[AArch64][GlobalISel] Legalize s16 G_FCONSTANT to avoid widening to G_CONSTANT

[HolyMolyCowMan]

When widening a G_FCONSTANT it is converted to a G_CONSTANT to avoid loss in accuracy (see #56454). This means that some folds such as G_FPEXT(G_FCONSTANT) fail to work when the scalar has been widened.

This PR legalizes s16s by default in line with how s16 G_CONSTANTs are treated.

[llvmbot]

@llvm/pr-subscribers-llvm-globalisel

@llvm/pr-subscribers-backend-aarch64

Author: Ryan Cowan (HolyMolyCowMan)

Changes

When widening a G_FCONSTANT it is converted to a G_CONSTANT to avoid loss in accuracy (see #56454). This means that some folds such as G_FPEXT(G_FCONSTANT) fail to work when the scalar has been widened.

This PR legalizes s16s by default in line with how s16 G_CONSTANTs are treated.

---

Full diff: https://github.com/llvm/llvm-project/pull/161205.diff

1 Files Affected:

• (modified) llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp (+2-2)

diff --git a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
index 7ee54c5932b15..1593f32d1fc6c 100644
--- a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
+++ b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
@@ -678,8 +678,8 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
       .widenScalarToNextPow2(0)
       .clampScalar(0, s8, s64);
   getActionDefinitionsBuilder(G_FCONSTANT)
-      .legalFor({s32, s64, s128})
-      .legalFor(HasFP16, {s16})
+      // Always legalize s16 to prevent G_FCONSTANT being widened to G_CONSTANT
+      .legalFor({s16, s32, s64, s128})
       .clampScalar(0, MinFPScalar, s128);
 
   // FIXME: fix moreElementsToNextPow2

[aemerson]

Yeah this needs to address the regbank regressions.

[HolyMolyCowMan]

By moving the matchFConstantToConstant to the post-legalizer lowering, patches such as #160902 can more easily fold operations involving G_FCONSTANTs, and still make use of zero registers when applicable.

[davemgreen]

LGTM