000000000
--- a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-5.c
+++ /dev/null
@@ -1,48 +0,0 @@
-/* { dg-do compile { target { aarch64*-*-linux* } } } */
-/* { dg-options "-O3 -march=armv8.2-a+sve -funroll-loops -ffast-math -static -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param target-variables=1 --param use-ref-group-index=1" } */
-
-#include <stdio.h>
-
-#define N 131590
-#define F 384477
-
-double diagPtr[N];
-double psiPtr[N];
-double ApsiPtr[N];
-int lPtr[F];
-int uPtr[F];
-double lowerPtr[F];
-double upperPtr[F];
-
-void
-AMUL (double *diagPtr, double *psiPtr, double *ApsiPtr, int *lPtr,
- int *uPtr, double *lowerPtr, double *upperPtr, int nCells, int nFaces)
-{
- for (int cell=0; cell<nCells; cell++)
- ApsiPtr[cell] = diagPtr[cell]*psiPtr[cell];
-
- for (int face=0; face<nFaces; face++)
- {
- ApsiPtr[uPtr[face]] += lowerPtr[face]*psiPtr[lPtr[face]];
- ApsiPtr[lPtr[face]] += upperPtr[face]*psiPtr[uPtr[face]];
- }
-}
-
-int
-main (int argc, char *argv[])
-{
- int nCells = N;
- int nFaces = F;
- int testIter = 2;
-
- for (int i=0; i<testIter; i++)
- AMUL (diagPtr,psiPtr,ApsiPtr,lPtr,uPtr,lowerPtr,upperPtr,nCells,nFaces);
-
- return 0;
-}
-
-/* { dg-final { scan-tree-dump-times "NOTICE: Prefetching target variable \""
- " psiPtr \"" 2 "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-not "runtime issue" "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-times "static issue" 2 "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-times "insert svprfd" 2 "llc_allocate" } } */
diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-6.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-6.c
deleted file mode 100644
index c0a6afe5b..000000000
--- a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-6.c
+++ /dev/null
@@ -1,47 +0,0 @@
-/* { dg-do compile { target { aarch64*-*-linux* } } } */
-/* { dg-options "-O3 -march=armv8.2-a+sve -funroll-loops -ffast-math -static -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param target-variables=3,a --param use-ref-group-index=1" } */
-
-#include <stdio.h>
-
-#define N 131590
-#define F 384477
-
-double diagPtr[N];
-double psiPtr[N];
-double ApsiPtr[N];
-int lPtr[F];
-int uPtr[F];
-double lowerPtr[F];
-double upperPtr[F];
-
-void
-AMUL (double *diagPtr, double *psiPtr, double *ApsiPtr, int *lPtr,
- int *uPtr, double *lowerPtr, double *upperPtr, int nCells, int nFaces)
-{
- for (int cell=0; cell<nCells; cell++)
- ApsiPtr[cell] = diagPtr[cell]*psiPtr[cell];
-
- for (int face=0; face<nFaces; face++)
- {
- ApsiPtr[uPtr[face]] += lowerPtr[face]*psiPtr[lPtr[face]];
- ApsiPtr[lPtr[face]] += upperPtr[face]*psiPtr[uPtr[face]];
- }
-}
-
-int
-main (int argc, char *argv[])
-{
- int nCells = N;
- int nFaces = F;
- int testIter = 2;
-
- for (int i=0; i<testIter; i++)
- AMUL (diagPtr,psiPtr,ApsiPtr,lPtr,uPtr,lowerPtr,upperPtr,nCells,nFaces);
-
- return 0;
-}
-
-/* { dg-final { scan-tree-dump-times "ERROR: not an unsigned integer" 1
- "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-not "runtime issue" "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-not "static issue" "llc_allocate" } } */
diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-7-null-var-name.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-7-null-var-name.c
deleted file mode 100644
index 4ad331626..000000000
--- a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-7-null-var-name.c
+++ /dev/null
@@ -1,52 +0,0 @@
