cmucl-cvs

cmucl-cvs@common-lisp.net

2904 discussions

[Git][cmucl/cmucl][issue-86-save-fpu-state-on-entry-to-alloc] Save just the xmm registers
by Raymond Toy 29 Aug '20

29 Aug '20

Raymond Toy pushed to branch issue-86-save-fpu-state-on-entry-to-alloc at cmucl / cmucl Commits: 17144e16 by Raymond Toy at 2020-08-28T16:23:59-07:00 Save just the xmm registers Instead of saving the entire FPU state, we really only need to save the xmm registers. - - - - - 1 changed file: - src/lisp/x86-assem.S Changes: ===================================== src/lisp/x86-assem.S ===================================== @@ -404,22 +404,39 @@ FUNCDEF(alloc_overflow_sse2) ENDFUNC(alloc_overflow_sse2) #else FUNCDEF(alloc_overflow_sse2) - # Need 512 bytes for the fpu save area, space to save ecx and edx, - # space for mxcsr, a temp, and one arg to pass to alloc. That's - # 512 + 20. But the save area needs to be 16-byte aligned, so - # allocate 512 + 32 bytes. The fpu area will be at offset 32. + # Need 8*16 bytes for the xmm registers, and space to save ecx + # and edx, space for mxcsr, a temp, and one arg to pass to alloc. + # That's 8*16 + 5*4 = 148 bytes. Might as well have a few + # more so the xmm0 area is 16-byte aligned. That makes it 160 + # bytes. # # Stack looks like: # - # +544 -> end - # +32 -> fpu save - # +20 -> unused + # +160 + # +144 -> xmm7 + # +128 -> xmm6 + # +112 -> xmm5 + # +96 -> xmm4 + # +80 -> xmm3 + # +64 -> xmm2 + # +48 -> xmm1 + # +32 -> xmm0 + # +20 -> unused # +16 -> temp # +12 -> mxcsr # + 8 -> save ecx # + 4 -> save edx # esp + 0 -> arg for alloc - STACK_PROLOGUE(32+512) + STACK_PROLOGUE(160) + movapd %xmm0, (32 + 0*16)(%esp) + movapd %xmm1, (32 + 1*16)(%esp) + movapd %xmm2, (32 + 2*16)(%esp) + movapd %xmm3, (32 + 3*16)(%esp) + movapd %xmm4, (32 + 4*16)(%esp) + movapd %xmm5, (32 + 5*16)(%esp) + movapd %xmm6, (32 + 6*16)(%esp) + movapd %xmm7, (32 + 7*16)(%esp) + movl %ecx, 8(%esp) # Save ecx and edx registers movl %edx, 4(%esp) stmxcsr 12(%esp) # Save MXCSR @@ -430,16 +447,21 @@ FUNCDEF(alloc_overflow_sse2) ldmxcsr 16(%esp) # Get new mxcsr value movl %eax, (%esp) # Put size on stack for first arg to alloc() - # Save all FPU regs because we don't know what's in use by lisp. - fxsave 32(%esp) - call GNAME(alloc) - fxrstor 32(%esp) - movl 4(%esp), %edx # Restore edx and ecx registers. eax has the return value. movl 8(%esp), %ecx ldmxcsr 12(%esp) + + movapd (32 + 0*16)(%esp), %xmm0 + movapd (32 + 1*16)(%esp), %xmm1 + movapd (32 + 2*16)(%esp), %xmm2 + movapd (32 + 3*16)(%esp), %xmm3 + movapd (32 + 4*16)(%esp), %xmm4 + movapd (32 + 5*16)(%esp), %xmm5 + movapd (32 + 6*16)(%esp), %xmm6 + movapd (32 + 7*16)(%esp), %xmm7 + STACK_EPILOGUE ret ENDFUNC(alloc_overflow_sse2) View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/commit/17144e16d4f7578644fac57… -- View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/commit/17144e16d4f7578644fac57… You're receiving this email because of your account on gitlab.common-lisp.net.

1 0

[Git][cmucl/cmucl][issue-86-save-fpu-state-on-entry-to-alloc] Remove stray #pragma
by Raymond Toy 28 Aug '20

28 Aug '20

Raymond Toy pushed to branch issue-86-save-fpu-state-on-entry-to-alloc at cmucl / cmucl Commits: e3aa51f3 by Raymond Toy at 2020-08-27T20:58:52-07:00 Remove stray #pragma Forgot to remove this; it's not needed anymore. - - - - - 1 changed file: - src/lisp/gencgc.c Changes: ===================================== src/lisp/gencgc.c ===================================== @@ -8527,7 +8527,6 @@ component_ptr_from_pc(lispobj * pc) return NULL; } -#pragma GCC optimize ("-O1") /* * Get lower and upper(middle) 28 bits of total allocation */ View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/commit/e3aa51f30a04d6f299a4c36… -- View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/commit/e3aa51f30a04d6f299a4c36… You're receiving this email because of your account on gitlab.common-lisp.net.

