Opt seems to report fewer stats when splitting the compilation to the several stages, rather than using clang


I am trying to recreate the compilation of using clang and -O2 by splitting the compilation to the several stages using the several llvm tools (clang as front-end, opt, llc…) but the optimiser seems to behave differently if called individually than when called by clang. I am using the latest llvm version. So I am calling the front end using:

clang -mllvm -stats -mllvm -debug-pass=Structure -target arm-none-eabi -isystem /usr/local/lpcxpresso_8.2.2_650/lpcxpresso/tools/arm-none-eabi/include -DHAVE_CONFIG_H -I. -I…/… -Wall -I …/…/i^Clude -I …/…/config/arm/boards/lpcxpresso_1769_xpresso/ -I …/…/config/arm/chips/lpc175x_6x/ “-DCALIB_SCALE=0” --static -mthumb -mcpu=cortex-m3 -D__USE_LPCOPEN -D__CODE_RED -D__REDLIB__ -DCHIP_LPC175X_6X -DCORE_M3 -ffunction-sections -fdata-sections -O2 -c -emit-llvm -o adpcm_optimized.bc adpcm.c

this also gives me the optimizer statistics and the structure of calling the passes.

then I am using

clang -mllvm -stats -mllvm -debug-pass=Structure -target arm-none-eabi -isystem /usr/local/lpcxpresso_8.2.2_650/lpcxpresso/tools/arm-none-eabi/include -DHAVE_CONFIG_H -I. -I…/… -Wall -I …/…/i^Clude -I …/…/config/arm/boards/lpcxpresso_1769_xpresso/ -I …/…/config/arm/chips/lpc175x_6x/ “-DCALIB_SCALE=0” --static -mthumb -mcpu=cortex-m3 -D__USE_LPCOPEN -D__CODE_RED -D__REDLIB__ -DCHIP_LPC175X_6X -DCORE_M3 -ffunction-sections -fdata-sections -O0 -c -emit-llvm -o adpcm_no_opt.bc adpcm.c

and then

opt -stats -O2 -S -debug-pass=Structure adpcm_no_opt.bc -o adpcm.ll

I was expecting the stats coming out of the opt call to be the same as calling clang with -O2. Instead they are very little:

18 basicaa - Number of times a GEP is decomposed

1 cgscc-passmgr - Maximum CGSCCPassMgr iterations on one SCC

104 globalopt - Number of globals marked unnamed_addr

6 globalsmodref-aa - Number of global vars without address taken

Am I doing something wrong or missing something? The Structure of calling the passes is still the same, and also the emitted code from the clang with -O2 and the opt with -O2 seems to be the same. So why is the optimiser called individually not reporting the full stats?

Thank you,


You need to invoke clang with -O2 -Xclang -disable-llvm-optzns instead of -O0

-O0 implies to clang to mark all functions as optnone in the generated IR

Thank you! This solved my issue!