-/* { dg-do compile { target { aarch64*-*-linux* } } } */
-/* { dg-options "-O3 -c -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param filter-kernels=0 --param target-variables=\"bb_16(D)->aux\"" } */
-
-#include <stdio.h>
-
-typedef struct stack_def
-{
- int top; /* index to top stack element */
- unsigned long reg_set; /* set of live registers */
- unsigned char reg[128]; /* register - stack mapping */
-} *stack;
-
-typedef struct block_info_def
-{
- struct stack_def stack_in; /* Input stack configuration. */
- struct stack_def stack_out; /* Output stack configuration. */
- unsigned long out_reg_set; /* Stack regs live on output. */
- int done; /* True if block already converted. */
- int predecessors; /* Number of predecessors that need
- to be visited. */
-} *block_info;
-
-typedef struct basic_block_def
-{
- void *aux;
-} *basic_block;
-
-unsigned char
-convert_regs_exit (basic_block bb, int value_reg_low, int value_reg_high)
-{
- stack output_stack;
-
- output_stack = &(((block_info) bb->aux)->stack_in);
- if (value_reg_low == -1)
- output_stack->top = -1;
- else
- {
- int reg;
- output_stack->top = value_reg_high - value_reg_low;
- for (reg = value_reg_low; reg <= value_reg_high; ++reg)
- {
- (output_stack->reg + 16)[value_reg_high - reg] = reg;
- output_stack->reg_set |= (unsigned long) 1 << reg;
- }
- }
- return output_stack->reg[0];
-}
-
-/* { dg-final { scan-tree-dump-not "Unrecognizable variable name"
- "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-not "runtime issue" "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-not "static issue" "llc_allocate" } } */
diff --git a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-8-tmp-var-name.c b/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-8-tmp-var-name.c
deleted file mode 100644
index 09a525ce1..000000000
--- a/gcc/testsuite/gcc.dg/llc-allocate/llc-tool-insertion-8-tmp-var-name.c
+++ /dev/null
@@ -1,54 +0,0 @@
-/* { dg-do compile { target { aarch64*-*-linux* } } } */
-/* { dg-options "-O3 -c -fllc-allocate -fdump-tree-llc_allocate-details-lineno --param filter-kernels=0 --param target-variables=tmp_var_0" } */
-
-#include <stdio.h>
-
-typedef struct stack_def
-{
- int top; /* index to top stack element */
- unsigned long reg_set; /* set of live registers */
- unsigned char reg[128]; /* register - stack mapping */
-} *stack;
-
-typedef struct block_info_def
-{
- struct stack_def stack_in; /* Input stack configuration. */
- struct stack_def stack_out; /* Output stack configuration. */
- unsigned long out_reg_set; /* Stack regs live on output. */
- int done; /* True if block already converted. */
- int predecessors; /* Number of predecessors that need
- to be visited. */
-} *block_info;
-
-typedef struct basic_block_def
-{
- void *aux;
-} *basic_block;
-
-unsigned char
-convert_regs_exit (basic_block bb, int value_reg_low, int value_reg_high)
-{
- stack output_stack;
-
- output_stack = &(((block_info) bb->aux)->stack_in);
- if (value_reg_low == -1)
- output_stack->top = -1;
- else
- {
- int reg;
- output_stack->top = value_reg_high - value_reg_low;
- for (reg = value_reg_low; reg <= value_reg_high; ++reg)
- {
- (output_stack->reg + 16)[value_reg_high - reg] = reg;
- output_stack->reg_set |= (unsigned long) 1 << reg;
- }
- }
- return output_stack->reg[0];
-}
-
-/* { dg-final { scan-tree-dump-not "Unrecognizable variable name"