1 0

[Git][cmucl/cmucl][issue-86-save-fpu-state-on-entry-to-alloc] Save FPU state in alloc_overflow_sse2
by Raymond Toy 28 Aug '20

28 Aug '20

Raymond Toy pushed to branch issue-86-save-fpu-state-on-entry-to-alloc at cmucl / cmucl Commits: 8b08b800 by Raymond Toy at 2020-08-27T20:39:07-07:00 Save FPU state in alloc_overflow_sse2 It's best to save the FPU state here instead of in alloc() because we can't know what the compiler might do. Remove the fpu save stuff from alloc(). gcc 9.3.1 builds lisp successfully. - - - - - 2 changed files: - src/lisp/gencgc.c - src/lisp/x86-assem.S Changes: ===================================== src/lisp/gencgc.c ===================================== @@ -8412,28 +8412,11 @@ gencgc_pickup_dynamic(void) void do_pending_interrupt(void); -//#pragma GCC optimize ("-O1") char * alloc(int nbytes) { -#if (defined(i386) || defined(__x86_64)) - /* - * Need to save and restore the FPU registers on x86, but only for - * sse2. See Trac ticket #61 - * (https://trac.common-lisp.net/cmucl/ticket/61) and gitlab - * ticket #86 - * (https://gitlab.common-lisp.net/cmucl/cmucl/-/issues/86). - * - * Not needed by sparc or ppc because we never call alloc from - * Lisp directly to do allocation. - */ - FPU_STATE(fpu_state); - - if (fpu_mode == SSE2) { - save_fpu_state(fpu_state); - } -#endif void *new_obj; + #if !(defined(sparc) || (defined(DARWIN) && defined(__ppc__))) /* * *current-region-free-pointer* is the same as alloc-tn (= @@ -8483,14 +8466,8 @@ alloc(int nbytes) } } -#if (defined(i386) || defined(__x86_64)) - if (fpu_mode == SSE2) { - restore_fpu_state(fpu_state); - } -#endif return new_obj; } -#pragma GCC optimize ("-O2") char * alloc_pseudo_atomic(int nbytes) ===================================== src/lisp/x86-assem.S ===================================== @@ -381,6 +381,7 @@ ENDFUNC(fastcopy16) * On exit: * %eax = address */ +#if 0 FUNCDEF(alloc_overflow_sse2) STACK_PROLOGUE(20) movl %ecx, 8(%esp) # Save ecx and edx registers @@ -401,7 +402,49 @@ FUNCDEF(alloc_overflow_sse2) STACK_EPILOGUE ret ENDFUNC(alloc_overflow_sse2) - +#else +FUNCDEF(alloc_overflow_sse2) + # Need 512 bytes for the fpu save area, space to save ecx and edx, + # space for mxcsr, a temp, and one arg to pass to alloc. That's + # 512 + 20. But the save area needs to be 16-byte aligned, so + # allocate 512 + 32 bytes. The fpu area will be at offset 32. + # + # Stack looks like: + # + # +544 -> end + # +32 -> fpu save + # +20 -> unused + # +16 -> temp + # +12 -> mxcsr + # + 8 -> save ecx + # + 4 -> save edx + # esp + 0 -> arg for alloc + STACK_PROLOGUE(32+512) + movl %ecx, 8(%esp) # Save ecx and edx registers + movl %edx, 4(%esp) + stmxcsr 12(%esp) # Save MXCSR + /* Clear the exceptions that might have occurred */ + movl 12(%esp), %edx + and $-64, %edx # Clear the exceptions + movl %edx, 16(%esp) + ldmxcsr 16(%esp) # Get new mxcsr value + movl %eax, (%esp) # Put size on stack for first arg to alloc() + + # Save all FPU regs because we don't know what's in use by lisp. + fxsave 32(%esp) + + call GNAME(alloc) + + fxrstor 32(%esp) + + movl 4(%esp), %edx # Restore edx and ecx registers. eax has the return value. + movl 8(%esp), %ecx + ldmxcsr 12(%esp) + STACK_EPILOGUE + ret +ENDFUNC(alloc_overflow_sse2) +#endif + #ifdef LINKAGE_TABLE /* Call into C code to resolve a linkage entry. The initial code in the View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/commit/8b08b800dc1c26d498fbc40… -- View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/commit/8b08b800dc1c26d498fbc40… You're receiving this email because of your account on gitlab.common-lisp.net.

