argument test failures in RISC-V with clang

[honggyukim]

I see some test failures in RISC-V as follows.

$ ./runtest.py -vdp -O0 -c clang '082|084|086|087'
        ...
t082_arg_many: diff result of clang -pg -O0
--- expect      2024-01-06 13:14:03.468976993 +0000
+++ result      2024-01-06 13:14:03.468976993 +0000
@@ -12,3 +12,3 @@
    } /* foo */
-   many(12, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144);
+   many(12, 1, 1, 2, 3, 5, 8, 13, -841301936, 1, 1, 2, 3);
    pass() {

082 arg_many            : NG
t084_arg_mixed: diff result of clang -pg -O0
--- expect      2024-01-06 13:14:03.408976075 +0000
+++ result      2024-01-06 13:14:03.408976075 +0000
@@ -4,3 +4,3 @@
    mixed_mul(-3.000000, 80000000000) = -240000000000;
-   mixed_div(4, -0.000002) = -2000000.000000;
+   mixed_div(4, 0.000000) = 0.000000;
    mixed_str("argument", 0.000000) = "return";

084 arg_mixed           : NG
t086_arg_stack: diff result of clang -pg -O0
--- expect      2024-01-06 13:14:03.478977146 +0000
+++ result      2024-01-06 13:14:03.478977146 +0000
@@ -12,3 +12,3 @@
    } /* foo */
-   many(8, 13, 21, 34, 55, 89, 144);
+   many(-953724848, 1, 1, 2, 3, 5, 8);
    pass() {

086 arg_stack           : NG
t087_arg_variadic: diff result of clang -pg -O0
--- expect      2024-01-06 13:14:03.428976381 +0000
+++ result      2024-01-06 13:14:03.428976381 +0000
@@ -1,3 +1,3 @@
 main() {
-   variadic("print %c %s %d %ld %lu %lld %f", 'a', "hello", 100, 1234, 5678, 9876543210, 3.141592) {
+   variadic("print %c %s %d %ld %lu %lld %f", 'a', "hello", 100, 1234, 5678, 9876543210, 0.000000) {
      vsnprintf(256, "print %c %s %d %ld %lu %lld %f");

087 arg_variadic        : NG

[honggyukim]

We know that the argument tests are broken in gcc, but it should be fine for clang.

[namhyung]

It seems arguments passed through stack are broken. I've looked at the binary and found that it adjusted stack pointers to save register values before calling _mcount. It'd be hard to know the exact location since it varies on each function.

I think we can only support stack arguments on (full) dynamic tracing.

[honggyukim]

it adjusted stack pointers to save register values before calling _mcount.

Do you mean the sp is adjusted as I marked as NOTE in the following code snippet?

00000000000107cc <bar>:
   107cc:       7179                    addi    sp,sp,-48               # NOTE: sp modified before mcount
   107ce:       f406                    sd      ra,40(sp)
   107d0:       f022                    sd      s0,32(sp)
   107d2:       1800                    addi    s0,sp,48
   107d4:       fcb43c23                sd      a1,-40(s0)
   107d8:       fca43823                sd      a0,-48(s0)
   107dc:       ef5ff0ef                jal     ra,106d0 <_mcount@plt>
        ...
0000000000010840 <foo>:
   10840:       7139                    addi    sp,sp,-64               # NOTE: sp modified before mcount
   10842:       fc06                    sd      ra,56(sp)
   10844:       f822                    sd      s0,48(sp)
   10846:       0080                    addi    s0,sp,64
   10848:       fca43823                sd      a0,-48(s0)
   1084c:       e85ff0ef                jal     ra,106d0 <_mcount@plt>
        ...
00000000000108ea <many>:
   108ea:       7159                    addi    sp,sp,-112              # NOTE: sp modified before mcount
   108ec:       f406                    sd      ra,40(sp)
   108ee:       f022                    sd      s0,32(sp)
   108f0:       1800                    addi    s0,sp,48
   108f2:       fca43823                sd      a0,-48(s0)
   108f6:       03143c23                sd      a7,56(s0)
   108fa:       03043823                sd      a6,48(s0)
   108fe:       f41c                    sd      a5,40(s0)
   10900:       f018                    sd      a4,32(s0)
   10902:       ec14                    sd      a3,24(s0)
   10904:       e810                    sd      a2,16(s0)
   10906:       e40c                    sd      a1,8(s0)
   10908:       dc9ff0ef                jal     ra,106d0 <_mcount@plt>
        ...
0000000000010990 <check>:
   10990:       7179                    addi    sp,sp,-48               # NOTE: sp modified before mcount
   10992:       f406                    sd      ra,40(sp)
   10994:       f022                    sd      s0,32(sp)
   10996:       1800                    addi    s0,sp,48
   10998:       fcb43823                sd      a1,-48(s0)
   1099c:       fca43c23                sd      a0,-40(s0)
   109a0:       d31ff0ef                jal     ra,106d0 <_mcount@plt>
        ...

Then you're right. There is no way to handle stack pushed arguments in this binary structure unless we disassemble the function prologue.

[honggyukim]

cc: @gichoel

[gichoel]

Based on the code snippet above, I understand the following two issues:
1. sp was manipulated before mcount() was called.
2. Each function manipulates sp using different values.

mcount.S reads values from RISC-V's argument registers, and I don't understand why this would be a problem with argument handling.

---

Apart from the above problem, I was analyzing the operations after the mcount_entry() function call.

Maybe when I am done with that, I will be able to understand the problem.