- "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-times "NOTICE: Prefetching target variable \""
- " bb_16(D)->aux \"" 1 "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-not "runtime issue" "llc_allocate" } } */
-/* { dg-final { scan-tree-dump-times "static issue" 1 "llc_allocate" } } */
diff --git a/gcc/tree-ssa-llc-allocate.c b/gcc/tree-ssa-llc-allocate.c
index 890f66e54..fa8979401 100644
--- a/gcc/tree-ssa-llc-allocate.c
+++ b/gcc/tree-ssa-llc-allocate.c
@@ -23,7 +23,6 @@ along with GCC; see the file COPYING3. If not see
#define INCLUDE_VECTOR
#define INCLUDE_LIST
#define INCLUDE_ALGORITHM
-#define INCLUDE_STRING
#include "system.h"
#include "coretypes.h"
#include "backend.h"
@@ -1866,7 +1865,10 @@ filter_and_sort_kernels (vector<class loop *> &sorted_kernels,
list<basic_block> walked_header_bb; /* Used to record nested loops. */
for (unsigned i = 0; i < kernels.size (); ++i)
- end_bb.insert (kernels[i]->header);
+ {
+ if (kernels[i]->inner == NULL)
+ end_bb.insert (kernels[i]->header);
+ }
dump_loop_headers ("kernels", kernels);
@@ -2380,30 +2382,6 @@ issue_builtin_prefetch (data_ref &mem_ref)
update_ssa (TODO_update_ssa_only_virtuals);
}
-/* Retrieve memory reference at the specific index. */
-
-data_ref
-get_data_ref_at_idx (ref_group &var_ref_group)
-{
- unsigned int mem_ref_size = static_cast<unsigned int>(
- var_ref_group.ref_scores.size ());
- if (strlen (param_mem_ref_index) == 0)
- return var_ref_group.first_use;
- else
- {
- /* Insert prefetch hint at highly-likely-used location with the given
- index. */
- if (var_ref_group.mem_ref_index >= mem_ref_size)
- {
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "WARNING: The target data_ref index is out "
- "of range. Use top index instead!\n");
- return var_ref_group.ref_scores[0].d_ref;
- }
- return var_ref_group.ref_scores[var_ref_group.mem_ref_index].d_ref;
- }
-}
-
/* Static form insertion and issue instruction. We may check the
determination of the ARM SVE architecture before SVE hint insertion. */
@@ -2415,7 +2393,7 @@ static_issue (vector<ref_group> &ref_groups, int num_issue_var)
for (int i = 0; i < num_issue_var; ++i)
{
- data_ref mem_ref = get_data_ref_at_idx (ref_groups[i]);
+ data_ref mem_ref = ref_groups[i].first_use;
if (mem_ref.vectorize_p)
{
enum internal_fn ifn_code = gimple_call_internal_fn
@@ -2591,10 +2569,7 @@ issue_llc_hint (vector<ref_group> &ref_groups)
}
if (param_force_issue)
{
- if (strlen (param_target_variables) > 0)
- static_issue (ref_groups, static_cast<int>(ref_groups.size ()));
- else
- static_issue (ref_groups, num_issue_var);
+ static_issue (ref_groups, num_issue_var);
return;
}
calc_type topn_calc_type = STATIC_CALC;
@@ -2626,224 +2601,6 @@ issue_llc_hint (vector<ref_group> &ref_groups)
}
/* ==================== phase entry ==================== */
-/* Check whether a string can be converted to an unsigned integer. */
-
-bool is_unsigned_int (const string &s)
-{
- if (s.empty () || s.size () > PREFETCH_TOOL_NUM_MAX_LEN)
- return false;
-
- for (unsigned int i = 0; i < s.size (); ++i)
- {
- if (s[i] < '0' || s[i] > '9')
- return false;
- }
- return true;
-}
-
-/* Parse a substring separated by comma. If the substring is valid and
- non-empty, store it as a parsed element. */
-
-bool
-parse_string_helper (const string &substr, vector<string>& str_elts,
- bool check_unsigned, size_t start, size_t end)
-{
- if (substr == "" && dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "WARNING: The input string from %lu to %lu is "
