Comparing floats with constant doubles promotes to double

[jrmuizel]

Bugzilla Link	9592
Resolution	FIXED
Resolved on	Mar 31, 2011 05:14
Version	trunk
OS	All
CC	@d0k

Extended Description

The following code:
float clamp_float(float a)
{
if (a > 1.)
return 1.;
else
return a;
}

compiles to:

    cvtss2sd        %xmm0, %xmm1
    ucomisd .LCPI0_0(%rip), %xmm1
    jbe     .LBB0_2

BB#1:

    movss   .LCPI0_1(%rip), %xmm0

.LBB0_2: # %entry
ret

GCC avoids the promotion to double.

[llvmbot]

Dragonegg compiles this to:

movss	.LCPI0_0(%rip), %xmm1
minss	%xmm0, %xmm1
movaps	%xmm1, %xmm0
ret

[d0k]

Thanks for the nice testcase.

This is another optzn gcc (and dragonegg) do in the frontend but LLVM misses. Fixed in r128625, clang now generates the same movss/minss/movaps sequence as gcc.