1 0

[Git][cmucl/cmucl][issue-86-save-fpu-state-on-entry-to-alloc] 3 commits: Update comments
by Raymond Toy 27 Aug '20

27 Aug '20

Raymond Toy pushed to branch issue-86-save-fpu-state-on-entry-to-alloc at cmucl / cmucl Commits: a95db7ba by Raymond Toy at 2020-08-26T23:30:54-07:00 Update comments - - - - - ad3862c9 by Raymond Toy at 2020-08-26T23:34:05-07:00 Clean up code - - - - - 01f8217b by Raymond Toy at 2020-08-26T23:41:36-07:00 Add -R flag to recompile lisp - - - - - 3 changed files: - .gitlab-ci.yml - src/lisp/gencgc.c - src/lisp/x86-arch.h Changes: ===================================== .gitlab-ci.yml ===================================== @@ -12,7 +12,7 @@ linux-runner: - mkdir snapshot - (cd snapshot; tar xjf ../cmucl-$version-linux.tar.bz2; tar xjf ../cmucl-$version-linux.extra.tar.bz2) script: - - bin/build.sh $bootstrap -C "" -o snapshot/bin/lisp + - bin/build.sh $bootstrap -R -C "" -o snapshot/bin/lisp - bin/make-dist.sh -I dist linux-4 - bin/run-tests.sh -l dist/bin/lisp 2>&1 | tee test.log @@ -24,6 +24,6 @@ osx-runner: - mkdir snapshot - (cd snapshot; tar xjf ../cmucl-$version-darwin.tar.bz2) script: - - bin/build.sh $bootstrap -C "" -o snapshot/bin/lisp + - bin/build.sh $bootstrap -R -C "" -o snapshot/bin/lisp - bin/make-dist.sh -I dist darwin-4 - bin/run-tests.sh -l dist/bin/lisp 2>&1 | tee test.log ===================================== src/lisp/gencgc.c ===================================== @@ -8416,10 +8416,13 @@ void do_pending_interrupt(void); char * alloc(int nbytes) { -#if 0 && (defined(i386) || defined(__x86_64)) +#if (defined(i386) || defined(__x86_64)) /* * Need to save and restore the FPU registers on x86, but only for - * sse2. See Ticket #61. + * sse2. See Trac ticket #61 + * (https://trac.common-lisp.net/cmucl/ticket/61) and gitlab + * ticket #86 + * (https://gitlab.common-lisp.net/cmucl/cmucl/-/issues/86). * * Not needed by sparc or ppc because we never call alloc from * Lisp directly to do allocation. @@ -8457,20 +8460,6 @@ alloc(int nbytes) set_current_region_free((lispobj) new_free_pointer); break; } else if (bytes_allocated <= auto_gc_trigger) { -#if 1 && (defined(i386) || defined(__x86_64)) - /* - * Need to save and restore the FPU registers on x86, but only for - * sse2. See Ticket #61. - * - * Not needed by sparc or ppc because we never call alloc from - * Lisp directly to do allocation. - */ - FPU_STATE(fpu_state); - - if (fpu_mode == SSE2) { - save_fpu_state(fpu_state); - } -#endif /* Call gc_alloc. */ boxed_region.free_pointer = (void *) get_current_region_free(); boxed_region.end_addr = @@ -8481,11 +8470,6 @@ alloc(int nbytes) set_current_region_free((lispobj) boxed_region.free_pointer); set_current_region_end((lispobj) boxed_region.end_addr); -#if 1 && (defined(i386) || defined(__x86_64)) - if (fpu_mode == SSE2) { - restore_fpu_state(fpu_state); - } -#endif break; } else { /* Run GC and try again. */ @@ -8499,7 +8483,7 @@ alloc(int nbytes) } } -#if 0 && (defined(i386) || defined(__x86_64)) +#if (defined(i386) || defined(__x86_64)) if (fpu_mode == SSE2) { restore_fpu_state(fpu_state); } ===================================== src/lisp/x86-arch.h ===================================== @@ -17,15 +17,13 @@ extern boolean os_support_sse2(void); #define FPU_STATE_SIZE 27 /* - * Need 512 byte area, aligned on a 16-byte boundary. So allocate - * 512+16 bytes of space and let the routine adjust the appropriate - * alignment. + * Need 512 byte area, aligned on a 16-byte boundary. */ #define SSE_STATE_SIZE 512 /* * Just use the SSE size for both x87 and sse2 since the SSE size is - * enough for either. + * enough for either. Make sure it's on a 16-byte boundary. */ #define FPU_STATE(name) u_int8_t name[SSE_STATE_SIZE] __attribute__((aligned(16))) View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/compare/4b80a6e595faa3e2343b62… -- View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/compare/4b80a6e595faa3e2343b62… You're receiving this email because of your account on gitlab.common-lisp.net.

