C2000 C/C++ Code Generation Tools v20.2.1.LTS Release Notes

Long-term Support Release
Compiler Downloads and Documentation
TI E2E Community
Defect Tracking Database
New Features:
Resolved defects
Known defects

Long-term Support Release

The C2000 CGT v20.2.0.LTS release is a long term support (LTS) release. This release will be supported for roughly 2 years with periodic bug fix updates.

Compiler Downloads and Documentation

Documentation for the “TI C2000 Optimizing Compiler User’s Guide” and the “TI C2000 Assembly Language User’s Guide” is available online at:

https://www.ti.com/tool/C2000-CGT

TI E2E Community

Questions concerning TI Code Generation Tools can be posted to the TI E2E Community forums. The “Development Tools” forum can be found at:

https://e2e.ti.com/support/tools

Defect Tracking Database

Compiler defect reports can be tracked at the Development Tools bug database at:

https://sir.ext.ti.com/

New Features:

Integer Division Unit (HWINTDIV) updates

The Integer Division Unit feature support has below updates. Feature enabled with --idiv_support=idiv0 and requires --abi=eabi and --float_support=fpu32/fpu64.

Improved performance for several HWINTDIV routines: In 18.12.0.LTS release, the initial HWINTDIV implementation had inefficiencies with directly loading from memory to fpu registers affecting almost all HWINTDIV sequences. The fpu register loads were being inefficiently loaded initially to non-fpu registers before final load to fpu registers resulting in unecessary register-to-register moves along with additional NOPs.

This was fixed in 19.6.0.STS
Improved performance for several HWINTDIV routines: The original HWINTDIV implementation in 18.12.0.LTS did not implement the most optimal sequences for i64/i32, i64/u32, and u64/u32 which instead leveraged slower 64/64 sequences.

Implemented the optimial HWINTDIV sequences for:
```
 i64/i32 traditional, euclidean, and modulo
 i64/u32 traditional, euclidean, and modulo
 u64/u32 traditional 
```
The original implementation had all intrinsics returning either ldiv_t, or lldiv_t (and unsigned equivalents). However, below intrinsics:
```
 i64/i32 traditional, euclidean, and modulo
 i64/u32 euclidean, and modulo (not traditional)
 u64/u32 traditional 
```
return llong quotient with long remainder which required updates to stdlib.h to include below new return types:
```
 typedef struct { long long quot; long rem; } __llldiv_t;
 typedef struct { unsigned long long quot; unsigned long rem; } __ullldiv_t; 
```
Assembler pipeline detection check HWINTDIV instructions cannot be placed in the delay slots for any FPU multicycle instructions. Assembler pipeline checks have been added to error if this is detected.
Build attributes Below HWINTDIV build attributes were added. View build attribute settings using the object file dump tool: ofd2000
```
 Tag_IDIV 1 (IDIV 0) 
```

Predefined macros for assembler/linker Below HWINTDIV predefined macros were added for use in assembling/linking:

 assembler: .TMS320C2800_IDIV set to 1 if --idiv_support=idiv0
 linker: __TMS320C28XX_IDIV__ set to 1 if --idiv_support=idiv0

Support inline memcpy operations larger than 255 words

Inlining of memcpy for operations less than or equal 255 words is already supported with use of the instruction: RPT #imm

Support has been added for inlining memcpy of operations larger than 255 words by using the RPT instruction with register operand.

Both memcpy inlining uses of RPT can be disabled with --rpt_threshold

FPU64 Run Time Support (RTS) library routine updates

Trig float routines: Below RTS library routines had implementations that included increased 32-bit precision if --float_support=fpu64 was available, however, these routines have been updated to use the same implementation for both fpu32 or fpu64 modes:
```
 sinf(float) cosf(float) tanf(float) cbrtf(float) 
```
For improved floating point precision with fpu64 mode, use the double routines.
sqrt(double): The RTS math library sqrtf(float) routine included use of the EISQRTF32 instruction if --float_support=fpu32 was enabled.
Added similar support for sqrt(double) to use the EISQRTF64 instruction if --float_support=fpu64 is enabled.

New diagnostic checks

Linker/assembler diagnostic for blocked vs non-blocked data access: New linker/assembler diagnostic checks were added to detect when data accesses that assume data is blocked are made to non-blocked definitions. This was for Jira feature ticket CODEGEN-1407.

For more background on data blocking see:
https://e2e.ti.com/blogs_/archives/b/toolsinsider/archive/2016/11/17/data-blocking-in-the-c2000-mcu-compiler-explained?keyMatch=data%20blocking&tisearch=Search-EN-Everything

The common case for this is a global CLA definition that is used in C28 code. However, the diagnostic is not limited to CLA definitions and instead more generically checks for blocked accesses with non-blocked definitions.

COFF and EABI use different implementations. See spru514 Compiler User guide for details on COFF vs EABI data page optimizations and data blocking.

COFF:
The COFF implementation is based on leveraging an unused symbol type.

