aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorLukas Larsson <[email protected]>2017-09-26 10:46:53 +0200
committerLukas Larsson <[email protected]>2017-09-27 13:44:43 +0200
commit8521b555ae3e956fd76d9de14c59f64ac13bc9a2 (patch)
tree85e8f8d867f0c0e36ac413fcb4d764ca585bcb45
parent3c23f2c6335d8586aca1efb9da6d20f00a8840fd (diff)
downloadotp-8521b555ae3e956fd76d9de14c59f64ac13bc9a2.tar.gz
otp-8521b555ae3e956fd76d9de14c59f64ac13bc9a2.tar.bz2
otp-8521b555ae3e956fd76d9de14c59f64ac13bc9a2.zip
erts: Add makefile target to check emu register allocation
-rw-r--r--erts/emulator/Makefile.in18
-rwxr-xr-xerts/emulator/utils/beam_emu_vars122
2 files changed, 139 insertions, 1 deletions
diff --git a/erts/emulator/Makefile.in b/erts/emulator/Makefile.in
index 02a9a9a93b..0cf8fee15d 100644
--- a/erts/emulator/Makefile.in
+++ b/erts/emulator/Makefile.in
@@ -499,7 +499,6 @@ ifeq ($(TARGET),win32)
RELEASE_INCLUDES += sys/$(ERLANG_OSTYPE)/erl_win_dyn_driver.h
endif
-
.PHONY: release_spec
ifdef VOID_EMULATOR
release_spec:
@@ -711,16 +710,27 @@ $(OBJDIR)/beams.$(RES_EXT): $(TARGET)/beams.rc
endif
+# We disable the implicit rule of .S -> .o so that the verbose asm
+# generate is not used for compiling erts. This is only a problem on
+# old solaris make
+%.o : %.S
+
# Usually the same as the default rule, but certain platforms (e.g. win32) mix
# different compilers
$(OBJDIR)/beam_emu.o: beam/beam_emu.c
$(V_EMU_CC) $(subst -O2, $(GEN_OPT_FLGS), $(CFLAGS)) $(INCLUDES) -c $< -o $@
+$(OBJDIR)/beam_emu.S: beam/beam_emu.c
+ $(V_EMU_CC) -S -fverbose-asm $(subst -O2, $(GEN_OPT_FLGS), $(CFLAGS)) $(INCLUDES) -c $< -o $@
+
$(OBJDIR)/%_pg.o: beam/%.c
$(V_CC) $(PROFILE_GENERATE) $(subst -O2, $(GEN_OPT_FLGS), $(CFLAGS)) $(INCLUDES) -c $< -o $@
$(OBJDIR)/%_pu.o: beam/%.c $(PROFILE_USE_DEPS)
$(V_CC) $(PROFILE_USE) $(subst -O2, $(GEN_OPT_FLGS), $(CFLAGS)) $(INCLUDES) -c $< -o $@
+$(OBJDIR)/%_pu.S: beam/%.c $(PROFILE_USE_DEPS)
+ $(V_CC) -S -fverbose-asm $(PROFILE_USE) $(subst -O2, $(GEN_OPT_FLGS), $(CFLAGS)) $(INCLUDES) -c $< -o $@
+
$(OBJDIR)/PROFILE: $(BINDIR)/$(PROFILE_EXECUTABLE)
$(V_at)echo " PROFILE ${PROFILE_EXECUTABLE}"
$(V_at)rm -f $(OBJDIR)/erl*.profraw
@@ -732,6 +742,8 @@ $(OBJDIR)/PROFILE: $(BINDIR)/$(PROFILE_EXECUTABLE)
-noshell -s estone_SUITE pgo -s init stop >> $(OBJDIR)/PROFILE_LOG
$(V_at)touch $@
+.SECONDARY: $(patsubst %.o,%_pu.gcda,$(PROFILE_OBJS))
+
$(OBJDIR)/%_pu.gcda: $(OBJDIR)/PROFILE
$(V_at)mv $(OBJDIR)/$*_pg.gcda $@
$(V_at)touch $@
@@ -1167,6 +1179,10 @@ DEP_FLAGS=-MM $(MG_FLAG) $(CFLAGS) $(INCLUDES) -Inifs/common -Idrivers/common -I
SYS_SRC=$(ALL_SYS_SRC)
endif
+.PHONY: check_emu_registers
+check_emu_registers: $(OBJDIR)/beam_emu$(PROFILE_MARKER).S
+ utils/beam_emu_vars -vars 'c_p E HTOP FCALLS I reg' $^
+
.PHONY: $(TARGET)/gen_git_version.mk
$(TARGET)/gen_git_version.mk:
# We touch beam/erl_bif.info.c if we regenerated the git version to force a
diff --git a/erts/emulator/utils/beam_emu_vars b/erts/emulator/utils/beam_emu_vars
new file mode 100755
index 0000000000..c798a4dada
--- /dev/null
+++ b/erts/emulator/utils/beam_emu_vars
@@ -0,0 +1,122 @@
+#!/usr/bin/perl -w
+use strict;
+
+# Analyse beam_emu.s and try to find out the registers
+# used for the important variables in process_main().
