clang crashes on the aws-crt-python project

[yurivict]

Project: https://github.com/awslabs/aws-crt-python
OS: FreeBSD 14-CURRENT

clang-crash.tgz

[MaskRay]

In aws_checksums_crc32c_hw, __asm__("loop_small_%=: CRC32B (%[in]), %[crc]" : "+c"(crc) : [ crc ] "c"(crc), [ in ] "r"(input));
causes TwoAddressInstructionPass.cpp:1493 to trigger an assertion failure.

DstIdx == 3 (killed %107:gr32(tied-def 3))
i == 5
MI->getOperand(3).getReg() == MI->getOperand(5).getReg() 

INLINEASM &"loop_small_${:uid}: CRC32B ($2), $1" [attdialect], $0:[regdef], implicit-def $ecx, $1:[reguse], killed $ecx, $2:[reguse:GR64], killed %106:gr64, $3:[reguse tiedto:$0], killed %107:gr32(tied-def 3), $4:[clobber], implicit-def dead early-clobber $df, $5:[clobber], implicit-def early-clobber $fpsw, $6:[clobber], implicit-def dead early-clobber $eflags, !8

Still investigating.

---

The : "+c"(crc) : [crc]"+c"(crc) pattern is suspicious. The following patch helps a bit, but Clang still crashes in one INLINEASM instruction.

@@ -6669 +6670 @@
-        : "+c"(crc)
+        : [crc]"+c"(crc)
@@ -6672 +6673 @@
-        : [ crc ] "c"(crc), [ in ] "d"(input)
+        : [ in ] "d"(input)
@@ -6755 +6756 @@
-        : "+c"(crc), "+d"(input)
+        : [crc]"+c"(crc), "+d"(input)
@@ -6759 +6760 @@
-        : [ crc ] "c"(crc), [ in ] "d"(input)
+        : [ in ] "d"(input)
@@ -6835 +6836 @@
-        : "+c"(crc), "+d"(input)
+        : [crc]"+c"(crc), "+d"(input)
@@ -6839 +6840 @@
-        : [ crc ] "c"(crc), [ in ] "d"(input)
+        : [ in ] "d"(input)
@@ -6871,0 +6873 @@
+      uint32_t crc1 = 0;
@@ -6873 +6875 @@
-            __asm__("loop_small_%=: CRC32B (%[in]), %[crc]" : "+c"(crc) : [ crc ] "c"(crc), [ in ] "r"(input));
+            __asm__("loop_small_%=: CRC32B (%[in]), %[crc]" : [crc] "+c"(crc) : [ in ] "r"(input));
@@ -6890 +6892 @@
-        __asm__("loop_leading_%=: CRC32B (%[in]), %[crc]" : "+c"(crc) : [ crc ] "c"(crc), [ in ] "r"(input));
+        __asm__("loop_leading_%=: CRC32B (%[in]), %[crc]" : [ crc ] "+c"(crc) : [ in ] "r"(input));
@@ -6920 +6922 @@
-        __asm__ __volatile__("loop_8_%=: CRC32Q (%[in]), %%rcx" : "+c"(crc) : [ crc ] "c"(crc), [ in ] "r"(input));
+        __asm__ __volatile__("loop_8_%=: CRC32Q (%[in]), %%rcx" : [crc]"+c"(crc) : [ in ] "r"(input));
@@ -6928,2 +6930 @@
-                             : "+c"(crc)
-                             : [ crc ] "c"(crc), [ in ] "r"(input));
+                             : [crc]"+c"(crc) : [ in ] "r"(input));

[DimitryAndric]

There are a lot of duplicates of this bug, and it's always some Amazon AWS code, e.g.:

• https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=234232 (2018)
• Assertion failed: (i == DstIdx || !MI->getOperand(i).isReg() || MI->getOperand(i).getReg() != RegA), function processTiedPairs, #40330 (originally bugzilla, filed for the above FreeBSD bug)
• Assertion failed in function processTiedPairs #39643 (even older bug from the FreeBSD aws-checksums port)
• Assertion failed TwoAddressInstructionPass.cpp:1492: i == DstIdx || !MI->getOperand(i).isReg() || MI->getOperand(i).getReg() != RegA' #54957
• clang: error: clang frontend command failed due to signal: gcc style inline asm #14802
• CodeGen assert on invalid inline assembly #9878

I think AWS is using some common CRC inline assembly that causes either assertions or backend errors with clang. Probably some impossible register constraint, which instead of generating a normal diagnostic causes an assertion failure or backend failure...

[llvmbot]

@llvm/issue-subscribers-backend-x86

Project: https://github.com/awslabs/aws-crt-python OS: FreeBSD 14-CURRENT

clang-crash.tgz