diff options
| author | Zheng Bao <fishbaozi@gmail.com> | 2022-08-03 17:57:47 +0800 |
|---|---|---|
| committer | Martin L Roth <gaumless@gmail.com> | 2022-10-12 16:55:58 +0000 |
| commit | d91f3a4eaf7f8399db83a9c2d5cba2452fefaf7a (patch) | |
| tree | fa32b8e3be93cf9c60bfa78e1f9f233c429d2d5d | |
| parent | dc522d22023a7ceb622e1f866a9a71c79343ae2c (diff) | |
lib/lzma: Build the source for decompression with flag -Ofast
The decompression is critical for speed of boot. So we sacrifice some
generated code size to optimize for speed.
This change speeds up the LZMA decompression between 3% and 6% at a
cost of just over 2k of additional code space.
BUG=b:223985641
TEST=Majolica
The test is done on Majolica and the result is listed below.
Time saved:
We tested the boot time with each flag for 10 times. The duration of
each decompression process is listed as below.
Load FSP-M Load ramstage Load payload
Ofast Os Ofast Os Ofast Os
------------------------------------------
62543 62959 20585 22458 9945 10626
62548 62967 20587 22461 9951 10637
62560 62980 20588 22478 9951 10641
62561 62988 20596 22478 9954 10643
62569 62993 20596 22479 9954 10643
62574 63000 20605 22492 9958 10647
62575 63026 20615 22495 9959 10647
62576 63038 20743 22614 9960 10647
62587 63044 20758 22625 9961 10647
62592 63045 20769 22637 9961 10647
-----------------------------------------
average 62568 63004 20644 22521 9955 10642
(unit: microseconds)
Size sacrificed:
The size of object file with -Os:
./build/ramstage/lib/lzmadecode.o: file format elf32-i386
4 .text.LzmaDecode 00000d84 00000000 00000000 00000076 2**0
CONTENTS, ALLOC, LOAD, READONLY, CODE
The size of object file with -Ofast:
./build/ramstage/lib/lzmadecode.o: file format elf32-i386
4 .text.LzmaDecode 00001719 00000000 00000000 00000080 2**4
CONTENTS, ALLOC, LOAD, READONLY, CODE
(Output by running "objdump -h ./build/ramstage/lib/lzmadecode.o")
We can see that size is increased from 3460 bytes to 5913 bytes, a
change of 2453 bytes or 171%.
Change-Id: Ie003164e2e93ba8ed3ccd207f3af31c6acf1c5e2
Signed-off-by: Zheng Bao <fishbaozi@gmail.com>
Reviewed-on: https://review.coreboot.org/c/coreboot/+/66392
Tested-by: build bot (Jenkins) <no-reply@coreboot.org>
Reviewed-by: Martin Roth <martin.roth@amd.corp-partner.google.com>
| -rw-r--r-- | src/lib/Kconfig | 7 | ||||
| -rw-r--r-- | src/lib/lzmadecode.c | 7 |
2 files changed, 14 insertions, 0 deletions
diff --git a/src/lib/Kconfig b/src/lib/Kconfig index 014230012d..ae96fc61ee 100644 --- a/src/lib/Kconfig +++ b/src/lib/Kconfig @@ -114,3 +114,10 @@ config CBFS_PRELOAD in the background before they are actually required. This feature depends on the read-only boot_device having a DMA controller to perform the background transfer. + +config DECOMPRESS_OFAST + bool + depends on COMPILER_GCC + default y + help + Compile the decompressing function in -Ofast instead of standard -Os diff --git a/src/lib/lzmadecode.c b/src/lib/lzmadecode.c index c45e131708..cb868290aa 100644 --- a/src/lib/lzmadecode.c +++ b/src/lib/lzmadecode.c @@ -19,6 +19,12 @@ to this file, however, are subject to the LGPL or CPL terms. */ +#if CONFIG(DECOMPRESS_OFAST) + #define __lzma_attribute_Ofast__ __attribute__((optimize("Ofast"))) +#else + #define __lzma_attribute_Ofast__ +#endif + #include "lzmadecode.h" #include <types.h> @@ -183,6 +189,7 @@ int LzmaDecodeProperties(CLzmaProperties *propsRes, #define kLzmaStreamWasFinishedId (-1) +__lzma_attribute_Ofast__ int LzmaDecode(CLzmaDecoderState *vs, const unsigned char *inStream, SizeT inSize, SizeT *inSizeProcessed, unsigned char *outStream, SizeT outSize, SizeT *outSizeProcessed) |