- "empty.\n", start, end);
- else if (check_unsigned && !is_unsigned_int (substr))
- {
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "ERROR: not an unsigned integer: %s\n",
- substr.c_str ());
- str_elts.clear ();
- return false;
- }
- else
- str_elts.push_back (substr);
- return true;
-}
-
-/* Parse a user input string, separated by comma. */
-
-void
-parse_string (const string &s, vector<string>& str_elts,
- bool check_unsigned = false)
-{
- string delim = ",";
- size_t start = 0;
- size_t end = s.find (delim);
- string substr = s.substr (start, end - start);
- while (end != string::npos)
- {
- if (!parse_string_helper (substr, str_elts, check_unsigned, start, end))
- return;
- start = end + delim.size ();
- end = s.find (delim, start);
- substr = s.substr (start, end - start);
- }
- parse_string_helper (substr, str_elts, check_unsigned, start, end);
-}
-
-/* Parse user input of target variables and memory indices and create a map
- that assigns a target variable to a memory index. */
-
-void
-parse_param_inputs (map<string, unsigned int> &var2mem_idx)
-{
- /* The user input length should have an input length limit. */
- if ((strlen (param_target_variables) >= PREFETCH_TOOL_INPUT_MAX_LEN
- || strlen (param_mem_ref_index) >= PREFETCH_TOOL_INPUT_MAX_LEN)
- && dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "INVALID INPUT: The user inputs for target variables "
- "and/or memory reference indices are too long for parsing.\n");
-
- vector<string> var_names;
- string target_variables = param_target_variables;
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "Start parsing target variables:\n");
- if (param_use_ref_group_index)
- parse_string (target_variables, var_names, true);
- else
- parse_string (target_variables, var_names, false);
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "Finish parsing target variables.\n\n");
-
- vector<string> var_mem_indices;
- string mem_indices = param_mem_ref_index;
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "Start parsing memory reference indices:\n");
- parse_string (mem_indices, var_mem_indices, true);
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "Finish parsing memory reference indices.\n\n");
-
- /* Construct a map of var_name: var_mem_index. */
- if (var_names.size () > 0)
- {
- if (var_mem_indices.size () < var_names.size ())
- {
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "WARNING: The number of provided memory "
- "reference indices is less than that of target "
- "variables.\nUse the top index for all variables "
- "instead.\n");
- for (string& var_name : var_names)
- var2mem_idx[var_name] = 0;
- }
- else
- {
- if (var_mem_indices.size () > var_names.size ()
- && dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "WARNING: The number of target variables is "
- "less than that of memory reference indices.\n");
- for (unsigned int i = 0; i < var_names.size (); ++i)
- {
- var2mem_idx[var_names[i]] = static_cast<unsigned int>(
- atoi (var_mem_indices[i].c_str ()));
- }
- }
- }
-}
-
-/* Filter reference groups by only selecting target variables from the user
- input. There are two options for prefetching target variables:
- 1. Specify variable name parsed by the pass, which you can double-check at
- "sorted ref_groups" section in the dump file.
- 2. Specify variable rank exhibited at "sorted ref_groups" section in the
- dump file.