1 0

[Git][cmucl/cmucl][issue-86-save-fpu-state-on-entry-to-alloc] 3 commits: Fix #85: Let each x86 configs set optimization level
by Raymond Toy 27 Aug '20

27 Aug '20

Raymond Toy pushed to branch issue-86-save-fpu-state-on-entry-to-alloc at cmucl / cmucl Commits: 29cac208 by Raymond Toy at 2020-08-27T00:36:27+00:00 Fix #85: Let each x86 configs set optimization level Add `COPT` variable in `Config.x86_common` to set the optimization level (defaulting to `-O2`). Then each `Config.x86` file can set `COPT` as desired if the default doesn't work. Thus, `Config.x86_linux` sets `COPT` to `-O1`, but others can use the default value. See issue #68. - - - - - d0b192cd by Raymond Toy at 2020-08-27T00:36:28+00:00 Merge branch 'issue-85-opt-level-set-in-x86-config' into 'master' Fix #85: Let each x86 configs set optimization level Closes #85 See merge request cmucl/cmucl!52 - - - - - 4b80a6e5 by Raymond Toy at 2020-08-26T23:26:12-07:00 Merge branch 'master' into issue-86-save-fpu-state-on-entry-to-alloc - - - - - 0 changed files: Changes: View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/compare/d51dabf0f0c6868834ba5c… -- View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/compare/d51dabf0f0c6868834ba5c… You're receiving this email because of your account on gitlab.common-lisp.net.

1 0

[Git][cmucl/cmucl] Pushed new branch issue-86-save-fpu-state-on-entry-to-alloc
by Raymond Toy 27 Aug '20

27 Aug '20

Raymond Toy pushed new branch issue-86-save-fpu-state-on-entry-to-alloc at cmucl / cmucl -- View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/tree/issue-86-save-fpu-state-o… You're receiving this email because of your account on gitlab.common-lisp.net.

1 0

[Git][cmucl/cmucl][master] 2 commits: Fix #85: Let each x86 configs set optimization level
by Raymond Toy 27 Aug '20

27 Aug '20

Raymond Toy pushed to branch master at cmucl / cmucl Commits: 29cac208 by Raymond Toy at 2020-08-27T00:36:27+00:00 Fix #85: Let each x86 configs set optimization level Add `COPT` variable in `Config.x86_common` to set the optimization level (defaulting to `-O2`). Then each `Config.x86` file can set `COPT` as desired if the default doesn't work. Thus, `Config.x86_linux` sets `COPT` to `-O1`, but others can use the default value. See issue #68. - - - - - d0b192cd by Raymond Toy at 2020-08-27T00:36:28+00:00 Merge branch 'issue-85-opt-level-set-in-x86-config' into 'master' Fix #85: Let each x86 configs set optimization level Closes #85 See merge request cmucl/cmucl!52 - - - - - 6 changed files: - src/lisp/Config.x86_common - src/lisp/Config.x86_darwin - src/lisp/Config.x86_linux - src/lisp/Config.x86_linux_clang - src/lisp/Config.x86_netbsd - src/lisp/Config.x86_solaris_sunc Changes: ===================================== src/lisp/Config.x86_common ===================================== @@ -45,10 +45,11 @@ endif CPPFLAGS := $(CPP_DEFINE_OPTIONS) $(CPP_INCLUDE_OPTIONS) CFLAGS += -Wstrict-prototypes -Wall -g -fno-omit-frame-pointer -# gcc 8.1.1 and 8.3.1 (and probably anything after 8.1.1?) won't -# produce a working lisp with -O2. Just use -O1. -CFLAGS += -O1 -ASFLAGS = -g +# Default optimization level. This can be changed in the individual +# configs. +COPT = -O2 + +ASFLAGS = -g ASSEM_SRC = x86-assem.S ARCH_SRC = x86-arch.c ===================================== src/lisp/Config.x86_darwin ===================================== @@ -6,6 +6,7 @@ include Config.x86_common # you have the SDK available. MIN_VER = -mmacosx-version-min=10.6 +CFLAGS += $(COPT) CPPFLAGS += -DDARWIN $(MIN_VER) -m32 CFLAGS += -g3 -mtune=generic ASFLAGS += -g3 $(MIN_VER) ===================================== src/lisp/Config.x86_linux ===================================== @@ -1,6 +1,10 @@ # -*- Mode: makefile -*- include Config.x86_common +# gcc 8.1.1 and 8.3.1 (and probably anything after 8.1.1?) won't +# produce a working lisp with -O2. Just use -O1. +COPT = -O1 +CFLAGS += $(COPT) CPPFLAGS += -m32 -D__NO_CTYPE -D_GNU_SOURCE CFLAGS += -rdynamic -march=pentium4 -mfpmath=sse -mtune=generic ===================================== src/lisp/Config.x86_linux_clang ===================================== @@ -3,6 +3,7 @@ include Config.x86_common CC = clang CPPFLAGS += -m32 -D__NO_CTYPE -D_GNU_SOURCE +CFLAGS += $(COPT) CFLAGS += -march=pentium4 -mfpmath=sse -mtune=generic UNDEFSYMPATTERN = -Xlinker -u -Xlinker & ===================================== src/lisp/Config.x86_netbsd ===================================== @@ -1,6 +1,7 @@ # -*- Mode: makefile -*- include Config.x86_common +CFLAGS += $(COPT) CPPFLAGS += -march=pentium4 -mfpmath=sse UNDEFSYMPATTERN = -Xlinker -u -Xlinker & ===================================== src/lisp/Config.x86_solaris_sunc ===================================== @@ -2,6 +2,7 @@ include Config.sparc_common CC = cc -xlibmieee -g +CFLAGS += $(COPT) CFLAGS += -Di386 CPP = cc -E DEPEND_FLAGS = -xM1 View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/compare/d1c5289eb069df2ecdbac3… -- View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/compare/d1c5289eb069df2ecdbac3… You're receiving this email because of your account on gitlab.common-lisp.net.

