[llvm] [AMDGPU] Limit DVGPR entry function max VGPRs to block size (PR #226242)

[Shilei Tian via llvm-commits]

================
@@ -636,9 +636,18 @@ unsigned GCNSubtarget::getBaseMaxNumVGPRs(
 unsigned GCNSubtarget::getMaxNumVGPRs(const Function &F) const {
   unsigned DynamicVGPRBlockSize = AMDGPU::getDynamicVGPRBlockSize(F);
   std::pair<unsigned, unsigned> Waves = getWavesPerEU(F);
-  return getBaseMaxNumVGPRs(
+
+  unsigned MaxNumVGPRs = getBaseMaxNumVGPRs(
       F, {getMinNumVGPRs(Waves.second, DynamicVGPRBlockSize),
           getMaxNumVGPRs(Waves.first, DynamicVGPRBlockSize)});
+
+  // In DVGPR mode, a wave launches with a single VGPR block allocated. Applied
+  // after getBaseMaxNumVGPRs so "amdgpu-num-vgpr" cannot raise it back up.
----------------
shiltian wrote:

In the long run, we need to remove amdgpu-num-vgpr

https://github.com/llvm/llvm-project/pull/226242