EABI:
The EABI implementation added a new relocation to indicate a blocked access relocation. The relocation is exactly the same but with suffix _BLKD

Both COFF/EABI implementations use a new $BLOCKED() built-in function in assembly for the compiler to mark accesses as blocked during code generation.

Example assembler warning:
```
 “test_blocked.asm”, WARNING! at line 59: [W0000] Definition for
 symbol _sym_block is
 non-blocked while
 references indicate
 blocked access with
 $BLOCKED()
 operator.
 MOV @$BLOCKED(_sym_block)+2,#0
```
Example linker warning:
```
 warning: Symbol, “sym_blocked”, referenced in “test.obj”, assumes that data 
 is blocked but is accessing non-blocked data in “test2.obj”. Runtime
 failures may result. 
```
Linker diagnostic for function calls between CLA and C28 code A new linker diagnostic check was added to detect when C28 or CLA code branches to code instead compiled for the CLA or C28.

COFF and EABI use different implementations.

COFF:
The COFF implementation is based on leveraging an unused symbol type.

EABI:
The EABI implementation added a new relocation to indicate both direct and indirect function calls. The relocation is exactly the same but with suffix _BR

The more common case to detect is for direct function calls. eg:
```
 LCR #func 
```
For indirect function calls, both COFF/EABI implementations use a new CALL() built-in function in assembly for the compiler to mark address loads that are used in subsequent indirect-function calls. eg:
```
 MOVL XAR7, #CALL(func)
 LCR *XAR7 
```
Example linker warning:
```
 warning: Invalid to call C28 function “_foo_c28" defined in “test.o” from CLA file “test_cla.o”
 warning: Invalid to call CLA function “_foo_cla" defined in “test_cla.o” from C28 file “test.o” 
```
Additional Assembler diagnostics added for FPU64: The existing assembler diagnostics for detecting FPU64 instructions/operands not supported in FPU32 mode (–float_support=fpu32) have been updated to emit errors for below cases:
```
 Use of RaL operand with below in fpu32 mode:
 MOVDD32, MOVXI and MOV32
 Use of Ra operand with below in fpu32 mode:
 MOVI32 and MOVF32
 Use of below instructions in fpu32 mode:
 F32DTOF64 and MOVIZF64 
```
NOTE: SWAPF, ZERO and MOVIZ use the same opcode for both fpu32 and fpu64 and both operands are accepted in either mode.
On fpu32-only hardware, the instructions only operate on 32-bits. On fpu64 hardware, the instructions only operate on 64-bits.
Additional advice with --advice:performance for EABI with fpu32 mode: With recent release of EABI support, floating point doubles have changed from being 32-bits in COFF ABI, to 64-bits for EABI.
As a migration aid, --advice:performance=all will now additionally emit below advice if double floating point operations are detected when compiling for EABI with --float_support=fpu32:
```
 (Performance) EABI double precision is 64-bits as opposed to 32-bits for COFF. 
 Consider changing doubles to floats for improved performance in FPU32-mode. 
```
Above advice can be disabled with either of below options:
```
 --advice:performance=none or --float_operations_allowed=all 
```

Hex utility --cmac option, supported added for --boot tables

The hex conversion utility’s Cipher-based Message Authentication Protocol (CMAC) feature now supports boot tables with CMAC tags.

To enable, use --cmac=key_file with --boot along with other required boot table options. The CMAC algorithm assumes a fill value of 1’s for gaps between boot table regions which requires setting --fill.

Hex utility --cmac option (cipher-based message authentication protocol)

The hex conversion utility supports the secure flash boot capability provided by TMS320F2838x devices, which have both C28 and ARM cores. The secure flash boot applies the Cipher-based Message Authentication Protocol (CMAC) algorithm to verify CMAC tags for regions of allocated memory.

In order to apply the CMAC algorithm to the appropriate regions in allocated memory, use the hex conversion utility as follows: --cmac=file where file contains a 128-bit hex CMAC key

NOTE: above additionally requires --image or --load_image options.

CMAC tag locations need to be defined in the user’s C code using below symbols along with LOCATION and RETAIN pragmas: secure boot regions: cmac_sb_1 cmac_sb_2 cmac_sb_3 or cmac_sb_4 user-specified region: cmac_all

The hex utility will then apply the CMAC algorithm to all the CMAC tags defined in the user’s C code.

See TMS320C28x Assembly Language Tools User Guide for details.

VCU register save/restore support for interrupt service routines

In prior releases, interrupt service routines (ISR) did not save/restore VCU registers.

Below new option generates VCU register save/restore to stack for ISR’s so vcu code can be re-entrant. If an ISR interrupts a VCU computation then it will not impact results.

 --isr_save_vcu_regs=on,off

NOTE: VCU save/restore are only emitted for ISR’s that call other functions since leaf ISR routines will not have any VCU register usage as there is no current VCU code generation aside from below save/restore.

