upmem / llvm Goto Github PK

@@ From Tuleap issue 1317, submitted by rjodin

When having register argument in an asm inline, we need a double % to specify a register ("move %%r18, saved_ktrace_args_in_registers")
But when not having register arguments, if we have a double %%, the asm inline is ignore. Is that case we need a single % to specify a register...

@@ Comments:
** @dfurodet ça marche?

The following code crashes clang:

#include <stdlib.h>
 #include <mram.h>

 #define TASKLETS_INITIALIZER TASKLET(main, 1024, 0)
 #include <rt.h>

 char __mram_keep toto[255];

 int main(void) {
     *((char *) toto) = 0x2a;

     return EXIT_SUCCESS;
 }

Because of the line *((char *) toto) = 0x2a;.

Clash report:

fatal error: error in backend: Cannot select: 0x563c48a23810: i32 = addrspacecast[255 -> 0] 0x563c48a23948
0x563c48a23948: i32 = DPUISD::Wrapper TargetGlobalAddress:i32<[255 x i8] addrspace(255)* @toto> 0
0x563c48a23740: i32 = TargetGlobalAddress<[255 x i8] addrspace(255)* @toto> 0
In function: main
clang: error: clang frontend command failed with exit code 70 (use -v to see invocation)
clang version 7.1.0 ([email protected]:upmem/clang.git 8e2c3063c7ce02f2b04eb24723e66f6d2d1dcf9f) ([email protected]:upmem/llvm.git 9dce7c1)
Target: dpu-upmem-dpurte
Thread model: posix
InstalledDir: /home/jchauzi/package-sdk-current/upmem-internal/usr/bin

@@ From Tuleap issue 701, submitted by jchauzi

https://dpu.dev/z/a5VR62

@@ Comments:
** First move and load could be merge as one load (load %r21, %zero, my_table)

@@ From Tuleap issue 1345, submitted by jchauzi

Example:

long attribute ((section(".mram.data"), address_space(255))) my_integer;

long foo() {
return my_integer;
}

We are supposed to use the soft cache mechanism to load my_integer (works when my_integer is a char, short or int), but it is not in that case and we fail to load what we wanted...

When compiling the following code :

#include <stdint.h>

uint32_t dup(uint8_t c) {
    return c + (c << 8);
}

I'm getting the following sequence:

lsl r1, r0, 8
or r0, r0, r1

which could be optimized in a single lsl_add

At the end of the day, I'm interested in getting a more optimized sequence for :

uint32_t memset32(int c) {
    uint8_t v = c;
    return v | (v << 8) | (v << 16) | (v << 24) ;
}

@@ From Tuleap issue 1317, submitted by rjodin

When having register argument in an asm inline, we need a double % to specify a register ("move %%r18, saved_ktrace_args_in_registers")
But when not having register arguments, if we have a double %%, the asm inline is ignore. Is that case we need a single % to specify a register...

@@ Comments:
** @dfurodet ça marche?

https://dpu.dev/z/0-Jytu

Ligne 38:

• in -g0: move r14, r18, true, .LBB0_4
• in -g3: move r14, r18; jump .LBB0_4

It seems like our MergeInstrPass is not correctly executed with debug symbols enabled.