1 0

[Git][cmucl/cmucl][issue-85-opt-level-set-in-x86-config] 2 commits: Fix typo
by Raymond Toy 27 Aug '20

27 Aug '20

Raymond Toy pushed to branch issue-85-opt-level-set-in-x86-config at cmucl / cmucl Commits: d1c5289e by Raymond Toy at 2020-08-26T17:12:40-07:00 Fix typo - - - - - 38372fd9 by Raymond Toy at 2020-08-26T17:14:30-07:00 Fix typo - - - - - 1 changed file: - src/lisp/Config.x86_freebsd Changes: ===================================== src/lisp/Config.x86_freebsd ===================================== @@ -3,7 +3,7 @@ include Config.x86_common # Set the path to your verison of GCC here. CC = gcc -m32 -CFLAGS += -O2 +CFLAGS += $(COPT) CPPFLAGS += -march=pentium4 -mfpmath=sse UNDEFSYMPATTERN = -Xlinker -u -Xlinker & View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/compare/449f8ec10cd560b5a1deab… -- View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/compare/449f8ec10cd560b5a1deab… You're receiving this email because of your account on gitlab.common-lisp.net.

1 0

[Git][cmucl/cmucl][master] Fix typo
by Raymond Toy 27 Aug '20

27 Aug '20

Raymond Toy pushed to branch master at cmucl / cmucl Commits: d1c5289e by Raymond Toy at 2020-08-26T17:12:40-07:00 Fix typo - - - - - 1 changed file: - src/lisp/Config.x86_freebsd Changes: ===================================== src/lisp/Config.x86_freebsd ===================================== @@ -3,6 +3,7 @@ include Config.x86_common # Set the path to your verison of GCC here. CC = gcc -m32 +CFLAGS += $(COPT) CPPFLAGS += -march=pentium4 -mfpmath=sse UNDEFSYMPATTERN = -Xlinker -u -Xlinker & View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/commit/d1c5289eb069df2ecdbac38… -- View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/commit/d1c5289eb069df2ecdbac38… You're receiving this email because of your account on gitlab.common-lisp.net.

1 0

[Git][cmucl/cmucl] Pushed new branch issue-85-opt-level-set-in-x86-config
by Raymond Toy 27 Aug '20

27 Aug '20

Raymond Toy pushed new branch issue-85-opt-level-set-in-x86-config at cmucl / cmucl -- View it on GitLab: https://gitlab.common-lisp.net/cmucl/cmucl/-/tree/issue-85-opt-level-set-in… You're receiving this email because of your account on gitlab.common-lisp.net.

1 0

Jump to page:

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

cmucl-cvs