For --vcu_support=vcu0/vcu2, below are added to ISR’s:

 VMOV32 SP++,VCRC 
 VMOV32 SP++,VSTATUS
 …
 VMOV32 VSTATUS,–SP 
 VMOV32 VCRC,–SP

For --vcu_support=vcrc, below are added to ISR’s:

 VMOV32 SP++,VCRC 
 VMOV32 SP++,VSTATUS
 VMOV32 SP++,VCRCPOLY 
 VMOV32 SP++,VCRCSIZE 
 …
 VMOV32 VCRCSIZE,–SP 
 VMOV32 VCRCPOLY,–SP 
 VMOV32 VSTATUS,–SP 
 VMOV32 VCRC,–SP

Default file extension for compiler generated object files

The default file extensions for object files created by the compiler have changed in order to prevent conflicts when C and C++ files have the same names. Object files generated from C source files have the .c.obj extension. Object files generated from C++ source files have the .cpp.obj extension.

Resolved defects

Resolved defects in v20.2.x.LTS:

ID	Summary
CODEGEN-7297	C28 –advice:performance parser segfault for tmu advice for indirect function calls
CODEGEN-7146	Compiler loses conversion to size_t in length argument to memcpy
CODEGEN-7144	Documentation of linker options –rom_model and –ram_model is inconsistent
CODEGEN-7141	Simple division expression optimized into incorrect complicated expression
CODEGEN-7117	Assembly tools manual mixes up RAM and ROM model description
CODEGEN-7098	Incorrect optimization of expression that mixes the types long and short
CODEGEN-7011	Compiler intermittently fails with INTERNAL ERROR: opt2000 experienced a segmentation fault
CODEGEN-6911	Floating point comparison that fits min/max-index pattern can produce incorrect code
CODEGEN-6787	C28x compiler manual does not document that types larger than 16-bits are aligned on 2-word address boundaries
CODEGEN-6783	Ensure C std headers do not issue MISRA warnings
CODEGEN-6682	Combination of many include paths and .cdecls causes assembler to segmentation fault
CODEGEN-6666	64-bit to 32-bit assignment with predecrement causes compiler crash with INTERNAL ERROR: no match for ASG
CODEGEN-6477	Using –vcu_support may lead to INTERNAL ERROR: Register allocation failed
CODEGEN-6456	Under eabi, the cinit entry for zero fill an of uninitialized section has a length that is 1/2 of what it should be
CODEGEN-6354	C28 Compiler should emit performance advice 2614 for atanf, et al
CODEGEN-6289	Pointer to VLA 2D array computation in IF condition is double-adjusted and garbled
CODEGEN-6132	pdd tool fails when program includes unused file
CODEGEN-5843	Use of –check_misra=14.10 causes non-MISRA related warning to be emitted
CODEGEN-5563	Checking for MISRA rule 20.1 causes macro redefinition to be ignored
CODEGEN-1678	Document #pragma once

Known defects

Known defects in v20.2.x.LTS:

ID	Summary
CODEGEN-7354	C28x compiler shell cl2000 incorrectly claims to support –analyze=callgraph
CODEGEN-6509	Compiler error: no instance of function template matches the argument list
CODEGEN-6070	Erroneous “redeclared with incompatible type” involving two tagless structs with same form
CODEGEN-5179	When a symbol is remapped, DW_TAG_TI_branch/DW_AT_name is not updated
CODEGEN-5078	Simple syntax error causes assembler to fail with INTERNAL ERROR
CODEGEN-4985	Typo on ULP Advisor message
CODEGEN-4943	MISRA 12.2 is incorrectly issued when one variable is used multiple times as an array index in the same expression
CODEGEN-4934	Incorrect issue MISRA diagnostic 10.5: If the bitwise operators ~ and << are applied
CODEGEN-4789	Compiler incorrectly issues MISRA diagnostic 12.7 Bitwise operators shall not be applied to operands whose underlying type is signed
CODEGEN-4712	No MISRA 10.1 and 10.6 diagnostics issued on initializations
CODEGEN-4346	MISRA 19.4 error fails to identify itself as MISRA diagnostic when problem is on the command line
CODEGEN-4304	C++ feature causes runtime failure on unaligned access on ARM9
CODEGEN-4298	Internal error when passing a temporary array of objects
CODEGEN-4297	Cannot take the address of std::ctype<char>::table_size
CODEGEN-4296	Undefined symbol isblank with -g or -ooff
CODEGEN-4290	wstring runtime failure - likely bug in swprintf
CODEGEN-4281	Unexpected type returned by bitset [] operator
CODEGEN-4276	std::multimap::clear is not noexcept
CODEGEN-4275	std::num_get does not parse floating-point strings correctly
CODEGEN-4259	noexcept(typeid(d)) runtime fail on polymorphic class type
CODEGEN-4258	deeply nested lambda functions hang the codegen
CODEGEN-4250	regex_constants::ECMAScript not expected to be 0