__fmax() __fmin() intrinsics in CGTools 5.2.10

These intrinsics should work with 5.2.10. Are you sure you are compiling with --float_support=fpu32?
I was able to generate the MINF32 instruction for the __fmin() intrinsic with a simple example. Can you attach an example that does not work for you including the compile options being used?

Sure, I made this quick example which has the same behavior as I have on my project.

#include <math.h>
int main(void)
{
    float a = 0.0, b = 1.0;
    double c = 0.0, d = 1.0, idx_c = 10, idx_d = 11;

    __fmax(a, b);
    __f32_max_idx(c, d, idx_c, idx_d);
}

Which gives me the following assembly
        ZERO      R0H                   ; |4|
        MOV32     *-SP[2],R0H
        MOVIZ     R0H,#16256            ; |4|
        MOV32     *-SP[4],R0H
        ZERO      R0H                   ; |5|
        MOV32     *-SP[6],R0H
        MOVIZ     R0H,#16256            ; |5|
        MOV32     *-SP[8],R0H
        MOVIZ     R0H,#16672            ; |5|
        MOV32     *-SP[10],R0H
        MOVIZ     R0H,#16688            ; |5|
        MOV32     *-SP[12],R0H
        MOV32     R1H,*-SP[10]
        MOV32     R3H,*-SP[8]
        MOV32     R0H,*-SP[6]
        MOV32     R2H,*-SP[12]
        MAXF32    R0H,R3H
       || MOV32  R1H,R2H               ; |8|
        MOV32     *-SP[10],R1H
        MOV32     *-SP[6],R0H
        MOVB      AL,#0
        SUBB      SP,#12
        LRETR

I only see the assembly code for the __f32_max_idx intrinsic. Line 7 is totally skipped.

Build options : -g -k -pdsw225 -fr"$(Proj_dir)\Debug" -d"_DEBUG" -d"LARGE_MODEL" -ml -v28 --float_support=fpu32

You need to be careful in devising test code for such things, because the compiler may optimize the generated assembly code.  For example, if the statement __fmax(a,b); doesn't require the contents of a or b to be modified, then the compiler can choose not to emit any corresponding assembly code.

You got a point. I simplified a lot in my example, maybe too much. To avoid pre-processing optimisation, I added some "weird" randomization of a and b which should be resolved at link time only. I expect that the compiler preprocessor will not be able to optimize the assembly code based on static analysis.

#include <math.h>
int main(void)
{
    float a = 0.0, b = 1.0;
    double c = 0.0, d = 1.0, idx_c = 10, idx_d = 11;

    a = (float) (unsigned long int) &a; // weird randomization of a
    b = (float) (unsigned long int) &b; // weird randomization of b
    __fmax(a, b);
    __f32_max_idx(c, d, idx_c, idx_d);
}

Still I only have the assembly code for __f32_max_idx() line 10, but nothing for __fmax() line 9.

        ZERO      R0H                   ; |4|
        MOV32     *-SP[2],R0H
        MOVIZ     R0H,#16256            ; |4|
        MOV32     *-SP[4],R0H
        ZERO      R0H                   ; |5|
        MOV32     *-SP[6],R0H
        MOVIZ     R0H,#16256            ; |5|
        MOV32     *-SP[8],R0H
        MOVIZ     R0H,#16672            ; |5|
        MOV32     *-SP[10],R0H
        MOVIZ     R0H,#16688            ; |5|
        MOV32     *-SP[12],R0H
        MOVZ      AR4,SP                ; |7|
        SUBB      XAR4,#2               ; |7|
        MOV32     R0H,XAR4
        NOP
        NOP
        NOP
        NOP
        UI32TOF32 R0H,R0H               ; |7|
        NOP
        MOV32     *-SP[2],R0H
        MOVZ      AR4,SP                ; |8|
        SUBB      XAR4,#4               ; |8|
        MOV32     R0H,XAR4
        NOP
        NOP
        NOP
        NOP
        UI32TOF32 R0H,R0H               ; |8|
        NOP
        MOV32     *-SP[4],R0H
        MOV32     R1H,*-SP[10]
        MOV32     R3H,*-SP[8]
        MOV32     R0H,*-SP[6]
        MOV32     R2H,*-SP[12]
        MAXF32    R0H,R3H
       || MOV32  R1H,R2H               ; |10|
        MOV32     *-SP[10],R1H
        MOV32     *-SP[6],R0H
        MOVB      AL,#0
        SUBB      SP,#12
        LRETR