[llvm] AMDGPU/GlobalISel: Fix broken exp10 lowering for f16 (PR #170708)

[Matt Arsenault via llvm-commits]

================
@@ -5916,8 +5907,11 @@ define half @v_exp10_f16(half %in) {
 ; SI-GISEL:       ; %bb.0:
 ; SI-GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; SI-GISEL-NEXT:    v_cvt_f32_f16_e32 v0, v0
-; SI-GISEL-NEXT:    v_mul_f32_e32 v0, 0x3fb8aa3b, v0
+; SI-GISEL-NEXT:    v_mul_f32_e32 v1, 0x3a2784bc, v0
+; SI-GISEL-NEXT:    v_mul_f32_e32 v0, 0x40549000, v0
+; SI-GISEL-NEXT:    v_exp_f32_e32 v1, v1
 ; SI-GISEL-NEXT:    v_exp_f32_e32 v0, v0
+; SI-GISEL-NEXT:    v_mul_f32_e32 v0, v0, v1
 ; SI-GISEL-NEXT:    v_cvt_f16_f32_e32 v0, v0
----------------
arsenm wrote:

This is the old f16 ABI on < gfx8 bug 

https://github.com/llvm/llvm-project/pull/170708