-*/
-
-void
-prefetch_variables (const vector<ref_group>& ref_groups,
- vector<ref_group>& reduced_ref_groups)
-{
- map<unsigned int, unsigned int> ref_group2mem_idx;
-
- map<string, unsigned int> var2mem_idx; /* externally defined. */
- parse_param_inputs (var2mem_idx);
-
- if (param_use_ref_group_index)
- {
- /* Use ref_group index at "sorted ref_groups" section to specify
- variable. */
- /* Collect the variables in "reduced_ref_group" only if their indices
- show up at "sorted ref_groups" section. */
- for (const pair<string, unsigned int> &var_mem_idx : var2mem_idx)
- {
- unsigned int var_idx = static_cast<unsigned int>(atoi (
- var_mem_idx.first.c_str ()));
- if (var_idx < ref_groups.size ())
- ref_group2mem_idx[var_idx] = var_mem_idx.second;
- else if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "WARNING: The index \"%u\" does not show "
- "up in the ref_groups.\n", var_idx);
- }
- }
- else
- {
- /* Use variable name shown up at "sorted ref_groups" section to specify
- variable:
- var2ref_group_idx + var2mem_idx -> ref_group2mem_idx. */
- /* Create a map that assigns the variable name to its corresponding
- ref_group index. */
- map<string, unsigned int> var2ref_group_idx; /* internally detected. */
- for (unsigned int i = 0; i < ref_groups.size (); ++i)
- {
- const ref_group &curr_ref_group = ref_groups[i];
- const int UINT_MAX_DIGIT = 10;
- /* Unrecognizable variable name related to ref_group. */
- if (!get_name (curr_ref_group.var))
- {
- /* If the variable name does not have a string representation,
- we can rename it by "tmp_var_" + <sorted_ref_group_index>. */
- char group_idx[UINT_MAX_DIGIT];
- sprintf (group_idx, "%u", i);
- string tmp_var_name = "tmp_var_" + std::string (group_idx);
- fprintf (dump_file, "Unrecognizable variable name at ref_group "
- "index %u.\nThe tree expression for variable is: ", i);
- print_generic_expr (dump_file, curr_ref_group.var, TDF_SLIM);
- fprintf (dump_file, "\n");
- var2ref_group_idx[tmp_var_name] = i;
- }
- else
- var2ref_group_idx[std::string (get_name (curr_ref_group.var))] = i;
- }
- /* Collect the variables in "reduced_ref_group" only if they show up in
- the ref_groups. */
- for (const pair<string, unsigned int> &var_mem_idx : var2mem_idx)
- {
- if (var2ref_group_idx.count (var_mem_idx.first))
- {
- unsigned int ref_group_idx = var2ref_group_idx[var_mem_idx.first];
- ref_group2mem_idx[ref_group_idx] = var_mem_idx.second;
- }
- else if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "WARNING: Target variable \" %s \" does "
- "not show up in the ref_groups. Check whether it needs "
- "temporary variable name.\n",
- var_mem_idx.first.c_str ());
- }
- }
-
- for (const pair<unsigned int, unsigned int> &ref_group_mem_idx :
- ref_group2mem_idx)
- {
- ref_group curr_ref_group = ref_groups[ref_group_mem_idx.first];
- curr_ref_group.mem_ref_index = ref_group_mem_idx.second;
- reduced_ref_groups.push_back (curr_ref_group);
- if (dump_file && (dump_flags & TDF_DETAILS))
- {
- fprintf (dump_file, "\nNOTICE: Prefetching target variable \" ");
- print_generic_expr (dump_file, curr_ref_group.var, TDF_SLIM);
- fprintf (dump_file, " \" at ref_group index %u and memory location "
- "index %u.\n", ref_group_mem_idx.first,
- ref_group_mem_idx.second);
- }
- }
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "\n\n");
-}
-
/* The LLC intelligent allocation consists of 6 steps. */
@@ -2869,17 +2626,7 @@ llc_allocate (void)
if (!record_and_sort_ref_groups (ref_groups, sorted_kernels, kernels_refs))
return;
- if (strlen (param_target_variables) > 0)
- {
- /* If "param_target_variables" is not empty, we will issue parsed target
- variables compulsorily. */
- param_force_issue = true;
- vector<ref_group> reduced_ref_groups;
- prefetch_variables (ref_groups, reduced_ref_groups);
- issue_llc_hint (reduced_ref_groups);
- }
- else
- issue_llc_hint (ref_groups);
+ issue_llc_hint (ref_groups);
}
/* Check whether the function is an operator reloading function. */
--
2.33.0
Reference in New Issue View Git Blame Copy Permalink