Fix: XCHAL_HAVE_ADDX config is virtually ineffective against GCC 10.1 or later

[jjsuwa-sys3175]

See esp8266/Arduino#7742

[earlephilhower]

This looks very nice, thanks! Did you build GCC with this machine definition change and verify it works, or do you need me to run a build pass on my server and try some trivial examples?

[jjsuwa-sys3175]

Did you build GCC with this machine definition change and verify it works, or do you need me to run a build pass on my server and try some trivial examples?

partially built (C compiler only, with neither C++ nor libraries), and the example:

/* ADDX[248]/SUBX[248] */
unsigned int test_addx2(unsigned int a) {
    return a * 3;
}
unsigned int test_addx4(unsigned int a) {
    return a * 5;
}
unsigned int test_addx8(unsigned int a) {
    return a * 9;
}
unsigned int test_subx2(unsigned int a, unsigned int b) {
    return a * 2 - b;
}
unsigned int test_subx4(unsigned int a, unsigned int b) {
    return a * 4 - b;
}
unsigned int test_subx8(unsigned int a) {
    return a * 7;
}

will be compiled to:

	.file	"test.c"
	.text
	.literal_position
	.align	4
	.global	test_addx2
	.type	test_addx2, @function
test_addx2:
	addx2	a2, a2, a2
	ret.n
	.size	test_addx2, .-test_addx2
	.literal_position
	.align	4
	.global	test_addx4
	.type	test_addx4, @function
test_addx4:
	addx4	a2, a2, a2
	ret.n
	.size	test_addx4, .-test_addx4
	.literal_position
	.align	4
	.global	test_addx8
	.type	test_addx8, @function
test_addx8:
	addx8	a2, a2, a2
	ret.n
	.size	test_addx8, .-test_addx8
	.literal_position
	.align	4
	.global	test_subx2
	.type	test_subx2, @function
test_subx2:
	subx2	a2, a2, a3
	ret.n
	.size	test_subx2, .-test_subx2
	.literal_position
	.align	4
	.global	test_subx4
	.type	test_subx4, @function
test_subx4:
	subx4	a2, a2, a3
	ret.n
	.size	test_subx4, .-test_subx4
	.literal_position
	.align	4
	.global	test_subx8
	.type	test_subx8, @function
test_subx8:
	subx8	a2, a2, a2
	ret.n
	.size	test_subx8, .-test_subx8
	.ident	"GCC: (GNU) 10.2.0"

[cmdline: ./gcc/xgcc -B./gcc/ -Os -fno-inline-functions -mlongcalls -mtext-section-literals -Wextra -S test.c ]

[earlephilhower]

-edit- Got the labels backwards. d'oh!

I just applied all 3 of your patches and rebuilt the chain, then did testing using my audio library (MP3 decoder).

Patched (all 3) Gcc10.2 + newlib 4.0:
Orig Gcc10.2 + newlib 4.0:

Executable segment sizes:
IROM   : 413020          - code in flash         (default or ICACHE_FLASH_ATTR) 
IRAM   : 28437   / 32768 - code in IRAM          (ICACHE_RAM_ATTR, ISRs...) 
DATA   : 1612  )         - initialized variables (global, static) in RAM/HEAP 
RODATA : 1420  ) / 81920 - constants             (global, static) in RAM/HEAP 
BSS    : 25888 )         - zeroed variables      (global, static) in RAM/HEAP 

Orig Gcc10.2 + newlib 4.0:
Patched (all 3) Gcc10.2 + newlib 4.0:

Executable segment sizes:
IROM   : 414060          - code in flash         (default or ICACHE_FLASH_ATTR) 
IRAM   : 28437   / 32768 - code in IRAM          (ICACHE_RAM_ATTR, ISRs...) 
DATA   : 1612  )         - initialized variables (global, static) in RAM/HEAP 
RODATA : 1420  ) / 81920 - constants             (global, static) in RAM/HEAP 
BSS    : 25888 )         - zeroed variables      (global, static) in RAM/HEAP 

While it's not dramatic, it is a decrease and I'm not finding any problems. Than you for the patches! I'll apply them w/the newlib 4 PR.

[jjsuwa-sys3175]

Thanks for merging @earlephilhower, i'll download that.

FYI: adding -free -fipa-pta to GCC options for the building target (Xtensa) saves a bit of IROM size: see arendst/Tasmota#9749 and arendst/Tasmota@e7cff92.

 # Environment variables for configure and building targets.  Only use $(call setenv,$@)
 ifeq ($(LTO),true)
-    CFFT := "-mlongcalls -flto -Wl,-flto -Os -g"
+    CFFT := "-mlongcalls -flto -Wl,-flto -Os -g -free -fipa-pta"
 else ifeq ($(LTO),false)
-    CFFT := "-mlongcalls -Os -g"
+    CFFT := "-mlongcalls -Os -g -free -fipa-pta"
 else
     $(error Need to specify LTO={true,false} on the command line)
 endif

-compiler.c.flags=-c {compiler.warning_flags} -std=gnu17 {build.stacksmash_flags} -Os -g -Wpointer-arith -Wno-implicit-function-declaration -Wl,-EL -fno-inline-functions -nostdlib -mlongcalls -mtext-section-literals -falign-functions=4 -MMD -ffunction-sections -fdata-sections {build.exception_flags} {build.sslflags} {build.waveform} {build.mmuflags} {build.non32xferflags}
+compiler.c.flags=-c {compiler.warning_flags} -std=gnu17 {build.stacksmash_flags} -Os -g -free -fipa-pta -Wpointer-arith -Wno-implicit-function-declaration -Wl,-EL -fno-inline-functions -nostdlib -mlongcalls -mtext-section-literals -falign-functions=4 -MMD -ffunction-sections -fdata-sections {build.exception_flags} {build.sslflags} {build.waveform} {build.mmuflags} {build.non32xferflags}

-compiler.cpp.flags=-c {compiler.warning_flags} {build.stacksmash_flags} -Os -g -mlongcalls -mtext-section-literals -fno-rtti -falign-functions=4 {build.stdcpp_level} -MMD -ffunction-sections -fdata-sections {build.exception_flags} {build.sslflags} {build.waveform} {build.mmuflags} {build.non32xferflags}
+compiler.cpp.flags=-c {compiler.warning_flags} {build.stacksmash_flags} -Os -g -free -fipa-pta -mlongcalls -mtext-section-literals -fno-rtti -falign-functions=4 {build.stdcpp_level} -MMD -ffunction-sections -fdata-sections {build.exception_flags} {build.sslflags} {build.waveform} {build.mmuflags} {build.non32xferflags}

[igrr]

@jcmvbkbc do you think this change should be applied upstream?

[jcmvbkbc]

@igrr I guess. I'm also curious what happened that the optimization that worked in 4.8 doesn't work any more, I'll have a look.

[jcmvbkbc]

Done.
I've fixed the original predicates and patterns instead of duplicating them.

[jjsuwa-sys3175]

thx for upstreaming @jcmvbkbc.

ps. how do you like the other two; #20 and #21?

[jcmvbkbc]

@jjsuwa-sys3175 Let me take a look. Also feel free to ping/cc me on future xtensa port improvements.