+#
+# Works for .s files from clang or gcc. For gcc, the -fverbose-asm
+# option must be used.
+#
+# Example:
+#
+# $ beam-emu-vars -vars 'c_p E HTOP FCALLS I reg freg' beam_emu.s
+# E: %r13
+# FCALLS: %rcx:98 %rax:88 16(%rsp):50 %rdi:6
+# HTOP: %r10:382 64(%rsp):88 72(%rsp):9 24(%rsp):7 %rcx:6 %r15:6 80(%rsp):3 88(%rsp):2
+# I: %rbx
+# c_p: %rbp
+# freg: 48(%rsp):11 %rcx:8 %rdi:5 %rax:4
+# reg: %r12
+#
+# That means that E, I, c_p, reg seems to be assigned to permanent registers.
+# HTOP seems to be assigned %r10, but it is saved to a scratch location
+# before any function calls. FCALLS and freg seems to be saved in a location on
+# the stack and loaded into a register when used.
+#
+# The exit status will be 0 if all variables are assigned to registers (most of
+# the time), and 1 if one or more variables are assigned to a stack location.
+
+my $vars = 'c_p E FCALLS freg HTOP I reg';
+
+while (@ARGV and $ARGV[0] =~ /^-(.*)/) {
+ $_ = $1;
+ shift;
+ ($vars = shift), next if /^vars/;
+ die "$0: Bad option: -$_\n";
+}
+
+my @vars = split(" ", $vars);
+my %vars;
+@vars{@vars} = @vars;
+
+my $inside;
+my %count;
+
+if (@ARGV != 1) {
+ usage();
+}
+
+while (<>) {
+ if (!$inside && /[.]globl\s*_?process_main/) {
+ $inside = 1;
+ } elsif ($inside && /[.]globl/) {
+ last;
+ }
+ if ($inside) {
+ if (/##DEBUG_VALUE:\s*process_main:([A-Za-z]*)\s*<-\s*(.*)/) {
+ # clang
+ my($var,$reg) = ($1,$2);
+ next if $reg =~ /^[-\d]+$/; # Ignore if number.
+ $count{$var}->{$reg}++ if $vars{$var};
+ next;
+ }
+
+ # Parse gcc verbose arguments. Comments are marked with
+ # one '#' (clang marks its comments with two '#').
+ my($src,$dst,$comment) = /movq\s+([^#]+), ([^#]+)#(?!#)\s*(.*)/;
+ next unless $comment;
+ $dst =~ s/\s*$//;
+ my($vsrc,$vdst) = split /,/, $comment, 2;
+ $vdst =~ s/^\s//;
+ update_count(\%count, $vsrc, $src);
+ update_count(\%count, $vdst, $dst);
+ if ($vars{$vdst} and $vsrc eq '%sfp') {
+ $count{$vdst}->{$src}++;
+ }
+ }
+}
+
+my @first;
+
+OUTER:
+for my $var (sort keys %count) {
+ my $total = 0;
+
+ foreach my $reg (keys %{$count{$var}}) {
+ $total += $count{$var}->{$reg}++;
+ }
+
+ foreach my $reg (keys %{$count{$var}}) {
+ if ($count{$var}->{$reg} > 0.9*$total) {
+ print "$var: $reg\n";
+ push @first, $var;
+ next OUTER;
+ }
+ }
+
+ my @r;
+ foreach my $reg (keys %{$count{$var}}) {
+ push @r, $reg;
+ }
+ @r = sort { $count{$var}->{$b} <=> $count{$var}->{$a} } @r;
+ @r = map { "$_:$count{$var}->{$_}" } @r;
+ push @first, $r[0];
+ print "$var: ", join(' ', @r), "\n";
+}
+
+foreach (@first) {
+ exit 1 if /%rsp/;
+}
+exit 0;
+
+sub update_count {
+ my($count_ref,$var,$reg) = @_;
+ return unless $vars{$var};
+ ${${$count_ref}{$var}}{$reg}++;
+}
+
+sub usage {
+ die qq[usage: beam_emu_vars [ -vars "var1 var2..." ] <filename>.s\n\n] .
+ "The exit status is 0 if all variables are assigned to registers,\n" .
+ "and 1 if one or more variables are allocated to a stack location.\n";
+}